A-Novel-Patent-Similarity-Measurement-Methodology

Document Similarity Prediction

• A-Novel-Patent-Similarity-Measurement-Methodology: Semantic Distance and Technological Distance

  • Yongmin Yoo, Cheonkam Jeong, Sanguk Gim, Junwon Lee, Deaho Seo

---

Dataset

• We used Google patent, randomly extract 420 pairs patents from 2019 to 2020, with which we perform expert validation
• The expert panel whom has expertise in Data Analytics, Data Mining, and Artificial Intelligence assesses how semantically similar two patents
• We used the score of law expert when there was a large difference in scores between experts.

Expert	Frequency
Expert panel	335
Law Expert	85
Total	420

---

Model Structure

• Semantic Similarity

  • We used Patent Bert

• Technical Similarity

  • $Intersection_{A,B}=Patent_{A} \cap Patent_{B}$
  • $Union_{A,B}=Patent_{A} \cup Patent_{B}$
  • $TD_{A,B}={Intersection_{A,B} \over Union_{A,B}}$

• Hybrid Similarity

  • $SDTD_{A,B}={(TD+1) \cdot SD \over 2}$

---

Result

Model	Pearson	Spearman
CNN	0.074951	0.035229
LSTM	0.05409	0.05967
Bi-LSTM	0.14646	0.185534
BERT	0.507742	0.542219
Proposed model	0.56935	0.644496