seq-to-mind / coref_dial_summ

One implementation of the paper "Coreference-Aware Dialogue Summarization".

Introduction

One implementation of the paper "Coreference-Aware Dialogue Summarization" in SIGDIAL 2021 (Best Paper Award).

Package Requirements

1. pytorch==1.7.1
2. transformers==4.8.2
3. click==7.1.2
4. sentencepiece==0.1.92
5. allennlp==2.6.0
6. allennlp-models==2.6.0

Dialogue Coreference Resolution

1. Download the off-the-shelf model from AllenNLP:
   allennlp-public-models/coref-spanbert-large-2021.03.10
2. You can obtain the coreference resolution from the model with the script:
   ./dialogue_coreference/dialogue_coreference.py
3. For dialogue coreference resolution post-processing, you can call the function in the file:
   ./dialogue_coreference/reading_and_writing_as_input_keep_SPAN.py
4. For end-to-end conversation samples construction with coreference information, please run or refer to the script:
   ./dialogue_coreference/end2end_build_data.py
5. Noted that the processed samples will be tokenized via the RoBERTa/BART sub-word tokenization.

Coref-Aware Summarization

1. The data after dialogue coreference resolution can be used to train the coref-aware summarizer.
2. You can read the samples in text format, then read the tokenized id/coreference information from each row.
   For instance, each row in the file train.source contains information as below:
   Text Tokens after BART tokenization ##### Token IDs after BART tokenization ##### Start Token ID of One Coreference-Linked Span ##### Target Token ID of One Coreference-Linked Span ##### Token Number after BART tokenization
3. For our implementation, you will need to replace the original generation_utils.py and modeling_bart.py in the Transformers library, with the corresponding files in this repo.
4. You can search the keyword 'coref' in our updated generation_utils.py and modeling_bart.py to see the implementation details.
5. See running configurations in the global_config.py file.
6. We provide the self-attention-layer and Transformer head manipulation methods to incorporate coreference information, which are computationally efficient.

Generated Summaries for SAMSum Corpus

See the predictions from the coreference-aware summarization model of SAMSum test set in ./model_outputs/

Citation

@inproceedings{liu-etal-2021-coreference,
    title = "Coreference-Aware Dialogue Summarization",
    author = "Liu, Zhengyuan and Shi, Ke and Chen, Nancy",
    booktitle = "Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue",
    month = jul,
    year = "2021",
    address = "Singapore and Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.sigdial-1.53",
    pages = "509--519",
}

coref_dial_summ's Issues

How to get the result of Coreference Resolution?

I've run dialogue_coreference.py and end2end_build_data.py.
However I get the result like this:
_# ĠHannah : ĠHey , Ġdo Ġyou Ġhave ĠBetty 's Ġnumber ? Ġ# ĠAmanda : ĠLem me Ġcheck Ġ# ĠHannah : Ġ< file _ gif > Ġ# ĠAmanda : ĠSorry , Ġcan 't Ġfind Ġit . Ġ# ĠAmanda : ĠAsk ĠLarry Ġ# ĠAmanda : ĠHe Ġcalled Ġher Ġlast Ġtime Ġwe Ġwere Ġat Ġthe Ġpark Ġtogether Ġ# ĠHannah : ĠI Ġdon 't Ġknow Ġhim Ġwell Ġ# ĠHannah : Ġ< file _ gif > Ġ# ĠAmanda : ĠDon 't Ġbe Ġshy , Ġhe 's Ġvery Ġnice Ġ# ĠHannah : ĠIf Ġyou Ġsay Ġso .. Ġ# ĠHannah : ĠI 'd Ġrather Ġyou Ġtexted Ġhim Ġ# ĠAmanda : ĠJust Ġtext Ġhim Ġ# ĠHannah : ĠUr gh .. ĠAlright Ġ# ĠHannah : ĠBye Ġ# ĠAmanda : ĠBye Ġbye ##### [10431, 11029, 35, 11468, 6, 109, 47, 33, 16666, 18, 346, 116, 849, 10641, 35, 13956, 1794, 1649, 849, 11029, 35, 28696, 21710, 1215, 47324, 15698, 849, 10641, 35, 19719, 6, 64, 75, 465, 24, 4, 849, 10641, 35, 12738, 6045, 849, 10641, 35, 91, 373, 69, 94, 86, 52, 58, 23, 5, 2221, 561, 849, 11029, 35, 38, 218, 75, 216, 123, 157, 849, 11029, 35, 28696, 21710, 1215, 47324, 15698, 849, 10641, 35, 1599, 75, 28, 9152, 6, 37, 18, 182, 2579, 849, 11029, 35, 318, 47, 224, 98, 7586, 849, 11029, 35, 38, 1017, 1195, 47, 25199, 123, 849, 10641, 35, 1801, 2788, 123, 849, 11029, 35, 9163, 4147, 7586, 41881, 849, 11029, 35, 36255, 849, 10641, 35, 36255, 15687] ##### [13, 19, 27, 34, 37, 42, 46, 56, 58, 62, 65, 73, 80, 85, 88, 93, 95, 98, 100, 102, 106, 108, 115, 119] ##### [6, 1, 13, 10, 27, 37, 8, 19, 56, 44, 58, 42, 62, 65, 73, 85, 93, 88, 80, 98, 100, 95, 108, 102] ##### 123

where is the result of Coreference Resolution?