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
Summarizing a conversation presents unique challenges compared to summarizing structured text like news articles. Dialogues involve multiple speakers with distinct roles, goals, and linguistic styles. For instance, in customer service interactions, a customer presents a problem while an agent provides a solution. Most existing summarization models treat dialogue as a monolithic block of text, failing to capture these speaker dynamics. Furthermore, the scarcity of high-quality, human-annotated "gold standard" dialogue summaries makes supervised learning difficult and expensive.
To address these issues, we introduce SuTaT (Summarization for Tete-a-Tetes), the first unsupervised abstractive summarization model specifically designed for one-on-one dialogues. SuTaT respects the interaction between participants by modeling their utterances separately while maintaining their inherent correlations.
Methodology
The SuTaT architecture is built upon two primary components designed to work without labeled summary data.
Conditional Generative Module
The core of SuTaT is a Sequential Interacting Variational Autoencoder (SIVAE) framework. It utilizes separate encoders and decoders for each speaker (e.g., customer and agent). A critical feature is the modeling of dependencies: the agent’s latent representation is conditioned on the customer’s. This reflects the real-world logic of a "tete-a-tete," where an agent's response is directly influenced by the customer's input.
Unsupervised Summarization and Copy Mechanism
To generate summaries, SuTaT employs a sentence-level self-attention mechanism. This allows the model to filter out uninformative "small talk" (e.g., "Hello," "You're welcome") and prioritize content-heavy utterances. Additionally, a Partial Copy Mechanism is implemented to ensure factual accuracy. This heuristic identifies factual tokens (like dates or reference numbers) in the source text and copies them into the summary when the decoder predicts a word from a predefined factual list.
The model is trained using a dual objective: a reconstruction loss to ensure the original dialogue can be recreated, and a similarity loss to keep the generated summary semantically aligned with the input.
Evaluation and Results
We evaluated SuTaT on two benchmarks: MultiWOZ-2.0 (goal-oriented dialogues like hotel bookings) and Taskmaster-1 (scenarios like ordering pizza). Since these datasets lack per-speaker summaries, we collected a small set of reference summaries via Mechanical Turk for evaluation purposes.
Automatic Evaluation
SuTaT-LSTM consistently outperformed other unsupervised baselines. As shown in the results table, it achieved significantly higher ROUGE scores across both datasets.
On MultiWOZ, SuTaT-LSTM reached a ROUGE-1 score of 28.51 for customer summaries, surpassing baselines like MeanSum (25.93) and Copycat (26.86). Interestingly, the LSTM-based variant generally outperformed the Transformer version, likely because Transformer decoders can sometimes ignore encoder inputs (posterior collapse).
Human Evaluation and Versatility
In human evaluations, SuTaT's "Correlation" score—measuring how well customer and agent summaries relate—reached 4.25/6.0, nearly matching human reference scores (4.52). Beyond summarization, the model's latent spaces proved effective for dialogue classification and single-turn conversation generation.
Discussion and Future Work
SuTaT demonstrates that modeling speaker interaction is key to effective dialogue summarization. By conditioning the agent's latent space on the customer's, the model learns the context of the interaction, leading to more coherent summaries.
However, limitations remain. Currently, SuTaT retains the original pronouns (1st/2nd person), whereas professional summaries typically use the 3rd person. Future work will explore automatic pronoun conversion and more robust, learned factual grounding mechanisms. Additionally, extending this architecture to multi-party dialogues (3+ speakers) remains an open challenge.
Summary updated