id
stringlengths 12
15
| title
stringlengths 8
162
| content
stringlengths 1
17.6k
| prechunk_id
stringlengths 0
15
| postchunk_id
stringlengths 0
15
| arxiv_id
stringlengths 10
10
| references
sequencelengths 1
1
|
|---|---|---|---|---|---|---|
1601.00257#40 | Modave Lectures on Applied AdS/CFT with Numerics | Sonâ shell = aif aev= G(VaF@)A [ee â hng FAs] + 1 ee â â af dey g®(DaD* â m?)® pes hngD*®®) + CC] _ 5 / d2/=Gi(BD â BD) A, â / BxV/â hngFâ Ay] â Lf. _ s(f aa â hngD%® + C.C.). (5.3) By the asymptotic expansion in (3.10) and (3.11), the divergence comes only from the last \-P6 z two boundary terms and can be read off as . So the holographic renormalization can be readily achieved by adding the boundary term â | 1601.00257#39 | 1601.00257#41 | 1601.00257 | [
"1510.02804"
] |
1601.00257#41 | Modave Lectures on Applied AdS/CFT with Numerics | J @Pa/â h|®|? to the original action. Whence we have 6S, ly ren op Gi") = Fa, 7 OSren a (O) = i =, (5.4) where jµ corresponds to the conserved particle current and the expectation value for the scalar operator O is interpreted as the condensate order parameter of superï¬ uid. If this scalar operator acquires a nonzero expectation value spontaneously in the situation where the source is turned oï¬ , the boundary system is driven into a superï¬ uid phase. # 5.3 Background solution, Free energy, and Phase transition With the assumption that the non-vanishing bulk matter ï¬ elds (Φ = zÏ , At, Ax) do not depend on the coordinate θ, the equations of motion can be explicitly written as 6Note that the outward normal vector is given by na = â z( â â z )a. â 13 â 0 = â 2 t Ï + (z + A2 +3z2â zÏ + (z3 â 1)â 2 t Ax â â tâ xAt â i(Ï â x Â¯Ï â Â¯Ï â xÏ ) + 2AxÏ Â¯Ï + 3z2â zAx + (z3 â 1)â 2 (5.6) 0 = â 2 0 = (z3 â 1)â 2 0 = â tâ zAt + i(Ï â z Â¯Ï â Â¯Ï â zÏ ) â â zâ xAx, z Ax, xAt + â tâ xAx + 2 Â¯Ï Ï At + i( Â¯Ï â tÏ â Ψâ t Â¯Ï ), z At + 3z2â zAt â â 2 (5.7) (5.8) | 1601.00257#40 | 1601.00257#42 | 1601.00257 | [
"1510.02804"
] |
1601.00257#42 | Modave Lectures on Applied AdS/CFT with Numerics | where the third one is the constraint equation and the last one reduces to the conserved equation for the boundary current when evaluated at the AdS boundary, i.e., # â tÏ = â â xjx. ap = â Onj®. (5.9) To specialize into the homogeneous phase diagram for our holographic model, we further make the following ansatz for our non-vanishing bulk matter ï¬ elds Ï = Ï (z), At = At(z). (5.10) Then the equations of motion for the static solution reduce to (5.11) | 1601.00257#41 | 1601.00257#43 | 1601.00257 | [
"1510.02804"
] |
1601.00257#43 | Modave Lectures on Applied AdS/CFT with Numerics | 0 = 3z2â zÏ + (z3 â 1)â 2 z Ï + (z â A2 0 = 2AtÏ Â¯Ï + 3z2â zAt + (z3 â 1)â 2 0 = Ï â z Â¯Ï â Â¯Ï â zÏ , t )Ï , z At, (5.12) (5.13) where the last equation implies that we can always choose a gauge to make Ï | 1601.00257#42 | 1601.00257#44 | 1601.00257 | [
"1510.02804"
] |
1601.00257#44 | Modave Lectures on Applied AdS/CFT with Numerics | real. It is not hard to see the above equations of motion have a trivial solution Ï = 0, At = µ, (5.14) which corresponds to the vacuum phase with zero particle density. On the other hand, to obtain the non-trivial solution dual to the superï¬ uid phase, we are required to resort to pseudo-spectral method. As a demonstration, we here plot the nontrivial proï¬ le for Ï and At at µ = 2 in Figure 5. | 1601.00257#43 | 1601.00257#45 | 1601.00257 | [
"1510.02804"
] |
1601.00257#45 | Modave Lectures on Applied AdS/CFT with Numerics | The variation of particle density and condensate with respect to the chemical potential is plotted in Figure 6, which indicates that the phase transition from the vacuum to a superï¬ uid occurs at µc = 1.715. It is noteworthy that such a phenomenon is rem- iniscent of the recently observed quantum critical behavior of ultra-cold cesium atoms in an optical lattice across the vacuum to superï¬ uid transition by tuning the chemical potential[36]. Moreover, the compactiï¬ ed dimension in the AdS soliton background can be naturally identi- ï¬ ed as the reduced dimension in optical lattices by the very steep harmonic potential as both mechanisms make the eï¬ ective dimension of the system in consideration reduced in the low energy regime. On the other hand, note that the particle density shows up at the same time as our superï¬ uid condensate, thus it is tempting to claim that this particle density Ï is simply the superï¬ uid density Ï s. | 1601.00257#44 | 1601.00257#46 | 1601.00257 | [
"1510.02804"
] |
1601.00257#46 | Modave Lectures on Applied AdS/CFT with Numerics | This claim is also consistent with the fact that we are working with â 14 â (5.5) (5.9) 2.0F 4 "ash 4 Figure 5: The bulk proï¬ le for the scalar ï¬ eld and time component of gauge ï¬ eld at the chemical potential µ = 2. | 3.0F 3 4 2.55 2.0 Po IkO>l ; 1.06 | 0.5 0 0.0 0 1 2 3 4 0 1 2 3 4 u u Figure 6: The variation of particle density and condensate with respect to the chemical potential, where we see the second order quantum phase transition take place at µc = 1.715. a zero temperature superï¬ uid where the normal ï¬ uid component should disappear. | 1601.00257#45 | 1601.00257#47 | 1601.00257 | [
"1510.02804"
] |
1601.00257#47 | Modave Lectures on Applied AdS/CFT with Numerics | As we will show later on by the linear response theory, this is actually the case. But to make sure that Figure 6 represents the genuine phase diagram for our holographic model, we are required to check whether the corresponding free energy density is the lowest in the grand canonical ensemble. By holography, the free energy density can be obtained from the renormalized on shell Lagrangian of matter ï¬ elds as follows7 1 _ _ F 5 / dz/â gi(®D°S â SDS) A, â Vâ hng ApFâ ¢|-=0] = Sno | de(aio)?, (5.15) where we have made use of the source free boundary condition for the scalar ï¬ eld at the AdS boundary. As shown in Figure 7, the superï¬ uid phase is the thermodynamically favored one 7Here we have used iSLorentzian = â SEuclidean and it = Ï with Ï the Euclidean time identiï¬ | 1601.00257#46 | 1601.00257#48 | 1601.00257 | [
"1510.02804"
] |
1601.00257#48 | Modave Lectures on Applied AdS/CFT with Numerics | ed as the inverse of temperature. â 15 â Figure 7: The diï¬ erence of free energy density for the superï¬ uid phase from that for the vacuum phase. compared to the vacuum phase when the chemical potential is greater than the critical value. So we are done. # 5.4 Linear response theory, Optical conductivity, and Superï¬ uid density Now let us set up the linear response theory for the later calculation of the optical conductivity of our holographic model. | 1601.00257#47 | 1601.00257#49 | 1601.00257 | [
"1510.02804"
] |
1601.00257#49 | Modave Lectures on Applied AdS/CFT with Numerics | To achieve this, we ï¬ rst decompose the ï¬ eld Ï into its real and imaginary parts as Ï = Ï r + iÏ i, (5.16) and assume that the perturbation bulk ï¬ elds take the following form Î´Ï r = Î´Ï r(z)eâ iÏ t+iqx, Î´Ï i = Î´Ï i(z)eâ iÏ t+iqx, δAt = δAt(z)eâ iÏ t+iqx, δAx = δAx(z)eâ iÏ t+iqx, (5.17) since the background solution is static and homogeneous. With this, the perturbation equa- tions can be simpliï¬ | 1601.00257#48 | 1601.00257#50 | 1601.00257 | [
"1510.02804"
] |
1601.00257#50 | Modave Lectures on Applied AdS/CFT with Numerics | ed as t )Î´Ï r â 2iÏ AtÎ´Ï i + q2Î´Ï r + 3z2â zÎ´Ï r + (z3 â 1)â 2 z Î´Ï r 0 = â Ï 2Î´Ï r + (z â A2 â 2AtÏ rδAt, 0 = â Ï 2Î´Ï i + (z â A2 t )Î´Ï i + 2iÏ AtÎ´Ï r + q2Î´Ï i + 3z2â zÎ´Ï i + (z3 â 1)â 2 z Î´Ï i +iÏ Ï rδAt + iqÏ rδAx, (5.19) (5.18) 0 = â Ï 2δAx â Ï qδAt + 3z2â zδAx + (z3 â 1)â 2 0 = (z3 â 1)â 2 z δAx + 2Ï 2 rδAx â 2iqÏ rÎ´Ï i, z δAt + 3z2â zδAt + q2δAt + Ï qδAx + 2Ï 2 rδAt + 4AtÏ rÎ´Ï r (5.20) +2iÏ Ï rÎ´Ï i, (5.21) 0 = â iÏ â zδAt â iqâ zδAx â 2(â zÏ rÎ´Ï i â Ï râ zÎ´Ï i), (5.22) where we have used Ï i = 0 for the background solution. | 1601.00257#49 | 1601.00257#51 | 1601.00257 | [
"1510.02804"
] |
1601.00257#51 | Modave Lectures on Applied AdS/CFT with Numerics | â 16 â Note that the gauge transformation A â A + â θ, Ï â Ï eiθ (5.23) with θ = 1 i λeâ iÏ t+iqx (5.24) induces a spurious solution to the above perturbation equations as δAt = â Î»Ï , δAx = λq, Î´Ï = Î»Ï . (5.25) We can remove such a redundancy by requiring δAt = 0 at the AdS boundary8. In addition, Î´Ï will also be set to zero at the AdS boundary later on. On the other hand, taking into account the fact that the perturbation equation (5.22) will be automatically satisï¬ ed in the whole bulk once the other perturbations are satisï¬ ed9, we can forget about (5.22) from now on. That is to say, we can employ the pseudo-spectral method to obtain the desired numerical solution by combining the rest perturbation equations with the aforementioned boundary conditions as well as the other boundary conditions at the AdS boundary, depending on the speciï¬ c problem we want to solve. In particular, to calculate the optical conductivity for our holographic model, we can simply focus on the q = 0 mode and further impose δAx = 1 at the AdS boundary. Then the optical conductivity can be extracted by holography as | 1601.00257#50 | 1601.00257#52 | 1601.00257 | [
"1510.02804"
] |
1601.00257#52 | Modave Lectures on Applied AdS/CFT with Numerics | Ï (Ï ) = â zδAx|z=0 iÏ (5.26) for any positive frequency Ï 10. According to the perturbation equations, the whole calculation is much simpliï¬ ed because δAx decouples from the other perturbation bulk ï¬ elds. We simply plot the imaginary part of the optical conductivity in Figure 8 for both vacuum and superï¬ uid phase, because the real part vanishes due to the reality of the perturbation equation and boundary condition for δAx. As it should be the case, the DC conductivity vanishes for the vacuum phase, but diverges for the superï¬ uid phase due to the 1 Ï behavior of the imaginary part of optical conductivity by the Krames-Kronig relation Im[o(w)] = =P [. a Rel) (5.27) T Joo Jâ W Ww | 1601.00257#51 | 1601.00257#53 | 1601.00257 | [
"1510.02804"
] |
1601.00257#53 | Modave Lectures on Applied AdS/CFT with Numerics | Furthermore, according to the hydrodynamical description of superï¬ uid, the superï¬ uid den- sity Ï s can be obtained by ï¬ tting this zero pole as Ï s ÂµÏ [39, 40, 41]. As expected, our numerics shows that the resultant superï¬ uid density is exactly the same as the particle density within â 17 â Figure 8: The left panel is the imaginary part of optical conductivity for the vacuum phase, and the right panel is for the superï¬ uid phase at µ = 6.5. our numerical accuracy. The other poles correspond to the gapped normal modes for δAx, which we are not interested in since we are focusing on the low energy physics. Let us come back to the equality between the particle density and superï¬ uid density. Although this numerical result is 100 percent reasonable from the physical perspective, it is highly non-trivial in the sense that the superï¬ uid density comes from the linear response theory while the particle density is a quantity associated with the equilibrium state. So it is better to have an analytic understanding for this remarkable equality. Here we would like to develop an elegant proof for this equality by a boost trick. To this end, we are ï¬ rst required to realize Ï s = â µâ zδAx|z=0 with Ï = 0. Such an Ï = 0 perturbation can actually be implemented by a boost 1 1 t= woe" vaâ ), x ize (a! â vtâ ) (5.28) acting on the superï¬ uid phase. | 1601.00257#52 | 1601.00257#54 | 1601.00257 | [
"1510.02804"
] |
1601.00257#54 | Modave Lectures on Applied AdS/CFT with Numerics | Note that the background metric is invariant under such a boost. As a result, we end up with a new non-trivial solution as follows 1 v = ¢,A,= A,, Al, = At. 5.29 g = 9,4, Vin ae Vice (5.29) We expand this solution up to the linear order in v as ¢! = $, A, = Ay, Al, = VAs, (5.30) which means that the linear perturbation δAx is actually proportional to the background solution At. So we have Ï s = Ï | 1601.00257#53 | 1601.00257#55 | 1601.00257 | [
"1510.02804"
] |
1601.00257#55 | Modave Lectures on Applied AdS/CFT with Numerics | immediately. 8The only exception is the Ï case, which can always be separately managed if necessary. 9This result comes from the following two facts. One is related to Bianchi â g â µ( identity 0 = â ava = z4 ) = 0 if the rest equations of motion hold. The other is special to our holographic model, in which the readers are encouraged to show that the z component of Maxwell equation turns out to be satisï¬ ed automatically at z = 1 if the rest equations hold there. 10Note that Ï (â Â¯Ï ) = Ï (Ï | 1601.00257#54 | 1601.00257#56 | 1601.00257 | [
"1510.02804"
] |
1601.00257#56 | Modave Lectures on Applied AdS/CFT with Numerics | ), so we focus only on the positive frequency here. â 18 â 1000 600} 4 400+ 4 200 4 | fu | A 0.0 05 1.0 15 2.0 3.0 Figure 9: The density plot of er | with g = 0.3 for the superfluid phase at pp = 6.5. The normal modes can be identified by the peaks, where the red one denotes the hydrodynamic normal mode wo = 0.209. 0.0 05 1.0 15 2.0 25 3.0 Figure 10: The spectral plot of ln |δ Ë Ï i(Ï , 1)| with q = 0.3 for the superï¬ uid phase at µ = 6.5, where the initial data are chosen as Î´Ï i = z with all the other perturbations turned oï¬ | 1601.00257#55 | 1601.00257#57 | 1601.00257 | [
"1510.02804"
] |
1601.00257#57 | Modave Lectures on Applied AdS/CFT with Numerics | . The normal modes can be identiï¬ ed by the peaks, whose locations are the same as those by the frequency domain analysis within our numerical accuracy. # 5.5 Time domain analysis, Normal modes, and Sound speed In what follows we shall use linear response theory to calculate the speed of sound by focusing solely on the hydrodynamic spectrum of normal modes of the gapless Goldstone from the spontaneous symmetry breaking, which is obviously absent from the vacuum phase. As such, the perturbation fields are required to have Dirichlet boundary conditions at the AdS boundary. Then we cast the linear perturbation equations and boundary conditions into the form £L(w)u = 0 with u the perturbation fields evaluated at the grid points by pseudo- spectral method. The normal modes are obtained by the condition det{[£(w)| = 0, which can be further identified by the density plot Feaena with the prime the derivative with respect to w. We demonstrate such a density plot in Figure 9, where the hydrodynamic mode is simply the closest mode to the origin, marked in red. Besides such a frequency domain | 1601.00257#56 | 1601.00257#58 | 1601.00257 | [
"1510.02804"
] |
1601.00257#58 | Modave Lectures on Applied AdS/CFT with Numerics | â 19 â 08 06+ 4 0.4 02+ 4 0.0 . rae . 0.0 0.2 04 0.6 08 1.0 1.2 14 Figure 11: The dispersion relation for the gapless Goldstone mode in the superï¬ uid phase at µ = 6.5, where the sound speed vs = 0.697 is obtained by ï¬ tting the long wave modes with Ï 0 = vsq. 0.7 | os! 05 0.4 Vs 035 Figure 12: | 1601.00257#57 | 1601.00257#59 | 1601.00257 | [
"1510.02804"
] |
1601.00257#59 | Modave Lectures on Applied AdS/CFT with Numerics | The variation of sound speed with respect to the chemical potential. When the chemical potential is much larger than the conï¬ ning scale, the conformality is restored and the sound speed approaches the predicted value 1â 2 analysis of spectrum of normal modes, there is an alternative called time domain analysis, which we would like to elaborate on below. We ï¬ rst cast the equations of motion into the following Hamiltonian formalism (5.31) â tÏ = iAtÏ + P, â tP = iAtP â (z + A2 â tAx = Î x + â xAt, â tÎ x = i(Ï â x Â¯Ï â Â¯Ï â xÏ ) â 2AxÏ Â¯Ï â 3z2â zAx + (1 â z3)â 2 x + iâ xAx)Ï â 2iAxâ xÏ + â 2 xÏ â 3z2â zÏ + (1 â z3)â 2 z Ï , (5.32) (5.33) z Ax, (5.34) 0 = (z3 â 1)â 2 z At + 3z2â zAt + â xÎ x â i( ¯P Ï â P Â¯Ï ), # â tâ zAt = â i(Ï â z Â¯Ï â Â¯Ï â zÏ ) + â zâ xAx. (5.36) Then with the assumption that the perturbation bulk ï¬ elds take the form as δ(t, z)eiqx, the | 1601.00257#58 | 1601.00257#60 | 1601.00257 | [
"1510.02804"
] |
1601.00257#60 | Modave Lectures on Applied AdS/CFT with Numerics | â 20 â (5.35) perturbation equations on top of the superï¬ uid phase is given by â tÎ´Ï r = â AtÎ´Ï i + δPr, â tÎ´Ï i = Ï rδAt + AtÎ´Ï r + δPi, â tδPr = AtÏ rδAt â AtδPi â (z + q2)Î´Ï r â 3z2â zÎ´Ï r + (1 â z3)â 2 â tδPi = â iqÏ rδAx + AtδPr â (z + q2)Î´Ï i â 3z2â zÎ´Ï i + (1 â z3)â 2 â tδAx = δΠx + iqδAt, â tδΠx = 2iqÏ rÎ´Ï i â 2Ï 2 0 = (z3 â 1)â 2 z Î´Ï r, z Î´Ï i, â tâ zδAt = 2â zÏ rÎ´Ï i â 2Ï râ zÎ´Ï i + iqâ zδAx. (5.44) As before, using the source free boundary conditions for all the perturbation ï¬ elds, we can obtain the temporal evolution of the perturbation ï¬ elds for any given initial data by Runge- Kutta method, where δAt is solved by the constraint equation (5.43). The normal modes can then be identiï¬ ed by the peaks in the Fourier transformation of the evolving data. We demonstrate such a spectral plot in Figure 10. As expected, such a time domain analysis gives rise to the same result for the locations of normal modes as that by the frequency domain analysis. Then the dispersion relation for the gapless Goldstone can be obtained and plotted in Figure 11, whereby the sound speed vs can be obtained by the ï¬ tting formula Ï 0 = vsq. As shown in Figure 12, the sound speed increases with the chemical potential and saturate to the predicted value 1â by conformal ï¬ eld theory when the chemical potential is much larger than 2 the conï¬ | 1601.00257#59 | 1601.00257#61 | 1601.00257 | [
"1510.02804"
] |
1601.00257#61 | Modave Lectures on Applied AdS/CFT with Numerics | ning scale[39, 40, 41], which is reasonable since it is believed that the conformality is restored in this limit. # 6. Concluding Remarks Like any other uniï¬ cation in physics, AdS/CFT correspondence has proven to be a unique tool for one to address various universal behaviors of near-equilibrium as well as far-from- equilibrium dynamics for a variety of strongly coupled systems, which otherwise would be hard to attack. During such an application, numerical computation has been playing a more and more important role in the sense that not only can numerics leave us with some conjectures to develop an analytic proof and some patterns to have an analytic understanding but also brings us to the regime where the analytic treatment is not available at all. In these lecture notes, we have touched only upon the very basics for the numerics in applied AdS/CFT. In addition, we work only with the probe limit in the concrete example we make use of to demonstrate how to apply AdS/CFT with numerics. The situation will become a little bit involved when the back reaction is taken into account. Regarding this, the readers are suggested to refer to [42] to see how to obtain the stationary inhomogeneous solutions to fully back reacted Einstein equation by Einstein-DeTurck method. | 1601.00257#60 | 1601.00257#62 | 1601.00257 | [
"1510.02804"
] |
1601.00257#62 | Modave Lectures on Applied AdS/CFT with Numerics | On the other â 21 â hand, the readers are recommended to refer to [43] to see how to evolve the fully back reacted dynamics, where with a black hole as the initial data it turns out that the Eddington like coordinates are preferred to the Schwarzschild like coordinates. # Acknowledgments H.Z. would like to thank the organizers of the Eleventh International Modave Summer School on Mathematical Physics held in Modave, Belgium, September 2015, where the lectures on which these notes are based were given. He is indebted to Nabil Iqbal for his valuable discussions at the summer school. H.Z. would also like to thank the organizers of 2015 International School on Numerical Relativity and Gravitational Waves held in Daejeon, Korea, July 2015, where these lectures were geared to the audience mainly from general relativity and gravity community. He is grateful to Keun-Young Kim, Kyung Kiu Kim, Miok Park, and Sang-Jin Sin for the enjoyable conversations during the school. H.Z. is also grateful to Ben Craps and Alex Sevrin for the fantastic infrastructure they provide at HEP group of VUB and the very freedom as well as various opportunities they oï¬ | 1601.00257#61 | 1601.00257#63 | 1601.00257 | [
"1510.02804"
] |
1601.00257#63 | Modave Lectures on Applied AdS/CFT with Numerics | er to him. M.G. is partially supported by NSFC with Grant Nos.11235003, 11375026 and NCET-12-0054. C.N. is supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Science, ICT & Future Planning(NRF- 2014R1A1A1003220) and the 2015 GIST Grant for the FARE Project (Further Advancement of Research and Education at GIST College). Y.T. is partially supported by NSFC with Grant No.11475179. H.Z. is supported in part by the Belgian Federal Science Policy Oï¬ ce through the Interuniversity Attraction Pole P7/37, by FWO-Vlaanderen through the project G020714N, and by the Vrije Universiteit Brussel through the Strategic Research Program â | 1601.00257#62 | 1601.00257#64 | 1601.00257 | [
"1510.02804"
] |
1601.00257#64 | Modave Lectures on Applied AdS/CFT with Numerics | High-Energy Physicâ . He is also an individual FWO Fellow supported by 12G3515N. # References [1] E. Witten, arXiv:hep-th/0106109. [2] A. Strominger, arXiv:hep-th/0106113. [3] M. Spradlin, A. Strominger, and A. Volovich, arXiv:hep-th/0110007. [4] R. Britto, F. Cachazo, B. Feng, and E. Witten, Phys. Rev. Lett. 94 (181602)(2005). [5] N. Arkani-Hamed, F. Cachazo, and J. Kaplan, JHEP 1009, 016(2010). [6] J. Maldacena, Adv. Theor. Math. Phys. 2, 231(1998). | 1601.00257#63 | 1601.00257#65 | 1601.00257 | [
"1510.02804"
] |
1601.00257#65 | Modave Lectures on Applied AdS/CFT with Numerics | # Cole ND [7] E. Witten, Adv. Theor. Math. Phys. 2, 253(1998). [8] S. Gubser, I. R. Klebanov, and A. M. Polyakov, Phys. Lett. B 428, 105(1998). [9] P. Breitenlohner and D. Z. Freedman, Annals Phys. 144, 249(1982). [10] P. Breitenlohner and D. Z. Freedman, Phys. Lett. B 115, 197(1982). [11] S. | 1601.00257#64 | 1601.00257#66 | 1601.00257 | [
"1510.02804"
] |
1601.00257#66 | Modave Lectures on Applied AdS/CFT with Numerics | Ryu and T. Takayanagi, Phys. Rev. Lett. 96, 181602(2006). â 22 â [12] V. E. Hubeny, M. Rangamani, and T. Takayanagi, JHEP 0707, 062(2007). [13] A. Lewkowycz and J. Maldacena, JHEP 08, 090(2013). [14] R. M. | 1601.00257#65 | 1601.00257#67 | 1601.00257 | [
"1510.02804"
] |
1601.00257#67 | Modave Lectures on Applied AdS/CFT with Numerics | Wald, Living. Rev. Rel. 4, 6(2001). [15] J. D. Brown and M. Henneaux, Commun. Math. Phys. 104, 207(1986). [16] A. Strominger, JHEP 02, 009(1998). [17] J. D. Brown and J. W. York, Phys. Rev. D 47, 1407(1993). [18] V. E. Hubeny, S. Minwalla, and M. | 1601.00257#66 | 1601.00257#68 | 1601.00257 | [
"1510.02804"
] |
1601.00257#68 | Modave Lectures on Applied AdS/CFT with Numerics | Rangamani, arXiv:1107.5780. [19] B. Swingle, Phys. Rev. D 86, 065007(2012). [20] X. L. Qi, arXiv:1309.6282. [21] J. Casalderrey-Solana, H. Liu, D. Mateos, K. Rajagopal, and U. A. Wiedemann, arXiv:1101.0618. [22] U. | 1601.00257#67 | 1601.00257#69 | 1601.00257 | [
"1510.02804"
] |
1601.00257#69 | Modave Lectures on Applied AdS/CFT with Numerics | Gursoy, E. Kiritsis, L. Mazzanti, G. Michalogiorgakis, and F. Nitti, Lect. Notes Phys. 828, 79(2011). [23] S. A. Hartnoll, Class. Quant. Grav. 26, 224002(2009). [24] J. McGreevy, Adv. High Energy Phys. 2010, 723105(2010). [25] C. P. Herzog, J. Phys. | 1601.00257#68 | 1601.00257#70 | 1601.00257 | [
"1510.02804"
] |
1601.00257#70 | Modave Lectures on Applied AdS/CFT with Numerics | A 42, 343001(2009). [26] G. T. Horowitz, arXiv:1002.1722. [27] N. Iqbal, H. Liu, and M. Mezei, arXiv:1110.3814. [28] W. J. Li, Y. Tian, and H. Zhang, JHEP 07, 030(2013). [29] N. Callebaut, B. Craps, F. Galli, D. C. Thompson, J. Vanhoof, J. Zaanen, and H. Zhang, JHEP 10, 172(2014). [30] B. | 1601.00257#69 | 1601.00257#71 | 1601.00257 | [
"1510.02804"
] |
1601.00257#71 | Modave Lectures on Applied AdS/CFT with Numerics | Craps, E. J. Lindgren, A. Taliotis, J. Vanhoof, and H. Zhang, Phys. Rev. D 90, 086004(2014). [31] R. Li, Y. Tian, H. Zhang, and J. Zhao, Phys. Lett. B 750, 520(2015). [32] Y. Du, C. Niu, Y. Tian, and H. Zhang, JHEP 12, 018(2015). [33] Y. Du, S. Q. Lan, Y. Tian, and H. Zhang, JHEP 01, 016(2016). [34] M. Guo, S. Q. Lan, C. Niu, Y. Tian, and H. Zhang, to appear. [35] T. Nishioka, S. Ryu, and T. Takayanagi, JHEP 1003, 131(2010). [36] X. Zhang, C. L. Hung, S. K. Tung, and C. Chin, Science 335, 1070(2012). [37] I. R. Klebanov and E. Witten, Nucl. Phys. B 556, 89(1999). [38] K. | 1601.00257#70 | 1601.00257#72 | 1601.00257 | [
"1510.02804"
] |
1601.00257#72 | Modave Lectures on Applied AdS/CFT with Numerics | Skenderis, Class. Quant. Grav.19, 5849(2002). [39] C. P. Herzog, P. K. Kovtun, and D. T. Son, Phys. Rev. D 79, 066002(2009). [40] A. Yarom, JHEP 0907, 070(2009). [41] C. P. Herzog and A. Yarom, Phys. | 1601.00257#71 | 1601.00257#73 | 1601.00257 | [
"1510.02804"
] |
1601.00257#73 | Modave Lectures on Applied AdS/CFT with Numerics | Rev. D 80, 106002(2009). [42] O. J. C. Dias, J. E. Santos, and B. Way, arXiv:1510.02804. [43] P. Chesler and L. G. Yaï¬ e, JHEP 07, 086(2014). â 23 â | 1601.00257#72 | 1601.00257 | [
"1510.02804"
] |
|
1512.05742#0 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 7 1 0 2 r a M 1 2 ] L C . s c [ 3 v 2 4 7 5 0 . 2 1 5 1 : v i X r a # A Survey of Available Corpora for Building Data-Driven Dialogue Systems Iulian Vlad Serban DIRO, Universit´e de Montr´eal 2920 chemin de la Tour, Montr´eal, QC H3C 3J7, Canada {IULIAN.VLAD.SERBAN} AT UMONTREAL DOT CA Ryan Lowe Department of Computer Science, McGill University 3480 University st, Montr´eal, QC H3A 0E9, Canada {RYAN.LOWE} AT MAIL DOT MCGILL DOT CA Peter Henderson Department of Computer Science, McGill University 3480 University st, Montr´eal, QC H3A 0E9, Canada {PETER.HENDERSON} AT MAIL DOT MCGILL DOT CA Laurent Charlin Department of Computer Science, McGill University 3480 University st, Montr´eal, QC H3A 0E9, Canada {LCHARLIN} AT CS DOT MCGILL DOT CA Joelle Pineau Department of Computer Science, McGill University 3480 University st, Montr´eal, QC H3A 0E9, Canada {JPINEAU} AT CS DOT MCGILL DOT CA Editor: David Traum # Abstract | 1512.05742#1 | 1512.05742 | [
"1511.06931"
] |
|
1512.05742#1 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | During the past decade, several areas of speech and language understanding have witnessed sub- stantial breakthroughs from the use of data-driven models. In the area of dialogue systems, the trend is less obvious, and most practical systems are still built through signiï¬ cant engineering and expert knowledge. Nevertheless, several recent results suggest that data-driven approaches are fea- sible and quite promising. To facilitate research in this area, we have carried out a wide survey of publicly available datasets suitable for data-driven learning of dialogue systems. We discuss im- portant characteristics of these datasets, how they can be used to learn diverse dialogue strategies, and their other potential uses. We also examine methods for transfer learning between datasets and the use of external knowledge. Finally, we discuss appropriate choice of evaluation metrics for the learning objective. | 1512.05742#0 | 1512.05742#2 | 1512.05742 | [
"1511.06931"
] |
1512.05742#2 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | # 1. Introduction Dialogue systems, also known as interactive conversational agents, virtual agents or sometimes chatterbots, are useful in a wide range of applications ranging from technical support services to language learning tools and entertainment (Young et al., 2013; Shawar and Atwell, 2007b). Large- 1 scale data-driven methods, which use recorded data to automatically infer knowledge and strategies, are becoming increasingly important in speech and language understanding and generation. Speech recognition performance has increased tremendously over the last decade due to innovations in deep learning architectures (Hinton et al., 2012; Goodfellow et al., 2015). Similarly, a wide range of data- driven machine learning methods have been shown to be effective for natural language processing, including tasks relevant to dialogue, such as dialogue act classiï¬ cation (Reithinger and Klesen, 1997; Stolcke et al., 2000), dialogue state tracking (Thomson and Young, 2010; Wang and Lemon, 2013; Ren et al., 2013; Henderson et al., 2013; Williams et al., 2013; Henderson et al., 2014c; Kim et al., 2015), natural language generation (Langkilde and Knight, 1998; Oh and Rudnicky, 2000; Walker et al., 2002; Ratnaparkhi, 2002; Stent et al., 2004; Rieser and Lemon, 2010; Mairesse et al., 2010; Mairesse and Young, 2014; Wen et al., 2015; Sharma et al., 2016), and dialogue policy learning (Young et al., 2013). We hypothesize that, in general, much of the recent progress is due to the availability of large public datasets, increased computing power, and new machine learning models, such as neural network architectures. To facilitate further research on building data-driven dialogue systems, this paper presents a broad survey of available dialogue corpora. Corpus-based learning is not the only approach to training dialogue systems. Researchers have also proposed training dialogue systems online through live interaction with humans, and ofï¬ | 1512.05742#1 | 1512.05742#3 | 1512.05742 | [
"1511.06931"
] |
1512.05742#3 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | ine using user simulator models and reinforcement learning methods (Levin et al., 1997; Georgila et al., 2006; Paek, 2006; Schatzmann et al., 2007; Jung et al., 2009; Schatzmann and Young, 2009; GaË si´c et al., 2010, 2011; Daubigney et al., 2012; GaË si´c et al., 2012; Su et al., 2013; Gasic et al., 2013; Pietquin and Hastie, 2013; Young et al., 2013; Mohan and Laird, 2014; Su et al., 2015; Piot et al., 2015; Cuay´ahuitl et al., 2015; Hiraoka et al., 2016; Fatemi et al., 2016; Asri et al., 2016; Williams and Zweig, 2016; Su et al., 2016). However, these approaches are beyond the scope of this survey. This survey is structured as follows. In the next section, we give a high-level overview of di- alogue systems. We brieï¬ y discuss the purpose and goal of dialogue systems. Then we describe the individual system components that are relevant for data-driven approaches as well as holistic end-to-end dialogue systems. In Section 3, we discuss types of dialogue interactions and aspects relevant to building data-driven dialogue systems, from a corpus perspective, as well as modalities recorded in each corpus (e.g. text, speech and video). We further discuss corpora constructed from both human-human and human-machine interactions, corpora constructed using natural versus un- natural or constrained settings, and corpora constructed using works of ï¬ | 1512.05742#2 | 1512.05742#4 | 1512.05742 | [
"1511.06931"
] |
1512.05742#4 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | ction. In Section 4, we present our survey over dialogue corpora according to the categories laid out in Sections 2-3. In particular, we categorize the corpora based on whether dialogues are between humans or between a human and a machine, and whether the dialogues are in written or spoken language. We discuss each corpus in turn while emphasizing how the dialogues were generated and collected, the topic of the dialogues, and the size of the entire corpus. In Section 5, we discuss issues related to: cor- pus size, transfer learning between corpora, incorporation of external knowledge into the dialogue system, data-driven learning for contextualization and personalization, and automatic evaluation metrics. We conclude the survey in Section 6. # 2. | 1512.05742#3 | 1512.05742#5 | 1512.05742 | [
"1511.06931"
] |
1512.05742#5 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Characteristics of Data-Driven Dialogue Systems This section offers a broad characterization of data-driven dialogue systems, which structures our presentation of the datasets. 2 # 2.1 An Overview of Dialogue Systems The standard architecture for dialogue systems, shown in Figure 1, incorporates a Speech Rec- ognizer, Language Interpreter, State Tracker, Response Generator, Natural Language Generator, and Speech Synthesizer. In the case of text-based (written) dialogues, the Speech Recognizer and Speech Synthesizer can be left out. While some of the literature on dialogue systems identiï¬ es only the State Tracker and Response Selection components as belonging inside the dialogue man- ager (Young, 2000), throughout this paper we adopt a broader view where language understanding and generation are incorporated within the dialogue system. This leaves space for the development and analysis of end-to-end dialogue systems (Ritter et al., 2011; Vinyals and Le, 2015; Lowe et al., 2015a; Sordoni et al., 2015b; Shang et al., 2015; Li et al., 2015; Serban et al., 2016; Serban et al., 2017b,a; Dodge et al., 2015; Williams and Zweig, 2016; Weston, 2016). We focus on corpus-based data-driven dialogue systems. That is, systems composed of machine learning solutions using corpora constructed from real-world data. These system components have variables or parameters that are optimized based on statistics observed in dialogue corpora. In particular, we focus on systems where the majority of variables and parameters are optimized. Such corpus-based data-driven systems should be contrasted to systems where each component is hand- crafted by engineers â for example, components deï¬ ned by an a priori ï¬ xed set of deterministic rules (e.g. Weizenbaum (1966); McGlashan et al. (1992)). These systems should also be contrasted with systems learning online, such as when the free variables and parameters are optimized directly based on interactions with humans (e.g. | 1512.05742#4 | 1512.05742#6 | 1512.05742 | [
"1511.06931"
] |
1512.05742#6 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | GaË si´c et al. (2011)). Still, it is worth noting that it is possible to combine different types of learning within one system. For example, some parameters may be learned using statistics observed in a corpus, while other parameters may be learned through interactions with humans. While there are substantial opportunities to improve each of the components in Figure 1 through (corpus-based) data-driven approaches, within this survey we focus primarily on datasets suitable to enhance the components inside the Dialogue System box. It is worth noting that the Natural Language Interpreter and Generator are core problems in Natural Language Processing with appli- cations well beyond dialogue systems. Automatic Speech Natural Language Dialogue State Recognizer Interpreter Tracker Text-To-Speech Natural Language Dialogue a P Response Synthesizer Generator fi Selection Dialogue System | 1512.05742#5 | 1512.05742#7 | 1512.05742 | [
"1511.06931"
] |
1512.05742#7 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Figure 1: Dialogue System Diagram 3 # 2.2 Tasks and objectives Dialogue systems have been built for a wide range of purposes. A useful distinction can be made between goal-driven dialogue systems, such as technical support services, and non-goal-driven dia- logue systems, such as language learning tools or computer game characters. Although both types of systems do in fact have objectives, typically the goal-driven dialogue systems have a well-deï¬ ned measure of performance that is explicitly related to task completion. Non-goal-driven Dialogue Systems. Research on non-goal-driven dialogue systems goes back to the mid-60s. It began, perhaps, with Weizenbaumâ s famous program ELIZA, a system based only on simple text parsing rules that managed to convincingly mimic a Rogerian psychotherapist by persistently rephrasing statements or asking questions (Weizenbaum, 1966). This line of research was continued by Colby (1981), who used simple text parsing rules to construct the dialogue system PARRY, which managed to mimic the pathological behaviour of a paranoid patient to the extent that clinicians could not distinguish it from real patients. However, neither of these two systems used data-driven learning approaches. Later work, such as the MegaHal system by Hutchens and Alder (1998), started to apply data-driven methods (Shawar and Atwell, 2007b). Hutchens and Alder (1998) proposed modelling dialogue as a stochastic sequence of discrete symbols (words) using 4â th order Markov chains. Given a user utterance, their system generated a response by following a two-step procedure: ï¬ rst, a sequence of topic keywords, used to create a seed reply, was ex- tracted from the userâ s utterance; second, starting from the seed reply, two separate Markov chains generated the words preceding and proceeding the seed keywords. This procedure produced many candidate responses, from which the highest entropy response was returned to the user. | 1512.05742#6 | 1512.05742#8 | 1512.05742 | [
"1511.06931"
] |
1512.05742#8 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Under the assumption that the coverage of different topics and general ï¬ uency is of primary importance, the 4â th order Markov chains were trained on a mixture of data sources ranging from real and ï¬ ctive dialogues to arbitrary texts. Unfortunately, until very recently, such data-driven dialogue systems were not applied widely in real-world applications (Perez-Marin and Pascual-Nieto, 2011; Shawar and Atwell, 2007b). Part of the reason for this might be due to their non-goal-driven nature, which made them hard to commercialize. Another barrier to commercialization might have been the lack of theoretical and empirical understanding of such systems. Nevertheless, in a similar spirit over the past few years, neural network architectures trained on large-scale corpora have been investigated. These models have demonstrated promising results for several non-goal-driven dialogue tasks (Rit- ter et al., 2011; Vinyals and Le, 2015; Lowe et al., 2015a; Sordoni et al., 2015b; Shang et al., 2015; Li et al., 2015; Serban et al., 2016; Serban et al., 2017b,a; Dodge et al., 2015; Williams and Zweig, 2016; Weston, 2016). | 1512.05742#7 | 1512.05742#9 | 1512.05742 | [
"1511.06931"
] |
1512.05742#9 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | However, they require having sufï¬ ciently large corpora â in the hundreds of millions or even billions of words â in order to achieve these results. Goal-driven Dialogue Systems. Initial work on goal-driven dialogue systems was primarily based on deterministic hand-crafted rules coupled with learned speech recognition models (e.g. off- the-shelf speech recognition software). One example is the SUNDIAL project, which was capable of providing timetable information about trains and airplanes, as well as taking airplane reserva- tions (Aust et al., 1995; McGlashan et al., 1992; Simpson and Eraser, 1993). Later, machine learn- ing techniques were used to classify the intention (or need) of the user, as well as to bridge the gap between text and speech (e.g. by taking into account uncertainty related to the outputs of the speech recognition model) (Gorin et al., 1997). Research in this area started to take off during the mid 1990s, when researchers began to formulate dialogue as a sequential decision making problem based on Markov decision processes (Singh et al., 1999; Young et al., 2013; Paek, 2006; Pieraccini | 1512.05742#8 | 1512.05742#10 | 1512.05742 | [
"1511.06931"
] |
1512.05742#10 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 4 et al., 2009). Unlike non-goal-driven systems, industry played a major role and enabled researchers to have access to (at the time) relatively large dialogue corpora for certain tasks, such as recordings from technical support call centres. Although research in the past decade has continued to push the ï¬ eld towards data-driven approaches, commercial systems are highly domain-speciï¬ c and heavily based on hand-crafted rules and features (Young et al., 2013). In particular, many of the tasks and datasets available are constrained to narrow domains. # 2.3 Learning Dialogue System Components Modern dialogue systems consist of several components, as illustrated in Figure 1. Several of the dialogue system components can be learned through so-called discriminative models, which aim to predict labels or annotations relevant to other parts of the dialogue system. Discriminative models fall into the machine learning paradigm of supervised learning. When the labels of interest are discrete, the models are called classiï¬ cation models, which is the most common case. When the labels of interest are continuous, the models are called regression models. One popular approach for tackling the discriminative task is to learn a probabilistic model of the labels conditioned on the available information P (Y |X), where Y is the label of interest (e.g. a discrete variable representing the user intent) and X is the available information (e.g. utterances in the conversation). Another popular approach is to use maximum margin classiï¬ ers, such as support vector machines (Cristianini and Shawe-Taylor, 2000). Although it is beyond the scope of this paper to provide a survey over such system components, we now give a brief example of each component. This will motivate and facilitate the dataset analysis. | 1512.05742#9 | 1512.05742#11 | 1512.05742 | [
"1511.06931"
] |
1512.05742#11 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Natural Language Interpreter. An example of a discriminative model is the user intent clas- siï¬ cation model, which acts as the Natural Language Interpreter. This model is trained to predict the intent of a user conditioned on the utterances of that user. In this case, the intent is called the label (or target or output), and the conditioned utterances are called the conditioning variables (or inputs). Training this model requires examples of pairs of user utterances and intentions. One way to obtain these example pairs would be to ï¬ rst record written dialogues between humans carrying out a task, and then to have humans annotate each utterance with its intention label. Depending on the complexity of the domain, this may require training the human annotators to reach a certain level of agreement between annotators. Dialogue State Tracker. A Dialogue State Tracker might similarly be implemented as a classi- ï¬ | 1512.05742#10 | 1512.05742#12 | 1512.05742 | [
"1511.06931"
] |
1512.05742#12 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | cation model (Williams et al., 2013). At any given point in the dialogue, such a model will take as input all the user utterances and user intention labels estimated by a Natural Language Interpreter model so far and output a distribution over possible dialogue states. One common way to represent dialogue states are through slot-value pairs. For example, a dialogue system providing timetable information for trains might have three different slots: departure city, arrival city, and departure time. Each slot may take one of several discrete values (e.g. departure city could take values from a list of city names). The task of the Dialogue State Tracker is then to output a distribution over every possible combination of slot-value pairs. | 1512.05742#11 | 1512.05742#13 | 1512.05742 | [
"1511.06931"
] |
1512.05742#13 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | This distribution â or alternatively, the K dialogue states with the highest probability â may then be used by other parts of the dialogue system. The Dialogue State Tracker model can be trained on examples of dialogue utterances and dialogue states labelled by humans. 5 Dialogue Response Selection. Given the dialogue state distribution provided by the Dialogue State Tracker, the Dialogue Response Selection component must select the correct system response (or action). This component may also be implemented as a classiï¬ cation model that maps dialogue states to a probability over a discrete set of responses. For example, in a dialogue system provid- ing timetable information for trains, the set of responses might include providing information (e.g. providing the departure time of the next train with a speciï¬ c departure and arrival city) and clariï¬ - cation questions (e.g. asking the user to re-state their departure city). | 1512.05742#12 | 1512.05742#14 | 1512.05742 | [
"1511.06931"
] |
1512.05742#14 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | The model may be trained on example pairs of dialogue states and responses. Natural Language Generator. Given a dialogue system response (e.g. a response providing the departure time of a train), the Natural Language Generator must output the natural language utterance of the system. This has often been implemented in commercial goal-driven dialogue systems using hand-crafted rules. Another option is to learn a discriminative model to select a natural language response. In this case, the output space may be deï¬ ned as a set of so-called surface form sentences (e.g. â | 1512.05742#13 | 1512.05742#15 | 1512.05742 | [
"1511.06931"
] |
1512.05742#15 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | The requested train leaves city X at time Yâ , where X and Y are placeholder values). Given the system response, the classiï¬ cation model must choose an appropriate surface form. Afterwards, the chosen surface form will have the placeholder values substituted in appropriately (e.g. X will be replaced by the appropriate city name through a database look up). As with other classiï¬ cation models, this model may be trained on example pairs of system responses and surface forms. Discriminative models have allowed goal-driven dialogue systems to make signiï¬ cant progress (Williams et al., 2013). With proper annotations, discriminative models can be evaluated automat- ically and accurately. Furthermore, once trained on a given dataset, these models may be plugged into a fully-deployed dialogue system (e.g. a classiï¬ cation model for user intents may be used as input to a dialogue state tracker). # 2.4 End-to-end Dialogue Systems Not all dialogue systems conform to the architecture shown in Figure 1. In particular, so-called end-to-end dialogue system architectures based on neural networks have shown promising results on several dialogue tasks (Ritter et al., 2011; Vinyals and Le, 2015; Lowe et al., 2015a; Sordoni et al., 2015b; Shang et al., 2015; Li et al., 2015; Serban et al., 2016; Serban et al., 2017b,a; Dodge et al., 2015). In their purest form, these models take as input a dialogue in text form and output a response (or a distribution over responses). We call these systems end-to-end dialogue systems because they possess two important properties. First, they do not contain or require learning any sub-components (such as Natural Language Interpreters or Dialogue State Trackers). Consequently, there is no need to collect intermediate labels (e.g. user intention or dialogue state labels). Second, all model parameters are optimized w.r.t. a single objective function. Often the objective function chosen is maximum log-likelihood (or cross-entropy) on a ï¬ | 1512.05742#14 | 1512.05742#16 | 1512.05742 | [
"1511.06931"
] |
1512.05742#16 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | xed corpus of dialogues. Although in the original formulation these models depended only on the dialogue context, they may be extended to also depend on outputs from other components (e.g. outputs from the speech recognition tracker), and on external knowledge (e.g. external databases). End-to-end dialogue systems can be divided into two categories: those that select deterministi- cally from a ï¬ xed set of possible responses, and those that attempt to generate responses by keeping a posterior distribution over possible utterances. Systems in the ï¬ rst category map the dialogue his- tory, tracker outputs and external knowledge (e.g. a database, which can be queried by the system) | 1512.05742#15 | 1512.05742#17 | 1512.05742 | [
"1511.06931"
] |
1512.05742#17 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 6 # to a response action: fθ : {dialogue history, tracker outputs, external knowledge} â action at, where at is the dialogue system response action at time t, and θ is the set of parameters that deï¬ nes f . Information retrieval and ranking-based systems â systems that search through a database of dialogues and pick responses with the most similar context, such as the model proposed by Banchs and Li (2012) â belong to this category. In this case, the mapping function fθ projects the dialogue history into a Euclidean space (e.g. using TF-IDF bag-of-words representations). The response is then found by projecting all potential responses into the same Euclidean space, and the response closest to the desirable response region is selected. The neural network proposed by Lowe et al. (2015a) also belongs to this category. In this case, the dialogue history is projected into a Euclidean space using a recurrent neural network encoding the dialogue word-by-word. Similarly, a set of can- didate responses are mapped into the same Euclidean space using another recurrent neural network encoding the response word-by-word. Finally, a relevance score is computed between the dialogue context and each candidate response, and the response with the highest score is returned. Hybrid or combined models, such as the model built on both a phrase-based statistical machine translation system and a recurrent neural network proposed by Sordoni et al. (2015b), also belong to this cate- gory. In this case, a response is generated by deterministically creating a ï¬ xed number of answers using the machine translation system and then picking the response according to the score given by a a neural network. Although both of its sub-components are based on probabilistic models, the ï¬ nal model does not construct a probability distribution over all possible responses.1 In contrast to a deterministic system, a generative system explicitly computes a full posterior probability distribution over possible system response actions at every turn: # Pθ(action at | dialogue history, tracker outputs, external knowledge). Systems based on generative recurrent neural networks belong to this category (Vinyals and Le, 2015). By breaking down eq. (2) into a product of probabilities over words, responses can be generated by sampling word-by-word from their probability distribution. | 1512.05742#16 | 1512.05742#18 | 1512.05742 | [
"1511.06931"
] |
1512.05742#18 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Unlike the deterministic response models, these systems are also able to generate entirely novel responses (e.g. by sampling word-by-word). Highly probable responses, i.e. the response with the highest probability, can fur- ther be generated by using a method known as beam-search (Graves, 2012). These systems project each word into a Euclidean space (known as a word embedding) (Bengio et al., 2003); they also project the dialogue history and external knowledge into a Euclidean space (Wen et al., 2015; Lowe et al., 2015b). Similarly, the system proposed by Ritter et al. (2011) belongs to this category. Their model uses a statistical machine translation model to map a dialogue history to its response. When trained solely on text, these generative models can be viewed as unsupervised learning models, because they aim to reproduce data distributions. In other words, the models learn to assign a prob- ability to every possible conversation, and since they generate responses word by word, they must learn to simulate the behaviour of the agents in the training corpus. Early reinforcement learning dialogue systems with stochastic policies also belong to this cat- egory (the NJFun system (Singh et al., 2002) is an example of this). In contrast to the neural network and statistical machine translation systems, these reinforcement learning systems typically | 1512.05742#17 | 1512.05742#19 | 1512.05742 | [
"1511.06931"
] |
1512.05742#19 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 1. Although the model does not require intermediate labels, it consists of sub-components whose parameters are trained with different objective functions. Therefore, strictly speaking, this is not an end-to-end model. 7 (1) (2) have very small sets of possible hand-crafted system states (e.g. hand-crafted features describing the dialogue state). The action space is also limited to a small set of pre-deï¬ ned responses. This makes it possible to apply established reinforcement learning algorithms to train them either online or ofï¬ ine, however it also severely limits their application area. As Singh et al. (Singh et al., 2002, p.5) remark: â We view the design of an appropriate state space as application-dependent, and a task for a skilled system designer.â # 3. Dialogue Interaction Types & Aspects This section provides a high-level discussion of different types of dialogue interactions and their salient aspects. The categorization of dialogues is useful for understanding the utility of various datasets for particular applications, as well as for grouping these datasets together to demonstrate available corpora in a given area. # 3.1 Written, Spoken & Multi-modal Corpora An important distinction between dialogue corpora is whether participants (interlocutors) interact through written language, spoken language, or in a multi-modal setting (e.g. using both speech and visual modalities). Written and spoken language differ substantially w.r.t. their linguistic properties. .Spoken language tends to be less formal, containing lower information content and many more pronouns than written language (Carter and McCarthy, 2006; Biber and Finegan, 2001, 1986). | 1512.05742#18 | 1512.05742#20 | 1512.05742 | [
"1511.06931"
] |
1512.05742#20 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | In particular, the differences are magniï¬ ed when written language is compared to spoken face-to- face conversations, which are multi-modal and highly socially situated. As Biber and Finegan (1986) observed, pronouns, questions, and contradictions, as well as that-clauses and if-clauses, appear with a high frequency in face-to-face conversations. Forchini (2012) summarized these differences: â ... studies show that face-to-face conversation is interpersonal, situation-dependent has no narrative concern or as Biber and Finegan (1986) put it, is a highly interactive, situated and immediate text type...â Due to these differences between spoken and written language, we will emphasize the distinction between dialogue corpora in written and spoken language in the following sections. Similarly, dialogues involving visual and other modalities differ from dialogues without these modalities (Card et al., 1983; Goodwin, 1981). When a visual modality is available â for example, when two human interlucators converse face-to-face â body language and eye gaze has a signiï¬ cant impact on what is said and how it is said (Gibson and Pick, 1963; Lord and Haith, 1974; Cooper, 1974; Chartrand and Bargh, 1999; de Kok et al., 2013). Aside from the visual modality, dialogue systems may also incorporate other situational modalities, including aspects of virtual environments (Rickel and Johnson, 1999; Traum and Rickel, 2002) and user proï¬ les (Li et al., 2016). # 3.2 Human-Human Vs. Human-Machine Corpora Another important distinction between dialogue datasets resides in the types of interlocutors â notably, whether it involves interactions between two humans, or between a human and a computer2. The distinction is important because current artiï¬ cial dialogue systems are signiï¬ cantly constrained. 2. Machine-machine dialogue corpora are not of interest to us, because they typically differ signiï¬ cantly from natural human language. Furthermore, user simulation models are outside the scope of this survey. | 1512.05742#19 | 1512.05742#21 | 1512.05742 | [
"1511.06931"
] |
1512.05742#21 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 8 These systems do not produce nearly the same distribution of possible responses as humans do under equivalent circumstances. As stated by Williams and Young (2007): (Human-human conversation) does not contain the same distribution of understanding errors, and humanâ human turn-taking is much richer than human-machine dialog. As a result, human-machine dialogue exhibits very different traits than human-human dia- logue (Doran et al., 2001; Moore and Browning, 1992). The expectation a human interlucator begins with, and the interface through which they interact, also affect the nature of the conversation (J. and D., 1988). For goal-driven settings, Williams and Young (2007) have previously argued against building data-driven dialogue systems using human-human dialogues: â ... using human-human conversation data is not appropriate because it does not contain the same distribution of understanding errors, and because human-human turn-taking is much richer than human-machine dialog.â This line of reasoning seems particularly applicable to spoken dialogue systems, where speech recognition errors can have a critical impact on performance and therefore must be taken into account when learning the dialogue model. The argument is also relevant to goal-driven dialogue systems, where an effective dialogue model can often be learned using reinforcement learning techniques. Williams and Young (2007) also argue against learning from corpora generated between humans and existing dialogue systems: â While it would be possible to use a corpus collected from an existing spoken dialogue system, supervised learning would simply learn to approximate the policy used by that spoken dialogue system and an overall performance improvement would therefore be unlikely.â Thus, it appears, for goal-driven spoken dialogue systems in particular, that the most effective strategy is learning online through interaction with real users. Nonetheless, there exists useful human-machine corpora where the interacting machine uses a stochastic policy that can generate sufï¬ cient coverage of the task (e.g. enough good and enough bad dialogue examples) to allow an effective dialogue model to be learned. In this case, the goal is to learn a policy that is eventually better than the original stochastic policy used to generate the corpus through a process known as bootstrapping. In this survey we focus on data-driven learning from human-human and human-machine di- alogue corpora. Despite the advantages of learning online through interactions with real users, learning based on human-human dialogue corpora may be more suitable for open domain dialogue systems because they reï¬ ect natural dialogue interactions. | 1512.05742#20 | 1512.05742#22 | 1512.05742 | [
"1511.06931"
] |
1512.05742#22 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | By natural dialogues, we mean conver- sations that are unconstrained and unscripted, e.g. between interlocutors who are not instructed to carry out a particular task, to follow a series of instructions, or to act out a scripted dialogue. In this setting, the dialogue process is relatively unaffected by researchers, e.g. the interlocutors are not in- terrupted by question prompts in the middle of a dialogue. As can be expected, such conversations include a signiï¬ cant amount of turn-taking, pauses and common grounding phenomena (Clark and Brennan, 1991). Additionally, they are more diverse, and open up the possibility for the model to learn to understand natural language. # 3.3 Natural Vs. Unnatural Corpora The way in which a dialogue corpus is generated and collected can have a signiï¬ cant inï¬ uence on the trained data-driven dialogue system. In the case of human-human dialogues, an ideal corpus should closely resemble natural dialogues between humans. Arguably, this is the case when conversations | 1512.05742#21 | 1512.05742#23 | 1512.05742 | [
"1511.06931"
] |
1512.05742#23 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 9 between humans are recorded and transcribed, and when the humans in the dialogue represent the true population of users with whom the dialogue system is intended to interact. It is even better if they are unaware of the fact that they are being recorded, but this is not always possible due to ethical considerations and resource constraints. Due to ethical considerations and resource constraints, researchers may be forced to inform the human interlocutors that they are being recorded or to setup artiï¬ cial experiments in which they hire humans and instruct them to carry out a particular task by interacting with a dialogue system. In these cases, there is no guarantee that the interactions in the corpus will reï¬ ect true interactions, since the hired humans may behave differently from the true user population. One factor that may cause behavioural differences is the fact that the hired humans may not share the same intentions and motivations as the true user population (Young et al., 2013). The unnaturalness may be further exacerbated by the hiring process, as well as the platform through which they interact. Such factors are becoming more prevalent as researchers increasingly rely on crowdsourcing platforms, such as Amazon Mechanical Turk, to collect and evaluate dialogue data (Jurcıcek et al., 2011). In the case of Wizard-of-Oz experiments (Bohus and Rudnicky, 2008; Petrik, 2004), a human thinks (s)he is speaking to a machine, but a human operator is in fact controlling the dialogue system. This enables the generation of datasets that are closer in nature to the dialogues humans may wish to have with a good AI dialogue system. Unfortunately, such experiments are expensive and time- consuming to carry out. Ultimately the impact of any unnaturalness in the dialogues depends on the task and context in which the dialogue system is deployed. # 3.4 Corpora from Fiction It is also possible to use artiï¬ cial dialogue corpora for data-driven learning. This includes cor- pora based on works of ï¬ ction such as novels, movie manuscripts and audio subtitles. However, unlike transcribed human-human conversations, novels, movie manuscripts, and audio subtitles de- pend upon events outside the current conversation, which are not observed. | 1512.05742#22 | 1512.05742#24 | 1512.05742 | [
"1511.06931"
] |
1512.05742#24 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | This makes data-driven learning more difï¬ cult because the dialogue system has to account for unknown factors. The same problem is also observed in certain other media, such as microblogging websites (e.g. Twitter and Weibo), where conversations also may depend on external unobserved events. Nevertheless, recent studies have found that spoken language in movies resembles spontaneous human spoken language (Forchini, 2009). Although movie dialogues are explicitly written to be spoken and contain certain artiï¬ cial elements, many of the linguistic and paralinguistic features contained within the dialogues are similar to natural spoken language, including dialogue acts such as turn-taking and reciprocity (e.g. returning a greeting when greeted). The artiï¬ cial differences that exist may even be helpful for data-driven dialogue learning since movie dialogues are more compact, follow a steady rhythm, and contain less garbling and repetition, all while still presenting a clear event or message to the viewer (Dose, 2013; Forchini, 2009, 2012). Unlike dialogues extracted from Wizard-of-Oz human experiments, movie dialogues span many different topics and occur in many different environments (Webb, 2010). They contain different actors with different intentions and relationships to one another, which could potentially allow a data-driven dialogue system to learn to personalize itself to different users by making use of different interaction patterns (Li et al., 2016). | 1512.05742#23 | 1512.05742#25 | 1512.05742 | [
"1511.06931"
] |
1512.05742#25 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 10 # 3.5 Corpus Size As in other machine learning applications such as machine translation (Al-Onaizan et al., 2000; G¨ulc¸ehre et al., 2015) and speech recognition (Deng and Li, 2013; Bengio et al., 2014), the size of the dialogue corpus is important for building an effective data-driven dialogue (Lowe et al., 2015a; Serban et al., 2016). There are two primary perspectives on the importance of dataset size for building data-driven dialogue systems. | 1512.05742#24 | 1512.05742#26 | 1512.05742 | [
"1511.06931"
] |
1512.05742#26 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | The ï¬ rst perspective comes from the machine learning literature: larger datasets place constraints on the dialogue model trained from that data. Datasets with few examples may require strong structural priors placed on the model, such as using a modular system, while large datasets can be used to train end-to-end dialogue systems with less a priori structure. The second comes from a statistical natural language processing perspective: since the statistical complexity of a corpus grows with the linguistic diversity and number of topics, the number of examples required by a machine learning algorithm to model the patterns in it will also grow with the linguistic diversity and number of topics. Consider two small datasets with the same number of dialogues in the domain of bus schedule information: in one dataset the conversations between the users and operator is natural, and the operator can improvise and chitchat; in the other dataset, the operator reads from a script to provide the bus information. Despite having the same size, the second dataset will have less linguistic diversity and not include chitchat topics. Therefore, it will be easier to train a data-driven dialogue system mimcking the behaviour of the operator in the second dataset, however it will also exhibit a highly pedantic style and not be able to chitchat. In addition to this, to have an effective discussion between any two agents, their common knowledge must be represented and understood by both parties. The process of establishing this common knowledge, also known as grounding, is especially critical to repair misunderstandings between humans and dialogue systems (Cahn and Brennan, 1999). Since the number of misunderstandings can grow with the lexical diversity and number of topics (e.g. misunderstanding the paraphrase of an existing word, or misunderstanding a rarely seen keyword), the number of examples required to repair these grow with linguistic diversity and topics. In particular, the effect of linguistic diversity has been observed in practice: Vinyals and Le (2015) train a simple encoder-decoder neural network on a proprietary dataset of technical support dialogues. Although it has a similar size and purpose as the Ubuntu Dialogue Corpus (Lowe et al., 2015a), the qualitative examples shown by Vinyals and Le (2015) are signiï¬ cantly superior to those obtained by more complex models on the Ubuntu Corpus (Serban et al., 2017a). This result may likely be explained in part due to the fact that technical support operators often follow a comprehensive script for solving problems. | 1512.05742#25 | 1512.05742#27 | 1512.05742 | [
"1511.06931"
] |
1512.05742#27 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | As such, the script would reduce the linguistic diversity of their responses. Furthermore, since the majority of human-human dialogues are multi-modal and highly am- biguous in nature (Chartrand and Bargh, 1999; de Kok et al., 2013), the size of the corpus may compensate for some of the ambiguity and missing modalities. If the corpus is sufï¬ ciently large, then the resolved ambiguities and missing modalities may, for example, be approximated using latent stochastic variables (Serban et al., 2017b). Thus, we include corpus size as a dimension of analysis. We also discuss the beneï¬ ts and drawbacks of several popular large-scale datasets in Section 5.1. | 1512.05742#26 | 1512.05742#28 | 1512.05742 | [
"1511.06931"
] |
1512.05742#28 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 11 # 4. Available Dialogue Datasets There is a vast amount of data available documenting human communication. Much of this data could be used â perhaps after some pre-processing â to train a dialogue system. However, cov- ering all such sources of data would be infeasible. Thus, we restrict the scope of this survey to datasets that have already been used to study dialogue or build dialogue systems, and to very large corpora of interactionsâ that may or may not be strictly considered dialogue datasetsâ which could be leveraged in the near future to build more sophisticated data-driven dialogue models. We restrict the selection further to contain only corpora generated from spoken or written English, and to cor- pora which, to the best of our knowledge, either are publicly available or will be made available in the near future. | 1512.05742#27 | 1512.05742#29 | 1512.05742 | [
"1511.06931"
] |
1512.05742#29 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | We ï¬ rst give a brief overview of each of the considered corpora, and later high- light some of the more promising examples, explaining how they could be used to further dialogue research.3 The dialogue datasets analyzed in this paper are listed in Tables 1-5. Column features indicate properties of the datasets, including the number of dialogues, average dialogue length, number of words, whether the interactions are between humans or with an automated system, and whether the dialogues are written or spoken. Below, we discuss qualitative features of the datasets, while statistics can be found in the aforementioned table. # 4.1 Human-Machine Corpora As discussed in Subsection 3.2, an important distinction between dialogue datasets is whether they consist of dialogues between two humans or between a human and a machine. Thus, we begin by outlining some of the existing human-machine corpora in several categories based on the types of systems the humans interact with: Restaurant and Travel Information, Open-Domain Knowledge Retrieval, and Other Specialized systems. Note, we also include human-human corpora here where one human plays the role of the machine in a Wizard-of-Oz fashion. # 4.1.1 RESTAURANT AND TRAVEL INFORMATION One common theme in human-machine language datasets is interaction with systems which provide restaurant or travel information. | 1512.05742#28 | 1512.05742#30 | 1512.05742 | [
"1511.06931"
] |
1512.05742#30 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Here weâ ll brieï¬ y describe some human-machine dialogue datasets in this domain. One of the most popular recent sources of such data has come from the datasets for structured dialogue prediction released in conjunction with the Dialog State Tracking Challenge (DSTC) (Williams et al., 2013). As the name implies, these datasets are used to learn a strategy for the Di- alogue State Tracker (sometimes called â belief trackingâ ), which involves estimating the intentions of a user throughout a dialog. State tracking is useful as it can increase the robustness of speech recognition systems, and can provide an implementable framework for real-world dialogue sys- tems. Particularly in the context of goal-oriented dialogue systems (such as those providing travel and restaurant information), state tracking is necessary for creating coherent conversational inter- faces. | 1512.05742#29 | 1512.05742#31 | 1512.05742 | [
"1511.06931"
] |
1512.05742#31 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | As such, the ï¬ rst three datasets in the DSTCâ referred to as DSTC1, DSTC2, and DSTC3 respectivelyâ are medium-sized spoken datasets obtained from human-machine interactions with 3. We form a live list of the corpora discussed in this work, along with links to downloads, at: http://breakend. github.io/DialogDatasets. Pull requests can be made to the Github repository (https://github. com/Breakend/DialogDatasets) hosting the website for continuing updates to the list of corpora. | 1512.05742#30 | 1512.05742#32 | 1512.05742 | [
"1511.06931"
] |
1512.05742#32 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 12 restaurant and travel information systems. All datasets provide labels specifying the current goal and desired action of the system. DSTC1 (Williams et al., 2013) features conversations with an automated bus information in- terface, where users request bus routes from the system and the system responds with clarifying queries or the desired information. DSTC2 introduces changing user goals in a restaurant booking system, while trying to provide a desired reservation(Henderson et al., 2014b). DSTC3 introduces a small amount of labelled data in the domain of tourist information. It is intended to be used in conjunction with the DSTC2 dataset as a domain adaptation problem (Henderson et al., 2014a). The Carnegie Mellon Communicator Corpus (Bennett and Rudnicky, 2002) also contains human-machine interactions with a travel booking system. It is a medium-sized dataset of interac- tions with a system providing up-to-the-minute ï¬ ight information, hotel information, and car rentals. Conversations with the system were transcribed, along with the userâ s comments at the end of the interaction. The ATIS (Air Travel Information System) Pilot Corpus (Hemphill et al., 1990) is one of the ï¬ rst human-machine corpora. It consists of interactions, lasting about 40 minutes each, between human participants and a travel-type booking system, secretly operated by humans. Unlike the Carnegie Mellon Communicator Corpus, it only contains 1041 utterances. In the Maluuba Frames Corpus (El Asri et al., 2017), one user plays the role of a conversa- tional agent in a Wizard-of-Oz fashion, while the other user is tasked with ï¬ | 1512.05742#31 | 1512.05742#33 | 1512.05742 | [
"1511.06931"
] |
1512.05742#33 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | nding available travel or vacation accommodations according to a pre-speciï¬ ed task. The Wizard is provided with a knowl- edge database which recorded their actions. Semantic frames are annotated in addition to actions which the Wizard performed on the database to accompany a line of dialogue. In this way, the Frames corpus aims to track decision-making processes in travel- and hotel-booking through natu- ral dialog. # 4.1.2 OPEN-DOMAIN KNOWLEDGE RETRIEVAL Knowledge retrieval and Question & Answer (QA) corpora are a broad distinction of corpora that we will not extensively review here. Instead, we include only those QA corpora which explicitly record interactions of humans with existing systems. The Ritel corpus (Rosset and Petel, 2006) is a small dataset of 528 dialogs with the Wizard-of-Oz Ritel platform. | 1512.05742#32 | 1512.05742#34 | 1512.05742 | [
"1511.06931"
] |
1512.05742#34 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | The projectâ s purpose was to integrate spoken language dialogue systems with open-domain information retrieval systems, with the end goal of allowing humans to ask general questions and iteratively reï¬ ne their search. The questions in the corpus mostly revolve around politics and the economy, such as â Who is currently presiding the Senate?â , along with some conversations about arts and science-related topics. Other similar open-domain corpora in this area include WikiQA Yang et al. (2015) and MS MARCO Nguyen et al. (2016), which compile responses from automated Bing searches and hu- man annotators. However, these do not record dialogs, but rather simply gather possible responses to queries. | 1512.05742#33 | 1512.05742#35 | 1512.05742 | [
"1511.06931"
] |
1512.05742#35 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | As such, we wonâ t discuss these datasets further, but rather mention them brieï¬ y as examples of other Open-Domain corpora in the ï¬ eld. 13 n o i t p i r c s e D # l a t o T # l a t o T # . g v A s c i p o T e p y T s d r o w f o s e u g o l a i d f o s n r u t f o m e t s y s n o i t a m r o f n i e d i r s u B M 7 3 . 0 0 0 , 5 1 6 5 . 3 1 s e l u d e h c s s u B n e k o p S ) 3 1 0 2 , . l a t e m e t s y s g n i k o o b t n a r u a t s e R K 2 3 4 0 0 0 , 3 8 8 . 7 s t n a r u a t s e R n e k o p S ) b 4 1 0 2 , . l a s t s i r u o t r o f n o i t a m r o f n I K 3 0 4 5 6 2 , 2 7 2 . 8 n o i t a m r o f n i t s i r u o T n e k o p S ) a 4 1 0 2 , . l a m e t s y s g n i k o o b d n a g n i n n a l p l e v a r T * M 2 1 8 4 , 5 1 7 6 . 1 1 l e v a r T n e k o p S s u p r o C ) 2 0 0 2 m e t s y s g n i k o o b d n a g n i n n a l p l e v a r T * K 4 1 1 . 1 4 4 . 5 2 l e v a r T n e k o p S n o i t s e u q n i a m o d - n e p o d e t a t o n n a n A k 0 6 2 8 5 * 3 . 9 s c i p o T e s r e v i D / d e t c i r t s e r n U n e k o p S m e t s y s e u g o l a i d m e t s y s r e t u p m o c h t i n e k o p s g n i r e w s n a w t c a r e t n i s n a m u H * K 7 8 . 6 6 2 1 s c i t a m e h t a M n e k o p S ) 6 0 0 2 g n i v o r p m e r o e h t l a c i t a m e h t a m o d o t ) 4 0 0 2 , . l a t e . s t n e m t n i o p p a g n i l u d e h c s r o f m e t s y s A * K 9 6 7 4 4 0 . 4 1 g n i l u d e h c S t n e m t n i o p p A n e k o p S s n o i t a t o n n a t c a e u g o l a i d s e d u l c n I ) 0 1 0 2 . s m e t s y s e u g o l a i d n e v i r d - l a o g r o F â | 1512.05742#34 | 1512.05742#36 | 1512.05742 | [
"1511.06931"
] |
1512.05742#36 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 9 6 3 1 5 1 n o i t a c a V & l e v a r T & A Q , t a h C s n o i t c a d n a d e l e b a l s e m a r f c i t n a m e S g n i k o o B n o i t a d n e m m o c e R . d e t a t o n n a e s a b - e g d e l w o n k a n o n e k a t . e c n a r e t t u r e p s d r o w f o r e b m u n e g a r e v a e h t n o d e s a b d e t a m i x o r p p a e r a s r e b m u n ) * ( d e r r a t S . s t e s a t a d e u g o l a i d . n a m u h a y b d e t a r e p o y l t e r c e s s i e n i h c a m e h t e r e h w , s e u g o l a i d z O - f o - d r a z i W e t a c i d n i ) â ( h t i e m a N s m a i l l i SWITI) # W ( # 1 C T S D # t e n o s r e d n e H ( # 2 C T S D # t e n o s r e d n e H ( # 3 C T S D r o t a c i n u m m o C U M C , y k c i n d u R d n a # t t e n n e B ( # â s u p r o C # t o l i P S I T A # t e l l i h p m e H ( # â s u p r o C # l e t i # R , l e t e P d n a # t e s s o R ( l a c i t a m e h t a # M G O L A D # I # a k s l o W ( # s f o o r P # â s u p r o C H C T A M | 1512.05742#35 | 1512.05742#37 | 1512.05742 | [
"1511.06931"
] |
1512.05742#37 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | # , . l a # t e a l i g r o e G ( â s e m a r F a b u u l a # M ) 7 1 0 2 , . l a # t e i r s A # l E ( e n i h c a m - n a m u H : 1 e l b a T 14 w d e k r a m s t e s a t a D n o i t p i r c s e D l a t o T # l a t o T # l a t o T s c i p o T h t g n e l s d r o w f o s e u g o l a i d f o y l l a b r e v e t a r o b a l l o c t s u m s r e k a e p s h c i h w n i k s a T P A L H m o r f s e u g o l a i D s r h 8 1 k 7 4 1 8 2 1 g n i c u d o r p e R - p a M . s r e h t o e h t n o d e t n i r p e t u o r a p a m s t n a p i c i t r a p e n o n o e c u d o r p e r o t k s a T . s n o i t a c o l n i a t r e c d n ï¬ | 1512.05742#36 | 1512.05742#38 | 1512.05742 | [
"1511.06931"
] |
1512.05742#38 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | o t e n o h p e l e t r e v o g n i t a r o b a l l o c e l p o e P s r h 3 3 * k 0 0 3 6 3 n o i t a c o L k s a T g n i d n i F e c n i v n o c o t s e i r t s g n i l e e f n e e r g - o r p g n o r t s ) y l e n i u n e g ( h t i w r e d a u s r e p A s r h 4 * k 5 3 8 e l y t s e f i L . s e l y t s e f i l n e e r g e r o m g n i t p o d a r e d i s n o c o t s e e d a u s r e p ) 7 0 0 2 , . l a d e n i a r t s n o c h c a e , s e t a b e d e l y t s - d r o f x O n i s c i p o t s u o i r a V * s r h 0 0 2 M 8 . 1 8 0 1 s e t a b e D . s e t a b e d - t s o p d n a - e r p d e d i v o r p s n o i n i p o e c n e i d u A . t c e j b u s e n o o t e s u o H e t i h W d n a s g n i t e e m y t l u c a f m o r f s n o i t c a r e t n I * s r h 0 2 2 M 2 0 0 2 n o i t a c u d E , s c i t i l o P n e k o p S . s e c n e r e f n o c s s e r p ) 0 0 0 2 , w o l r a B a d n a , c i p o t l a c i t i l o p a n o n o i s s u c s i d a : s t n e m i r e p x e o w T s r h 1 1 * k 0 0 1 4 5 s e m a G , s c i t i l o P . e m a g g n i y a l p - e l o r s n o i t a t o n n a h t i w e m a g g n i y a l p - e l o r f l o w e r e W f o g n i d r o c e r A s r h 7 * k 0 6 5 1 e m a G g n i y a l P - e l o R s u p r o C . s s e r g o r p e m a g o t d e t a l e r ) 0 1 0 2 r o t a r e p o n a h t i w s n o i t a s r e v n o c g n i d l o h e l i h w d e d r o c e r e r e w s r e s U s r h 0 5 * k 0 5 4 0 0 1 l a n o i t o m E . s n o i t c a e r l a n o i t o m e e k o v e o t d e n g i s e d s e l o r s t p o d a o h w s n o i t a s r e v n o C . e p y k S r e v o e g n a h c x e n o i t a m r o f n i t s i r u o T s r h 1 2 k 3 7 2 5 3 t s i r u o T ) 6 1 0 2 . s n o r t a p d n a s n a i r a r b i l n e e w t e b s n o i t c a r e t n i e n o h p e l e T * 0 4 1 K 1 2 2 8 s e i r i u q n I y r a r b i L , ) s t i n u e s r u o c s i d ( s e m a r f , s c i p o t n o i s s u c s i d , s t c a e u g o l a i d d e t a t o n n A ) 4 1 0 2 . s r i a p r e w s n a - n o i t s e u q , f l e s t i t c e j o r p s u p r o c e h t : e d u l c n i s c i p o T . s g n i t e e m I S C I f o s g n i d r o c e R s r h 2 7 * K 1 1 5 7 s g n i t e e M I S C I f o s e i r o e h t d n a g n i s s e c o r p e g a u g n a l l a r u t a n , n o i t i n g o c e r h c e e p s c i t a m o t u a . s t o p s t o h d n a , s r i a p r e w s n a - n o i t s e u q , s t c a e u g o l a i D . e g a u g n a l . s e t u o r t h g i e r f d a o r l i a r f o g n i n n a l p e v i t a r o b a l l o C s r h 5 . 6 K 5 5 8 9 t h g i e r F d a o r l i a R g n i n n a l P e t u o R ) 5 9 9 1 . t c e j o r p l i b o m b r e V e h t r o f d e t c e l l o c a t a d h c e e p s s u o e n a t n o p S s r H 8 3 K 0 7 2 6 2 7 t n e m t n i o p p A . e s e n a p a J d n a , n a m r e G , h s i l g n E n i s i s u p r o c l l u F g n i l u d e h c S . s c i t s i t a t s h s i l g n E w o h s y l n o e W h c e e p s h s i l g n E f o e t a r e g a r e v a e h t n o d e s a b s e t a m i t s e e r a s r e b m u n ) * ( d e r r a t S . ) l m t h . y t i l a u q / l a i r o t u t / d o r p e c i o v / s l a i r o t u t / s v c n / g r o . s v c n . w w w ( h c e e p S d n a e c i o V r o f r e t n e C | 1512.05742#37 | 1512.05742#39 | 1512.05742 | [
"1511.06931"
] |
1512.05742#39 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | e m a N s u p r o C k s a T p a M C R C H # , . l a t e n o s r e d n A ( s u p r o C d n u o r A g n i k l a # W # e h T ) 3 1 0 2 , . l a # t e n a n n e r B ( e s a b a t a D e v i s a u s r e P n e e r G e i w o C - s a l g u o D ( s e t a b e D d e r a u q S e c n e g i l l e t n I ) 6 1 0 2 , . l a # t e g n a h Z ( l a n o i s s e f o r P f o # s u p r o C e h T ( h s i l g n E n a c i r e m A e s a b a t a D y r c i m M B O N H A M # i ) 1 1 0 2 # , . l a # t e # n u S ( # f l o W P A D # I # I # e h T , n a j n a r a t t i h C d n a # g n u H ( # s u p r o c E N A M E S # I ) 0 1 0 2 , . l a t e n w o e K c M ( a r o p r o C 5 C T S D / 4 C T S D , 5 1 0 2 # , . l a # t e m K # i ( s u p r o C e u g o l a i D i u q o L , r a h c a S d n a u a e n n o s s a P ( s u p r o C A D R M ) 4 0 0 2 # , . l a # t e g r e b i r h S ( s u p r o C s e u g o l a i D 3 9 S N A R T # I # , n e l l # â UaTTY A d n a # n a m e e H ( | 1512.05742#38 | 1512.05742#40 | 1512.05742 | [
"1511.06931"
] |
1512.05742#40 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | # s u p r o C l i b o m b r e V ) 0 0 0 2 , . l a # t e # r e g r u B ( 15 . s t e s a t a d e u g o l a i d n e k o p s d e n i a r t s n o c n a m u h - n a m u H : 2 e l b a T m o r f n o i t p i r c s e D l a t o T # l a t o T # l a t o T s c i p o T h t g n e l s d r o w f o s e u g o l a i d f o s c i p o t d e ï¬ | 1512.05742#39 | 1512.05742#41 | 1512.05742 | [
"1511.06931"
] |
1512.05742#41 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | i c e p s - e r p n o s n o i t a s r e v n o c e n o h p e l e T * s r h 0 0 3 M 3 0 0 4 , 2 s c i p o T l a u s a C ) 2 9 9 1 , . l a t e y e r f d o G ( r o s s e n i s u b l a m r o f m o r f , s t x e t n o c y n a m s e u g o l a i d h s i t i r B * s r h 0 0 0 , 1 M 0 1 4 5 8 s c i p o T l a u s a C ) C N B ( s u p r o C . s n i - e n o h p d n a s w o h s o i d a r o t s g n i t e e m . s d n e i r f e s o l c r o s r e b m e m y l i m a f n e e w t e b s n o i t a s r e v n o c e n o h p e l e T s r h 0 6 * k 0 4 5 0 2 1 s c i p o T l a u s a C ) 7 9 9 1 , . l a t e n r e h t u o S a h t i w s n a c i r e m A n e e w t e b s n o i t a s r e v n o c e n o h p e l e T s r h 0 2 * k 0 8 1 0 6 s c i p o T l a u s a C ) 6 9 9 1 . 3 9 9 1 n i d e d r o c e r k l a t e g a n e e t s u o e n a t n o p S s r h 5 5 k 0 0 5 0 0 1 d e t c i r t s e r n U n o d n o L f o . y l t e r c e s d e d r o c e r e r e w s n o i t a s r e v n o C ) 5 9 9 1 , s a h c u s , s t x e t n o c l a m r o f n i f o y t e i r a v e d i w m o r f s e u g o l a i d h s i t i r B * s r h 0 5 5 M 5 â | 1512.05742#40 | 1512.05742#42 | 1512.05742 | [
"1511.06931"
] |
1512.05742#42 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | s c i p o T l a u s a C m a h g n i t t o N d n a . c t e , s t n a r u a t s e r , s n o l a s r i a h h s i l g n E n i ) 8 9 9 1 e l p o e p f o p u o r g a n e e w t e b n o i t c a r e t n i l a r u t a n f o s r u o h l a r e v e S s r h 8 * k 0 7 2 d e t c i r t s e r n U s u p r o C n o i t a s r e v n o C ) 3 1 0 2 . s g n i d r o c e r g n i t e e m e c a f - o t - e c a F s r h 0 0 1 * k 0 0 9 5 7 1 s g n i t e e M , s n o i t a s r e v n o c l a r u t a n d e t p i r c s n u h t i w e s a b a t a d l a u s i v - o i d u A n i m 0 5 1 * k 0 2 0 3 d e t c i r t s e r n U . s n o i t a t o n n a l a u s i v g n i d u l c n i ) 3 1 0 2 , . l a t e o e d i v D 3 h t i w b D C C e h t f o n o i s r e v A n i m 7 1 * k 5 . 2 7 1 d e t c i r t s e r n U e s a b a t a D n o i t a s r e v n o C ) 5 1 0 2 , . l a t e r e t n e v e d n a V ( c i l b u p d n a , e n o h p e l e t , e c a f - o t - e c a f f o n o i t c e l e S * s r h 0 8 k 0 0 8 0 8 2 s c i p o T l a u s a C f o . n i a t i r B m o r f e u g o l a i d n o i s s u c s i d ) 6 0 0 2 , s i l l a , s e i r o m e m r i e h t t u o b a g n i k l a t e v o b a r o 0 6 d e g a e l p o e p f o e u g o l a i D s r h 0 6 k 0 0 8 4 1 3 s c i p o T l a u s a C e h t f o y r u t n e c a m o r f e d i s y r t n u o c e h t f o e r o l k l o f e h t d n a k r o w , s e i l i m a f ) 9 9 9 1 e h t r o f d e z i n a g r o e s a b a t a d l a n o i t a n r e t n I * s r h 0 0 0 , 1 M 0 1 K 1 1 d e t c i r t s e r n U . n o i t i s i u q c a e g a u g n a l d n o c e s d n a t s r ï¬ | 1512.05742#41 | 1512.05742#43 | 1512.05742 | [
"1511.06931"
] |
1512.05742#43 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | f o y d u t s ) 5 8 9 1 , e v i t a t n e s e r p e r s w e i v r e t n i d n a s n o i t a s r e v n o c , s e v i t a r r a N * s r h 2 K 0 2 5 9 s c i p o T l a u s a C d n a . a n i l o r a C h t r o N , y t n u o C g r u b n e l k c e M f o s t n e d i s e r e h t f o ) C C N C ) 4 0 0 2 i t s e e r a s r e b m u n ) * ( d e r r a t S . s t e s a t a d e u g o l a i d n e k o p s h s i l g n E n a c i r e m A D N E I R F L L A C h s i l g n E n a c i r e m A E M O H L L A C e s a b a t a D n o i t a s r e v n o C h s i l g n E n e k o p S y a D s t c e l a i D h s i l g n E f o a t a D e g a u g n a L d l i h C e h T ( n o i t c e l l o C n o i t a s r e v n o C w o n S d n a y e n n i h W m ¨o r t s n e t S d n a d u r e l s a H e v i t a r r a N e t t o l r a h C e h T s u p r o C c i n o r h c a i D e h T t c e l a i D n r e h t u o S - n o N s u p r o C g n i t e e s u p r o C n e k o p S e h T s u p r o C n e g r e B e h T | 1512.05742#42 | 1512.05742#44 | 1512.05742 | [
"1511.06931"
] |
1512.05742#44 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | e g a u g n a L e g a n e e T m e t s y S e g n a h c x E # n a v a n a C , t t o c S d n a # y e r b u A # [epowumny|nyl l a n o i t a N h s i t i r e g d i r b m a C e h T ) 2 9 9 1 W d n a d r a o b h c t i # Aeq-jussarg # ) b D C C D 4 ( , y h t r a C c M f f i d r a C D 4 i t l u M 4 6 D # t e # f o ( # t e # n a v a n a C ( t n e s e r P # ) b D C C s l a n e R # M , h c e e L ( # h c e e p S # s u p r o C f f i d r a C # l e t r e O # y e v r u S e r a e B s t r a A # oe) e m a N c a # I # M A # w S # M # B ( ( ( ( ( ( ( ( ( 16 , e d I d n a # n e p p e R ( s u o e n a t n o p s n a m u h - n a m u H : 3 e l b a T m o r f n o i t p i r c s e D # l a t o T # l a t o T # l a t o T # l a t o T s c i p o T s d r o w f o s k r o w f o s e u g o l a i d f o s e c n a r e t t u f o . s m l ï¬ | 1512.05742#43 | 1512.05742#45 | 1512.05742 | [
"1511.06931"
] |
1512.05742#45 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | n a c i r e m A f o s t p i r c s e i v o M M 6 3 5 7 K 2 3 1 k 4 6 7 e i v o M s e u g o l a i d e m o c o t d e r e t l ï¬ e r a h c i h w s e c n a r e t t u f o s e l p i r T M 3 1 4 1 6 K 5 4 2 k 6 3 7 e i v o M . s e l p i r t X Y X m o r f - - s e u g o l a i d s t p i r c s f o s t e s b u s o w T * M 6 1 0 0 5 , 1 â K 3 6 2 * M 1 e i v o M . ) s m l ï¬ n a c i r e m A / h s i t i r B d e x i m 0 0 5 d n a s m l ï¬ n a c i r e m A 0 0 0 1 ( s t p i r c s d e t a t o n n a , s t p i r c s m l ï¬ m o r f s n o i t a s r e v n o c t r o h S * M 9 7 1 6 K 0 2 2 K 5 0 3 e i v o M . a t a d a t e m r e t c a r a h c h t i w s e u g o l a i d ) 1 1 0 2 , e e L d n a e m o c o t d e r e t l ï¬ | 1512.05742#44 | 1512.05742#46 | 1512.05742 | [
"1511.06931"
] |
1512.05742#46 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | e r a h c i h w s e c n a r e t t u f o s e l p i r T * M 2 6 8 7 , 1 K 7 8 k 3 7 1 e i v o M . s e l p i r t X Y X m o r f - - s e u g o l a i d . s a r e p o p a o s n a c i r e m A f o s t p i r c s n a r T M 0 0 1 0 0 0 , 2 2 â M 2 . 1 * M 0 1 w o h s V T s t p i r c s d n a ) y r o e h T g n a B g i B ( y d e m o c a m o r f s t p i r c s V T * k 0 0 6 1 9 1 â | 1512.05742#45 | 1512.05742#47 | 1512.05742 | [
"1511.06931"
] |
1512.05742#47 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | K 0 1 * k 0 6 w o h s V T c i t s i u g n i l r o f . w o h s ) s e n o r h T f o e m a G d e t a t o n n a , b D S M I m o r f ( a m a r d s t p i r c S M 6 9 . 2 6 8 K 1 5 1 k 4 6 6 s t p i r c s e i v o M . s e p y t e h c r a r e t c a r a h c d n a s e r u t c u r t s s t p i r c s m o r f s r i a p e s n o p s e r - n o i t c a r e t n i d e n g i l A M 0 2 4 8 1 , 6 M 5 3 . 3 M 7 . 6 e i v o M . s e l t i t b u s e i v o m s e l t i t b u s . d e n g i l a - r e k a e p s t o n e r a h c i h w s e l t i t b u s e i v o M B 1 7 0 9 , 7 0 2 â M 6 3 * M 0 4 1 e i v o M s e l t i t b u s ) 0 6 7 1 â 0 6 5 1 ( m o r f s k r o w l a n o i t c ï¬ d e t p i r c s s u o i r a V M 2 1 . 7 7 1 â â | 1512.05742#46 | 1512.05742#48 | 1512.05742 | [
"1511.06931"
] |
1512.05742#48 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | s k r o W n e t t i r W . s g n i d e e c o r p l a i r t t r u o c s a l l e w s a s g n i d e e c o r P l a i r T & r e p s e u g o l a i d f o r e b m u n e g a r e v a n o d e s a b s e t a m i t s e e t a c i d n i ) â ( h t i w d e t o n e d s e i t i t n a u Q . s t e s a t a d e s e h t n i d e t a r a p e s y l t i c i l p x e e b t o n y a m s e u g o l a i D . s u p r o c e h t n i s k r o w r o s t p i r c s f o r e b m u n e h t d n a ) 2 1 0 2 . ) s e t u n i m 6 3 ( e m i t n u r w o h s V T e g a r e v a o t ) s e t u n i m 2 1 1 ( e m i t n u r m l ï¬ e g a r e v a f o o i t a r e h t n o d e s a b d e t s u j d a d e s a b d e t a m i t s e e r a s e i t i t n a u q ) * ( d e r r a t S ( . ) s e c a f r e t n i / m o c . b d m i . w w w / / : p t t h ( e s a b a t a d D B M I e h t m o r f d e p a r c s | 1512.05742#47 | 1512.05742#49 | 1512.05742 | [
"1511.06931"
] |
1512.05742#49 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | . s t e s a t a d e u g o l a i d d e t p i r c s n a m u h - n a m u H , e e L d n a s u p r o C e u g o l a i D - e i v o M ) a 2 1 0 2 # s u p r o C , s h c n a B ) 3 1 0 2 l i z i # s u p r o C s e i r e S e n i l n O s t p i r c S m # m # Wy ) 6 0 0 2 , r e k l a # l i ) b 2 1 0 2 M - u c s e l u c i N - u c s e n a D # F m o r f a r e p O p a o S n a c i r e m A # , . l a , r u e h o C d n a t p i r c S e i v o M d e r e t l i F ) 6 1 0 2 ) 2 1 0 2 , n n a m e d e i T ( ( ) 0 6 7 1 â 0 6 5 1 ( ) b 4 1 0 2 # t e ) 4 1 0 2 , . l a e i v o m , s e i v a D | 1512.05742#48 | 1512.05742#50 | 1512.05742 | [
"1511.06931"
] |
1512.05742#50 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Joyyema) o[AIg Koy) r e k l a # e l y t ) 2 1 0 2 , s h c n a B # s u p r o C e l T b u S s e l p i r T - e i v o M s e l t i t b u S n e p O # , . l a W d n a # s u p r o C D V T S r e t c a r a h C C D - e i v o M # W # , . l a # t e # i a x i e m A ( ( : 4 # n a b r e S ( # l l e n r o C # t e # s u p r o C # s u p r o C # t e e m a N # ¨o t y K # orhy) # D E C e l b a T # y o R # WILY # o i N # l i F ( ( ( ( ( ( 17 e r e w s t e s a t a d w o h s V T s a w a t a d s i h T i r e m a T # e h t m o r f d e v i r e d s e t a m # sayeuNsy i t s E # . s w o h s V T d n a # s # m ï¬ # l # f o # s h t g n e l e g a r e v a # e h t d n a , # m ï¬ # l # r e p s e c n a r e t t u d n a # s d r o w # f o # r e b m u n e g a r e v a # e h t n o . ) l m t h . s t n u o c d r o w / t a m r o f / m o c . i r e m a t . w w w / / : p t t h ( # s r e t i r # W # r o f e d i u G SOLUNOWOS ae splOM YsI[suq Jodoid â e1od109 SWS pue Dy] Sv Yyons â eJodso9 urey99 JO nq â | 1512.05742#49 | 1512.05742#51 | 1512.05742 | [
"1511.06931"
] |
1512.05742#51 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | (aoeds ev Aq pomo][o} pue papacaid gouonbes v aq JSNUI pJOM B â 9'1) Sadeds UO paseq S]UNOD pJOM SuIsn paynduios ore saturn (,.) paueyg â sjaseyep onsoyeip usyM â so[din o8pa[Mouy sv vyepry}oU! SIAOUT UONepUdUTWIODOY Sapnjouy â suIa}SAs ONSoyeIp USALIP-[BOS 10] WSS8I ANTE ee SOTAOJ RB WO â TeyD *(pajoe.yxe sanSoyerp ou) sso] DUI wo}skg wos podeios weans yD gx 49901 QI8EE | Suneiodg mungy yy â OUI Uo weeqs eyo nyuNqA wayskg Wo} pajoe.nxe sansoyeiq WOOL MOE ILL Sunesodg munqn, yey â sordoy MATAIOVUT pure â Teonttod â [e1ouad noge suOTIES.AUOD ws al OzS syse} [R1I00g yeyo â suomtsod Jesoul 10 yeontyod aytoads jnoge sayeqoq WEL MII SPS sontog wni0, â \uaWIdeIesIp Jo (aserydesed *3'9) yuaUIIaIse JO (SIOZ ad4q Jey payejouuy â sUOTRSIOAUOD WINIO} aJeqQaq AkIID Wr'l MOI TZ poloLnsoiuy wmni0j srayeqag â parejouur [ag] pue od) juoweasy â spkary} WINJOJ SuOISsnosIq IpedryIAA pur [eUINOsOATT SOIT TR TZ poloLnsoiuy wniog | saseg yey PLIOM spre, Surdeyd siakvjd usaayjoqg suoles1aAuoZ MTZ 9971 Ise SULIO) OUD yey â | 1512.05742#50 | 1512.05742#52 | 1512.05742 | [
"1511.06931"
] |
1512.05742#52 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | uRIe_ JO SIOTMES, OWS oY) UT sioke|d usamjaq suoTRsIoAUOZD - IZ S6 SULIO} OWED yeyo â yey [eJoUNs Jo â sppoiqns asnge oNsaulop JoyjIe WoL s}sod Jppoy WEOI-W6I â ¬ELIZ â ¬S'LT djoy asnqy wni0, sndiog â JIPP2Y SSO1OV SJUOUTUIOD G/T] - - - poornsomuy) wmni0, â siskjeue Sumy UIA â siasn OM] U99MI9qQ PI}d9]]OI SaSeSSoUl SIS #899â 08S ME SI poornsomuy) SaSeSSoU SINS (6002 ssunsod winioj JONAS) aL 4098LP L89 poornsomuy) SO[QOINI JONIML Woy payoexe sojdin y-g-V PS) TCP â ¬ poioinsoiup SoO]qoIs 1A, JONIME WO pojoesyxo sar[dar pure sjoom], INSTI Wel id palornsezu SO[QOIOIA, (L007 *SWIOOL JeYD OUTTUO DyTIads-a8v WO] SIsOg WOO! 1ST POL palornsezu yyo SPIOM JO sonsoyerp Jo | sun jo uondiiosaq # [RIOL # [RIOL # SAV soidoy, odAL, e m a N # s u p r o C # t a h C S P N ) 7 0 0 2 , l l e t r a # â Tee M d n a # h t y s r o F ( # s u p r o C ) 0 1 0 2 # , . l a # t e # r e t t i Jona) Jon, # r e t t i # w T # R ( s u p r o C e l p i r T r e t t i # w T ) b 5 1 0 2 , . l a # t e | 1512.05742#51 | 1512.05742#53 | 1512.05742 | [
"1511.06931"
] |
1512.05742#53 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | i n o d r o S ( # s u p r o C # t e N e s U , y r u b t s e # â Ainqisep W d n a l u o a h S ( # s u p r o C S M S S U N ) 3 1 0 2 , n a K d n a # n e h C ( t i d d e R s u p r o C e s u b A c i t s e m o D t i d d e R ) 5 1 0 2 , . l a t e g n i d a r h c S ( # n a t a C # f o # s r e l t t e S ) 2 1 0 2 , . l a # t e s o n e t n a f A ( s u p r o C s d r a C ) 2 1 0 2 , . l a # t e i l a l a j D ( s e g a P k l a T a i d e p i k i W n i t n e m e e r g A ) 2 1 0 2 , . l a # t e s a e r d n A ( s r e t a b e D e t a e r C y b t n e m e e r g A , n w o e K c M d n a # l a h t n e s o R ( # s u p r o C # t n e m u g r A # t e n r e t n I ) b 2 1 0 2 , . l a # t e r e k l a # JOyTRAA) # W ( # s u p r o C C P M ) 0 1 0 2 , . l a # t e h k i a h S ( s u p r o C e u g o l a i D u t n u b U ) a 5 1 0 2 # , . l a # t e # e w o L ( # s u p r o C t a h C u t n u b U ) 3 1 0 2 # , a h A d n a # s u h t U ( | 1512.05742#52 | 1512.05742#54 | 1512.05742 | [
"1511.06931"
] |
1512.05742#54 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | t e s a t a D g o l a i D e i v o M ) 5 1 0 2 , . l a # t e # e g d o D ( 18 w n a m u h - n a m u H : 5 e l b a T s r e t c a r a h c # f o d e t a m # poyeutysa i t s e e c n a r e t t u r e p s d r o w e g a r e v a g n i s u d e t u p m o c s d n u o b r e p p u d n a # r e w o l s e t a c i d n i ) (,) ( # e l g n a i r T . e g a s u g n a l s o t e u d r e h t e g o t d e t a n e t a c n o c | 1512.05742#53 | 1512.05742#55 | 1512.05742 | [
"1511.06931"
] |
1512.05742#55 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | M 1 2 . t a h t # e t o N # . s u p r o c # e h t # d e d r o c e r # f o s k c o l b s u o u g i t n o c # f o t r a p h s i l g n E # e h t n o # y l n o e r a ) â ( # y b d e t a c i d n i s g o l a i D d e s a b s e t a m # soyeUNse i t s e s e t a c i d n i (,,) . s r i a p A Q d e t a l u m i s f o m r o f e r a u q S # e h t # n i e r a . ) 5 1 0 2 ( ) (,) ( t e s a t a d g n i d a r h c S g o l a i D e i v o M # s u p r o c t i d d e R r a l i # seyTUs m i s # a n o # e h t m o r f s e u g o l a i d # e h t s a # s n r u t e g a r e v a # e h t e t a l u c l a c d n a # s p u o r g s w e n # f o # r e b m u n # l a t o t # e h t # e t o n # e w # , t e N e s U # f o e s a c # e h t # n I . t a h c t n a p i c i t r a p - i t l u m a # n i n o i t a s r e v n o c # l l e w s a # s n e k o t o t s r e f e r d n a e z i s r a l i # Ie[TUIIS m i s f o t e s a t a d r e t t i w T a n o d e s a b e t a m i t s e n a s e t a c i d n i ) â ¡ ( | 1512.05742#54 | 1512.05742#56 | 1512.05742 | [
"1511.06931"
] |
1512.05742#56 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | . p u o r g s w e n r e p d e t c e l l o c s t s o p f o r e b m u n e g a r e v a _ t i d d e r _ e l b a l i a v a _ y l c i l b u p _ y r e v e _ e v a h _ i / 7 g l x b 3 / s t n e m m o c / s t e s a t a d / r / m o c . t i d d e r . w w w / / : s p t t h # : o t # s r e f e r ) â ¦ ( . s d r o w s a / t n e m m o c # 4.1.3 OTHER | 1512.05742#55 | 1512.05742#57 | 1512.05742 | [
"1511.06931"
] |
1512.05742#57 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | The DIALOG mathematical proof dataset (Wolska et al., 2004) is a Wizard-of-Oz dataset in- volving an automated tutoring system that attempts to advise students on proving mathematical theorems. This is done using a hinting algorithm that provides clues when students come up with an incorrect answer. At only 66 dialogues, the dataset is very small, and consists of a conglomeration of text-based interactions with the system, as well as think-aloud audio and video footage recorded by the users as they interacted with the system. The latter was transcribed and annotated with simple speech acts such as â signaling emotionsâ or â self-addressingâ . | 1512.05742#56 | 1512.05742#58 | 1512.05742 | [
"1511.06931"
] |
1512.05742#58 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | The MATCH corpus (Georgila et al., 2010) is a small corpus of 447 dialogues based on a Wizard-of-Oz experiment, which collected 50 young and old adults interacting with spoken dialogue systems. These conversations were annotated semi-automatically with dialogue acts and â Informa- tion State Updateâ (ISU) representations of dialogue context. The corpus also contains information about the usersâ cognitive abilities, with the motivation of modeling how the elderly interact with dialogue systems. # 4.2 Human-Human Spoken Corpora Naturally, there is much more data available for conversations between humans than conversations between humans and machines. Thus, we break down this category further, into spoken dialogues (this section) and written dialogues (Section 4.3). The distinction between spoken and written dia- logues is important, since the distribution of utterances changes dramatically according to the nature of the interaction. As discussed in Subsection 3.1, spoken dialogues tend to be more colloquial and generally well-formed as the user speaks in train-of-thought manner; they also tend to use shorter words and phrases. Conversely, in written communication, users have the ability to reï¬ ect on what they are writing before they send a message. Written dialogues can also contain spelling errors or abbreviations, though, which are generally not transcribed in spoken dialogues. # 4.2.1 SPONTANEOUS SPOKEN CORPORA | 1512.05742#57 | 1512.05742#59 | 1512.05742 | [
"1511.06931"
] |
1512.05742#59 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | We ï¬ rst introduce datasets in which the topics of conversation are either casual, or not pre-speciï¬ ed in any way. We refer to these corpora as spontaneous, as we believe they most closely mimic spontaneous and unplanned spoken interactions between humans. Perhaps one of the most inï¬ uential spoken corpora is the Switchboard dataset (Godfrey et al., 1992). This dataset consists of approximately 2,500 dialogues from phone calls, along with word- by-word transcriptions with about 500 total speakers. A computer-driven robot operator system introduced a topic for discussion between two participants, and recorded the resulting conversation. About 70 casual topics were provided, of which about 50 were frequently used. The corpus was originally designed for training and testing various speech processing algorithms; however, it has since been used for a wide variety of other tasks, including the modeling of dialogue acts such as â | 1512.05742#58 | 1512.05742#60 | 1512.05742 | [
"1511.06931"
] |
1512.05742#60 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | statementâ , â questionâ , and â agreementâ (Stolcke et al., 2000). Another important dataset is the British National Corpus (BNC) (Leech, 1992), which contains approximately 10 million words of dialogue. These were collected in a variety of contexts ranging from formal business or government meetings, to radio shows and phone-ins. Although most of the conversations are spoken in nature, some of them are also written. BNC covers a large number of sources, and was designed to represent a wide cross-section of British English from the late twentieth century. The corpus also includes part-of-speech (POS) tagging for every word. The vast | 1512.05742#59 | 1512.05742#61 | 1512.05742 | [
"1511.06931"
] |
1512.05742#61 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 19 array of settings and topics covered by this corpus renders it very useful as a general-purpose spoken dialogue dataset. Other datasets have been collected for the analysis of spoken English over the telephone. The CALLHOME American English Speech Corpus (Canavan et al., 1997) consists of 120 such conversations totalling about 60 hours, mostly between family members or close friends. Similarly, the CALLFRIEND American English-Non-Southern Dialect Corpus (Canavan and Zipperlen, 1996) consists of 60 telephone conversations lasting 5-30 minutes each between English speakers in North America without a Southern accent. It is annotated with speaker information such as sex, age, and education. The goal of the project was to support the development of language identiï¬ cation technologies, yet, there are no distinguishing features in either of these corpora in terms of the topics of conversation. An attempt to capture exclusively teenage spoken language was made in the Bergen Corpus of London Teenager Language (COLT) (Haslerud and Stenstr¨om, 1995). Conversations were recorded surreptitiously by student â recruitsâ , with a Sony Walkman and a lapel microphone, in order to obtain a better representation of teenager interactions â in-the-wildâ . | 1512.05742#60 | 1512.05742#62 | 1512.05742 | [
"1511.06931"
] |
1512.05742#62 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | This dataset has been used to identify trends in language evolution in teenagers (Stenstr¨om et al., 2002). The Cambridge and Nottingham Corpus of Discourse in English (CANCODE) (McCarthy, 1998) is a subset of the Cambridge International Corpus, containing about 5 million words collected from recordings made throughout the islands of Britain and Ireland. It was constructed by Cam- bridge University Press and the University of Nottingham using dialogue data on general topics between 1995 and 2000. It focuses on interpersonal communication in a range of social contexts, varying from hair salons, to post ofï¬ ces, to restaurants. This has been used, for example, to study language awareness in relation to spoken texts and their cultural contexts (Carter, 1998). In the dataset, the relationships between speakers (e.g. roommates, strangers) is labeled and the interac- tion type is provided (e.g. professional, intimate). Other works have attempted to record the physical elements of conversations between humans. To this end, a small corpus entitled d64 Multimodal Conversational Corpus (Oertel et al., 2013) was collected, incorporating data from 7 video cameras, and the registration of 3-D head, torso, and arm motion using an Optitrack system. | 1512.05742#61 | 1512.05742#63 | 1512.05742 | [
"1511.06931"
] |
1512.05742#63 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | Signiï¬ cant effort was made to make the data collection process as non-intrusiveâ and thus, naturalisticâ as possible. Annotations were made to attempt to quantify overall group excitement and pairwise social distance between participants. A similar attempt to incorporate computer vision features was made in the AMI Meeting Cor- pus (Renals et al., 2007), where cameras, a VGA data projector capture, whiteboard capture, and digital pen capture, were all used in addition to speech recordings for various meeting scenarios. As with the d64 corpus, the AMI Meeting Corpus is a small dataset of multi-participant chats, that has not been disentangled into strict dialogue. The dataset has often been used for analysis of the dynamics of various corporate and academic meeting scenarios. In a similar vein, the Cardiff Conversation Database (CCDb) (Aubrey et al., 2013) is an audio- visual database containing unscripted natural conversations between pairs of people. The original dataset consisted of 30 ï¬ ve minute conversations, 7 of which were fully annotated with transcrip- tions and behavioural annotations such as speaker activity, facial expressions, head motions, and smiles. The content of the conversation is an unconstrained discussion on topics such as movies. While the original dataset featured 2D visual feeds, an updated version with 3D video has also been derived, called the 4D Cardiff Conversation Database (4D CCDb) (Vandeventer et al., 2015). | 1512.05742#62 | 1512.05742#64 | 1512.05742 | [
"1511.06931"
] |
1512.05742#64 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | 20 This version contains 17 one-minute conversations from 4 participants on similarly un-constrained topics. The Diachronic Corpus of Present-Day Spoken English (DCPSE) (Aarts and Wallis, 2006) is a parsed corpus of spoken English made up of two separate datasets. It contains more than 400,000 words from the ICE-GB corpus (collected in the early 1990s) and 400,000 words from the London- Lund Corpus (collected in the late 1960s-early 1980s). ICE-GB refers to the British component of the International Corpus of English (Greenbaum and Nelson, 1996; Greenbaum, 1996) and contains both spoken and written dialogues from English adults who have completed secondary education. The dataset was selected to provide a representative sample of British English. The London-Lund Corpus (Svartvik, 1990) consists exclusively of spoken British conversations, both dialogues and monologues. It contains a selection of face-to-face, telephone, and public discussion dialogues; the latter refers to dialogues that are heard by an audience that does not participate in the dialogue, in- cluding interviews and panel discussions that have been broadcast. The orthographic transcriptions of the datasets are normalised and annotated according to the same criteria; ICE-GB was used as a gold standard for the parsing of DCPSE. The Spoken Corpus of the Survey of English Dialects (Beare and Scott, 1999) consists of 1000 recordings, with about 0.8 million total words, collected from 1948-1961 in order to document various existing English dialects. People aged 60 and over were recruited, being most likely to speak the traditional â | 1512.05742#63 | 1512.05742#65 | 1512.05742 | [
"1511.06931"
] |
1512.05742#65 | A Survey of Available Corpora for Building Data-Driven Dialogue Systems | uncontaminatedâ dialects of their area and encouraged to talk about their memories, families, work, and their countryside folklore. The Child Language Data Exchange System (CHILDES) (MacWhinney and Snow, 1985) is a database organized for the study of ï¬ rst and second language acquisition. The database contains 10 million English words and approximately the same number of non-English words. It also contains transcripts, with occasional audio and video recordings of data collected from children and adults learning both ï¬ rst and second languages, although the English transcripts are mostly from children. This corpus could be leveraged in order to build automated teaching assistants. The expanded Charlotte Narrative and Conversation Collection (CNCC), a subset of the ï¬ rst release of the American National Corpus (Reppen and Ide, 2004), contains 95 narratives, conver- sations and interviews representative of the residents of Mecklenburg County, North Carolina and its surrounding communities. The purpose of the CNCC was to create a corpus of conversation and conversational narration in a â | 1512.05742#64 | 1512.05742#66 | 1512.05742 | [
"1511.06931"
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.