Using Language Models and Latent Semantic Analysis to Characterise the N400m Neural Response

Using Language Models and Latent Semantic Analysis to Characterise the N400m Neural Response Mehdi Parviz, Mark Johnson Department of Computing Macquarie University Sydney, Australia {mehdi.parviz,mark.johnson} @mq.edu.au Blake Johnson, Jon Brock Macquarie Centre for Cognitive Science Macquarie University Sydney, Australia {blake.johnson,jon.brock} @mq.edu.au Abstract The N400 is a human neuroelectric response to semantic incongruity in on-line sentence processing, and implausibility in context has been identified as one of the factors that influ- ence the size of the N400. In this paper we in- vestigate whether predictors derived from La- tent Semantic Analysis, language models, and Roark’s parser are significant in modeling of the N400m (the neuromagnetic version of the N400). We also investigate significance of a novel pairwise-priming language model based on the IBM Model 1 translation model. Our experiments show that all the predictors are significant. Moreover, we show that predictors based on the 4-gram language model and the pairwise-priming language model are highly correlated with the manual annotation of con- textual plausibility, suggesting that these pre- dictors are capable of playing the same role as the manual annotations in prediction of the N400m response. We also show that the pro- posed predictors can be grouped into two clus- ters of significant predictors, suggesting that each cluster is capturing a different character- istic of the N400m response. 1 Introduction There is increasing interest in using computational models to help understand on-line sentence process- ing in humans. New experimental techniques in psycholinguistics and neurolinguistics are produc- ing rich data sets that are difficult to interpret us- ing standard techniques, and it is reasonable to ask if the statistical models developed in computational linguistics can be helpful here (Keller, 2010). The N400 is a human brain response to se- mantic incongruity or implausibility that has been widely studied in psycholinguistics and neurolin- guistics. A large set of factors has been shown to influence the strength of the N400, including intra- and extra-sentential context (Kutas and Fed- ermeier, 2000; Van Petten and Kutas, 1990). Here we study the strength of the N400 as measured by magnetoencephalography (MEG) (so the signal we study is sometimes called the N400m) on sentence- final words in a variety of “constraining” and “non- constraining” sentential contexts (Kalikow et al., 1977). For example, Her entry should win first prize is an example of a constraining-context sen- tence, while We are speaking about the prize is a non-constraining context sentence (target words are underlined in this paper). This paper shows that language models of the kind developed in computational linguistics can be used to help identify the factors that determine the strength of the N400. We investigate a number of different kinds of predictors constructed from a va- riety of language models and Latent Semantic Anal- ysis (LSA) to determine how well they describe the N400. The first set of predictors is derived from LSA, which is a method for analysing relationships between a set of documents and the terms they con- tain (Mitchell et al., 2010). LSA has been success- …

Content truncated to 500 words