Senti17 at SemEval-2017 Task 4: Ten Convolutional Neural Network Voters for Tweet Polarity Classification

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 700–703, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics Senti17 at SemEval-2017 Task 4: Ten Convolutional Neural Network Voters for Tweet Polarity Classification Hussam Hamdan Labex Observatoire de la vie littraire (OBVIL) Laboratoire d’Informatique de Paris 6 (LIP6), Pierre and Marie Curie University, UMR 7606 4 place Jussieu, 75005, Paris, France Hussam.Hamdan@lip6.fr Abstract This paper presents Senti17 system which uses ten convolutional neural networks (Con- vNet) to assign a sentiment label to a tweet. The network consists of a convolutional layer followed by a fully-connected layer and a Soft- max on top. Ten instances of this net- work are initialized with the same word em- beddings as inputs but with different initial- izations for the network weights. We com- bine the results of all instances by selecting the sentiment label given by the majority of the ten voters. This system is ranked fourth in SemEval-2017 Task4 over 38 systems with 67.4% average recall. 1 Introduction Polarity classification is the basic task of sentiment analysis in which the polarity of a given text should be classified into three categories: positive, negative or neutral. In Twitter where the tweet is short and written in informal language, this task needs more attention. SemEval has proposed the task of Mes- sage Polarity Classification in Twitter since 2013, the objective is to classify a tweet into one of the three polarity labels (Rosenthal et al., 2017). We can remark that in 2013, 2014 and 2015 most best systems were based on a rich feature extraction process with a traditional classifier such as SVM (Mohammad et al., 2013) or Logistic regression (Hamdan et al., 2015). In 2014, Kim (2014) pro- posed to use one convolutional neural network for sentence classification, he fixed the size of the in- put sentence and concatenated its word embeddings for representing the sentence, this architecture has been exploited in many later works. Severyn and Moschitti (2015) adapted the convolutional network proposed by Kim (2014) for sentiment analysis in Twitter, their system was ranked second in SemEval- 2015 while the first system (Hagen et al., 2015) com- bined four systems based on feature extraction and the third ranked system used logistic regression with different groups of features (Hamdan et al., 2015). In 2016, we remark that the number of participa- tions which use feature extraction systems were de- graded, and the first four systems used Deep Learn- ing, the majority used a convolutional network ex- cept the fourth one (Amir et al., 2016). Despite of that, using Deep Learning for sentiment analysis in Twitter has not yet shown a big improvement in comparison to feature extraction, the fifth and sixth systems (Hamdan, 2016) in 2016 which were built upon feature extraction process were only (3 and 3.5% respectively) less than the first system. But We think that Deep Learning is a promising direction in sentiment analysis. Therefore, we proposed to use convolutional networks for Twitter polarity classifi- …

Content truncated to 500 words