Choi et al, EMNLP 2006

From ScribbleWiki: Analysis of Social Media

Joint Extraction of Entities and Relations for Opinion Recognition

Choi et al, EMNLP 2006 Paper

In this work, Choi et al. combine the tasks of opinion and source (of opinion) extraction.

Choi et al. perform source tagging using linear Conditional Random Fields (CRFs), which is an approach they had used in the previous work (Choi el al. 2005). Additionally, in this work, they perform opinion detection as a similar sequence tagging task. Than they link these potential sources and opinions thus obtained in a binary link-detection problem- i.e. decide if a link should be placed between a source and an opinion in the given sentence.

Finally, global constraints for the sentence are encoded using Integer Linear Programming (ILP). This step is crucial in enforcing the coherence of the sentence. Some of the coherence constraints are “an opinion must have a source expressing it”, “a source should be linked to at least one and at most two opinions”. The opinion detection module, as well as the source tagging module from the previous stages retrieve n-best answers. This module decides which of these entity-opinion pairs to keep, and which to discard. The objective function for the ILP takes into account the probability weights given to the sequence in by the previous modules, while incorporating sentence coherence constraints.

In this work opinions refer to Subjectivity as defined by Wiebe et al. The sources are the agents who hold the opinion (private state). Accordingly, they use the MPQA data from this group. This data has been annotated with subjective expressions and their sources, amongst other things.

Other research that find sources of opinions are Kim and Hovy,(2005) and Bethard et al.(2004).

Irrespective of the approach (sequence tagging/ classification/ ranking) most of these systems use semantic class information as an important knowledge source for Source extraction. This is because only a subset on entities in the sentence ( like persons, or organizations) are capable of holding opinions. Another important source of information is the dependency and path between the entity and the opinion expression, as given in the syntactic parse of the sentence.

The source extractor from Choi et al. 2005 has been implemented in the OpinionFinder System that is available for free download from http://www.cs.pitt.edu/mpqa/opinionfinderrelease/

--Swapna 18:04, 9 October 2007 (CDT)