Abstract

In the field of opinion mining, considerable work has gone into the creation of word lists annotated for subjectivity, i.e. marking words such as “positive, beautiful, crank” as subjective and others such as “chair, woman, read” as (mostly) objective. These word lists are then used in the automatic identification of opinions at the sentence or document level.

However, a variety of words are subjectivity-ambiguous, i.e. they have at least one objective and one subjective sense as is shown by the two example senses of “positive” below.

(1) OBJECTIVE : positive, electropositive—-having a positive electric charge; protons are positive (2) SUBJECTIVE plus, positive—-involving advantage or good; a plus (or positive) factor

In this talk, I concentrate on this latter problem by automatically creating lists of subjective senses, instead of subjective words, via adding subjectivity labels for senses to electronic lexica, using the example of WordNet. Specifically, I will discuss the following results:

(1) Human annotation experiments which show that assigning subjectivity to word senses is a well-defined task. These experiments also show that subjectivity-ambiguity is frequent.

(2) A semi-supervised approach based on minimum cuts that assigns subjectivity labels to word senses in lexical relation graphs.

This algorithm outperforms supervised graph and non-graph based algorithms significantly, reducing the error rate by up to 40%. In addition, the semi-supervised approach achieves the same results as the supervised framework with less than 20% of the training data.

[This talk has been postponed from 22 January.]