More or less supervised supersense tagging of Twitter

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 1–11, Dublin, Ireland, August 23-24 2014. More or less supervised supersense tagging of Twitter Anders Johannsen, Dirk Hovy, H´ector Mart´ınez Alonso, Barbara Plank, Anders Søgaard Center for Language Technology University of Copenhagen, Denmark Njalsgade 140 ajohannsen@hum.ku.dk, dirk@cst.dk, alonso@hum.ku.dk plank@cst.dk, soegaard@hum.ku.dk Abstract We present two Twitter datasets annotated with coarse-grained word senses (super- senses), as well as a series of experiments with three learning scenarios for super- sense tagging: weakly supervised learn- ing, as well as unsupervised and super- vised domain adaptation. We show that (a) off-the-shelf tools perform poorly on Twitter, (b) models augmented with em- beddings learned from Twitter data per- form much better, and (c) errors can be reduced using type-constrained inference with distant supervision from WordNet. 1 Introduction Supersense tagging (SST, Ciaramita and Altun, 2006) is the task of assigning high-level ontolog- ical classes to open-class words (here, nouns and verbs). It is thus a coarse-grained word sense dis- ambiguation task. The labels are based on the lexi- cographer file names for Princeton WordNet (Fell- baum, 1998). They include 15 senses for verbs and 26 for nouns (see Table 1). While WordNet also provides catch-all supersenses for adjectives and adverbs, these are grammatically, not seman- tically motivated, and do not provide any higher- level abstraction (recently, however, Tsvetkov et al. (2014) proposed a semantic taxonomy for ad- jectives). They will not be considered in this paper. Coarse-grained categories such as supersenses are useful for downstream tasks such as question- answering (QA) and open relation extraction (RE). SST is different from NER in that it has a larger set of labels and in the absence of strong orthographic cues (capitalization, quotation marks, etc.). More- over, supersenses can be applied to any of the lex- ical parts of speech and not only proper names. Also, while high-coverage gazetteers can be found for named entity recognition, the lexical resources available for SST are very limited in coverage. Twitter is a popular micro-blogging service, which, among other things, is used for knowledge sharing among friends and peers. Twitter posts (tweets) announce local events, say talks or con- certs, present facts about pop stars or program- ming languages, or simply express the opinions of the author on some subject matter. Supersense tagging is relevant for Twitter, be- cause it can aid e.g. QA and open RE. If someone posts a message saying that some LaTeX module now supports “drawing trees”, it is important to know whether the post is about drawing natural objects such as oaks or pines, or about drawing tree-shaped data representations. This paper is, to the best of our knowledge, the first work to address the problem of SST for Twit- ter. While there exist corpora of newswire and literary texts that are annotated with supersenses, e.g., SEMCOR (Miller et al., 1994), no data is available for microblogs or related domains. This paper introduces two new data sets. Furthermore, most, if not all, of previous …

Content truncated to 500 words