AI Learns to Summarize Conversations Without Annotated Examples Using Bridge Domains

Dialogue summarization has become one of the most consequential capabilities in modern natural language processing. Every customer service interaction, medical consultation, virtual meeting, and technical support exchange generates a conversation that someone, eventually, must condense into a usable summary. End-to-end neural summarization models perform this task impressively well, but their performance rests on a hidden foundation: large collections of dialogues paired with human-written reference summaries. Building such annotated datasets is expensive and slow, and for many specialized domains—clinical dialogues, legal consultations, industrial support calls—the data simply does not exist in annotated form. A new study published in Scientific Reports by Ye Wang, Jian Jiang, Yichen Qi, Dongxue Zhang, Xudong Niu, Lei Sheng, and Hongwei Lou of the Advanced Computing and Digital Engineering Research Center at the Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, tackles this bottleneck head-on with an unsupervised domain adaptation framework that teaches a summarization model to handle dialogues even when no annotated dialogue summaries are available at all.

The core insight of the research is that the gap between two kinds of text is not a binary chasm but a graded transition. On one side sits the source domain: abundant, richly annotated data from a related text type, such as news articles, where summaries are plentiful and the writing is formal, structured, and succinct. On the other side sits the target domain: dialogues, which are informal, repetitive, riddled with disfluencies, and shaped by multiple speakers talking over and around each other. A summarization model trained on news learns to compress formal prose, but when confronted with a sprawling multi-party conversation it stumbles over turn-taking, overlapping references, and the loose coherence of spoken language. Direct transfer from source to target fails because the statistical properties of the two domains differ too sharply. The researchers’ answer is to stop treating the transfer as a single leap and instead construct a sequence of intermediate domains that bridge the gap incrementally.

The technical heart of the method is a ratio-adaptive mixup approach for data augmentation. Mixup, in its classic form, blends two training examples together, typically by interpolating their representations or their text, to create a synthetic example that lies somewhere between them. Applied to domain adaptation, mixup can generate pseudo-examples that mix the stylistic features of news text with those of dialogue text. The innovation here is the adjective: ratio-adaptive. Rather than generating a single blended domain or a fixed set of blends, the method dynamically adjusts the mixing ratio, producing multiple inferred transitional domains that span the continuum from pure source to pure target. At one end of this continuum, examples are dominated by the formal, succinct characteristics of the source; at the other end, they increasingly resemble the informal, repetitive, multi-party texture of real dialogues. By controlling these ratios, the framework effectively manufactures a curriculum of synthetic domains.

This curriculum matters because of how neural networks learn. A model fine-tuned abruptly on data from a very different distribution tends to forget the competencies it acquired during pretraining and source-domain training, a phenomenon related to catastrophic forgetting. When the target domain is not merely different but also unlabeled, the problem compounds: the model has no ground-truth summaries to anchor its adaptation. Intermediate domains soften the distribution shift at each step. The model first adapts from news to a slightly dialogue-flavored hybrid, then to a more dialogue-like hybrid, and so on, with each transition small enough that the knowledge accumulated at the previous stage remains useful. Domain knowledge thus flows gradually from source to target, rather than being forced through a discontinuity that destroys it.

Guiding this gradual transfer requires a way to judge where a given example sits on the source-to-target spectrum and how well the model is handling it. The framework incorporates heuristic representations—structured signals that characterize the stylistic and structural properties of each domain—to steer the augmentation process. Alongside these, the authors employ a language model discriminator, a component trained to distinguish between text distributions. The discriminator provides feedback on how convincing the synthetic intermediate examples are and how far the adapted model has moved toward the target distribution. This adversarial or discriminative signal acts as a compass during training, preventing the adaptation process from drifting into degenerate outputs that are neither good summaries nor plausible dialogue-derived text.

A particularly elegant feature of the proposed system is the development of two complementary models: a source-dominant model and a target-dominant model. The source-dominant variant leans more heavily on the annotated source data, preserving the strong summarization competence that comes from abundant supervision, while the target-dominant variant pushes further toward the unlabeled target dialogues, capturing their distinctive conversational characteristics. Because the two models emphasize different ends of the domain continuum, they exhibit complementary strengths and complementary failure modes. Where the source-dominant model produces fluent, well-structured summaries but may miss conversational nuances, the target-dominant model is more attuned to the idiosyncrasies of dialogue but risks drifting from the summarization conventions learned under supervision. Combining or selecting between them allows the system to balance fidelity to the summarization task against sensitivity to the target domain.

The evaluation of the framework was conducted on three widely used dialogue summarization benchmarks: SAMSum, DIALOGSum, and TODSum. These datasets cover distinct conversational settings. SAMSum consists of chat-like dialogues with manually annotated summaries, DIALOGSum offers a broader and more diverse collection of everyday conversations, and TODSum focuses on task-oriented dialogues, such as those occurring in service and booking scenarios, where the summary must capture goals, slots, and outcomes rather than narrative content. Testing across all three is significant because it probes whether the method generalizes across different flavors of conversational data rather than overfitting to a single benchmark’s style. According to the authors, extensive experiments on these benchmarks demonstrate the superiority of their approach over the alternatives they compared against, in the challenging setting where annotated target-domain summaries are unavailable and target data itself is limited.

The practical implications extend well beyond benchmark scores. Consider a hospital system that wants automatic summaries of doctor-patient conversations, a call center that wants condensed records of support tickets, or a collaboration platform that wants meeting recaps. In each case, collecting thousands of dialogues and paying professionals to write reference summaries is prohibitively costly, and privacy regulations may further restrict data sharing. The unsupervised adaptation paradigm means an organization can start from a publicly available annotated summarization dataset in a resource-rich domain, such as news, and adapt the resulting model to its conversational data without any annotation effort on the target side. The requirement that even unlabeled target data may be limited makes the approach more realistic still, since many organizations hold only modest amounts of domain text rather than the massive corpora typically assumed by modern machine learning.

The work also contributes to a broader conceptual shift in how researchers think about domain adaptation for generative tasks. Much of the classic domain adaptation literature addresses classification, where the goal is to align feature representations so that a decision boundary transfers between domains. Summarization is harder: the output is a full text sequence, and the notion of what makes a good summary may itself shift between domains. A news summary emphasizes facts, entities, and events; a dialogue summary must track who said what, resolve pronouns across speaker turns, and distill intentions and outcomes from a messy exchange. By explicitly modeling the transition between these output regimes with augmented intermediate domains, the method acknowledges that adaptation for generation is a journey through changing task definitions, not merely a realignment of input features.

There are, of course, open questions that the study leaves for future work. The quality of synthetic intermediate domains ultimately depends on the mixup procedure and the heuristic representations that guide it, and one can ask how the approach scales when the source and target domains are even more distant than news and dialogue—for example, when adapting from written text to highly technical or multilingual conversations. The interplay between the source-dominant and target-dominant models also invites further study into optimal combination strategies. Nevertheless, the research, published open access under a Creative Commons license and supported by the Innovative Projects international funds of the Changchun Institute of Optics, Fine Mechanics and Physics and the Changguang-Fudan Collaborative Fund Project, marks a meaningful step toward summarization systems that can be deployed wherever conversations happen, regardless of whether anyone has ever labeled a single example there. As intelligent dialogue systems spread through medicine, commerce, and daily life, techniques that break the dependence on costly annotation may prove as important as the models they enable.

Subject of Research: Unsupervised domain adaptation for dialogue summarization using augmented heuristic intermediate domains

Article Title: Unsupervised domain adaptation for dialogue summarization with augmented heuristic intermediate domains

Article References: Wang, Y., Jiang, J., Qi, Y., Zhang, D., Niu, X., Sheng, L., & Lou, H. (2026). Unsupervised domain adaptation for dialogue summarization with augmented heuristic intermediate domains. Scientific Reports. https://doi.org/10.1038/s41598-026-74855-z

Image Credits: AI Generated

DOI: 10.1038/s41598-026-74855-z

Keywords: dialogue summarization, unsupervised domain adaptation, natural language processing, data augmentation, mixup, intermediate domains, language model discriminator, SAMSum, DIALOGSum, TODSum, transfer learning, machine learning