ICASSP2023 MUG Challenge Track1 - Topic Segmentation (TS)

Who this competition fits

This suits researchers and practitioners with NLP experience on Mandarin meeting transcripts, especially those able to model topic boundaries in long spoken-language documents. It is open to academia and industry, but Alibaba employees and Zhejiang University faculty/students are excluded.

From registration to a first submission

1. 01

   Mandarin text processing

2. 02

   topic segmentation

3. 03

   long-document modeling

4. 04

   Transformer fine-tuning

5. 05

   ModelScope baseline usage

6. 06

   meeting transcript preprocessing

Before you commit: The core challenge is finding coherent topic boundaries in long, noisy spoken transcripts: the text highlights disfluencies, redundancies, grammar errors, coreference, colloquial language, and fragmented utterances, while sessions may contain several thousand words. The track is also constrained: participants may use only the specified corpus, permitted public pre-trained models, and listed public corpora.