Supervised Ranking of Co-occurrence Profiles for Acquisition of Continuous Lexical Attributes

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2172–2183, Dublin, Ireland, August 23-29 2014. Supervised Ranking of Co-occurrence Profiles for Acquisition of Continuous Lexical Attributes Julian Brooke Department of Computer Science University of Toronto jbrooke@cs.toronto.edu Graeme Hirst Department of Computer Science University of Toronto gh@cs.toronto.edu Abstract Certain common lexical attributes such as polarity and formality are continuous, creating chal- lenges for accurate lexicon creation. Here we present a general method for automatically placing words on these spectra, using co-occurrence profiles, counts of co-occurring words within a large corpus, as a feature vector to a supervised ranking algorithm. With regards to both polarity and formality, we show this method consistently outperforms commonly-used alternatives, both with respect to the intrinsic quality of the lexicon and also when these newly-built lexicons are used in downstream tasks. 1 Introduction Lexicon acquisition represents one key way that the information in large corpora and other resources can be leveraged in various NLP tasks, particularly when the range of lexical items involved in a particular phenomenon is much more diverse than can typically be captured in manually-built resources. Another property of the lexicon which might limit a manual approach is the fact that certain attributes are not discrete, instead falling on a continuous spectrum; although there are manually-built dictionaries which contain fine-grained judgments of spectra—an example is the MRC psychological database (Coltheart, 1980)—these tend to be very low in coverage, reflecting the difficulty in collecting this information. Within computational linguistics, the continuous lexical attribute that has received the most attention is undoubtedly the positive-negative spectrum, otherwise known as semantic orientation (SO) or polarity. Much of the work focused on acquisition of this attribute at the lexical level has involved simplification to a binary (positive-negative) or ternary (positive-neutral-negative) distinction (Hatzivassiloglou and McKeown, 1997; Takamura et al., 2005; Kaji and Kitsuregawa, 2007; Rao and Ravichandra, 2009; Mohammad et al., 2009; Volkova et al., 2013) but other work explicitly offers a continuous quantification (Turney, 2002; Turney and Littman, 2003; Baccianella et al., 2010). Another spectrum with a prominent role in the lexicon is formality (Brooke et al., 2010; Lahiri et al., 2011), which includes colloquial words at one end, socially-distancing words at the other, and common vocabulary in the middle. In this paper, we will focus on these two spectra; the method presented, however, is intended to be general, and as such could be easily applied to other spectra such as those in the MRC database, e.g. abstractness (Turney et al., 2011), and other kinds of variation captured in, for instance, Osgood’s semantic differential (Osgood et al., 1957). The typical approach to this problem involves semi-supervised methods using vector space and/or graph representations and a set of seed terms. Our method is novel in that it uses fully supervised SVM ranking of co-occurrence profiles, i.e. normalized counts of instances of binary text co-occurrence between the target word and a large set of profiling words, selected on the basis of their frequency, in a publicly-available …

Content truncated to 500 words