计算机科学
人工智能
自然语言处理
特征(语言学)
语言学
匹配(统计)
期限(时间)
钥匙(锁)
集合(抽象数据类型)
透视图(图形)
作者
Futa Tajima,Yuto Funatsu,Runa Takeuchi,Kei Kimura,Kouta Kameoka,Takafumi Nakanishi
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-12-15
卷期号:: 265-268
标识
DOI:10.1109/iiai-aai-winter69777.2025.00056
摘要
This study proposes a novel information retrieval framework that allows users to discover novels by simply inputting a free-form "imagined story summary." Unlike conventional approaches that embed full texts or rely on keyword matching, the proposed method constructs a search space using review-style summaries generated from novel texts. These summaries, produced via Gemini and embedded with BERT, capture emotional tone, narrative structure, and reader-oriented impressions rather than surface-level lexical features. This enables user queries—often vague, abstract, or emotionally expressive—to be mapped into a semantically aligned vector space.The system embeds both AI-generated novel summaries and user-entered imaginative summaries, calculates cosine similarity, and presents both the most similar and most contrasting novels. This dual-direction search supports creative exploration by revealing not only works aligned with the user’s intent but also those with opposing thematic structures. Experiments conducted on 352 Aozora Bunko works demonstrated that the system successfully retrieved novels sharing psychological or ethical structures, even when expressions differed lexically. For instance, queries inspired by Akutagawa’s Rashōmon not only identified the original work but also surfaced thematically resonant novels such as Kyōka Izumi’s Polytheism and The Saint of Kōya. This study highlights that review-style summaries serve as effective content proxies for creative works, forming a feature space closer to human interpretation. While limitations include possible abstraction and homogenization in AI-generated summaries, results show that the approach effectively supports human-centered, emotion-oriented literary search. The framework introduces a new direction for creative information retrieval based on imaginative language and semantic embedding..
科研通智能强力驱动
Strongly Powered by AbleSci AI