File size: 4,530 Bytes
0472cff
2c328f8
 
 
 
 
 
a54afc8
2c328f8
 
a54afc8
 
 
 
0472cff
 
a54afc8
 
 
0472cff
a54afc8
0472cff
 
 
 
 
 
 
 
 
a54afc8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18b997c
0472cff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4529707
0472cff
 
 
 
 
4529707
0472cff
 
 
 
 
4529707
0472cff
 
 
 
 
4529707
0472cff
 
 
 
 
4529707
0472cff
 
 
 
 
 
 
 
 
 
 
a54afc8
0472cff
 
a54afc8
 
 
 
 
 
 
 
 
0472cff
 
 
 
 
 
 
 
a54afc8
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
---
language:
- as
- brx
- mr
- ne
- sa
license: mit
size_categories:
- 10M<n<100M
task_categories:
- token-classification
- other
pretty_name: CLASSER
---

# CLASSER: Cross-lingual Annotation Projection enhancement through Script Similarity for Fine-grained Named Entity Recognition

**CLASSER** is a framework for cross-lingual annotation projection with script-similarity-based refinement to create high-quality fine-grained named entity recognition datasets. It is part of the [AWED-FiNER](https://github.com/PrachuryyaKaushik/AWED-FiNER) ecosystem.

[**Paper**](https://huggingface.co/papers/2601.10161) | [**GitHub**](https://github.com/PrachuryyaKaushik/AWED-FiNER) | [**Interactive Demo**](https://huggingface.co/spaces/prachuryyaIITG/AWED-FiNER)

Utilizing CLASSER, fine-grained named entity recognition dataset is created in five languages: Assamese (as), Bodo (brx), Marathi (mr), Nepali (ne) and Sanskrit (sa).

## CLASSER Framework Overview

<img src="CLASSER_framework.png" alt="CLASSER Framework Overview" width="450"/>

*Figure: Overview of the CLASSER framework.*

## Sample Usage

You can use the AWED-FiNER agentic tool to interact with expert models trained using this framework. Below is an example using the `smolagents` library:

```python
from smolagents import CodeAgent, HfApiModel
from tool import AWEDFiNERTool

# Initialize the expert tool
ner_tool = AWEDFiNERTool()

# Initialize the agent (using a model of your choice)
agent = CodeAgent(tools=[ner_tool], model=HfApiModel())

# The agent will automatically use AWED-FiNER for specialized NER
# Case: Processing a vulnerable language (Bodo)
agent.run("Recognize the named entities in this Bodo sentence: 'बिथाङा दिल्लियाव थाङो।'")
```

## CLASSER Dataset Statistics

<table>
<thead>
    <tr>
      <th rowspan="2">Language</th>
      <th colspan="3">Train set</th>
      <th colspan="3">Development set</th>
      <th colspan="4">Test set</th>
    </tr>
    <tr>
      <th>Sentences</th><th>Entities</th><th>Tokens</th>
      <th>Sentences</th><th>Entities</th><th>Tokens</th>
      <th>Sentences</th><th>Entities</th><th>Tokens</th><th>IAA&nbsp;(κ)</th>
    </tr>
  </thead>
  <tbody align="center">
    <tr>
      <td><b>Assamese (as)</b></td>
      <td>140,257</td><td>204,611</td><td>1,972,697</td>
      <td>15,585</td><td>15,763</td><td>219,114</td>
      <td>1,000</td><td>1,407</td><td>14,270</td><td><b>0.901</b></td>
    </tr>
    <tr>
      <td><b>Bodo (brx)</b></td>
      <td>212,835</td><td>302,713</td><td>2,958,455</td>
      <td>23,649</td><td>33,808</td><td>329,145</td>
      <td>1,000</td><td>1,423</td><td>14,082</td><td><b>0.875</b></td>
    </tr>
    <tr>
      <td><b>Marathi (mr)</b></td>
      <td>611,902</td><td>889,217</td><td>8,135,813</td>
      <td>67,990</td><td>97,943</td><td>948,020</td>
      <td>1,000</td><td>1,443</td><td>13,996</td><td><b>0.887</b></td>
    </tr>
    <tr>
      <td><b>Nepali (ne)</b></td>
      <td>414,561</td><td>617,957</td><td>5,531,683</td>
      <td>46,062</td><td>64,098</td><td>642,489</td>
      <td>1,000</td><td>1,436</td><td>14,142</td><td><b>0.882</b></td>
    </tr>
    <tr>
      <td><b>Sanskrit (sa)</b></td>
      <td>265,114</td><td>378,287</td><td>3,488,871</td>
      <td>29,458</td><td>40,589</td><td>377,306</td>
      <td>1,000</td><td>1,412</td><td>12,925</td><td><b>0.861</b></td>
    </tr>
  </tbody>
</table>

*Note: IAA (Inter-Annotator Agreement) scores are represented using Cohen's κ.*

## Citation

If you use this dataset, please cite the following papers:

```bibtex
@misc{kaushik2026awedfiner,
  title        = {AWED-FiNER: Agents, Web Applications, and Expert Detectors for Fine-grained Named Entity Recognition across 36 Languages for 6.6 Billion Speakers},
  author       = {Kaushik, Prachuryya and Anand, Ashish},
  year         = {2026},
  note         = {arXiv preprint, submitted},
  archivePrefix= {arXiv},
  eprint       = {submit/7163987}
}

@inproceedings{kaushik2025classer,
  title     = {{CLASSER}: Cross-lingual Annotation Projection enhancement through Script Similarity for Fine-grained Named Entity Recognition},
  author    = {Kaushik, Prachuryya and Anand, Ashish},
  booktitle = {Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics},
  year      = {2025},
  publisher = {Association for Computational Linguistics},
  note      = {Main conference paper}
}
```