Scientometrics Прикладные компьютерные науки
A metaphor-aware framework for sleeping patent status prediction using large language models and hierarchical feature tokenizer transformer
Аннотация
Sleeping patents are granted patents that remain idle, neither exploited by assignees nor commercialized through market transactions, representing inefficiencies in technology management. Forecasting which patents are likely to be awakened is critical for portfolio management and innovation resource allocation. Existing studies mainly rely on patent intrinsic attributes or transaction records, which capture commercialization outcomes rather than early-stage demand. This study proposes a metaphor-aware framework that incorporates demand-side signals from online reviews as an additional predictive source. Non-expert consumers evaluating complex products frequently express dissatisfaction through metaphorical expressions, which share little lexical overlap with patent texts and are therefore overlooked by conventional methods. A fine-tuned large language model (LLM) is used to extract explicit pain points, metaphorical pain points, and decode metaphorical pain points into standardized technical-demand descriptions. Four feature sets are constructed: patent intrinsic attributes (A), explicit pain point features (B), metaphorical pain point features (C), and demand-gap features (D). The demand-gap features are treated as structural complements that describe the asymmetry between explicit and metaphorical demand signals, rather than as primary demand-content features. These features are processed by a Hierarchical Feature Tokenizer Transformer (HFT-Transformer). The empirical results show that incorporating demand-side signals into patent intrinsic attributes improves prediction performance, and that metaphorical pain point features provide additional and non-redundant information beyond explicit pain point features. Demand-gap features provide a structural context for interpreting the asymmetry between explicit and metaphorical demand signals, and their value is more evident when they are incorporated through the proposed hierarchical gating mechanism. Integrating patent intrinsic attributes with explicit and metaphor-aware demand-side signals allows earlier and more informed identification of patents with high awakening potential.