Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,105 @@
|
|
1 |
---
|
2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
tags:
|
3 |
+
- sentence-summarization
|
4 |
+
- multilingual
|
5 |
+
- nlp
|
6 |
+
- indicnlp
|
7 |
+
datasets:
|
8 |
+
- ai4bharat/IndicSentenceSummarization
|
9 |
+
language:
|
10 |
+
- as
|
11 |
+
- bn
|
12 |
+
- gu
|
13 |
+
- hi
|
14 |
+
- kn
|
15 |
+
- ml
|
16 |
+
- mr
|
17 |
+
- or
|
18 |
+
- pa
|
19 |
+
- ta
|
20 |
+
- te
|
21 |
+
license:
|
22 |
+
- mit
|
23 |
+
widget:
|
24 |
+
- जम्मू एवं कश्मीर के अनंतनाग जिले में शनिवार को सुरक्षाबलों के साथ मुठभेड़ में दो आतंकवादियों को मार गिराया गया। </s> <2hi>
|
25 |
+
|
26 |
+
|
27 |
+
|
28 |
---
|
29 |
+
|
30 |
+
# MultiIndicSentenceSummarization
|
31 |
+
|
32 |
+
This repository contains the [IndicBART](https://huggingface.co/ai4bharat/IndicBART) checkpoint finetuned on the 11 languages of [IndicSentenceSummarization](https://huggingface.co/datasets/ai4bharat/IndicSentenceSummarization) dataset. For finetuning details,
|
33 |
+
see the [paper](https://arxiv.org/abs/2203.05437).
|
34 |
+
<ul>
|
35 |
+
<li >Supported languages: Assamese, Bengali, Gujarati, Hindi, Marathi, Odiya, Punjabi, Kannada, Malayalam, Tamil, and Telugu. Not all of these languages are supported by mBART50 and mT5. </li>
|
36 |
+
<li >The model is much smaller than the mBART and mT5(-base) models, so less computationally expensive for decoding. </li>
|
37 |
+
<li> Trained on large Indic language corpora (431K sentences). </li>
|
38 |
+
<li> All languages, have been represented in Devanagari script to encourage transfer learning among the related languages. </li>
|
39 |
+
</ul>
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
## Using this model in `transformers`
|
44 |
+
|
45 |
+
```
|
46 |
+
from transformers import MBartForConditionalGeneration, AutoModelForSeq2SeqLM
|
47 |
+
from transformers import AlbertTokenizer, AutoTokenizer
|
48 |
+
tokenizer = AutoTokenizer.from_pretrained("ai4bharat/MultiIndicSentenceSummarization", do_lower_case=False, use_fast=False, keep_accents=True)
|
49 |
+
# Or use tokenizer = AlbertTokenizer.from_pretrained("ai4bharat/MultiIndicSentenceSummarization", do_lower_case=False, use_fast=False, keep_accents=True)
|
50 |
+
model = AutoModelForSeq2SeqLM.from_pretrained("ai4bharat/MultiIndicSentenceSummarization")
|
51 |
+
# Or use model = MBartForConditionalGeneration.from_pretrained("ai4bharat/MultiIndicSentenceSummarization")
|
52 |
+
|
53 |
+
# Some initial mapping
|
54 |
+
bos_id = tokenizer._convert_token_to_id_with_added_voc("<s>")
|
55 |
+
eos_id = tokenizer._convert_token_to_id_with_added_voc("</s>")
|
56 |
+
pad_id = tokenizer._convert_token_to_id_with_added_voc("<pad>")
|
57 |
+
|
58 |
+
# To get lang_id use any of ['<2as>', '<2bn>', '<2en>', '<2gu>', '<2hi>', '<2kn>', '<2ml>', '<2mr>', '<2or>', '<2pa>', '<2ta>', '<2te>']
|
59 |
+
# First tokenize the input. The format below is how IndicBART was trained so the input should be "Sentence </s> <2xx>" where xx is the language code. Similarly, the output should be "<2yy> Sentence </s>".
|
60 |
+
inp = tokenizer("जम्मू एवं कश्मीर के अनंतनाग जिले में शनिवार को सुरक्षाबलों के साथ मुठभेड़ में दो आतंकवादियों को मार गिराया गया। </s> <2hi>", add_special_tokens=False, return_tensors="pt", padding=True).input_ids
|
61 |
+
|
62 |
+
# For generation. Pardon the messiness. Note the decoder_start_token_id.
|
63 |
+
|
64 |
+
model_output=model.generate(inp, use_cache=True,no_repeat_ngram_size=3, num_beams=5, length_penalty=0.8, early_stopping=True, pad_token_id=pad_id, bos_token_id=bos_id, eos_token_id=eos_id, decoder_start_token_id=tokenizer._convert_token_to_id_with_added_voc("<2hi>"))
|
65 |
+
|
66 |
+
# Decode to get output strings
|
67 |
+
decoded_output=tokenizer.decode(model_output[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
68 |
+
print(decoded_output) # जम्मू एवं कश्मीरः अनंतनाग में सुरक्षाबलों के साथ मुठभेड़ में दो आतंकवादी ढेर
|
69 |
+
|
70 |
+
# Note that if your output language is not Hindi or Marathi, you should convert its script from Devanagari to the desired language using the Indic NLP Library.
|
71 |
+
|
72 |
+
```
|
73 |
+
# Note:
|
74 |
+
If you wish to use any language written in a non-Devanagari script, then you should first convert it to Devanagari using the <a href="https://github.com/anoopkunchukuttan/indic_nlp_library">Indic NLP Library</a>. After you get the output, you should convert it back into the original script.
|
75 |
+
|
76 |
+
## Benchmarks
|
77 |
+
|
78 |
+
Scores on the `IndicSentenceSummarization` test sets are as follows:
|
79 |
+
|
80 |
+
Language | Rouge-1 / Rouge-2 / Rouge-L
|
81 |
+
---------|----------------------------
|
82 |
+
as | 46.77 / 59.29 / 71.17
|
83 |
+
bn | 34.91 / 49.29 / 53.97
|
84 |
+
gu | 29.97 / 45.92 / 47.46
|
85 |
+
hi | 28.11 / 45.34 / 49.29
|
86 |
+
kn | 70.03 / 77.32 / 83.37
|
87 |
+
ml | 54.42 / 66.42 / 66.45
|
88 |
+
mr | 26.98 / 46.50 / 45.75
|
89 |
+
or | 23.66 / 43.65 / 45.93
|
90 |
+
pa | 37.20 / 52.22 / 54.12
|
91 |
+
ta | 38.97 / 56.83 / 57.85
|
92 |
+
te | 35.28 / 53.44 / 55.33
|
93 |
+
|
94 |
+
|
95 |
+
## Citation
|
96 |
+
|
97 |
+
If you use this model, please cite the following paper:
|
98 |
+
```
|
99 |
+
@inproceedings{Kumar2022IndicNLGSM,
|
100 |
+
title={IndicNLG Suite: Multilingual Datasets for Diverse NLG Tasks in Indic Languages},
|
101 |
+
author={Aman Kumar and Himani Shrotriya and Prachi Sahu and Raj Dabre and Ratish Puduppully and Anoop Kunchukuttan and Amogh Mishra and Mitesh M. Khapra and Pratyush Kumar},
|
102 |
+
year={2022},
|
103 |
+
url = "https://arxiv.org/abs/2203.05437"
|
104 |
+
}
|
105 |
+
```
|