# Dial2vec: Self-Guided Contrastive Learning of Unsupervised Dialogue Embeddings

**Authors:** Che Liu, Rui Wang, Junfeng Jiang, Yongbin Li, Fei Huang  
**arXiv:** [2210.15332](https://arxiv.org/abs/2210.15332)  
**Date:** 2022-10-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.15332v1/

## Key points

- Authors introduce the task of learning unsupervised dialogue embeddings
- Existing approaches often ignore conversational interactions, leading to poor performance
- Authors propose a self-guided contrastive learning approach called dial2vec
- Dial2vec captures conversational interaction patterns between interlocutors to guide embedding learning
- Dialogue embedding is obtained by aggregating embeddings from all interlocutors
- Comprehensive benchmark with six dialogue datasets and three evaluation tasks: domain categorization, semantic relatedness, and dialogue retrieval
- Dial2vec outperforms strongest baseline on all three tasks with average improvements in purity, Spearman's correlation, and mean average precision (MAP)
- Analysis shows informative and discriminative embeddings for both interlocutors by leveraging conversational interactions
- Best performance achieved through an interlocutor-level pooling strategy for aggregation of embeddings
- Codes and data available publicly at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/dial2vec
- Overall, dial2vec is a novel approach for learning unsupervised dialogue embeddings that considers conversational interactions between interlocutors

## Summary

In the paper "Dial2vec: Self-Guided Contrastive Learning of Unsupervised Dialogue Embeddings," authors Che Liu, Rui Wang, Junfeng Jiang, Yongbin Li, and Fei Huang introduce the task of learning unsupervised dialogue embeddings. They highlight that existing approaches often ignore the conversational interactions between interlocutors, resulting in poor performance. To address this issue, the authors propose a self-guided contrastive learning approach called dial2vec. Dial2vec treats a dialogue as an information exchange process and captures the conversational interaction patterns between interlocutors to guide the learning of embeddings corresponding to each interlocutor. The dialogue embedding is obtained by aggregating the embeddings from all interlocutors. To evaluate their approach, the authors establish a comprehensive benchmark consisting of six widely-used dialogue datasets. They consider three evaluation tasks: domain categorization, semantic relatedness, and dialogue retrieval. Dial2vec outperforms the strongest baseline on all three tasks with average improvements of 8.7 points in terms of purity, 9.0 points in Spearman's correlation, and 13.8 points in mean average precision (MAP). Further analysis reveals that dial2vec obtains informative and discriminative embeddings for both interlocutors by leveraging conversational interactions. The best performance is achieved when aggregating these embeddings through an interlocutor-level pooling strategy. The authors provide all codes and data publicly at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/dial2vec . Overall, this paper introduces dial2vec as a novel approach for learning unsupervised dialogue embeddings that considers conversational interactions between interlocutors. Experimental results demonstrate its effectiveness across various evaluation tasks and datasets.