In recent years, a number of models of lexical access based on attractor networks have appeared. These models reproduce a number of effects seen in psycholinguistic experiments, but all suffer from unrealistic representations of lexical semantics. In an effort to improve this situation we are looking at techniques developed in the information retrieval literature that use the statistics found in large corpora to automatically produce vector representations for large numbers of words. This paper concentrates on the problem of transforming the real-valued cooccurrence vectors produced by these statistical techniques into the binary- or bipolar-valued vectors required by attractor network models, while maintaining the important inter-vector distance relationships. We describe an algorithm we call discrete multidimensional scaling which accomplishes this, and present the results of a set of experiments using this algorithm.

In this paper we investigate representational and methodological issues in a attractor network model of the mapping from orthography to semantics based on [Plaut, 1995]. We find that, contrary to psycholinguistic studies, the response time to concrete words (represented by more 1 bits in the output pattern) is slower than for abstract words. This model also predicts that response times to words in a dense semantic neighborhood will be faster than words which have few semantically similar neighbors in the language. This is conceptually consistent with the neighborhood effect seen in the mapping from orthography to phonology [Seidenberg & McClelland, 1989, Plaut etal., 1996] in that patterns with many neighbors are faster in both pathways, but since there is no regularity in the random mapping used here, it is clear that the cause of this effect is different than that of previous experiments. We also report a rather distressing finding. Reaction time in this model is measured by the time it takes the network to settle after being presented with a new input. When the criterion used to determine when the network is "settled" is changed to include testing of the hidden units, each of the results reported above change the direction of effect - abstract words are now slower, as are words in dense semantic neighborhoods. Since there are independent reasons to exclude hidden units from the stopping criterion, and this is what is done in common practice, we believe this phenomenon to be of interest mostly to neural network practitioners. However, it does provide some insight into the interaction between the hidden and output units during settling.