README.md

---
license: apache-2.0
datasets:
- allenai/mslr2022
language:
- en
pipeline_tag: summarization
---

# Longformer for biomedical extractive summarization

## Description
Work done for my [Bachelor's thesis](https://amslaurea.unibo.it/id/eprint/29686).

Longformer fine-tuned on [MS^2](https://github.com/allenai/mslr-shared-task) for extractive summarization.\
The model architecture is similar to [BERTSum](https://github.com/nlpyang/BertSum).\
Training code is available at [biomed-ext-summ](https://github.com/NotXia/biomed-ext-summ).

## Usage
```python
summarizer = pipeline("summarization",
  model = "NotXia/longformer-bio-ext-summ",
  tokenizer = AutoTokenizer.from_pretrained("NotXia/longformer-bio-ext-summ"),
  trust_remote_code = True,
  device = 0
)

sentences = ["sent1.", "sent2.", "sent3?"]
summarizer({"sentences": sentences}, strategy="count", strategy_args=2)
>>> (['sent1.', 'sent2.'], [0, 1])
```

### Strategies
Strategies to summarize the document:
- `length`: summary with a maximum length (`strategy_args` is the maximum length).
- `count`: summary with the given number of sentences (`strategy_args` is the number of sentences).
- `ratio`: summary proportional to the length of the document (`strategy_args` is the ratio [0, 1]).
- `threshold`: summary only with sentences with a score higher than a given value (`strategy_args` is the minimum score).