Common critiques of natural language processing (NLP)
methods cite their lack of multimodal sensory information,
claiming an inability to learn situated, action-oriented relations
through language alone. Barsalou’s (1983) theory of
ad hoc categories, which are formed from to achieve goals in
real-world scenarios, correspond theoretically to those types
of relations with which language models ought to have great
difficulty. Recent NLP models have developed dynamic approaches
to word representations, where the same word can
have different encodings depending on the context in which it
appears. Testing these models using categorization tasks with
human response data demonstrates that situated properties may
be partially captured through semantic analysis. We discuss
possible ways in which different notions of situatedness may
be distinguished for future development and testing of NLP
models.