PP-Attachment: A Committee Machine Approach

PP-Attachment: A Committee Machine Approach Martha A. Alegre Departament de LSI Universitat Polithcnica de Catalunya malegre©lsi, upc. es Agusti Lloberas Josep M. Sopena Lab. de Neurocomputaci6 Universitat de Barcelona pep~axon, psi. ub. es Lab. de Neurocomputacid Universitat de Barcelona agust i¢axon, psi. ub. es Abstract In this paper we use various methods for multiple neural network combination in tasks of prepo- sitional phrase attachment. Experiments with aggregation functions such as unweighted and weighted average, OWA operator, Choquet inte- gral and stacked generalization demonstrate that combining multiple networks improve the esti- mation of each individual neural network. Using the Ratnaparkhi data set (the complete training set and the complete test set) we obtained an ac- curacy score of 86.08%. In spite of the high cost in computational time of neural net training, the response time in test mode is faster than others methods. 1 Introduction Structural ambiguity is one of the most seri- ous problems that Natural Language Processing (NLP) systems face. This ambiguity takes place because the syntactic information alone does not suffice to make an assignment decision. Con- structions such as Prepositional Phrase (PP), co- ordination, or relative clauses are affected. An exhaustive study about the information needed to deal with this particular structural ambiguity has not been carried out as of yet; nevertheless, in the current literature we can find several pro- posals. • In certain cases, it seems that the informa- tion needed to solve the attachment comes from the general context. (1.a) John has a telescope. (1.b) He saw the girl with the telescope. In this particular case, a correct attachment would require a model representing the sit- uation in which the different entities are in- volved. If this were true for all of the cases, determining PP assignment would require highly complex computation. • In some other cases, the information de- termining the PP attachment seems to be local. Some works [Woods et al, 1972], [Boguraev, 1979], [Marcus et al. 1993] sug- gested several strategies that based their 231 decision-making on the relationships ex- isting between predicates and arguments- what [Katz and Fodor, 1963] called selec- tional restrictions. Cases belonging to this group seem to be easier to handle computa- tionally than the former ones. Regarding these different cases we can speak of two kinds of disambiguation mechanisms. One that can be called a low level mechanism which uses mainly information regarding selectional re- strictions between predicates and arguments. This mechanism uses a local context in order to solve syntactic disambiguation: that which is constituted by the predicate and its arguments. The second mechanism uses higher level infor- mation such as situation models. If the low level mechanism does not solve the ambiguity, the high level mechanism, which would be acti- vated later, should be able to do it. There are empirical data that seem to support the fact that human beings use these two mechanisms both for word sense disambiguation and syntactic disam- biguation. For a review see [Sopena et al. 1998]. 1.1 Local …

Content truncated to 500 words