Family lexicon: Using language models to encode memories of personally familiar and famous people and places in the brain

Knowledge about personally familiar people and places is extremely rich and varied, involving pieces of semantic information connected in unpredictable ways through past autobiographical memories. In this work, we investigate whether we can capture brain processing of personally familiar people and places using subject-specific memories, after transforming them into vectorial semantic representations using language models. First, we asked participants to provide us with the names of the closest people and places in their lives. Then we collected open-ended answers to a questionnaire, aimed at capturing various facets of declarative knowledge. We collected EEG data from the same participants while they were reading the names and subsequently mentally visualizing their referents. As a control set of stimuli, we also recorded evoked responses to a matched set of famous people and places. We then created original semantic representations for the individual entities using language models. For personally familiar entities, we used the text of the answers to the questionnaire. For famous entities, we employed their Wikipedia page, which reflects shared declarative knowledge about them. Through whole-scalp time-resolved and searchlight encoding analyses, we found that we could capture how the brain processes one's closest people and places using person-specific answers to questionnaires, as well as famous entities. Overall encoding performance was significant in a large time window (200-800ms). Using spatio-temporal EEG searchlight, we found that we could predict brain responses significantly better than chance earlier (200-500ms) in bilateral temporo-parietal electrodes and later (500-700ms) in frontal and posterior central electrodes. We also found that XLM, a contextualized (or large) language model, provided superior encoding scores when compared with a simpler static language model as word2vec. Overall, these results indicate that language models can capture subject-specific semantic representations as they are processed in the human brain, by exploiting small-scale distributional lexical data.