Disambiguating Noun and Verb Senses Using Automatically Acquired Selectional Preferences

Disambiguating Noun and Verb Senses Using Automatically Acquired Selectional Preferences* J udita Preiss Computer Laboratory Diana McCarthy and John Carroll Cognitive & Computing Sciences University of Sussex Brighton BNl 9QH, UK {dianam,johnca}©cogs.susx.ac.uk University of Cambridge, JJ Thomson Avenue Cambridge CB3 OFD, UK Judita.Preiss©cl.cam.ac.uk Abstract Our system for the SENSEVAL-2 all words task uses automatically acquired selectional prefer- ences to sense tag subject and object head nouns, along with the associated verbal pred- icates. The selectional preferences comprise probability distributions over WordN et nouns, and these distributions are conditioned on WordNet verb classes. The conditional distri- butions are used directly to disambiguate the head nouns. We use prior distributions and Bayes rule to compute the highest probability verb class, given a noun class. We also use anaphora resolution and the 'one sense per dis- course' heuristic to cover nouns and verbs not occurring in these relationships in the target text. The selectional preferences are acquired without recourse to sense tagged data so our system is unsupervised. 1 Introduction In the first SENSEVAL, we used automati- cally acquired selectional preferences to disam- biguate head nouns occurring in specific gram- matical relationships (Carroll and McCarthy, 2000). The selectional preference models pro- vided co-occurrence behaviour between Word- Net synsets1 in the noun hyponym hierarchy and verbal predicates. Preference scores, based on mutual information, were attached to the classes in the models. These scores were condi- tioned on the verbal context and the grammat- ical relationship in which the nouns for training had occurred. The system performed compara- • This work was supported by UK EPSRC projects GR/153175 'PSET: Practical Simplification of English Text' and GR/N36462/93 'Robust Accurate Statistical Parsing (RASP)'. 1 We will hereafter refer to WordN et synsets as classes. bly to the other system using selectional prefer- ences alone. The work here is an extension of this earlier work, this time applied to the English all words task. We use probability distributions rather than mutual information to quantify the prefer- ences. The preference models are modifications of the Tree Cut Models (TCMs) originally pro- posed by Li and Abe (1995; 1998). A TCM is a set of classes cutting across the WordNet noun hypernym hierarchy which covers all the nouns of WordNet disjointly, i.e. the classes in the set are not hyponyms of one another. The set of classes is associated with a probability distri- bution. In our work, we acquire TCMs condi- tioned on a verb class, rather than a verb form. We then use Bayes rule to obtain probability estimates for verb classes conditioned on co- occurring noun classes. Using selectional preferences alone for disam- biguation enables us to investigate the situa- tions when they are useful, as well as cases when they are not. However, this means we loose out in cases where preferences do not provide the necessary information and other complemen- tary information would help. Another disad- vantage of using selectional preferences alone for disambiguation is that the preferences only ap- ply to the grammatical slots for …

Content truncated to 500 words