Cross-Task Performance of Domain-Adapted Embeddings for Intellectual Property Applications

Effective analysis of patent literature is crucial for engineers and designers to identify prior art and track innovation trends. However, general-purpose text embeddings often miss the domain-specific terminology and structural nuances of patents, limiting their usefulness for intellectual-property (IP) tasks. While fine-tuning on specialized patent data improves performance on targeted tasks such as similarity search, it remains unclear how well such domain-adapted embeddings generalize across different IP tasks. In this study, we fine-tune pre-trained models from the Hugging Face ecosystem on the PatentMatch dataset. This dataset provides expert-verified claim/prior-art pairs extracted from the official European Patent Office (EPO) search reports, ensuring real-world relevance and high semantic label quality. The fine-tuning aims to pull semantically similar sentences closer in embedding space to improve retrieval accuracy. To assess the cross-task robustness, we compare the pre-trained and fine-tuned models with a strong proprietary baseline model (text-embedding-3-small model from OpenAI) on two tasks: (i) sentence-similarity retrieval on PatentMatch and (ii) IPC subclass classification based on EPO patent abstracts. Our experiments reveal a critical trade-off: while fine-tuning substantially enhances performance on sentence similarity, these same embedding models show a significant performance degradation in the classification task compared to both the original non-fine-tuned models and the baseline model. We hypothesize that specializing on sentence-level semantics causes the models to lose the broader, thematic information essential for document-level classification. These findings caution that optimizing embeddings for one IP task may impair others, suggesting task-specific rather than universal patent embeddings. These insights are particularly relevant for the development of intelligent design tools that utilize AI to effectively extract and apply technical knowledge from patents. They underscore the necessity of tailoring embedding models to specific design tasks to enhance efficiency and quality in the data-driven design process.