Investigation Of Categorical Semantic Information Processing In The Brain And Natural Language Processing Models

Semantic representations stored in the nervous system in the human brain have been a hot topic in the field of cognitive neural research.It has been found that semantic representations in the human brain are stored in categories according to feature information.In the field of computers,various computer language models for natural language processing have also attempted to simulate semantic representations closest to the human brain using different algorithms.Previously,our group research has initially found roughly the similarity between computer semantic representations and human brain semantic representations.However,at the level of specific semantic features(e.g.,category semantic information),the representations in the dual complex system of brain and computer,the computer mechanism and the relationship between them are not very clear.Can we explore which of the different computer language models is closest to the semantic representation of the human brain,and which representations of categorical semantic information can be best predicted by computer models,using categorical semantic information as the main reference for analysis?To explore these questions,in this study we further delve into the semantic representation storage aspects of the human brain and computer models from the perspective of category affiliation.First,we selected 11 category words and 20 example words under each category,and two-word pairs under both correlated and uncorrelated conditions that category words and example words can form.Based on these wordpairs,we designed and executed a set of EEG semantic priming experiments.For data processing,we chose different time windows to obtain different EEG components,i.e.,we could obtain a characterization index of the distance of different categories of semantic information in a human brain,where the N400 component was the most relevant for semantic analysis.Then,we chose three computer semantic models,and each word-pair can find the corresponding relative semantic distance metrics in the corresponding computer model-characterized by the cosine value between the respective word vectors of two words,and by the same calculation as the EEG data,we obtained the metrics for each computer model characterizing each category of semantic information.Finally,we combined the computer model metrics with the EEG data metrics to compare the fit of the three computer models' representations of the semantic information of the categories and the differences in the similarity of the neurosemantic representations of the semantic information of the different categories between the two systems.The results show that the CBOW and CWE models can better characterize the category semantic information in the human brain,and the Glo Ve model predicts a poorer performance.Further analysis on different word categories revealed that different computer language models and EEG data showed different performance patterns.Semantic categories such as illness,seasonality,emotion and sports were better fitted by computer models and human brain,but semantic categories such as organ,occupation and dress were more different.Combined with the results of the questionnaire data conducted to obtain the experimental stimuli,it is speculated that it is the frequency of daily use of different exemplar words and the probability of simultaneous occurrence of category words and exemplar words in the corpus that produce the specific experimental results.This study links two complex systems,the human brain and the computer,using different categories of semantic information as mediators,deepening previous research and expanding new ideas for evaluating computer semantic models.