Thread Structure Learning on Online Health Forums with Partially Labeled Data

Yunzhong Liu, Jinhe Shi, Yi Chen

Research output: Contribution to journalArticlepeer-review

1 Scopus citations

Abstract

Thread structures, the reply relationships between posts, in online forums are very important for readers to understand the thread content, and for improving the effectiveness of automated forum information retrieval, expert findings, and so on. However, most online forums only have partially labeled structures, which means that some reply relationships are known while the others are unknown. To address this problem, studies have been performed to learn and predict thread structures. However, existing work does not leverage the partially available thread structures to learn the complete thread structure. We have also observed that many online health forums are a type of person-centric forums, where persons are mentioned across posts, providing hints about the reply relationships between posts. In this article, we first proposed to learn the complete thread structures by leveraging the partially known structures based on a statistical machine learning model - thread conditional random fields (threadCRFs). Then, we proposed to use person resolution, the process of identifying the same person mentioned in different contexts, together with threadCRF for thread structure learning. We have empirically verified the effectiveness of the proposed approaches.

Original languageEnglish (US)
Article number8886425
Pages (from-to)1273-1282
Number of pages10
JournalIEEE Transactions on Computational Social Systems
Volume6
Issue number6
DOIs
StatePublished - Dec 2019

All Science Journal Classification (ASJC) codes

  • Modeling and Simulation
  • Social Sciences (miscellaneous)
  • Human-Computer Interaction

Keywords

  • Thread conditional random fields (threadCRF)
  • thread structure learning

Fingerprint Dive into the research topics of 'Thread Structure Learning on Online Health Forums with Partially Labeled Data'. Together they form a unique fingerprint.

Cite this