DRAFTSpeech and Language Processing: An introduction to natural language processing,computational linguistics, and speech recognition. Daniel Jurafsky & James H. Martin.Copyrightc! 2006, All rights reserved. Draft of November 6, 2006. Do not citewithout permission.19COMPUTATIONALLEXICAL SEMANTICSHe was white and shaken, like a dry martini.P.G. Wodehouse, Cocktail Time (1958)This chapter introduces a series of topics related to computing with word mea nings,or computational lexical semantics. Roughly in parallel with the sequence of topicsin Ch. 18, we’ll introduce computational tasks associated with word senses, relationsamong words, and the thematic structure of predicate-bearing words.We begin with word sense disambiguation, the ta sk of examining word tokensin con text and determinin g which sense o f e ach word is being used. WSD is a taskwith a long history in computational linguistics, and as we will see, is a non-trivial un-dertaking given the somewhat elusive na ture of many word senses. Nevertheless, thereare robust algorithms that can achieve high levels of accuracy g iven certain reasonableassumptions.We next look at algorithm s for computing r elationships between words, witha particular focus on word similarity, and the hypernym, hyponym, and meronymWordNet relations introdu ced in Ch. 18. Relations such as these play important roles inapplications such as question answering, natural language g e neration, automatic essaygrading and plagiarism detec tion.Finally, we describe algorithms for semantic role labeling, also known as caserole or thema tic role assignment. These algorithm s generally use information ex-tracted from parses of the kind introduced in Ch. 12 to assign semantic roles such asAGENT, THEME and INSTRUMENT to the phrases in a sentence with respect to particu-lar predicates.19.1 WORD SENSE DISAMBIGUATION: OVERVIEWOur discussion of compositional semantic analyzers in Ch. 17 pretty much ignoredthe issue of lexical ambigu ity. It sho uld be clear by now that this is an unre a sonableapproa c h. Without some means of selecting correct senses for the words in an input,the enormous am ount of homonymy and polysemy in the lexicon would quickly over-whelm a ny approach in an avalanche of competing interpretations.The task of selecting the correct sen se for a word is c a lled word sense disam-DRAFT2 Chapter 19. Computational Lexical Semanticsbiguation, or WSD. Disambiguating word senses h as the potential to improve manyWORD SENSEDISAMBIGUATIONWSDnatural language processing tasks. As we’ll see in Ch. 24, ma c hine translation isone area where word sense ambiguities c an cause severe problems; others includequestion-answering, information retrieval, and text classification. The way thatWSD is exploited in these and other applications varies widely based on the p a rticu-lar needs of the application. The discussion presented here ig nores these applica tion-specific differences and focuses on the implementation and evaluation of WSD systemsas a stand-alone task.In their most basic form, WSD algorithms take as input a word in context alongwith a fixed inventory of potential word senses, and return the correct word sense forthat use. Both the nature of the in put, and the inventory of senses depends on thetask. For machine translatio n from E nglish to Spanish, the sense tag inventory for anEnglish word might be the set of different Spanish translations. If speech synthesis isour task, the inventory might be restricted to homographs with differing pronunciationssuch as bass and bow. If our task is automatic indexing of medical articles, the sensetag inventory might be the set of Me SH (Medical Subject Headings) thesaurus entries.When we are evaluating WSD in isolation, we can use the set of senses from a d ic tio-nary/thesaurus resource like WordNet or LDOCE. Fig. 19.1 shows an example for thethe word bass, which can refer to a musical in strument or a kind of fish.1WordNet Spanish RogetSense Translation Category Target Word in Contextbass4lubina FISH/INSECT . . . fish as Pacific salmon and striped bass and. . .bass4lubina FISH/INSECT . . . produce filets of smoked bass or sturgeon. . .bass7bajo MUSIC . . . exciting jazz bass player since Ray Brown. . .bass7bajo MUSIC . . . play bass because he doesn’t have to solo. . .Figure 19.1 Possible definitions for the inventory of sense tags for bass.It is useful to distinguish two variants of the generic WSD task. In the lexicalsample task, a small pre-selecte d set of target words is chosen, along with an inventoryLEXICAL SAMPLEof senses for each word from some lexicon . Since the set of words and the set ofsenses is small, supervised machine learning approaches are often used to handlelexical sample tasks. For each word , a number of corpus instance s (context sentenc es)can be selected and hand-labeled with the correct sense of the target word in each.Classifier systems can then be trained using these labeled example s. Unlabeled targetwords in context can then be labeled using such a trained classifier. Early work inword sense disambiguation focused solely on lexical sample ta sks o f this sort, buildingword-specific algorithms for disam biguating single words like line, interest, or plant.In contrast, in the all-words task sy stems are given entire texts and a lexicon withALL-WORDSan inventory of senses for each entry, and are required to disamb iguate every contentword in the text. The all-words task is very similar to part-of-speech tagging, exceptwith a much larger set of tags, since each lemma has its own set. A consequence of this1The WordNet database includes 8 senses; we have arbitrarily selected two for this example; we havealso arbitrarily selected one of the many possible Spanish names for fishes which could be used to translateEnglish sea-bass.DRAFTSection 19.2. Supervised Word Sense Disambiguation 3larger set of tags is a serious data sparseness problem; there is unlikely to be adequatetraining data fo r every word in the test set. Moreover, given the number of polysemouswords in reasonab ly-sized lexicons, approaches based on training o ne classifier perterm are unlikely to be practical.In the fo llowing sections we explore the application of various mach ine learningparadigms to word sense disambiguation. We begin with supervised learning, followedby a section on how systems are standardly evaluated. We then turn to a variety ofmethods for dealing with the lack of sufficient fully-supervised
View Full Document
Unlocking...