SLIDE - a Sentiment Lexicon of Common Idioms

SLIDE – a Sentiment Lexicon of Common Idioms Charles Jochim, Francesca Bonin, Roy Bar-Haim, Noam Slonim IBM Research {charlesj,fbonin}@ie.ibm.com, {roybar,noams}@il.ibm.com Abstract Idiomatic expressions are problematic for most sentiment analysis approaches, which rely on words as the basic linguistic unit. Compositional solutions for phrase sentiment are not able to handle idioms correctly because their sentiment is not derived from the sentiment of the individual words. Previous work has explored the importance of idioms for sentiment analysis, but has not addressed the breadth of idiomatic expressions in English. In this paper we present an approach for collecting sentiment annotation of idiomatic multiword expressions using crowdsourcing. We collect 10 annotations for each idiom and the aggregated label is shown to have good agreement with expert annotations. We describe the resulting publicly available lexicon and how it captures sentiment strength and ambiguity. The Sentiment Lexicon of IDiomatic Expressions (SLIDE) is much larger than previous idiom lexicons. The lexicon includes 5,000 frequently occurring idioms, as estimated from a large English corpus. The idioms were selected from Wiktionary, and over 40% of them were labeled as sentiment-bearing. Keywords: Idiom, Lexicon, Sentiment Analysis 1. Introduction Multiword expressions (MWE) are a key challenge in Natu- ral Language Processing (Sag et al., 2002). Among MWEs, idioms are often defined as non-compositional multiword expressions, the meaning of which cannot be deduced from the literal meaning of constituent words (Nunberg et al., 1994). Sentiment analysis systems typically consider words as the basic sentiment units. Word sentiments are either learned from the training data or looked up in a sentiment lexicon. Text sentiment is then derived by means of aggregation over word sentiments, often with some treatment of com- positional phenomena such as valence shifters (Polanyi and Zaenen, 2004) and mixed sentiment (Kiritchenko and Mo- hammad, 2016). Other approaches are based on bottom-up sentiment composition, starting at the word level and com- puting the sentiment of each phrase based on the seman- tics and sentiment of its daughter phrases, according to the syntactic structure of the sentence (Moilanen and Pulman, 2007; Nakagawa et al., 2010; Socher et al., 2013). Due to their non-compositionality, idioms are often not handled correctly by current sentiment analysis systems (Balahur et al., 2010). Word-level sentiment analysis would miss the positive sentiment in two thumbs up, and on the other hand, we might incorrectly assign positive sentiment to as well as, because of the positive sentiment of well. Similarly, we would like to know that it is not good if some- thing bites the dust, while we would be happy to hear that our handing of idioms was dead on. Ignoring idioms over- looks an important signal of the sentiment of the text, as fig- urative and idiomatic language often directs sentence polar- ity (Rentoumi et al., 2012). For the above reasons, idioms have begun to receive some attention in recent sentiment analysis literature (Williams et al., 2015; Liu et al., 2017). Yet, robust treatment of idioms is hindered by their limited coverage in current sentiment datasets and …

Content truncated to 500 words