teragram: Rule-based detection of sentiment phrases using SAS Sentiment Analysis

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 513–519, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics teragram: Rule-based detection of sentiment phrases using SAS Sentiment Analysis Hilke Reckman, Cheyanne Baird, Jean Crawford, Richard Crowell, Linnea Micciulla, Saratendu Sethi, and Fruzsina Veress SAS Institute 10 Fawcett Street Cambridge, MA 02138, USA hilke.reckman@sas.com Abstract For SemEval-2013 Task 2, A and B (Sen- timent Analysis in Twitter), we use a rule- based pattern matching system that is based on an existing ‘Domain Independent’ sentiment taxonomy for English, essentially a highly phrasal sentiment lexicon. We have made some modifications to our set of rules, based on what we found in the annotated training data that was made available for the task. The resulting system scores competitively, espe- cially on task B. 1 Introduction SAS taxonomies for sentiment analysis are primar- ily topic-focused. They are designed to track sen- timent around brands, entities, or other topics and subtopics in a domain (Lange and Sethi, 2011; Lakkaraju and Sethi, 2012; Albright and Lakkaraju, 2011). Domain-independent taxonomies have a second function. In addition to performing topic- focused tasks, they can be set up to perform senti- ment analysis at the document level, classifying the whole document as positive, negative, or neutral. In this task all sentiment expressions are taken into ac- count, rather than only those which are related to the tracked topic. This second function is becom- ing increasingly important. It allows for a broader perspective that is complementary to topic-focused opinion mining. We participated in both subtask A and B of SemEval-2013 Task 2: Sentiment Analysis in Twit- ter (Wilson et al., 2013) with an adaptation of our existing system. For task B, identifying the overall sentiment of a tweet, our taxonomy mainly needed some fine-tuning to specifically accommodate Twit- ter data. (Normally tweets only make up a small part of the data we work with.) We also made a few adaptations to focus entirely on document level sentiment, whereas originally the main focus of our system was on tracking sentiment around products. For task A, identifying the sentiment of ambiguous phrases in a tweet, a few more modifications were needed. Our system is entirely rule-based, and the rules are hand-written. In some cases, statistical text min- ing approaches are used for the discovery of topics and terms to facilitate rule writing. Our sentiment analysis software does offer a statistical component, but our experience is that purely rule-based models work better for our typical sentiment analysis tasks. Advantages of rules are that problems observed in the output can be targeted directly, and the model can become more and more refined over time. Also, they allow for simple customization. In our brand- centered work, we customize our taxonomies for one or more brands that we want to track. When we build a taxonomy for a new domain, we build upon work we have done before in other domains. The assignment of …

Content truncated to 500 words