Abstract High-quality dialogue-summary paired data is expensive to produce and domain-sensitive, making abstractive dialogue summarization a challenging task. In this work, we propose the first unsupervised abstractive dialogue summarization model for tete-a-tetes (SuTaT). Unlike standard text summarization, a dialogue summarization method should consider the multi-speaker scenario where the speakers have different roles, goals, and language styles. In a tete-a-tete, such as a customer-agent conversation, SuTaT aims to summarize for each speaker by modeling the customer utterances and the agent utterances separately while retaining their correlations. SuTaT consists of a conditional generative module and two unsupervised summarization modules. The conditional generative module contains two encoders and two decoders in a variational autoencoder framework where the dependencies between two latent spaces are captured. With the same encoders and decoders, two unsupervised summarization modules equipped with sentence-level self-attention mechanisms generate summaries without using any annotations. Experimental results show that SuTaT is superior on unsupervised dialogue summarization for both automatic and human evaluations, and is capable of dialogue classification and single-turn conversation generation.
Introduction
Dialogue summarization is inherently difficult due to two primary factors. First, high-quality "dialogue-summary" pairs are rare and expensive to create, causing supervised models to struggle when applied to new domains. Second, dialogues are interactions between speakers with distinct goals, styles, and vocabularies. For example, in a customer service interaction, the customer describes a problem while the agent provides a solution. Standard unsupervised summarization tools often ignore these speaker dynamics, leading to summaries that lose the "thread" of the conversation or conflate speaker roles.
To address these issues, we introduce SuTaT (Summarization for Tete-a-Tetes), the first unsupervised abstractive summarization model designed specifically for one-on-one dialogues. By utilizing a dual-latent space Variational Autoencoder (VAE) and sentence-level self-attention, SuTaT learns to capture the distinct roles and dependencies between speakers to generate coherent, role-specific summaries.
Methods
SuTaT consists of three main components designed to handle the nuances of two-party interactions:
- Conditional Generative Module: Built on a Sequential Integrated Variational Autoencoder (SIVAE) framework, this module uses separate encoders and decoders for each speaker. It models the dependency between them: the agent’s latent representation is conditioned on the customer’s, mimicking real-world logic where an agent's response depends on the customer's request.
- Sentence-Level Self-Attention: To filter out non-informative utterances (e.g., "Hello"), SuTaT uses a self-attention mechanism to assign higher weights to key utterances, creating a weighted "summary representation" in the latent space.
- Partial Copy Mechanism: To prevent factual "hallucinations," the model employs a heuristic that replaces predicted entity-like words with the most likely factual words from the original source text.
Results and Evaluation
The model was evaluated on two goal-oriented dialogue datasets: MultiWOZ 2.0 (hotel/restaurant bookings) and Taskmaster-1 (pizza orders/movie tickets). Training was entirely unsupervised, using only raw dialogues.
SuTaT-LSTM achieved the highest ROUGE scores across both datasets, significantly outperforming baselines like LexRank and other unsupervised abstractive models like Copycat. For instance, on MultiWOZ, it achieved a ROUGE-1 score of 28.51 for customer summaries, compared to 26.86 for the next best baseline.
Human evaluations confirmed these results, with judges ranking SuTaT higher in terms of Informativeness and Correlation (the alignment between customer and agent summaries). Additionally, the generated summaries proved effective for downstream tasks, achieving an AUC of 0.85 in dialogue classification on MultiWOZ without seeing any domain labels during training.
Conclusion
SuTaT demonstrates that role-specific modeling and speaker dependency are crucial for effective dialogue summarization. While currently limited to two-speaker interactions and first-person perspectives, it provides a robust foundation for unsupervised summarization in goal-oriented domains. Future work will focus on extending the architecture to multi-party conversations and improving factual consistency.
Summary updated