File size: 9,401 Bytes
8337f9f
 
 
 
 
cf1cf45
8337f9f
cf1cf45
8337f9f
 
 
 
 
 
 
 
 
 
 
4948419
949f4f5
bca982b
949f4f5
 
 
 
 
 
 
 
 
ff897a1
 
949f4f5
 
 
 
ff897a1
 
 
 
 
 
949f4f5
 
bca982b
949f4f5
 
bca982b
949f4f5
a4ba8c9
bca982b
a4ba8c9
bca982b
 
 
 
 
 
 
 
 
 
 
 
8337f9f
 
 
 
 
 
 
a24471f
8337f9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
949f4f5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
---
annotations_creators:
- machine-generated
language_creators:
- found
language:
- hi
license:
- mit
multilinguality:
- monolingual
size_categories:
- 10K<n<100K
source_datasets:
- extended|bbc__hindi_news_classification
task_categories:
- text-classification
task_ids:
- natural-language-inference
pretty_name: BBC Hindi NLI Dataset
dataset_info:
  config_name: bbc hindi nli
  features:
  - name: premise
    dtype: string
  - name: hypothesis
    dtype: string
  - name: label
    dtype:
      class_label:
        names:
          '0': not-entailment
          '1': entailment
  - name: topic
    dtype:
      class_label:
        names:
          '0': india
          '1': news
          '2': international
          '3': entertainment
          '4': sport
          '5': science
  splits:
  - name: train
    num_bytes: 2990064
    num_examples: 15552
  - name: validation
    num_bytes: 496800
    num_examples: 2580
  - name: test
    num_bytes: 494424
    num_examples: 2592
  download_size: 309124
  dataset_size: 3981288
configs:
- config_name: bbc hindi nli
  data_files:
  - split: train
    path: bbc hindi nli/train-*
  - split: validation
    path: bbc hindi nli/validation-*
  - split: test
    path: bbc hindi nli/test-*
  default: true
---

# Dataset Card for BBC Hindi NLI Dataset

## Table of Contents
- [Dataset Description](#dataset-description)
  - [Dataset Summary](#dataset-summary)
  - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
  - [Languages](#languages)
- [Dataset Structure](#dataset-structure)
  - [Data Instances](#data-instances)
  - [Data Fields](#data-fields)
  - [Data Splits](#data-splits)
- [Dataset Creation](#dataset-creation)
  - [Curation Rationale](#curation-rationale)
  - [Source Data](#source-data)
  - [Annotations](#annotations)
  - [Personal and Sensitive Information](#personal-and-sensitive-information)
- [Considerations for Using the Data](#considerations-for-using-the-data)
  - [Social Impact of Dataset](#social-impact-of-dataset)
  - [Discussion of Biases](#discussion-of-biases)
  - [Other Known Limitations](#other-known-limitations)
- [Additional Information](#additional-information)
  - [Dataset Curators](#dataset-curators)
  - [Licensing Information](#licensing-information)
  - [Citation Information](#citation-information)
  - [Contributions](#contributions)

## Dataset Description

- **Repository:** [GitHub](https://github.com/midas-research/hindi-nli-data)
- **Paper:** [Aclweb](https://www.aclweb.org/anthology/2020.aacl-main.71)
- **Point of Contact:** [GitHub](https://github.com/midas-research/hindi-nli-data)

### Dataset Summary

- Dataset for Natural Language Inference in Hindi Language. BBC Hindi Dataset consists of textual-entailment pairs.
- Each row of the Datasets if made up of 4 columns - Premise, Hypothesis, Label and Topic.
- Context and Hypothesis is written in Hindi while Entailment_Label is in English.
- Entailment_label is of 2 types - entailed and not-entailed.
- Dataset can be used to train models for Natural Language Inference tasks in Hindi Language.
[More Information Needed]

### Supported Tasks and Leaderboards

- Natural Language Inference for Hindi

### Languages

Dataset is in Hindi

## Dataset Structure

- Data is structured in TSV format. 
- Train and Test files are in seperate files


### Dataset Instances

An example of 'train' looks as follows.

```
{'hypothesis': 'यह खबर की सूचना है|', 'label': 'entailed', 'premise': 'गोपनीयता की नीति', 'topic': '1'}

```
### Data Fields

- Each row contatins 4 columns - Premise, Hypothesis, Label and Topic.

### Data Splits

- Train : 15553
- Valid : 2581
- Test : 2593

## Dataset Creation

- We employ a recasting technique from Poliak et al. (2018a,b) to convert publicly available BBC Hindi news text classification datasets in Hindi and pose them as TE problems
- In this recasting process, we build template hypotheses for each class in the label taxonomy
- Then, we pair the original annotated sentence with each of the template hypotheses to create TE samples.
- For more information on the recasting process, refer to paper "https://www.aclweb.org/anthology/2020.aacl-main.71"

### Source Data

Source Dataset for the recasting process is the BBC Hindi Headlines Dataset(https://github.com/NirantK/hindi2vec/releases/tag/bbc-hindi-v0.1)

#### Initial Data Collection and Normalization

-  BBC Hindi News Classification Dataset contains 4, 335 Hindi news headlines tagged across 14 categories: India, Pakistan,news, International, entertainment, sport, science, China, learning english, social, southasia, business, institutional, multimedia
-  We processed this dataset to combine two sets of relevant but low prevalence classes.
- Namely, we merged the samples from Pakistan, China, international, and southasia as one class called international.
- Likewise, we also merged samples from news, business, social, learning english, and institutional as news.
- Lastly, we also removed the class multimedia because there were very few samples.

#### Who are the source language producers?

Pls refer to this paper: "https://www.aclweb.org/anthology/2020.aacl-main.71"

### Annotations

#### Annotation process

Annotation process has been described in Dataset Creation Section.

#### Who are the annotators?

Annotation is done automatically.

### Personal and Sensitive Information

No Personal and Sensitive Information is mentioned in the Datasets.

## Considerations for Using the Data

Pls refer to this paper: https://www.aclweb.org/anthology/2020.aacl-main.71

### Discussion of Biases

Pls refer to this paper: https://www.aclweb.org/anthology/2020.aacl-main.71

### Other Known Limitations

No other known limitations

## Additional Information

Pls refer to this link: https://github.com/midas-research/hindi-nli-data

### Dataset Curators

It is written in the repo : https://github.com/avinsit123/hindi-nli-data that 
- This corpus can be used freely for research purposes.
- The paper listed below provide details of the creation and use of the corpus. If you use the corpus, then please cite the paper.
- If interested in commercial use of the corpus, send email to midas@iiitd.ac.in.
- If you use the corpus in a product or application, then please credit the authors and Multimodal Digital Media Analysis Lab - Indraprastha Institute of Information Technology, New Delhi appropriately. Also, if you send us an email, we will be thrilled to know about how you have used the corpus.
- Multimodal Digital Media Analysis Lab - Indraprastha Institute of Information Technology, New Delhi, India disclaims any responsibility for the use of the corpus and does not provide technical support. However, the contact listed above will be happy to respond to queries and clarifications.
- Rather than redistributing the corpus, please direct interested parties to this page
- Please feel free to send us an email:
  - with feedback regarding the corpus.
  - with information on how you have used the corpus.
  - if interested in having us analyze your data for natural language inference.
  - if interested in a collaborative research project.


### Licensing Information

Copyright (C) 2019 Multimodal Digital Media Analysis Lab - Indraprastha Institute of Information Technology, New Delhi (MIDAS, IIIT-Delhi).
Pls contact authors for any information on the dataset.

### Citation Information

```
    @inproceedings{uppal-etal-2020-two,
    title = "Two-Step Classification using Recasted Data for Low Resource Settings",
    author = "Uppal, Shagun  and
      Gupta, Vivek  and
      Swaminathan, Avinash  and
      Zhang, Haimin  and
      Mahata, Debanjan  and
      Gosangi, Rakesh  and
      Shah, Rajiv Ratn  and
      Stent, Amanda",
    booktitle = "Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing",
    month = dec,
    year = "2020",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.aacl-main.71",
    pages = "706--719",
    abstract = "An NLP model{'}s ability to reason should be independent of language. Previous works utilize Natural Language Inference (NLI) to understand the reasoning ability of models, mostly focusing on high resource languages like English. To address scarcity of data in low-resource languages such as Hindi, we use data recasting to create NLI datasets for four existing text classification datasets. Through experiments, we show that our recasted dataset is devoid of statistical irregularities and spurious patterns. We further study the consistency in predictions of the textual entailment models and propose a consistency regulariser to remove pairwise-inconsistencies in predictions. We propose a novel two-step classification method which uses textual-entailment predictions for classification task. We further improve the performance by using a joint-objective for classification and textual entailment. We therefore highlight the benefits of data recasting and improvements on classification performance using our approach with supporting experimental results.",
}
```

### Contributions

Thanks to [@avinsit123](https://github.com/avinsit123) for adding this dataset.