Contextual Summarization of Email Threads

Abstract

Email has become the primary medium for official communication, enabling the exchange of text, files, and attachments among recipients. A single email thread often encapsulates extensive information shared among multiple participants across diverse topics, resulting in complex and lengthy discussions. Effective summarization of email threads is crucial for generating concise and precise summaries without omitting critical details. This study presents a novel methodology for generating contextual summaries of email threads. The proposed framework begins with constructing a custom dataset comprising university event related email threads (2 < thread length < 6) annotated for summarization tasks. The summarization process involves a three step approach: (1) clustering semantically similar sentences using K-Means and Agglomerative Hierarchical Clustering, (2) extracting contextual information from clusters using Latent Dirichlet Allocation (LDA) and Key Phrase Extraction, and (3) generating abstractive summaries for each contextual cluster using pre-trained transformer models (BART and T5). The proposed approach was systematically evaluated at each stage using standard automatic evaluation metrics, demonstrating its effectiveness in identifying and condensing essential information from lengthy threads. The results highlight the potential of this methodology to streamline email thread summarization. Future work will explore advanced techniques and models to enhance the quality and applicability of contextual summarization further.