Poster #3 - From distributional to semantic categories

• In Event: 3-003 - Poster Session 09
  In Poster Session: PS 09 Section - Attention, Learning, Memory

Integrative Statement

In recent years, there has been a renewal of interest in the idea that semantic similarities can be derived from distributional patterns in which words (or objects) tend to occur. An apple and a banana may be considered similar because they share a number of relevant features, but they also may be considered similar because they co-occur with similar sets of objects and words (apple – eat; banana - eat). Thus, tracking patterns of co-occurrence may potentially be a useful strategy for learning about object similarities and for developing taxonomic relations. In the light of this account, young children’s preference for co-occurrence relations may be understood as an important aid to the subsequent taxonomic relations development.
Here we investigated what types of relations 4-year-old children and adults track when exposed to rich distributional information in linguistic input. Additionally, we tested the usefulness of this information in semantic categorization. Participants were exposed to novel words embedded in familiar sentence frames. Novel words replaced either adjectives or nouns in sentences. Each novel adjective was consistently paired with one novel noun and one familiar noun (e.g. foobly mipp, foobly apple), forming a “triad”. Within each triad, the adjective and nouns could be linked based on reliable co-occurrence. On the other hand, nouns from the same triad, never occurred in the same sentence. However, participants who acquired adjective-noun co-occurrence information could link nouns based on the shared context (shared adjective), making a “same class” link. After they received extensive training, participants were first tested in a free association task. Subsequently, they performed a labeling task in which they were asked to use novel nouns to label familiar objects. These familiar objects belonged to the same semantic categories as did familiar nouns from the triads. In the free association task, both age groups relied exclusively on co-occurrence links. However, in the labeling task, adults reliably mapped newly learned nouns to the semantic category of the corresponding familiar noun (i.e. same-triad noun). In contrast, only one third of 4-year-olds responded in accordance with the “same-class” link.
Our findings suggest that taxonomic relations can be learned quite well from word co-occurrence patterns. However, there is also evidence that even when co-occurrence links are very strong, they may not give an instant rise to taxonomic links, as these regularities may be harder to track.

Authors

• Olivera Savic, The Ohio State University
  Presenting Author

• Vladimir M. Sloutsky, The Ohio State University
  Non-Presenting Author