PRHLT: Combination of Deep Autoencoders with Classification and Regression Techniques for SemEval-2015 Task 11

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 689–693, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics PRHLT: Combination of Deep Autoencoders with Classification and Regression Techniques for SemEval-2015 Task 11 Parth Gupta PRHLT Research Center Universitat Polit`ecnica de Val`encia Camino de Vera, s/n 46022 Valencia, SPAIN pgupta@dsic.upv.es Jon Ander G´omez PRHLT Research Center Universitat Polit`ecnica de Val`encia Camino de Vera, s/n 46022 Valencia, SPAIN jon@dsic.upv.es Abstract This paper presents the system we developed for Task 11 of SemEval 2015. Our system had two stages: The first one was based on deep autoencoders for extracting features to com- pactly represent tweets. The next stage con- sisted of a classifier or a regression function for estimating the polarity value assigned to a given tweet. We tested several techniques in order to choose the ones with the highest accu- racy. Finally, three regression techniques re- vealed as the best ones for assigning the polar- ity value to tweets. We presented six runs cor- responding to three regression different tech- niques in combination with two variants of the autoencoder, one with input as bags of words and another with input as bags of character 3- grams. 1 Introduction Sentiment Analysis from texts is a growing field of research due to its social and economic relevance. Task 11 of SemEval-2015 (Semantic Evaluation Ex- ercises) was proposed to the research community in order to foster the development of systems and tech- niques for Sentiment Analysis (Ghosh et al., 2015). We faced this challenging task with a system based on deep autoencoders in combination with classification and regression techniques. We used deep autoencoders to extract features from tweets by means of two ways of splitting text: i) words and ii) character 3-grams. The training of autoencoders was unsupervised. The extracted features (10) and a few manually added features (5) were used for train- ing classifiers or regression functions to estimate the tweet’s polarity value. The rest of the paper is organized as follows. Sec- tion 2 describes the proposed system. Section 3 presents the obtained results on the test set. Finally, conclusions are discussed in Section 4. 2 System Description Our system consists of two stages: (1) Dimension- ality reduction by means of deep autoencoders. (2) Polarity value assignment by using different classi- fication and regression techniques. The text of tweets was preprocessed before being used as input to the autoencoders. The autoencoders take as input a representation of each tweet. Two different representations were used: bags of words and bags of character 3-grams. In both cases the output of an autoencoder was a vector of 10 real values. Optionally we added other features in order to improve the polarity assignment, these additional features are binary features indicating whether some symbols or hash tags appear in the tweet. The idea behind adding these extra features is to set a con- text for learning under their influence. The different subsets of used features are described in subsection 2.3. The step of …

Content truncated to 500 words