pytorch implementation of Get To The Point: Summarization with Pointer-Generator Networks
- Train with pointer generation and coverage loss enabled
- Training with pointer generation enabled
- How to run training
- Papers using this code
After training for 100k iterations with coverage loss enabled (batch size 8)
ROUGE-1:
rouge_1_f_score: 0.3907 with confidence interval (0.3885, 0.3928)
rouge_1_recall: 0.4434 with confidence interval (0.4410, 0.4460)
rouge_1_precision: 0.3698 with confidence interval (0.3672, 0.3721)
ROUGE-2:
rouge_2_f_score: 0.1697 with confidence interval (0.1674, 0.1720)
rouge_2_recall: 0.1920 with confidence interval (0.1894, 0.1945)
rouge_2_precision: 0.1614 with confidence interval (0.1590, 0.1636)
ROUGE-l:
rouge_l_f_score: 0.3587 with confidence interval (0.3565, 0.3608)
rouge_l_recall: 0.4067 with confidence interval (0.4042, 0.4092)
rouge_l_precision: 0.3397 with confidence interval (0.3371, 0.3420)
After training for 500k iterations (batch size 8)
ROUGE-1:
rouge_1_f_score: 0.3500 with confidence interval (0.3477, 0.3523)
rouge_1_recall: 0.3718 with confidence interval (0.3693, 0.3745)
rouge_1_precision: 0.3529 with confidence interval (0.3501, 0.3555)
ROUGE-2:
rouge_2_f_score: 0.1486 with confidence interval (0.1465, 0.1508)
rouge_2_recall: 0.1573 with confidence interval (0.1551, 0.1597)
rouge_2_precision: 0.1506 with confidence interval (0.1483, 0.1529)
ROUGE-l:
rouge_l_f_score: 0.3202 with confidence interval (0.3179, 0.3225)
rouge_l_recall: 0.3399 with confidence interval (0.3374, 0.3426)
rouge_l_precision: 0.3231 with confidence interval (0.3205, 0.3256)
- Follow data generation instruction from https://github.com/abisee/cnn-dailymail
- Run start_train.sh, you might need to change some path and parameters in data_util/config.py
- For training run start_train.sh, for decoding run start_decode.sh, and for evaluating run run_eval.sh
Note:
-
In decode mode beam search batch should have only one example replicated to batch size https://github.com/atulkum/pointer_summarizer/blob/master/training_ptr_gen/decode.py#L109 https://github.com/atulkum/pointer_summarizer/blob/master/data_util/batcher.py#L226
-
It is tested on pytorch 0.4 with python 2.7
-
You need to setup pyrouge to get the rouge score
- Automatic Program Synthesis of Long Programs with a Learned Garbage Collector NeuroIPS 2018 https://github.com/amitz25/PCCoder
- Automatic Fact-guided Sentence Modification AAAI 2020 https://github.com/darsh10/split_encoder_pointer_summarizer
- Resurrecting Submodularity in Neural Abstractive Summarization
- StructSum: Summarization via Structured Representations EACL 2021
- Concept Pointer Network for Abstractive Summarization EMNLP'2019 https://github.com/wprojectsn/codes
- VAE-PGN based Abstractive Model in Multi-stage Architecture for Text Summarization INLG2019
- Clickbait? Sensational Headline Generation with Auto-tuned Reinforcement Learning EMNLP'2019 https://github.com/HLTCHKUST/sensational_headline
- Abstractive Spoken Document Summarization using Hierarchical Model with Multi-stage Attention Diversity Optimization INTERSPEECH 2020
- Nutribullets Hybrid: Multi-document Health Summarization NAACL 2021
- A Corpus of Very Short Scientific Summaries CoNLL 2020
- Towards Faithfulness in Open Domain Table-to-text Generation from an Entity-centric View AAAI 2021
- CDEvalSumm: An Empirical Study of Cross-Dataset Evaluation for Neural Summarization Systems Findings of EMNLP2020
- A Study on Seq2seq for Sentence Compression in Vietnamese PACLIC 2020
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent