Tweester at SemEval-2017 Task 4: Fusion of Semantic-Affective and pairwise classification models for sentiment analysis in Twitter

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 675–682, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics Tweester at SemEval-2017 Task 4: Fusion of Semantic-Affective and pairwise classification models for sentiment analysis in Twitter Athanasia Kolovou 2,3, Filippos Kokkinos 1, Aris Fergadis 1,3, Pinelopi Papalampidi 1 Elias Iosif 1,3, Nikolaos Malandrakis 4, Elisavet Palogiannidi 1, Harris Papageorgiou 3 Shrikanth Narayanan 4, Alexandros Potamianos 1,3 1School of ECE, National Technical University of Athens, Zografou 15780, Athens, Greece 2 Department of Informatics, University of Athens, Athens, Greece 3 “Athena” Research and Innovation Center, Maroussi 15125, Athens, Greece 4 Signal Analysis and Interpretation Laboratory (SAIL), USC, Los Angeles, CA 90089, USA akolovou@di.uoa.gr, el11142@mail.ntua.gr fergadis@central.ntua.gr, el12003@central.ntua.gr iosife@central.ntua.gr, malandra@usc.edu epalogiannidi@gmail.com, xaris@ilsp.athena-innovation.gr shri@sipi.usc.edu, potam@central.ntua.gr Abstract In this paper, we describe our submission to SemEval2017 Task 4: Sentiment Anal- ysis in Twitter. Specifically the proposed system participated both to tweet polar- ity classification (two-, three- and five class) and tweet quantification (two and five-class) tasks. The submitted system is based on “Tweester” (Palogiannidi et al., 2016) that participated in last year’s Sen- timent analysis in Twitter Tasks A and B. Specifically it comprises of multiple independent models such as neural net- works, semantic-affective models and af- fective models inspired by topic modeling that are combined in a late fusion scheme. 1 Introduction Tweets are short length pieces of text, usually written in informal style that contain abbrevia- tions, misspellings and creative syntax (like emoti- cons, hashtags etc). The challenging nature of sentiment analysis in Twitter motivated the or- ganization of numerous tasks within the Seman- tics Evaluation (SemEval) workshop. In this pa- per, we show how our sentiment analysis frame- work called “Tweester” (winner of Subtask B in SemEval-2016 (Palogiannidi et al., 2016)), can be applied to all subtasks, namely Subtask A (message polarity classification), Subtask B (tweet classification according to a two-point scale), Sub- task C (sentiment conveyed by a tweet towards the topic on a five-point scale), Subtask D (esti- mate the distribution of the tweets across a two- point scale), Subtask E (estimate the distribution of the tweets across a five-point scale ). The sys- tem achieved high performance ranking 5th, 3rd, 4th, 5th and 8th for Subtasks A, B, C, D, and E, respectively (S.Rosenthal et al., 2017). In Section 2 we provide more details on the individual sys- tems as well as the fusion scheme. Experimental procedure is described in Section 3 and some con- cluding remarks as well as an outlook on future work are presented in Section 4. 2 System Description The submitted system is based on the fusion of several systems. Specifically the system con- sists of: 1) the semantic-affective system (sub- mitted to the SemEval 2016 Task 4 (Palogiannidi et al., 2016)) that incorporates affective, semantic- similarity, sarcasm/irony and topic modeling fea- tures, 2) a single and a two-step convolutional neu- ral network, 3) a system based on word embed- dings, 4) a “stacking” based system that trans- forms the 3-class …

Content truncated to 500 words