인공지능 언어 모델의 절 연결 능력에 대한 일고찰: 터보 및 다빈치 모델을 대상으로

초록

This study examines whether AI language models can aptly connect two clauses using ‘-ese’ and ‘-nikka’ representing reason. To achieve this goal, we used two language models of ChatGPT. Employing the gpt-3.5-turbo-0301 and text-davinci-003 models, we generated 800 examples following up the given clauses marked by ‘-ese’, ‘-essese’, ‘-nikka’ and ‘-essunukka’ representing reason. The generated examples were analyzed as follows. The proportion of appropriate examples was 59.48% in the turbo-model and 56.75% in the davinci-model. Both models exhibited the lowest proportion of appropriate examples when the given clause was marked by ‘-essunikka’. Among the aptly-connected examples, declarative sentences were overwhelmingly the most common sentence type, and the modality of the main clause was often associated with the meanings of ‘impossibility’, ‘wish’, ‘obligation’ and ‘intention’ in both models. The two models showed different outcomes regarding the frequency of the past tense marker ‘-ess-’ in the main clause in conjunction with the type of the dependent clause.

키워드