Distributional Term Set Expansion

Distributional Term Set Expansion Amaru Cuba Gyllensten, Magnus Sahlgren RISE SICS Box 1263, SE-164 29 Kista {amaru.cuba.gyllensten, magnus.sahlgren}@ri.se Abstract This paper is a short empirical study of the performance of centrality and classification based iterative term set expansion methods for distributional semantic models. Iterative term set expansion is an interactive process using distributional semantics models where a user labels terms as belonging to some sought after term set, and a system uses this labeling to supply the user with new, candidate, terms to label, trying to maximize the number of positive examples found. While centrality based methods have a long history in term set expansion (Sarmento et al., 2007; Pantel et al., 2009), we compare them to classification methods based on the the Simple Margin method, an Active Learning approach to classification using Support Vector Machines (Tong and Koller, 2002). Examining the performance of various centrality and classification based methods for a variety of distributional models over five different term sets, we can show that active learning based methods consistently outperform centrality based methods. Keywords: Term Set Expansion, Lexicon Acquisition, Distributional Semantics, Word Embeddings, Active Learning 1. Introduction One of the most commonly used resources in Natural Lan- guage Processing is the term set: a set of, optionally, la- beled words. It is a standard approach to sentiment, sub- jectivity, and stance detection: compile lists of terms rep- resenting the categories in question, and then calculate the occurrence of these terms in data. A text is assigned to the category whose terms are most prevalent in the text. This approach – often referred to as lexicon-based classification – is simplistic, but surprisingly powerful, and often pro- vides useful results in the absence of supervised classifiers (Eisenstein, 2017). Another closely related use case is topic monitoring in social media, in which case the frequency of topic-related terms over time can be used to gauge public interest in those topics. The performance of such lexicon-based approaches obvi- ously depends on the quality of the lexicon being used. A common approach is to use distributional models (word embeddings) to populate the lexicon on the basis of a small set of manually selected seed terms (e.g. “bad” and “sub- par” as seed terms for negative sentiment, and “good” and “ace” as seed terms for positive sentiment). The seed terms are used as probes into the distributional model with the goal of finding other terms that are (distributionally) simi- lar to the seed terms. Iterative term set expansion is the it- erated, interactive, version of this procedure: An annotator defines an initial, incomplete, term set. This term set is fed to the term set expansion method, generating new candidate terms. The annotator labels these as belonging to, or not be- longing to, the term set, which is updated accordingly. The new updated term set is then fed to the expansion method, and the process is repeated indefinitely. In this way, a small set of manually defined seed terms can be (semi-) automat- ically expanded into …

Content truncated to 500 words