Reproduce “A Survey on Sentence Embedding Models Performance for Patent Analysis” on OpenResearch

About this paper

Hamid Bekamiri, Daniel S. Hain, Roman Jurowetzki

Researchers from Aalborg University Business School conducted a survey on sentence embedding models for patent analysis, finding that domain-adapted models like PatentSBERTa and Bert-for-patents achieved higher accuracy in multi-label patent classification compared to general-purpose models, with optimal performance varying across technical domains. The work also proposes a standard library and dataset to support future research in this area.

Key insights

• Domain-specific pre-training or fine-tuning is crucial for NLP models in patent analysis, as adapted models significantly outperform general-purpose models without such specialization.
• The optimal embedding model for patent analysis is not universal; performance varies considerably depending on the specific technical domain or CPC section.
• Even simpler, well-adapted non-parameterized models (like TF-IDF weighted Word2Vec) can sometimes outperform complex general-purpose transformer architectures if the latter are not adequately prepared for the patent domain.

Summary by alphaXiv.