Extracting Phrases Describing Problems with Products and Services from Twitter Messages

Abstract

Social media contain many types of
information useful to businesses. In this paper we
discuss a trigger-target based approach to extract
descriptions of problems from Twitter data. It is important
to note that the descriptions of problems are factual
statements as opposed to subjective opinions about
products/services. We first identify the problem tweets
i.e. the tweets containing descriptions of problems.
We then extract the phrases that describe the problem.
In our approach such descriptions are extracted as a
combination of trigger and target phrases. Triggers
are mostly domain independent verb phrases and are
identified by using hand crafted lexical and syntactic
patterns. Targets on the other hand are domain specific
noun phrases syntactically related to the triggers. We
frame the problem of finding target phrase corresponding
to a trigger phrase as a ranking problem and show the
results of experiments with maximum entropy classifiers
and voted perceptrons. Both approaches outperform the
rule based approach reported before.