Conhecimento sintatico-semantico e processamento de sentenças em rede neural recorrente simples / Syntactic-semantic knowledge and sentence processing in a simple recurrent neural network

AUTOR(ES)
DATA DE PUBLICAÇÃO

2006

RESUMO

This dissertation begins with a retlection about the interdisciplinary roots of psycholinguistics, which is taken as the research field concemed with the investigation of the cognitive processes subjacent to the verbal behavior. This investigation employs different methods and techniques, given the complexity inherent to the study of processes that take place inside the organism, and therefore are not directly observable. In view of this, we introduce the methodology adopted in this dissertation - modeling. The usage of models is initially dealt with in general terms, and then is analysed in the context of the study of cognitive processes. At this point, it is crucial to notice that modeling such processes implies making assumptions about how they are. The general hypothesis underlying the models we are interested in is that cognition is computation, that is, cognitive processes are computational. This hypothesis has given birth to two different paradigms of modeling - one based on the fundamentals of digital computing, and other based on the principIes of neurocomputing. Afier a brief discussion of their main differences, the later is presented in details. Once the theoretical-methodological basis of the neurocomputing framework is established, we replicate Elman s (1990) experiment in which a simple recurrent neural network is trained to process words in sequences, forming simple sentences. It should be noted that the sequences reflect grammatical information that can be described in syntactic (subjects, verbs and objects) and semantic (agents, patients and themes) terms. Our intent was to investigate whether and how such information is leamed by the system. Analyses of this experiment have shown that the processing is guided by lexical information, that is, linguistic information at the sentence leveI is not used. In the next step we have modified this experiment by training the network with the saroe set of sentences, just introducing an end-of-period mark between them. The analyses have shown that, in this new experiment, the processing is guided by both syntactic and semantic information, such as thematic roles. Taking these results as starting point, we devised a new experiment - a similar network was trained with a set of sentences in which word sequences were main1y led by semantic information like thematic role; the end-of-period mark between sentences was used. The intent was to controllinguistic information so as to increase our confidence in the results. Results consistent with the fonner experiments were obtained, that is, the processing is indeed guided by both syntactic and semantic infonnation at the sentence level. The consistency of the results achieved a1lows us to state that there are both syntactic and semantic knowledge operating during sentence processing. This result is important because it extends the already known simple recurrent network ability in coding syntactic information, showing that SRNs also are capable of handling semantic infonnation like thematic roles

ASSUNTO(S)

psycholinguistics psicolinguistica cognição conexionismo cognition connectionism

ACESSO AO ARTIGO

Documentos Relacionados