Revisiting the automatic prediction of lexical errors in Mandarin

Speech errors provide cues for explaining the process of word retrieval. For example, speech errors are less likely to occur with high-frequency words since these words already receive a high level of activation. The current analysis further develops existing findings in two ways. First, instead of considering the overall frequency of the words in the entire corpora, we consider the gap in frequency between sequential pairs of words. We hypothesize that speech errors are more likely to occur if the target has a much lower frequency than its preceding word. Second, we use word embedding methods to quantify the semantic distance between sequential pairs of words. We hypothesize that speech errors are more likely to occur with words that have a large semantic distance from their preceding context. We also consider the potential effects of phonetic distance between sequential pairs of words and position-in-utterance of words in utterances. The results from a Mandarin corpus of speech errors show that word frequency and semantic distance between sequential pairs of words can be used to predict the occurrence of speech errors with an accuracy above the majority baseline.