Interactive Expansion-Based Summarization with Deep Reinforcement Learning

A deep RL based solution for the interactive expansion-based summarization task. Two subtasks are addressed: query-assisted summarization and suggested-queries extraction. Each subtask has an implemented model here.

Resources for the paper: Interactive Query-Assisted Summarization via Deep Reinforcement Learning

Our deep RL models are based on the RL-MMR system, borrowing code from their repository.

We use the evaluation framework from here to collect user sessions via controlled crowdsourcing and evaluate the sessions.

General Info

Purpose of the task:

• Explore a document set interactively.
• An InterExp system provides an initial summary, and then expands on it in response to queries submitted by the user.
• See demo.

Query-assisted summarization:

• Input:
  • document set on a topic
  • textual query (can be empty for a generic summary)
  • history of texts seen (can be empty)
  • length of summary to output
• Ouput: query-biased summary on the topic, accounting for the history, with the requested length
  • if the query and history are empty, the summary is a standard generic summary

Suggested-queries extraction:

• Input:
  • document set on a topic
  • history of texts seen (can be empty)
  • number of suggestions to output
• Output: list of suggested queries (short phrases) for the topic, accounting for the history, with the requested number of suggestions

Code and Folders in the Project

Summarization

Query-assisted summarization model code.

• To train: python train_full_rl.py <args>
  • Example: python train_full_rl.py --reward_summ r1r --reward_query_every 2 --query_encoding_in_input true --ignore_queries_summ_reward true --beta 0.5
• To test: there are several ways to test a trained model (see individual scripts for required arguments)
  • Command line interaction application: test_interaction.py
  • Full dataset test (validation or test sets): test_full.py
  • Test on sessions collected previously (like simulations): test_session_data.py

SuggestedQueries

Suggested queries extraction model code.

• To train: python train_full_rl.py <args>
  • Example: python train_full_rl.py --phrasing_method posregex --max_num_steps 20 --reward_func tf --weight_initial 0.5 --weight_preceding 0.5 --stem True --remove_stop True --beta 0.5
• To test a trained model: test_full.py

QFSE

Based on the application from the InterExp respository, we implemented the summarizer and suggested query extractor modules for our models. Some minor code modifications were made to original files to accomodate changes for our models.

WebApp

Also based on the application from the InterExp respository, we updated the application to support updating suggested queries, since the original method only used a static list of suggested queries initialized once.

Crowdsourcing

See the InterExp respository for the controlled crowdsourcing resources. Here there are collected sessions:

• BaselineOriginal: the sessions from the InterExp project, used for simulation testing at Summarization/test_session_data.py
• Collected: the sessions we collected in this project, comparing two variants of our system to a baseline from the InterExp project.
  • experiment1 was run with summarization saved_model Apr21_12-00-53 (only uses the input query at inference time in the qMMR function)
  • experiment1 was run with summarization saved_model Apr12_23-23-47 (injects the query at train time in all components, and also at inference time)
  • both experiments used suggested queries saved_model kp_Jul07_17-56-44 (reward tf, gamma1=0.5, gamma2=0.9, stemming and stop word removal)

Citation

If any resource from this repository is used, please cite:

@inproceedings{shapira-etal-2022-interactive,
    title = "Interactive Query-Assisted Summarization via Deep Reinforcement Learning",
    author = "Shapira, Ori  and Pasunuru, Ramakanth and Bansal, Mohit and Dagan, Ido and Amsterdamer, Yael",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.184",
    doi = "10.18653/v1/2022.naacl-main.184",
    pages = "2551--2568",
}