Pattern Recognition and Image Analysis Computer Vision and Pattern Recognition

Contextual Learning Strategies for Dialogue Summarization Using Large Language Models

Penyasov, D. E., Turalchuk, K. A.
Springer Nature · 2026-06-01 · DOI 10.1134/s1054661826700252

Abstract

Abstract Summarization of dialogues is a challenging subproblem in natural language processing that requires accounting for fragmented, multi-party dialogue, and frequent changes of topics and dependencies inherent in conversational data. Despite advances in automatic text summarization, most modern models are trained on monologue corpora, such as news papers, which significantly differ from dialogues in both structure and content. As a result, these models often exhibit low transferability to conversational summarization problems. An additional complicating factor is the scarcity of high-quality labeled conversational datasets, which makes it difficult to adapt effectively models. In this paper, an alternative, data-driven approach that uses large language models as annotators in a context-aware learning framework is studied. A method for selecting examples for contextual learning is proposed using a retrieval-augmented approach, in which the selection of examples is based on their semantic similarity to the input dialogue. In a series of controlled experiments, the effect of the quality of demonstration examples and their selection strategy on the efficiency of summarization was estimated. Experimental results show that context-aware example selection based on semantic similarity significantly improves the quality of automatic dialogue summarization and reduces the need for costly manual annotation.

Read at publisher →