WarwickDCS: From Phrase-Based to Target-Specific Sentiment Recognition

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 657–663, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics WarwickDCS: From Phrase-Based to Target-Specific Sentiment Recognition Richard Townsend University of Warwick richard@sentimentron.co.uk Adam Tsakalidis University of Warwick A.Tsakalidis@warwick.ac.uk Yiwei Zhou University of Warwick Yiwei.Zhou@warwick.ac.uk Bo Wang University of Warwick Bo.Wang@warwick.ac.uk Maria Liakata University of Warwick M.Liakata@warwick.ac.uk Arkaitz Zubiaga University of Warwick A.Zubiaga@warwick.ac.uk Alexandra Cristea University of Warwick acristea@dcs.warwick.ac.uk Rob Procter University of Warwick rob.procter@warwick.ac.uk Abstract We present and evaluate several hybrid sys- tems for sentiment identification for Twit- ter, both at the phrase and document (tweet) level. Our approach has been to use a novel combination of lexica, traditional NLP and deep learning features. We also analyse tech- niques based on syntactic parsing and token- based association to handle topic specific sen- timent in subtask C. Our strategy has been to identify subphrases relevant to the designated topic/target and assign sentiment according to our subtask A classifier. Our submitted subtask A classifier ranked fourth in the Se- mEval official results while our BASELINE and µPARSE classifiers for subtask C would have ranked second. 1 Introduction Twitter holds great potential for analyses in the so- cial sciences both due to its explosive popularity, in- creasing accessibility to large amounts of data and its dynamic nature. For sentiment analysis on twit- ter the best performing approaches (Mohammad et al., 2013; Zhu et al., 2014) have used a set of rich lexical features. However, the development of lex- ica can be time consuming and is not always suit- able when shifting between domains, which exam- ine new topics and user populations (Thelwall and Buckley, 2013). Excitingly, the state of the art has recently shifted toward novel semi-supervised tech- niques such as the incorporation of word embed- dings to represent the context of words and concepts (Tang et al., 2014b). Moreover, it is important to be able to identify sentiment in relation to particular entities, topics or events (aspect-based sentiment). We have followed a hybrid approach which incor- porates traditional lexica, unigrams and bigrams as well as word embeddings using word2vec (Mikolov et al., 2013) to train classifiers for subtasks A and B. For subtask C, sentiment targeted towards a par- ticular topic, we have developed a set of different strategies which use either syntactic dependencies or token-level associations with the topic word in com- bination with our A classifier to produce sentiment annotations. 2 Phrase-Based Sentiment Analysis (Subtask A) as a Means to an End (subtask C) Phrase-based sentiment analysis (subtask A) in tweets is a long standing task where the goal is to classify the sentiment of a designated expression within the tweet as either positive, negative or neutral. The state of the art for subtask A achieves high performance usually based on methodologies employing features obtained from either manually or automatically generated lexica (Mohammad et al., 2013; Zhu et al., 2014). However, lexica by definition lack contextual information and are oftain domain dependent. Recent work (Tang et …

Content truncated to 500 words