A New Corpus of Lexical Substitution and Word Blend Errors: Probing the Semantic Structure of Lemma Access Failures

Abstract

Models of lemma access in language production predict occasional mis-selection of lemmas linked to highly similar concepts (synonyms) and concepts standing in a set-superset relation (subsumatives). It is unclear, however, if such errors occur in spontaneous speech, and if they do, whether humans can detect them given their minimal impact on sentence meaning. This data report examines a large corpus of English spontaneous speech errors and documents a low but non-negligible occurrence of these categories. The existence of synonym and subsumative errors is documented in a larger open access data set that supports a range of new investigations of the semantic structure of lexical substitution and word blend speech errors.