Dr. Patrick Krauss: Analysis of language processing in artificial neural networks and the human brain
Gastvortrag von Dr. Patrick Krauss an der FU Berlin
Redner: Dr. Patrick Krauss (Universitätsklinikum Erlangen)
Titel: Analysis of language processing in artificial neural networks and the human brain
In contemporary linguistics, there exist a number of competing theories of grammar that can be divided into two main classes, namely generative grammar and construction grammar. Generative grammar regards grammar as a system of rules (syntax) that when applied to words (lexicon) generates exactly those combinations of words that form grammatical sentences in a given language. Hence, according to this theory, syntax and lexicon are clearly distinguished modules. In contrast, according to construction grammar, instead of being discrete modules and thus subject to very different processes, syntax and lexicon form the extremes of a continuum (from regular to idiosyncratic).
In order to decide between the competing linguistic theories, a combination of neuroimaging and computational approaches will be used to identify the mechanisms underlying language processing. At a first step, state-of-the-art deep neural networks will be trained on natural language data (text corpora) and tested with new, previously not learned input, i.e. parts of language of different complexity. In addition, MEG/EEG and fMRI measurements on human participants will be performed using the same stimuli as for the computational models.
So far, we developed a new statistical method for analyzing and comparing neuroimaging data obtained from EEG/MEG measurements in humans, which allows quantifying the similarity/dissimilarity of different cortical activation patterns. Recently, this method has been generalized and now also enables to open the Black Box of artificial neural networks, i.e. to analyze and compare the emerged representations. Using these methods, the similarity/dissimilarity of internal representations in the computational models corresponding to the inputs will be compared with human temporal and spatial cortical dynamics.
The constructed computational models, together with the experimental results, may serve to decide between competing linguistic theories, or to develop modified theories of grammar, respectively
Zeit & Ort
17.04.2019 | 16:00 - 18:00
Habelschwerdter Allee 45 JK 31/122