pytorch implementation of Get To The Point: Summarization with Pointer-Generator Networks

1. Train with pointer generation and coverage loss enabled
2. Training with pointer generation enabled
3. How to run training
4. Papers using this code

Train with pointer generation and coverage loss enabled

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)

Training with pointer generation enabled

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)

How to run training:

1. Follow data generation instruction from https://github.com/abisee/cnn-dailymail
2. Run start_train.sh, you might need to change some path and parameters in data_util/config.py
3. 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

Papers using this code:

1. Automatic Program Synthesis of Long Programs with a Learned Garbage Collector https://github.com/amitz25/PCCoder
2. Automatic Fact-guided Sentence Modification https://github.com/darsh10/split_encoder_pointer_summarizer
3. Resurrecting Submodularity in Neural Abstractive Summarization
4. StructSum: Incorporating Latent and Explicit Sentence Dependencies for Single Document Summarization
5. Concept Pointer Network for Abstractive Summarization https://github.com/wprojectsn/codes
6. VAE-PGN based Abstractive Model in Multi-stage Architecture for Text Summarization