Statistical learning of Korean phonotactics

다국어 초록 (Multilingual Abstract)

As in many languages, Korean distinguishes phonotactically impossible sound sequences from phonotactically possible, but rare sequences. This paper examines three such cases in Korean: post-obstruent tensing, diphthong restrictions, labial-[?] sequences ([p?] but *[w?]). Learning simulations of the Korean phonotactics were conducted using the phonotactics learning model by Hayes and Wilson (2008). It turns out that the resulting grammar sometimes fails to distinguish possible but rare sequences from impossible sequences, that is, rare possible forms sometimes scored worse than impossible sequences. Since the Hayes and Wilson model uses markedness constraints only, learning simulations that employs both markedness and faithfulness constraints (Goldwater and Johnson 2003) were also carried out. This seemingly does better, but the success of learning with faithfulness constraints is due to the proper level of generality provided by human linguists.

목차 (Table of Contents)

• 1. Introduction
• 2. Acceptability judgments in Korean
• 3. Learning Algorithm
• 4. Learning simulation
• 5. The solutions
• 6. Conclusion
• REFERENCES