|
--- |
|
language: en |
|
tags: |
|
- abstractive summarization |
|
model-index: |
|
- name: kubershahi/pegasus-inshorts |
|
results: |
|
- task: |
|
type: abstractitive summarization |
|
name: abstractive summarization |
|
dataset: |
|
name: inshorts |
|
type: inshorts |
|
config: inshorts |
|
split: train |
|
metrics: |
|
- name: ROUGE-1 |
|
type: rouge |
|
value: 4.2525 |
|
verified: true |
|
- name: ROUGE-2 |
|
type: rouge |
|
value: 4.2525 |
|
verified: true |
|
- name: ROUGE-L |
|
type: rouge |
|
value: 17.4469 |
|
verified: true |
|
- name: ROUGE-LSUM |
|
type: rouge |
|
value: 18.8907 |
|
verified: true |
|
- name: loss |
|
type: loss |
|
value: 3.0317161083221436 |
|
verified: true |
|
- name: gen_len |
|
type: gen_len |
|
value: 20.3122 |
|
verified: true |
|
--- |
|
|
|
|
|
# Problem Statment: |
|
|
|
Given a news article, generate a summary of two-to-three sentences and a headline for the article. The summary should be abstractive rather than extractive. |
|
In abstractive summarization, new sentences are generated as part of the summary and the sentences in the summary might not be present in the news article. |
|
|
|
|
|
# Model Description |
|
|
|
This model builds on the [google/pegasus-large](https://huggingface.co/google/pegasus-large) model by finetuning it on a custom summary-headline dataset called [inshorts](https://github.com/kubershahi/ashoka-aml/blob/master/dataset/news_headline.csv). |
|
After finetuning, to generate an appropriate headline of an article, get the summary of the article first from the pegasus-large model and then pass the summary through this model. |
|
The two-way approach was taken to get apt headline from summary rather then generating the headline from the pegasus-large itself. |
|
|
|
|
|
For more details about the project, click [here](https://github.com/kubershahi/ashoka-aml). |
|
|
|
|
|
|
