Adventures with Age, Animacy, and Object-Similarity Embeddings
As we age, we adjust how we associate objects with each other. We examine aspects of age, animacy, and object similarity by using an interpretable computational model trained to perform an odd-one-out-among-three task and the dataset of human responses used to train that model. The trained model contains a vector embedding for each object used in the task; that vector is used for relating that object with other objects, where each vector dimension corresponds with a human-identifiable category (like “body-part related”). First, we use this model to select questions for experimentally examining differences between child (age 6) and adult prioritization of taxonomic and thematic features in performing that odd-one-out task. Second, we examine what information is encoded when the model is trained to have very few dimensions. Finally, we compare which features of the model best explain the responses for adult respondents of different age groups.
Tuesday, June 21, 2022
10:55 EDT
7:55 PDT, 8:55 MDT, 11:55 ADT
15:55 BST
Dawn McKnight
Master’s Student
University of Alberta
Dawn is a third-year master’s student studying NLP under the supervision of Alona Fyshe. They’re a big fan of video-game music and languages, and in their free time, you’ll oftentimes find them reading about games, participating in/hosting game-music-listening parties and guessing games, and reading about the history of English and other languages.
