Phrase similarity in humans and machines.

Abstract

Computational models of semantics have emerged as\npowerful tools for natural language processing. Recent\nwork has developed models to handle compositionality,\nbut these models have typically been evaluated on large,\nuncontrolled corpora. In this paper, we constructed\na controlled set of phrase pairs and collected phrase\nsimilarity judgments, revealing novel insights into hu-\nman semantic representation. None of the computa-\ntional models that we considered were able to capture\nthe pattern of human judgments. The results of a sec-\nond experiment, using the same stimuli with a trans-\nformational judgment task, support a transformational\naccount of similarity, according to which the similarity\nbetween phrases is inversely related to the number of ed-\nits required to transform one mental model into another.\nTaken together, our results indicate that popular mod-\nels of compositional semantics do not capture important\nfacets of human semantic representation.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 29 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.