Grounding co-occurrence: Identifying features in a lexical co-occurrence model of semantic memory

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• Lexical co-occurrence models represent word meaning using high-dimensional vectors derived from text usage.
• These automated models are advantageous over human-judgment-based semantic models.
• A key criticism is that co-occurrence models lack grounded representations, with concepts existing only relative to each other.

Purpose of the Study:

• To ground lexical co-occurrence models by connecting vector representations to empirical human data.
• To develop a method for retrieving semantic features from word vectors.
• To test the generalizability of this grounding method to novel concepts.

Main Methods:

• Training a feed-forward neural network using backpropagation.
• Mapping co-occurrence vectors to feature norms collected from human subjects.
• Evaluating the network's accuracy in retrieving and generalizing concept features.

Main Results:

• The neural network accurately retrieved semantic features from co-occurrence vectors.
• The model demonstrated high accuracy in generalizing feature retrieval to novel concepts.
• This establishes a link between abstract vector representations and concrete feature norms.

Conclusions:

• Lexical co-occurrence models can be effectively grounded by mapping them to human-perceived features.
• Neural networks provide a viable mechanism for bridging the gap between computational semantics and human cognition.
• This research advances the development of more robust and interpretable models of semantic memory.