SegNSP: Revisiting Next Sentence Prediction for Linear Text Segmentation

José Isidro, Filipe Cunha, Purificação Silvano, Alípio Jorge, Nuno Guimarães, Sérgio Nunes, Ricardo Campos · 2026

Read SegNSP: Revisiting Next Sentence Prediction for Linear Text Segmentation by José Isidro, Filipe Cunha, Purificação Silvano, Alípio Jorge, Nuno Guimarães, Sérgio Nunes, Ricardo Campos (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-03474.txt.

Description

José Isidro and six co-authors revive an old training task, next sentence prediction, to split long texts into topical sections. Splitting text well helps summarization, search, and question answering, but topic boundaries are fuzzy and writing styles vary. Next sentence prediction was dropped from modern pretraining, yet it directly models whether one sentence follows another, which is exactly what boundary detection needs.

Their system, SegNSP, predicts whether the next sentence stays on topic without needing topic labels. A segmentation-aware loss and harder negative examples sharpen its sense of discourse flow. The design avoids extra task-specific supervision that other recent methods require. Clean segmentation is quiet infrastructure: many language tools work better once long documents are cut into coherent parts.

Catalog record

CREATORJosé Isidro, Filipe Cunha, Purificação Silvano, Alípio Jorge, Nuno Guimarães, Sérgio Nunes, Ricardo Campos
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-03474

Browse the full library · About the Tower of Babel