Enhancing General Sentiment Lexicons for Domain-Specific Use

Proceedings of the 27th International Conference on Computational Linguistics, pages 1056–1064 Santa Fe, New Mexico, USA, August 20-26, 2018. 1056 Enhancing General Sentiment Lexicons for Domain-Specific Use Tim Kreutz Walter Daelemans CLiPS - Computational Linguistics Group Department of Linguistics University of Antwerp {tim.kreutz,walter.daelemans}@uantwerpen.be Abstract Lexicon based methods for sentiment analysis rely on high quality polarity lexicons. In recent years, automatic methods for inducing lexicons have increased the viability of lexicon based methods for polarity classification. SentProp is a framework for inducing domain-specific po- larities from word embeddings. We elaborate on SentProp by evaluating its use for enhancing DuOMan, a general-purpose lexicon, for use in the political domain. By adding only top senti- ment bearing words from the vocabulary and applying small polarity shifts in the general-purpose lexicon, we increase accuracy in an in-domain classification task. The enhanced lexicon performs worse than the original lexicon in an out-domain task, showing that the words we added and the polarity shifts we applied are domain-specific and do not translate well to an out-domain setting. 1 Introduction Work in sentiment analysis can roughly be divided into two approaches: corpus based techniques and lexicon based techniques (Taboada et al., 2011). Corpus based techniques use supervised learning on large target domain corpora to learn classification models. Currently, implementations using deep learn- ing make up the state of the art in binary polarity classification on Twitter (Rosenthal et al., 2017). Lexicon based techniques rely on high quality polarity lexicons. Document-level polarity can be predicted by looking at occurrences of lexical units in the document and doing some form of averaging on their respective polarity values. Despite not yielding the state of the art results of corpus based methods, lexicon based methods still see frequent use since they do not rely on big (labelled) data and allow for more interpretability. In recent years, polarity lexicons have been improved by adding new phrases from web data (Velikovich et al., 2010) or by automatically learning domain-specific polarities (Hamilton et al., 2016). Sentprop is a framework for inducing domain-specific polarities (Hamilton et al., 2016). SentProp has been shown to accurately reproduce domain-specific lexicons from in-domain word embeddings. Additionally, it has been used to examine interesting historical and community-specific polarity shifts. We elaborate on SentProp by evaluating its use for domain adaptation of a general-purpose lexicon. Starting from a Dutch general polarity lexicon and a corpus of political forum posts, we automatically expand and enhance the general lexicon entries. We show that the enhanced lexicon outperforms the general-purpose lexicon on an in-domain classification task and we further add to the existing body of work by evaluating the enhanced lexicon on an out-domain task. In the out-domain task the general purpose-lexicon yields better accuracy, showing that our proposed method learns domain-specific polar- ities that do not translate well to other domains. The paper is organized as follows. Section 2 will embed SentProp in a larger context of semi- supervised lexicon expansion research. In section 3 we describe the corpus used for inducing polarities, …

Content truncated to 500 words