Parameter Estimation for LDA-Frames

Proceedings of NAACL-HLT 2013, pages 482–486, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Parameter Estimation for LDA-Frames Jiˇr´ı Materna Centre for Natural Language Processing Faculty of Informatics, Masaryk University Botanick´a 68a, 602 00, Brno, Czech Republic xmaterna@fi.muni.cz Abstract LDA-frames is an unsupervised approach for identifying semantic frames from semanti- cally unlabeled text corpora, and seems to be a useful competitor for manually created databases of selectional preferences. The most limiting property of the algorithm is such that the number of frames and roles must be pre- defined. In this paper we present a modifi- cation of the LDA-frames algorithm allowing the number of frames and roles to be deter- mined automatically, based on the character and size of training data. 1 Introduction Semantic frames and valency lexicons are useful lexical sources capturing semantic roles valid for a set of lexical units. The structures of linked seman- tic roles are called semantic frames. Linguists are using them for their ability to describe an interface between syntax and semantics. In practical natural language processing applications, they can be used, for instance, for the word sense disambiguation task or in order to resolve ambiguities in syntactic analy- sis of natural languages. The lexicons of semantic frames or verb valencies are mainly created manually or semi-automatically by highly trained linguists. Manually created lex- icons involve, for instance, a well-known lexi- con of semantic frames FrameNet (Ruppenhofer et al., 2006) or a lexicon of verb valencies VerbNet (Schuler, 2006). These and other similar lexical re- sources have many promising applications, but suf- fer from several disadvantages: • Creation of them requires manual work of trained linguists which is very time-consuming and expensive. • Coverage of the resources is usually small or limited to some specific domain. • Most of the resources do not provide any information about relative frequency of us- age in corpora. For instance, both patterns [Person] acquire [Physical object] and [Person] acquire [Disease] reflect correct usage of verb acquire, but the former is much more frequent in English. • Notion of semantic classes and frames is sub- jectively biased when the frames are created manually without corpus evidence. In order to avoid those problems we proposed a method for creating probabilistic semantic frames called LDA-frames (Materna, 2012). The main idea of LDA-frames is to generate the set of semantic frames and roles automatically by maximizing pos- terior probability of a probabilistic model on a syn- tactically annotated training corpus. A semantic role is represented as probability distribution over all its realizations in the corpus, a semantic frame as a tu- ple of semantic roles, each of them connected with some grammatical relation. For every lexical unit (a verb in case of computing verb valencies), a prob- ability distribution over all semantic frames is gen- erated, where the probability of a frame corresponds to the relative frequency of usage in the corpus for a given lexical unit. An example of LDA-frames 482 computed on the British National Corpus is avail- …

Content truncated to 500 words