This repository maintains the code, data for the paper ArQuSumm: Argument-aware Quantitative Summarization of Online Conversations (AAAI2026).
We explored a new task, Argument-aware Quantitative Summarization to generate reveal and summarize the claim-reason structure of arguments in conversations.
It is recommended to set up the environment and install required libraries using conda. It is also recommended that the machine should have GPUs to perform inference at a reasonable time.
conda create --name arqusumm python=3.9
conda activate arqusummconda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidiaconda install pytorch torchvision torchaudio cpuonly -c pytorchconda install pytorch::pytorch torchvision torchaudio -c pytorchFor other versions, please visit: https://pytorch.org/get-started/locally/
We need some additional packages to run the code. The list of packages is listed in requirements.txt. On the main directory of the repository, run:
pip install -r requirements.txt
We propose a pipeline framework for Argument-aware Quantitative Summarization, namely ArQuSumm
for generating claim-reason structured summary.
To reveal the underlying argument structure within sentences, ArQuSumm leverages LLM few-shot learning grounded in the argumentation theory to identify propositions within sentences and their claim-reason relationships.
For quantitative summarization, ArQuSumm employs argument structure-aware clustering algorithms to aggregate arguments and quantify their support.
ArQuSumm consists of 3 components/stages:
- Stage 1: Argument Proposition Extraction
- Stage 2: Argument-aware Clustering
- Stage 3: Structured Summary Generation
We released data from the NYT, Reddit and StackExchange domains used in our evaluation tasks (dimensions) of claim-reason structured summaries.
The data derives from the test set of the ConvoSumm dataset from the corresponding domains, which contains 250 examples per domain.
Datasets can be accessed under the data/ folder,
following the nyt/, reddit/, and stack/ subdirectories for each dataset.
Files in each folder:
raw_input.csv: Raw input comments and crowdsourced summaries collected from ConvoSumm dataset.source_comments_arg_extracted.csv: Comments already analyzed for ABSA predictions (Stage 1)claim_clustered.csv: clusters of claims extracted from user comments (Stage 2)claim_reason_clustered.csv: clusters of reasons supporting claim clusters (Stage 2 )summaries.csv: Claim-reason structured summary generated from claim and reason clusters (Stage 3 )gold_summaries_arg_extracted.csv: Claims, reasons extracted from crowdsourced summaries (for automatic evaluation)