---
title: "对话式论辩研究综述 — 魏忠钰 2023"
description: "Computational argumentation has mostly studied monological texts such as essays and speeches."
authors: ["魏忠钰", "丁佳玙", "沈晨晨", "高源", "梁敬聪", "纪程炜", "林嘉昱", "黄萱菁"]
year: 2023
venue: "中文信息学报 (Journal of Chinese Information Processing)"
source: "http://jcip.cipsc.org.cn/article/id/zwxxxb_3617"
licence: "Free access (jcip.cipsc.org.cn), no open licence stated — readonly; © 中文信息学报"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/papers/wei-2023-dialogical-argumentation/"
---
[Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/wei-2023-dialogical-argumentation/)

[home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2023](https://md.research.elamur.ai/free/papers/2023/page/7/) / 对话式论辩研究综述

# 对话式论辩研究综述

**Authors:** 魏忠钰, 丁佳玙, 沈晨晨, 高源, 梁敬聪, 纪程炜, 林嘉昱, 黄萱菁 · **Year:** 2023 · **Venue:** 中文信息学报 (Journal of Chinese Information Processing) · **Licence:** Free access (jcip.cipsc.org.cn), no open licence stated — readonly; © 中文信息学报

**Links:** [Publisher's page](http://jcip.cipsc.org.cn/article/id/zwxxxb_3617) · [PDF](http://jcip.cipsc.org.cn/cn/article/pdf/preview/zwxxxb_3617.pdf)

## Summary
Computational argumentation has mostly studied monological texts such as essays and speeches. Debates with several participants, like online forums, courtrooms and classroom discussions, have received much less systematic attention. This survey from Fudan University maps the emerging field of dialogical argumentation. It describes a typical analysis pipeline. The input is an original text and a reply. Claims are extracted from each, and then the pipeline matches interacting arguments, labels how they interact (support, attack, rebuttal) and judges argument quality. The review also covers stance analysis, generation tasks (summary-style and targeted argument generation), debating applications, public datasets and evaluation practice. For each task it traces how models developed, from feature-based classifiers to discrete VAEs, knowledge-graph encoders, BERT-based matchers and cross-encoding architectures. The authors conclude that the field is still at an early stage. They name multimodal dialogue analysis and knowledge-infused argument generation as promising next steps. For teams building argument-mining resources in new languages or domains, the paper is a compact overview of task definitions and baselines.

## Key points
- **Interactive argument pair identification** was introduced by Ji et al. (2019) using ChangeMyView (CMV) data. Quoted sentences (marked with ">") and the reply sentence that follows them gave 13,046 positive pairs. Negatives were sampled 1:4 from the same reply, so the task is to pick the right reply among five candidates. Ji et al. solved it with a discrete variational autoencoder.
- **Follow-up work on argument pairs:**
  - Yuan et al. built a CMV argumentation knowledge graph with over 200k nodes and over 800k edges, and encoded reasoning paths with a Transformer.
  - BERT outperformed the original model.
  - Wu et al. added Mixup data augmentation and context modelling.
- **Argument Pair Extraction (APE)** pulls interacting argument pairs from two long texts, for example on the Review-Rebuttal peer-review dataset. Proposed methods include:
  - Bao et al.'s mutual-guidance framework with a sentence relation graph.
  - Cheng et al.'s attention-guided multi-layer multi-cross encoding, which treats prediction as table filling.
- **Stance analysis** spans several lines of work:
  - Supervised systems such as Sobhani et al., who performed strongly at SemEval-2016 Task 6.
  - Methods for unknown targets (Augenstein; Wachsmuth's route kernels; Kobbe's consequence patterns).
  - Cross-target transfer (SEKT by Zhang et al. 2020; Hardalov et al. 2021, using mixture-of-experts with domain-adversarial training).
- **Argument quality** is described along three dimensions: logical, rhetorical and dialectical. Assessment can be absolute (a score or class) or relative (pairwise comparison, such as predicting debate winners or comparing SVM vs Bi-LSTM convincingness models).

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/shelves/linguistics/) · [Natural language processing](https://md.research.elamur.ai/free/shelves/nlp/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@article{2023,
  title = {{对话式论辩研究综述}},
  author = {魏忠钰 and 丁佳玙 and 沈晨晨 and 高源 and 梁敬聪 and 纪程炜 and 林嘉昱 and 黄萱菁},
  journal = {中文信息学报 (Journal of Chinese Information Processing)},
  year = {2023},
  url = {http://jcip.cipsc.org.cn/article/id/zwxxxb_3617}
}
```

- [← previous 基于语言模型增强的中文关系抽取方法](https://md.research.elamur.ai/free/papers/xue-2023-language-model-enhanced/)
- [next → 常识问答研究综述](https://md.research.elamur.ai/free/papers/fan-2023-commonsense-question-answering/)
- [↑ back to papers of 2023](https://md.research.elamur.ai/free/papers/2023/page/7/)