Text Matching to Measure Patent Similarity
The document describes a new text-based measure of patent similarity that uses the title and abstract to identify unique keywords for each patent. It then calculates the Jaccard index between patents to measure similarity. This approach is validated by expert assessment and by comparing text-matched patents to those matched by the US Patent Classification System (USPC). The results show the text-based measure outperforms USPC by having fewer false positive and false negative matches. The type I and II errors of the USPC are estimated to be 12-22% and 20-52% respectively depending on the matching method used.
1. Text Matching to Measure Patent Similarity
Sam Arts
Faculty of Business and Economics
KU Leuven
sam.arts@kuleuven.be
Bruno Cassiman
IESE Business School, KU Leuven
bcassiman@iese.edu
Juan Carlos Gomez
University of Guanajuato
jc.gomez@ugto.mx
OECD Blue Sky Conference 2016
2. 2
The United States Patent Classification System (USPC)
• Prior and current research relies on patent classification
(USPC)
– To identify similar patents (counterfactual control)
– e.g., Jaffe, Trajtenberg, and Henderson, 1993; Almeida, 1996; Agrawal, Cockburn, and Rosell,
2010
– To measure similarity between patents and patent portfolios
– e.g., Argyres, 1996; Ahuja, 2000; Rosenkopf and Almeida, 2003; Makri, Hitt, and Lane, 2010
• USPC
– Too broad
– Changes over time (patents are reclassified)
– Manually assigned
– e.g. Thompson and Fox-Kean, 2005; Belenzon and Schankerman, 2013; …
3. 3
• Unclear what the bias
– Type I: false positive (dissimilar patents, same USPC)
– Type II: false negative (similar patents, different USPC)
• No alternatives
– Using subclasses instead of classes
– e.g. Thompson and Fox-Kean, 2005
– Using all classes instead of primary
– e.g. Benner and Waldfogel, 2008
• Unclear how alternatives affect Type I or Type II bias
The United States Patent Classification System (USPC)
4. 4
• Title and abstracts from all US utility patents granted
between 1976-2013 (4.4 million)
• Concatenate title and abstract, lowercase, eliminate stop
words (SMART system >600 words), words<2 characters,
numbers, words which appear only once
• Each patent collection of unique keywords
• 526,561 keywords; avg 37 per patent
• Drop patents with less than 10 keywords (0.3% of sample)
Text-based measure of similarity
5. 5
• Simple Jaccard index
– Range 0-1
• For each of 4.4 million patents, select closest text-matched
patent within same year (cfr JHT 1993)
– Min Jaccard of 0.05 (0.5% drop)
– More drop when matching on USPC!
• Avg Jaccard 0.24
– 14 common keywords for 2 patents with 37 keywords
• As a baseline, select distant text-match patent within same
year (Jaccard=0, closest filing date)
Text matching (instead of USPC)
6. 6
Validation: closest text-matched patents in same year
Patent pairs with a larger Jaccard are more like to belong to same patent family (docdb), inventor(s),
assignee(s), and are more likely to cite each other
7. Validation: expert assessment
7
• 5 independent R&D scientists
– Semiconductor devices, chemical engineering, power plants, genetics, and
optical inspection systems
• For each expert
– Randomly select 10 baseline patents
– For each baseline patent one random patent with Jaccard
– 0.00
– 0.05-0.25,
– 0.25-0.50,
– 0.50-0.75,
– 0.75 onwards
– Randomize order and ask experts to rate similarity 1-7
9. 9
Estimate bias related to USPC
• For each of the 4.4 million patents select three USPC
matched patents
• Three common ways of matching, approximate filing date
and …
– Primary class
– e.g. Jaffe et al. 1993
– No match for 2% of patents
– Primary class and subclass (nested)
– e.g., Almeida 1996
– No match for 20% of patents
– All classes and subclasses
– Jaccard overlap in subclasses
– e.g. Agrawal et al. 2010
– No match for 4% of patents
10. 10
Type I error – false positive matches
• Dissimilar patents, same USPC
• Low similarity
– Primary class: 0.054
– Primary class and subclass (nested): 0.092
– All classes and subclasses: 0.097
• Lower bound: % USPC matches with Jaccard=0
– Primary class: 12%
– Primary class and subclass (nested): 4.3%
– All classes and subclasses: 4.0%
11. 11
Type II error – false negative matches
• Similar patents, different USPC
• Lower bound: % different USPC among patents with Jaccard index of 1
– Primary class: 22.4%
– Primary class and subclass (nested): 52.3%
– All classes and subclasses: 20.0%
12. Validation: superiority text-matching over USPC
12
Text-matched patents are more like to belong to same patent family (docdb), inventor(s), assignee(s),
and are more likely to cite each other
14. 14
Conclusions
• Text mining
– To measure patent similarity and select counterfactual control patents
– Outperforms USPC
• Fine-grained
• Does not rely on human classification
• No changes over time
– Measure similarity between portfolio’s, aggregate keywords at portfolio level
• Bias related to USPC
– Matching on primary subclass instead of class reduces Type I but increases
Type II
– Matching on all subclasses instead of primary reduces both Type I and Type II
– Unexpected large share of Type I and particularly Type II errors remain
present
• Code and data publically available
– JAVA standard libraries, csv files with cleaned words and 200 closest matches.
15. 15
• Develop new measure of patent similarity based on text
• Validate new measure
– Same patent family, assignee, inventors, cite each other
– Expert assessments
• Estimate bias related to USPC
• Validate superiority over USPC
– Patent family, assignee, inventors, cite each other
– Expert assessments
Text mining
17. 17
• Title + abstract: Process for amplifying, detecting, and/or-cloning nucleic acid
sequences, The present invention is directed to a process for amplifying and
detecting any target nucleic acid sequence contained in a nucleic acid or mixture
thereof. The process comprises treating separate complementary strands of the
nucleic acid with a molar excess of two oligonucleotide primers, extending the
primers to form complementary primer extension products which act as
templates for synthesizing the desired nucleic acid sequence, and detecting the
sequence so amplified. The steps of the reaction may be carried out stepwise or
simultaneously and can be repeated as often as desired. In addition, a specific
nucleic acid sequence may be cloned into a vector by using primers to amplify
the sequence, which contain restriction sites on their non-complementary ends,
and a nucleic acid fragment may be prepared from an existing shorter fragment
using the amplification process
• 52 unique keywords: acid act addition amplification amplified amplify
amplifying carried cloned complementary comprises contained desired
detecting directed ends excess existing extending extension form fragment
invention mixture molar non-complementary nucleic oligonucleotide prepared
present primer primers process products reaction repeated restriction separate
sequence sequencesthe shorter simultaneously sites specific steps stepwise
strands synthesizing target templates treating vector
Text-based measure of similarity