File size: 1,919 Bytes
99a86c5 e3ad01c 99a86c5 2c507de 99a86c5 2c507de 6a82177 2c507de 99a86c5 6a82177 | 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 | ---
language: en
license: apache-2.0
base_model: allenai/scibert_scivocab_uncased
pipeline_tag: text-classification
library_name: transformers
tags:
- scientific-text
- citation-intent
datasets:
- allenai/multicite
- allenai/scicite
widget:
- text: "Deep learning has achieved remarkable success in many NLP tasks [1, 2]."
example_title: "Background citation"
- text: "We adopt the evaluation protocol proposed by Smith et al. (2020)."
example_title: "Non-background citation"
---
# SciBERT Citation-Background Classifier
[`allenai/scibert_scivocab_uncased`](https://huggingface.co/allenai/scibert_scivocab_uncased) fine-tuned as a binary classifier of citation intent: does a citation sentence cite prior work as background (`BACKGROUND`) or for any other reason (`NOT_BACKGROUND`)?
## Labels
| id | label | meaning |
|----|----------------|--------------------------------------------------------------|
| 0 | NOT_BACKGROUND | citation used for method, comparison, extension, motivation… |
| 1 | BACKGROUND | citation provides background/context for the citing paper |
## How to use
```python
from transformers import pipeline
clf = pipeline("text-classification", model="hongccccccc/scibert-citation-background-classifier")
clf("Deep learning has achieved remarkable success in many NLP tasks [1, 2].")
# [{'label': 'BACKGROUND', 'score': ...}]
```
## Training
- **Base model:** SciBERT (uncased, scivocab; `BertForSequenceClassification`, single-label, 2 classes)
- **Data:** the combined annotated citation contexts of [SciCite](https://github.com/allenai/scicite) (Cohan et al., 2019) and [MultiCite](https://github.com/allenai/multicite) (Lauscher et al., 2022) — 27,052 instances, 11,635 (43%) labeled `BACKGROUND`
- **Fine-tuned:** August 2023, `transformers` 4.32.0
## Evaluation
F1 score of **0.81** on a held-out test set (80-20 train-test split of the combined dataset).
|