[Paper Review] Text-based Technological Signatures and Similarities: How to create them and what to do with them.

Daniel S. Hain, Roman Jurowetzki+2 more|arXiv (Cornell University)|Mar 27, 2020

Computational Drug Discovery Methods92 references4 citations

TL;DR

This paper introduces a method to generate text-based technological signatures for patents using NLP embeddings, enabling efficient computation of technological similarities across the entire patent universe via approximate nearest neighbor techniques. The approach supports large-scale technological network analysis, with validated applications in measuring knowledge flows, patent quality, and technological change.

ABSTRACT

This paper describes a new approach to measure technological similarity between patents by leveraging their textual description. Using embedding techniques from natural language processing, we represent their description as a high dimensional numerical vector, thus capturing their technological signature. Deploying an almost near linear-scaling approximate nearest neighbor matching techniques, we are able to compute technological similarity scores for all existing patents. This enables us to represent the whole patent universe as a technological network. We validate both technological signature and similarity in various ways, and demonstrate their usefulness to create patent quality indicators, measure knowledge flows, and map technological change.

Motivation & Objective

• To develop a scalable method for measuring technological similarity between patents using textual descriptions.
• To create compact, high-dimensional technological signatures from patent text using NLP embedding techniques.
• To enable efficient similarity computation across the entire patent universe using near-linear scaling algorithms.
• To validate the technological signatures and similarity scores for reliability and practical utility.
• To demonstrate applications in patent quality assessment, knowledge flow tracking, and mapping technological evolution.

Proposed method

• Represent each patent's textual description as a high-dimensional numerical vector using pre-trained NLP embeddings.
• Apply approximate nearest neighbor (ANN) algorithms to efficiently compute pairwise technological similarity scores at scale.
• Construct a technological network by modeling patents as nodes and similarity scores as edges.
• Validate the technological signatures through consistency checks and benchmarking against known technological relationships.
• Use the resulting similarity scores to derive indicators for patent quality, knowledge diffusion, and technological change.
• Leverage the embedding space to analyze semantic and structural patterns in technological development.

Experimental results

Research questions

• RQ1 How can patent text be transformed into a stable, informative technological signature using NLP techniques?
• RQ2 To what extent can approximate nearest neighbor methods enable scalable similarity computation across the entire patent universe?
• RQ3 How reliable are the generated technological signatures and similarity scores in capturing real-world technological relationships?
• RQ4 What practical applications emerge from the technological network constructed using these signatures?
• RQ5 Can the method effectively measure knowledge flows and track technological change over time?

Key findings

• The method successfully generates stable and informative technological signatures from patent text using NLP embeddings.
• Approximate nearest neighbor techniques enable scalable similarity computation across the entire patent universe with near-linear time complexity.
• The resulting technological network captures meaningful relationships between patents, validated through consistency and benchmarking.
• The approach enables the creation of novel patent quality indicators based on technological proximity and innovation context.
• The framework effectively maps knowledge flows and tracks patterns of technological change across industries and time.
• The study demonstrates that text-based technological signatures provide a robust foundation for large-scale innovation analytics.