Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,102 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: cc
|
3 |
+
task_categories:
|
4 |
+
- text-classification
|
5 |
+
language:
|
6 |
+
- en
|
7 |
+
tags:
|
8 |
+
- math
|
9 |
+
---
|
10 |
+
|
11 |
+
TLDR: MIT OCW Math Lectures with Student Questions
|
12 |
+
|
13 |
+
# SIGHT: A Large Annotated Dataset on Student Insights Gathered from Higher Education Transcripts
|
14 |
+
|
15 |
+
[Project Page](https://rosewang2008.github.io/sight/), [Paper](https://arxiv.org/pdf/2306.09343.pdf), [Poster](assets/poster.pdf)
|
16 |
+
|
17 |
+
Authors: Rose E. Wang*, Pawan Wirawarn*, Noah Goodman and Dorottya Demszky
|
18 |
+
|
19 |
+
*= Equal contributions
|
20 |
+
|
21 |
+
In the Proceedings of Innovative Use of NLP for Building Educational Applications 2023
|
22 |
+
|
23 |
+
If you find our work useful or interesting, please consider citing it!
|
24 |
+
|
25 |
+
```
|
26 |
+
@inproceedings{wang2023sight,
|
27 |
+
title={SIGHT: A Large Annotated Dataset on Student Insights Gathered from Higher Education Transcripts},
|
28 |
+
author={Wang, Rose E and Wirawarn, Pawan and Goodman, Noah and Demszky, Dorottya},
|
29 |
+
year={2023},
|
30 |
+
month = jun,
|
31 |
+
booktitle = {18th Workshop on Innovative Use of NLP for Building Educational Applications},
|
32 |
+
month_numeric = {6}
|
33 |
+
}
|
34 |
+
```
|
35 |
+
|
36 |
+

|
37 |
+
|
38 |
+
## Motivation
|
39 |
+
Lectures are a learning experience for both students and teachers.
|
40 |
+
Students learn from teachers about the subject material, while teachers learn from students about how to refine their instruction.
|
41 |
+
Unfortunately, online student feedback is unstructured and abundant, making it challenging for teachers to learn and improve. We take a step towards tackling this challenge.
|
42 |
+
First, we contribute a dataset for studying this problem: SIGHT is a large dataset of 288 math lecture transcripts and 15,784 comments collected from the Massachusetts Institute of Technology OpenCourseWare (MIT OCW) YouTube channel.
|
43 |
+
Second, we develop a rubric for categorizing feedback types using qualitative analysis.
|
44 |
+
Qualitative analysis methods are powerful in uncovering domain-specific insights, however they are costly to apply to large data sources.
|
45 |
+
To overcome this challenge, we propose a set of best practices for using large language models (LLMs) to cheaply classify the comments at scale.
|
46 |
+
We observe a striking correlation between the model's and humans' annotation:
|
47 |
+
Categories with consistent human annotations (>$0.9$ inter-rater reliability, IRR) also display higher human-model agreement (>$0.7$), while categories with less consistent human annotations ($0.7$-$0.8$ IRR) correspondingly demonstrate lower human-model agreement ($0.3$-$0.5$).
|
48 |
+
These techniques uncover useful student feedback from thousands of comments, costing around $\$0.002$ per comment.
|
49 |
+
We conclude by discussing exciting future directions on using online student feedback and improving automated annotation techniques for qualitative research.
|
50 |
+
|
51 |
+
|
52 |
+
## Repository structure
|
53 |
+
|
54 |
+
Scripts are in `run_analysis.sh` for replicating the paper analysis. Please refer to the `prompts` directory for replicating the annotations.
|
55 |
+
|
56 |
+
The repo structure:
|
57 |
+
|
58 |
+
```
|
59 |
+
.
|
60 |
+
├── data
|
61 |
+
├── annotations # Sample (human) and full SIGHT annotations
|
62 |
+
├── comments # Per-video comments
|
63 |
+
├── metadata # Per-video metadata like playlist ID or video name
|
64 |
+
└── transcripts # Per-video transcript, transcribed with Whisper Large V2
|
65 |
+
├── prompts # Prompts used for annotation
|
66 |
+
├── results # Result plots used in paper
|
67 |
+
├── scripts # Python scripts for analysis
|
68 |
+
├── requirements.txt # Install requirements for running code
|
69 |
+
├── run_analysis.sh # Complete analysis script
|
70 |
+
├── LICENSE
|
71 |
+
└── README.md
|
72 |
+
```
|
73 |
+
|
74 |
+
## Installation
|
75 |
+
|
76 |
+
To install the required libraries:
|
77 |
+
|
78 |
+
```
|
79 |
+
conda create -n sight python=3
|
80 |
+
conda activate sight
|
81 |
+
pip install -r requirements.txt
|
82 |
+
```
|
83 |
+
|
84 |
+
## Experiments
|
85 |
+
|
86 |
+
TLDR: Running `source run_analysis.sh` replicates all the results we report in the paper.
|
87 |
+
|
88 |
+
|
89 |
+
Plots (e.g., the IRR comparison in Figure 3) are saved under `results/` as PDF files.
|
90 |
+
Numbers (e.g., sample data information in Table 2 or IRR values in Table 3) are printed out under `results/` as txt files.
|
91 |
+
|
92 |
+
|
93 |
+
## Annotations
|
94 |
+
|
95 |
+
The automated annotations provided in this GitHub repository have been scaled on categories with high inter-rater reliability (IRR) scores.
|
96 |
+
While we have made efforts to ensure the reliability of these annotations, it is important to note that the automated annotations may not be completely error-free.
|
97 |
+
We recommend using these annotations as a starting point and validating them through additional human annotation or other means as necessary.
|
98 |
+
By using these annotations, you acknowledge and accept the potential limitations and inherent uncertainties associated with automated annotation methods, like annotating at scale with GPT-3.5.
|
99 |
+
|
100 |
+
We welcome any contributions to improve the quality of the annotations in this repository!
|
101 |
+
If you have made improvements to the annotations or expanded the annotations, feel free to submit a pulll request with your changes.
|
102 |
+
We appreciate all efforts to make these annotations more useful for the education and NLP community!
|