Choi, ACL 2010

From Cohen Courses
Jump to navigationJump to search

Citation

Choi, Y., and C. Cardie. Hierarchical Sequential Learning for Extracting Opinions and their Attributes. ACL-2010

Online version

[1]

Summary

This paper apply a hierarchical parameter sharing technique using Conditional Random Field (CRF) to fine grained opinion analysis. It aim to jointly identify the boundaries of opinion expression as well as to determine two of their key attributes - polarity and intensity.

Brief description of the method

They used following features:

  • Prior-polarity & Prior-intensity
  • EXP-Polarity, Exp-Intensity & EXP-Span: Expression level attributes
  • Per-Token Features
    • Common Per-Token Features
    • Polarity Per-Token Features
    • Intensity Per-Token Features
  • Transition Features
    • Polarity Transition Features
    • Intensity Transition Features

Experimental Result

The method was evaluated on the MultiPerspective Question Answering (MPQA) corpus. Joint with Hierarchy method works best while baseline which cascade two separately trained models performs worst. The performance for negative class is much higher than positive class. It is mainly because data for negative class is double than the positive class.