Simple Algorithms For Sentiment Analysis On Sentiment Rich, Data Poor Domains.

Proceedings of the 27th International Conference on Computational Linguistics, pages 3424–3435 Santa Fe, New Mexico, USA, August 20-26, 2018. 3424 Simple Algorithms For Sentiment Analysis On Sentiment Rich, Data Poor Domains. Prathusha K Sarma University of Wisconsin-Madison kameswarasar@wisc.edu William A Sethares University of Wisconsin-Madison sethares@wisc.edu Abstract Standard word embedding algorithms learn vector representations from large corpora of text doc- uments in an unsupervised fashion. However, the quality of word embeddings learned from these algorithms is affected by the size of training data sets. Thus, applications of these algorithms in domains with only moderate amounts of available data is limited. In this paper we introduce an algorithm that learns word embeddings jointly with a classifier. Our algorithm is called SWESA (Supervised Word Embeddings for Sentiment Analysis). SWESA leverages document label in- formation to learn vector representations of words from a modest corpus of text documents by solving an optimization problem that minimizes a cost function with respect to both word em- beddings and the weight vector used for classification. Experiments on several real world data sets show that SWESA has superior performance on domains with limited data, when compared to previously suggested approaches to word embeddings and sentiment analysis tasks. 1 Introduction Word embedding algorithms learn vector representations for words that are useful to quantify semantic relationships between words in a given text. Additionally, word embeddings are used to initialize several algorithms for sentiment analysis, sentence encoding etc. Currently popular embedding algorithms such as word2vec (Mikolov et al., 2013b; Le and Mikolov, 2014), GloVe (Pennington et al., 2014), are based off neural network methods and have achieved tremendous success in various word, sentence/document level evaluation tasks. These algorithms thrive on the large volumes of training data sets when learning high quality word embeddings. However, there is an increase in application domains where getting large amounts of data is not always possible. Furthermore, in some of these domains, representing words from off-the-shelf word embeddings such as ones obtained from training word2vec, GloVe on Wikipedia or common-crawl may not be efficient. This is because the sentiment expressed by a word in such datasets could be somewhat different from the sentiment expressed by the same word when it appears in Wikipedia, common-crawl. A concrete example of such domains is given in the next paragraph. In such cases off-the-shelf word embeddings are not very useful and better techniques are required to learn word embeddings and classifiers for sentiment analysis. The goal of this paper is to build classification algorithms for sentiment analysis on small, domain specific data sets that are sentiment rich. Such an algorithm is not limited to sentiment classification and easily generalizes to other text classification problems as well. An example of a small sized and sentiment rich data set is the Substance Use Disorder (SUD) data set (Mohr et al., 2013), (Moore et al., 2011) obtained from digital health intervention treatments. These treatments aim to predict relapse risk by analyzing the content of participants’ text messages. Though forum moderators can monitor and …

Content truncated to 500 words