This paper proposes an approach to distinguish between and link complaints and requests within a complaint survey dataset.The mixture of complaints and requests can obscure the intended opinion, while conveying either a complaint or a request alone is also insufficient. Expressing only a complaint leaves the specific course of action unclear, while expressing only a request obscures the origin of that request. The proposed method addresses this issue by distinguishing between complaints and requests and then organizing them by corresponding complaints with their potential solutions. The method extracts complaints and requests by referencing sentence-end expressions, clue words, and morphological features. Corresponding complaints with requests, unlike tasks that match similar content, requires capturing a causal relationship.To accomplish the corresponding, this paper compares combinations of elemental technologies, such as document vector similarity and interpretation by Large Language Models (LLMs), to determine an effective corresponding method.The results demonstrated that combining the elemental technologies yielded superior precision compared to applying each technology in isolation.
