Discriminative learning predicts human recognition of English blend sources Strict compositionality in morphological theory is problematic for explaining how language-users comprehend phenomena like the partial yet non-decomposable forms in phonaesthemes and in blends

Strict compositionality in morphological theory is problematic for explaining how language-users comprehend phenomena like the partial yet non-decomposable forms in phonaesthemes and in blends like edutainment. An alternative account, based on discriminative learning, proposes that language-users associate linguistic cues (e.g., short segment or letter strings) with multiple simultaneous possible lexical and grammatical meanings [1,2]. Language-users acquire associations with different strengths based on exposure to these cues in conjunction with different meanings. Association strengths are derived from the probability and frequency with which each individual cue occurs together with each meaning and with other cues [3,4]. Discriminative learning predicts that when language-users see a set of orthographic cues, the meanings that are strongly associated with those cues should be activated. If this is how readers parse complex words, then wordforms that have cues which are better associated with the appropriate word meanings should be recognized faster and more reliably. This account, implemented as the naïve discriminative reading (NDR) model, has suggestive evidence from simulation-based analyses of online reading times [1,2], but has yet to be tested directly against human readers' comprehension abilities. In this paper, we evaluate the NDR on off-line identifications of English blend source meanings, hypothesizing that readers' ability to parse out source meanings from written blend forms should be correlated with how strongly the model associates the cues in each form with the correct source meanings. An NDR model was trained on the wordforms and frequency data in the English CELEX database using the naive discriminative learning R package [5]. Following [1], letter unigrams and bigrams were taken as cues to word meanings. The meanings associated with each word included the lemma (lexical meaning) and the inflectional meanings listed in CELEX. For example, the meanings associated with the form geese were GOOSE and PLURAL. Two experiments were conducted, using (1) attested English blends and (2) novel blends. In Experiment 1, 89 attested noun, adjective, and verb blends were taken from [6] and [7]; blends were excluded if they contained a partial form listed as productive in [8]. 100 participants, recruited via Mechanical Turk, were asked to guess the two source words that each blend was constructed from. In Experiment 2, novel blends were generated using 20 pairs of co-hyponyms (e.g., sting and insult) taken from WordNet. Four possible blends for each pair were constructed by splitting on inter-syllable or onset-rime boundaries (insting, stingsult). 100 participants were shown one …