[ACL 2026] From Documents to Segments: Revisiting Topic Modeling via Segment-Based Topic Assignment

Existing topic modeling has primarily used methods that assign a single representative topic to an entire document. However, when a single document contains multiple mixed subjects, compressing them into one topic causes Topic Contamination.

For example, if a review mentioning food, service, and price is classified solely under the topic of 'food', information about service and price is included in the food topic, distorting overall statistics. This is a core cause of reduced corpus-level analysis and search accuracy.

To solve this problem, researchers from LG AI Research proposed a method that assigns topics to Segments, which are smaller semantic units within a document, rather than to the entire document. This restructures the unit of topic modeling from documents to segments, enabling more refined topic analysis.