Subjects: Psychology >> Social Psychology submitted time 2023-03-28 Cooperative journals: 《心理科学进展》
Abstract： How semantics are represented in human brains is a central issue in cognitive neuroscience. Previous studies typically detect semantic information by manipulating the properties of stimuli or task demands, or by asking a group of participants to judge the stimuli according to several given dimensions or features. Despite having brought valuable insights into the neurobiology of language, these approaches have some limitations. First, the experimental approach may only provide a coarse depiction of semantic properties, while human judgment is time-consuming and the results may vary substantially across subjects. Second, the conventional approach has difficulty quantifying the effect of context on word meaning. Third, the conventional approach is unable to extract the topic information of discourses, the semantic relations between the different parts of a discourse, or the semantic distance between discourses. The recently-developed natural language processing (NLP) techniques provide a useful tool that may overcome the above-mentioned limitations. Grounded on the distributional hypothesis of semantics, NLP models represent meanings of words, sentences, or documents in the form of computable vectors, which can be derived from word-word or word-document co-occurrence relationships, and neural networks trained on language tasks. Recent studies have applied NLP techniques to model the semantics of stimuli and mapped the semantic vectors onto brain activities through representational similarity analyses or linear regression. Those studies have mainly examined how the brain (i) represents word semantics; (ii) integrates context information and represents sentence-level meanings; and (iii) represents the topic information and the semantic structure of discourses. Besides, a few studies have applied NLP to untangle sentences’ syntactic and semantic information and looked for their respective neural representations. A consistent finding across those studies is that, the representation of semantic information of words, sentences and discourses, as well as the syntactic information, seems to recruit a widely distributed network covering the frontal, temporal, parietal and occipital cortices. This observation is in contrast to the results from conventional imaging studies and lesions studies which typically report localized neural correlates for language processing. One possible explanation for this discrepancy is that NLP language models trained on large-scale text corpus may have captured multiple aspects of semantic information, while the conventional experimental approach may selectively activate a (or several) specific aspects of semantics and therefore only a small part of the brain can be detected. Though NLP techniques provide a powerful tool to quantify semantic information, they still face some limitations when being applied to investigate semantic representations in the brain. Firstly, embeddings from NLP models (especially those from a deep neural network) are uninterpretable. Secondly, models differ from each other in training material, network architecture, amount of parameters, training tasks and so on, which may lead to potential discrepancies among research results. Finally, model training procedures differ from how humans learn language and semantics, and the inner computational and processing mechanism may also be fundamentally different between NLP models and the human brain. Therefore, researchers need to select a proper model based on research questions, test the validity of models with experimental designs, and interpret results carefully. In the future, it is promising to (i) adopt more informational semantic representation methods such as knowledge-graph and multimodal models; (ii) apply NLP models to assess the language ability of patients; (iii) improve the interpretability and performance of models taking advantages of cognitive neuroscience findings about how human process language.
Subjects: Psychology >> Social Psychology submitted time 2023-03-28 Cooperative journals: 《心理科学进展》
Abstract： As a fundamental technique in natural language processing (NLP), word embedding quantifies a word as a low-dimensional, dense, and continuous numeric vector (i.e., word vector). This process is based on machine learning algorithms such as neural networks, through which semantic features of a word can be extracted automatically. There are two types of word embeddings: static and dynamic. Static word embeddings aggregate all contextual information of a word in an entire corpus into a fixed vectorized representation. The static word embeddings can be obtained by predicting the surrounding words given a word or vice versa (Word 2Vec and FastText) or by predicting the probability of co-occurrence of multiple words (GloVe) in large-scale text corpora. Dynamic or contextualized word embeddings, in contrast, derive a word vector based on a specific context, which can be generated through pre-trained language models such as ELMo, GPT, and BERT. Theoretically, the dimensions of a word vector reflect the pattern of how the word can be predicted in contexts; however, they also connote substantial semantic information of the word. Therefore, word embeddings can be used to analyze semantic meanings of text. In recent years, word embeddings have been increasingly applied to study human psychology. In doing this, word embeddings have been used in various ways, including the raw vectors of word embeddings, vector sums or differences, absolute or relative semantic similarity and distance. So far, the Word Embedding Association Test (WEAT) has received the most attention. Based on word embeddings, psychologists have explored a wide range of topics, including human semantic processing, cognitive judgment, divergent thinking, social biases and stereotypes, and sociocultural changes at the societal or population level. Particularly, the WEAT has been widely used to investigate attitudes, stereotypes, social biases, the relationship between culture and psychology, as well as their origin, development, and cross-temporal changes. As a novel methodology, word embeddings offer several unique advantages over traditional approaches in psychology, including lower research costs, higher sample representativeness, stronger objectivity of analysis, and more replicable results. Nonetheless, word embeddings also have limitations, such as their inability to capture deeper psychological processes, limited generalizability of conclusions, and dubious reliability and validity. Future research using word embeddings should address these limitations by (1) distinguishing between implicit and explicit components of social cognition, (2) training fine-grained word vectors in terms of time and region to facilitate cross-temporal and cross-cultural research, and (3) applying contextualized word embeddings and large pre-trained language models such as GPT and BERT. To enhance the application of word embeddings in psychological research, we have developed the R package “PsychWordVec”, an integrated word embedding toolkit for researchers to study human psychology in natural language.
Hits
Hits
Downloads
Comment