Algorithms of Anaphora: Human vs ChatGPT Strategies in Processing Korean Bridging References

This study examines ChatGPT’s use of bare and demonstrative anaphoric NPs in Korean bridging references, comparing its performance with Korean speakers’ judgments through two tasks: rating and cloze task. Both tasks were constructed around two conditions, bridging relations (Whole-Part and Product-Producer) and subject types (topic and nominative subjects), for a detailed comparison. The results indicate a marked difference between ChatGPT’s algorithmic and Korean speakers’ intuitive judgments. ChatGPT consistently rated demonstrative NPs with ku higher, showing a preference for explicit anaphoric references, unlike Korean speakers, who favored the demonstrative ku more selectively in Product-Producer contexts and topic-marked sentences. In the cloze task, both groups showed a preference for the demonstrative ku in Product-Producer contexts, though differences in the choice of anaphoric markers emerged across bridging relations. Scatter plot analyses highlighted both convergence and divergence in the judgments of ChatGPT and human participants. The discrepancies are likely due to ChatGPT’s English-centric training, which favors explicit markers like the anaphoric marker the, and the unique morphological distinctions in Korean anaphoric references. Korean speakers demonstrated efficient cognitive processing by adapting anaphoric markers to various contexts. This study underscores the need to consider linguistic diversity and cognitive strategies in AI model development for multilingual applications.