CLaC-SentiPipe: SemEval2015 Subtasks 10 B,E, and Task 11

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 479–485, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics CLaC-SentiPipe: SemEval2015 Subtasks 10 B,E, and Task 11 Canberk ¨Ozdemir and Sabine Bergler CLaC Labs, Concordia University 1455 de Maisonneuve Blvd West Montreal, Quebec, Canada, H3G 1M8 ozdemir.berkin@gmail.com, bergler@cse.concordia.ca Abstract CLaC Labs participated in two shared tasks for SemEval2015, Task 10 (subtasks B and E) and Task 11. The underlying system con- figuration is nearly identical and consists of two major components: a large Twitter lex- icon compiled from tweets that carry cer- tain selected hashtags (assumed to guaran- tee a sentiment polarity) and then inducing that same polarity for the words that occur in the tweets. We also use standard senti- ment lexica and combine the results. The lex- ical sentiment features are further differenti- ated according to some linguistic contexts in which their triggers occur, including bigrams, negation, modality, and dependency triples. We studied feature combinations comprehen- sively for their interoperability and effective- ness on different datasets using the exhaustive feature combination technique of (Shareghi and Bergler, 2013a; Shareghi and Bergler, 2013b). For Subtask 10B we used a SVM, and a decision tree regressor for Task 11. The re- sulting systems ranked ninth for Subtask 10B, fourth for Subtask 10E, and first for Task 11. 1 Introduction The field of Sentiment Analysis is in its second phase: initially, the task was defined, annotation standards, corpora, and feature resources were iden- tified and provided to the research community (see (Pang and Lee, 2008)). Now, we have regular com- munity challenges such as the SemEval Twitter Sen- timent shared tasks which allow us to compare dif- ferent feature choice and combination across re- search labs and across successive data sets. We de- scribe here the systems we submitted to SemEval15 for Twitter Sentiment Analysis at the tweet level (Task 10B) and Figurative Language in Twitter (Task 11). The tasks and the design of the datasets is described in detail in (Rosenthal et al., 2015) for Task 10 and in (Ghosh et al., 2015) for Task 11. We also submitted a sentiment lexicon transformed from our in-house lexical resource for Task 10E. Our system is based on a pipeline design in 5 major phases, described below. Following standard text preprocessing, we use Stanford dependencies (De Marneffe et al., 2006) and linguistic features negation, modality and their scope in connection with standard sentiment lexica from the literature and an in-house lexical resource compiled with the technique used for the NRC lexicon (Mohammad et al., 2013). These features were successful in both Task 10B (rank 9 on 40 for Twitter 2015 data, sev- enth on 40 for Twitter 2015 sarcasm data) and Task 11 (rank 1 of 35 runs by 15 teams). Our sentiment lexicon submitted to Task 10E ranked fourth of ten. 2 Pipeline Design CLaCSentiPipe is a pipeline system that attempts to test the interoperability of different sentiment lexica and a selected set of linguistic annotations. The lexical …

Content truncated to 500 words