Microsoft PowerPoint _KAIST_NTCIR5_Patent-DigitalPoster.ppt

Size: px

Start display at page:

Download "Microsoft PowerPoint _KAIST_NTCIR5_Patent-DigitalPoster.ppt"

みさえののした
5 years ago
Views:

1 Patent Document Retrieval and Classification at KAIST KAIST CS Dept. / BOLA Jae-Ho Kim, Jin-Xia Huang, Ha-Yong Jung, Key-Sun Choi

2 Introduction Tasks Document retrieval subtask Theme categorization subtask Our Approach Patent retrieval: find most important parts Patent document is structured Re-organize patents according to detailed descriptions Patent categorization: noisy elimination K-NN based: eliminate un-important documents and unimportant words in categorization MEM based: only eliminate un-important words in categorization By using re-organized patents! 2

3 Patent Retrieval - Finding the most important components of patent - Perform IR by component-by-component comparison 3

4 Patent Retrieval Find most important parts Patent document is structured Re-organize patents according to detailed descriptions Component-by-component comparison Two documents are similar, if the two documents are in the same technical classes and have the same (or similar) problem and solution (method). 4

5 Patent Retrieval: Find most important parts Phase 1: Document Indexing Target patent documents Re-organization Field 1 Field 2 Field n Keyword extraction Indexing Phase 2: Document Retrieval Query patent document Re-organization Field 1 Field 2 Field n Keyword extraction Query expansion Index 1 Index 2 Index n Query 1 Query 2 Query n Retrieval result 1 Retrieval result 2 Retrieval result n Merging & Re-arranging Document list 5

6 Patent document is structured Normative section <Bibliography> [publication date] [title of invention] <Abstract> [purpose] [composition] <Claims> [claim1] [claim2] <Description> [industrial application field] [problem to be solved] [means of solving problems] [operation] [embodiment examples] [effects of invention] <Explanation of Drawings> <Drawings> <DOCNO>PATENT-JA-UPA </DOCNO> <SDO BIJ> (43) 公開日平成 7 年 (1995)1 月 6 日 (54) 発明の名称スラリ散布を行う土壌作業機... <SDO ABJ> 目的スラリの処理と土壌作業を同時に行うことで構成トラクタとスラリを積載したバキウムカ - との間に <SDO DEJ> 産業上の利用分野本発明はスラリ散布を行う土壌作業機に関し発明が解決しようとする課題このようなスラリを圃場に供給する課題を解決するための手段上述のような目的を達成するために作用本発明のスラリ散布を行う土壌作業機は実施例以下本発明を採用した土壌作業機について添付した図面に発明の効果以上の説明から明らかなように <SDO EDJ> 図 1 本発明のスラリ散布を行う土壌作業機の側面図である <SDO DRJ> 図 1 Detailed component <SDO CLJ> 請求項 1 バキウムカ-を牽引して土壌心土に対して作業を行い Applicant- 請求項 2 トラクタに対して3 点リンクを介して装着される defined tags 6

7 Re-organization of Patent Documents <DOCNO>PATENT-JA-UPA </DOCNO> <SDO BIJ> (43) 公開日平成 7 年 (1995) (54) 発明の名称スラリ散布を行う... <SDO ABJ> 目的スラリの処理と土壌作業を構成トラクタとスラリを積載した [Publication Date] [Title of Invention] [Purpose] [Composition] <Bibliography> <Abstract> <DOCNO>PATENT-JA-UPA </DOCNO> <Technical Field> <Purpose> <SDO CLJ> 請求項 1 バキウムカ - を牽引して請求項 2 トラクタに対して [Claim1] [Claim2] <Scope of Claim> <Method> <SDO DEJ> 産業上の利用分野本発明はスラリ発明が解決しようとする課題このようなスラリを圃場に供給する課題を解決するための手段上述のような目的を達成するために作用本発明のスラリ散布を行う土壌作業機は実施例以下本発明を採用した土壌作業機について添付した図面に発明の効果以上の説明から明らかな <Detailed Explanation> [Application Field] [Problem to be solved] [Means of solving Problems] [Operation] [Embodiment Example] [Effects of Invention] <Claim> <Explanation> <Example> 図 1 本発明のスラリ散布を行う... <SDO EDJ> <Explanation of Drawings> [Figure1] 図 1... <SDO DRJ> [Figure1] <Drawings> 7

8 Find Most Important Parts in Patent Documents Patent Retrieval Throw 6 queries to 6 indexes The pairs of same fields are compared Using keywords and BM 25 Okapi (by using Lemur Toolkit) Use weighted summation for final result Query Document Document set Technological field Technological field w t Purpose Purpose w p Query Document Document set Method Method w m w c Result Full text Full text Claim Explanation Claim Explanation w exp w exa Example Example 8

9 Experiment Results in Document Retrieval Subtask Baseline Retrieval using full documents Post processing for increasing precision Re-retrieval from retrieved 6,000 documents by the comparison of Noun-Verb pairs score = score original + β score re retrieval Query Document Using re-organized patents showed better results! Full text Document set Full text RunId Condition a.ntc4 b.ntc4 Topics a.ntc5 b.ntc5 baseline β = d0010 β = d0011 β = d0012 d0013 d0014 d0015 β = 0.15 β = 0.20 β = 0.25 β =

10 Patent Categorization - Using the most important components of patent - Perform categorization by k-nn based and MEM based approaches: different in noisy elimination 10

11 Patent Categorization Use only the most important components Use re-organized patents, which has been proved helpful in patent retrieval Different noisy elimination approaches K-NN based: classifies a given patent into the theme codes of k documents similar to it Eliminate un-important documents by patent retrieval Eliminate un-important words by keyword extraction MEM based: classifies by similarity calculation between different document vectors Eliminate un-important words by using weighing functions and threshold 11

12 knn-based Patent Categorization Phase 1: Document Indexing Target patent documents Re-organization Field 1 Field 2 Field n Keyword extraction Indexing Index 1 Index 2 Index n Phase 2: Document Retrieval Query patent document Re-organization Field 1 Field 2 Field n Keyword extraction Query expansion Query 1 Query 2 Query n Phase 3: Document Categorization Retrieval result 1 Retrieval result 2 Retrieval result n Category list Category score calculation Merging & Re-arranging Most similar k Documents 12

13 MEM-based Patent Categorization Feature selection to reduce the feature size Vector: term vector with a word bag Weighting function: TF/IDF and TF/ICF CF(w): the category frequency of a word w TFICF ( w ( d ) i Threshold for selection C ) = TF ( wi, d) ICF ( wi ) ICF ( wi ) = log( ) CF( w ) Fixed threshold = { t t [0, n] } Floating threshold = { t t [avg(tficf(w(d)))-m, avg(tficf(w(d)))+m] m, n is an integer got from empirical experiments Classifier: MAXENT Toolkit i 13

14 Document Categorization Subtask knn-based Theme classification Training documents: 2 years documents (1993, 1997) Equally weighted merging (technological field, technological field) (purpose, purpose) (method, method) RunID ft001 ft002 ft003 ft004 ft005 Condition k=10 k=20 k=30 k=50 k=100 MAP

15 Experiments on MEM-based Classification K-NN-based approach showed better MAP than MEM based one Approach (ID) Condition Training data size (GB) MAP Top Map of k-nn MEM (ft006) MEM (ft007) TFICF, avg-2 TFICF, avg MEM was more fit to F-Term classification than Theme one MEM is relatively better for small category classification Run Theme F-term ID dt001 dt002 df001 ffx01 Condition TFICF, avg-1 TFIDF, avg TFICF, avg-1 TFICF, avg-2 MAP Top MAP among all teams/runs (dry run) (formal run) (dry run) (formal run) 15

16 Conclusions Component-by-component comparison Re-organization by using structural information of patent Precise comparison of components among document K-NN based approach for patent categorization Elimination of un-important documents for noisy reduction Using similar k documents for categorization 16

1 1 tf-idf tf-idf i

14 A Method of Article Retrieval Utilizing Characteristics in Newspaper Articles 1055104 2003 1 31 1 1 tf-idf tf-idf i Abstract A Method of Article Retrieval Utilizing Characteristics in Newspaper Articles