File size: 2,807 Bytes
1cd59e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9664e81
1cd59e3
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
language:
- he
---# Coreference Project by

DDRND (Mafat) as part of the Israeli national NLP program (see our GitHub at https://nnlp-il.mafat.ai/#Our-Github)
and the Israeli Association of Human Language Technologies (https://www.iahlt.org)

## Introduction

The coreference corpus is an extension of IAHLT's named entities dataset for
Hebrew and Arabic. This project is a work in progress, such that a subset of
articles (full-doc level) that are already annotated for entities are being
further annotated for (named) entity coreference.

The corpus consists of 1 apc articles from Youtube transcripts (0%); 201.
arb articles from the Kul al-Arab news organisation (96%), the All Rights.
entitlements organisation (0%), Weizmann popular science articles (2%);.
657 heb articles from Bagatz court decisions (3%), Davar news organisation.
(75%), Israel Hayom news organisation (3%), Knesset protocols (1%),.
Weizmann popular science articles (4%), Hebrew Wikipedia entries (11%);.

The corpus, 1 paragraphs (apc), 2811 paragraphs (arb) and 9610 paragraphs
(heb), has been annotated with morpheme-level mention spans, assembled
into coreference clusters with entity types.

## Data set

The current release includes the following files:

Annotated documents (.jsonl):

1. data/coref-4-rc7-heb-all -- heb articles
2. data/coref-4-rc7-heb-unique -- heb articles, each annotated once 
3. data/coref-4-rc7-heb-iaa -- heb articles, used for IAA

Additionally, all files are provided in a human-readable form (readable_data/*).

## Format

Each article is a single json record. Some articles have been doubly-annotated
for the purposes of inter-annotator agreement study, their articles appear
multiple times.

The jsonl structure is:
```
{ text: str, 
  user: str,
  metadata: { source: str, doc_id: str, ... },
  clusters: [ {
    metadata: { name: str, entity: str }, 
    mentions: [ (int, int, dict) ]
  } ]
}
```

The `text` field contains the raw text of the original article. The top-level
`metadata` dictionary provides document-level metadata, minimally `source` and
`doc_id`.

The `clusters` field is a list of JSON cluster records each containing a
`metadata` and `mentions` field. The cluster-level `metadata` field has a name
for the cluster and its entity type. The `mentions` field is a list of triples:
the span indices of the text plus a metadata dictionary. We provide no
mention-level metadata in this release.

Not all clusters have been annotated for entity type; this will be completed in
a future release.

## Acknowledgments

We would like to thank all the people who contributed to this corpus:

Amir Cohen 
Amjad Aliat
Emmanuel Kowner
Israel Landau
Mutaz Ayesh
Nick Howell
Noam Ordan
Omer Strass
Shahar Adar
Shira Wigderson
Yifat Ben Moshe
amirejmail
hiba_ammash