Identifying generalizable knowledge from the distribution of tonotactic accidental gaps in Mandarin

This study investigates tonotactic accidental gaps (unattested syllable-tone combinations) in Mandarin Chinese. In a corpus study, we found that, independent of syllable type, T2 (rising) and T3 (falling-rising) gaps were over-represented, whereas T1 (high level) and T4 (falling) gaps were under-represented. We also observed fewer T1 gaps with voiceless onsets and more T2 and T3 gaps with voiceless onsets, a pattern that is consistent with cross-linguistic observations. While these trends were generally reflected in a wordlikeness rating experiment by Mandarin listeners, their judgements of these gaps, similar to those of real words, were also guided by neighborhood density. Furthermore, T2 gaps with real-word T3 counterparts were rated as more wordlike, a result attributed to the T3 sandhi in Mandarin Chinese. Finally, we used harmonic scores generated from the UCLA Phonotactic Learner to explicitly test the role of lexical knowledge and markedness constraints in modeling speakers’ tonotactic knowledge reflected in the wordlikeness ratings. We found that grammars induced from lexical data were the most successful at predicting wordlikeness ratings of gaps and lexical syllables combined. However, when focused on the ratings of tonotactic gaps, grammars with markedness constraints informed by cross-linguistic observations were more successful even without the constraints being weighted on lexical data. The results show how lexical knowledge and universal markedness, which is not entirely learnable from the lexicon, may account for some tonotactic generalizations.