Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: cc-by-3.0
|
3 |
+
language:
|
4 |
+
- he
|
5 |
+
size_categories:
|
6 |
+
- 100K<n<1M
|
7 |
+
---
|
8 |
+
# HebNLI - A Natural Language Inference Dataset in Hebrew
|
9 |
+
|
10 |
+
## Summary
|
11 |
+
HebNLI is a Hebrew dataset for natural language inference (NLI) tasks.
|
12 |
+
|
13 |
+
## Introduction
|
14 |
+
This dataset is the first of its kind in the Hebrew language and aims to serve as training data for NLI tasks.
|
15 |
+
HebNLI is based on MultiNLI, a large crowd-sourced corpus of sentences from varied genres and writing styles in the English language.
|
16 |
+
MultiNLI was originally built by collecting hundreds of thousands of base sentences from which different taggers derived follow-up sentences that stand in one of 3 logical relations to the base sentences: entailment, contradiction or neutral.
|
17 |
+
Different taggers were then given paired sentences - base sentence and a derived sentence. The logical relation between them was determined by the majority vote, and each pair of sentences was labled according to the determined logical relation.
|
18 |
+
In HebNLI we used machine translation (Google Gemini) to translate the English corpus to Hebrew, such that each base sentence and its compiled derivative sentences appear in Hebrew.
|
19 |
+
|
20 |
+
## Genres/Sources in HebNLI
|
21 |
+
HebNLI comprises 7 of the original 10 genres/sources that appeared in MultiNLI:
|
22 |
+
1. Nine eleven - Written protocols from a commitee investigating the events of 9/11.
|
23 |
+
2. Government - Reports, speeches and press releases published on U.S.A government websites.
|
24 |
+
3. Letters - A database of letters written in the late 90's and early 2000's.
|
25 |
+
4. OUP (Oxford University Press) - Publications about the textile industry and about child development.
|
26 |
+
5. Slate - Pop-culture articles published in Slate magazine.
|
27 |
+
6. Travel - Travel guides by Berlitz press.
|
28 |
+
7. Fiction - Texts extracted from modern works of literature.
|
29 |
+
|
30 |
+
The remaining three sources were found to either be too English-oriented to be properly translated to Hebrew by machine translation ("Verbatim" magazine source), or included too many broken sentences and filler-words to be properly translated to Hebrew by machine translation (face-to-face conversations and telephone conversations sources).
|
31 |
+
|
32 |
+
## Dataset Statistics
|
33 |
+
The table below shows the distribution of each source corpus within HebNLI (how many setences exist in the dataset from each source).
|
34 |
+
|
35 |
+
| Genre/Source | HebNLI Corpus |
|
36 |
+
|------------------|------------------|
|
37 |
+
| Nine eleven | 1878 |
|
38 |
+
| Government | 76953 |
|
39 |
+
| Letters | 1974 |
|
40 |
+
| OUP | 1986 |
|
41 |
+
| Slate | 71082 |
|
42 |
+
| Travel | 75776 |
|
43 |
+
| Fiction | 73734 |
|
44 |
+
|
45 |
+
Total # of sentences = 303,383.
|
46 |
+
|
47 |
+
### HebNLI Blog Post
|
48 |
+
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
|
49 |
+
|
50 |
+
### Original MultiNLI Paper
|
51 |
+
https://cims.nyu.edu/~sbowman/multinli/paper.pdf
|
52 |
+
|
53 |
+
## Contributors
|
54 |
+
HebNLI was translated and checked for quality by Webiks for MAFAT, as part of the National Natural Language Processing Plan of Israel.
|
55 |
+
|
56 |
+
Contributors: Hilla Merhav Fine (Webiks), Yaniv Maylik (Webiks), Carrine Cherf (Webiks), Tal Geva (MAFAT).
|
57 |
+
|
58 |
+
## Acknowledgments
|
59 |
+
We would like to express our gratitude to Adina Williams, Nikita Nangia and Samuel R. Bowman, the creators of [the original NLI dataset MultiNLI](https://huggingface.co/datasets/nyu-mll/multi_nli).
|