Method of identification of patent trends based on descriptions of technical functions

Transcription

1 Journal of Physcs: Conference Seres PAPER OPEN ACCESS Method of dentfcaton of patent trends based on descrptons of techncal functons To cte ths artcle: D M Korobkn et al 08 J. Phys.: Conf. Ser Vew the artcle onlne for updates and enhancements. Ths content was downloaded from IP address on 0/0/08 at 00:6

2 Internatonal Conference Informaton Technologes n Busness and Industry 08 IOP Publshng IOP Conf. Seres: Journal of Physcs: Conf. Seres (08) do :0.088/ /05/3/03065 Method of dentfcaton of patent trends based on descrptons of techncal functons D M Korobkn, S A Fomenkov, A B Golovanchkov Volograd State Techncal Unversty, 8, Lenna Ave., Volgograd, , Russa E-mal: dkorobkn80@mal.ru Abstract. The use of the global patent space to determne the scentfc and technologcal prortes for the techncal systems development (dentfyng patent trends) allows one to forecast the drecton of the techncal systems development and, accordngly, select patents of prorty techncal subects as a source for updatng the techncal functons database and physcal effects database. The authors propose an orgnal method that uses as trend terms not ndvdual ungrams or n-gram (usually for exstng methods and systems), but structured descrptons of techncal functons n the form Subect-Acton-Obect (), whch n the authors' opnon are the bass of the nventon.. Introducton Due to a modern quck change of the global market and a wdespread occurrence of nnovatons, the lfe cycle of a new techncal soluton can be very short. In ths stuaton, t s actual to forecast the drecton of techncal systems development (trend) based on montorng technology changes n the global patent space (patent trend analyss). The scentsts are addressng the problem of dentfyng the patent trends: R. Fretsch [], H. Bronwyn [], B. Yoon [3], Y. Park, A. Kontostaths [4], etc. Based on the analyss of patent trends, t s useful to select the prorty techncal drectons as a source of nformaton for the actualzaton of physcal effects and techncal functons databases. At present, the reecton of usng the Internatonal Patent Classfcaton (IPC) n the process of patent trends determnaton s actual. As the bass of the developed method, t s proposed to use descrptons of techncal functons n the ( Subect, Acton, Obect ) form to dentfy both nterclass patent trends and trends wthn IPC classes. s a semantc structure that s used to represent the techncal functons extracted from textual data, ncludng the patent texts. Obect and Subect are the words or phrases that are related to the text semantcs. Actons are verbs, whch represent the operatons that connect the obects and subects. The development of natural language processng technologes allows structures to contan complete semantc nformaton about the techncal functons descrbed n the text. Currently, the problem of the patent databases analyss for the extracton of techncal functons s actvely studed by scentsts H. Park [5], D.U. Yufeng [6], S. Cho [7], K. Km, J.Y. Lee, J.Hu, J. Yoon, J. Guo, S. Fang, and others. They apply the techncal functons at the ntal stages of new techncal systems desgn and use the representaton of techncal functons n the form (Subect- Acton-Obect). Content from ths work may be used under the terms of the Creatve Commons Attrbuton 3.0 lcence. Any further dstrbuton of ths work must mantan attrbuton to the author(s) and the ttle of the work, ournal ctaton and DOI. Publshed under lcence by IOP Publshng Ltd

3 Internatonal Conference Informaton Technologes n Busness and Industry 08 IOP Publshng IOP Conf. Seres: Journal of Physcs: Conf. Seres (08) do :0.088/ /05/3/ Method for determnng scentfc and technologcal prortes of techncal systems development The authors propose an orgnal method that uses as trend terms not ndvdual ungrams or n-gram (usually for exstng methods and systems), but structured descrptons of techncal functons n the form Subect-Acton-Obect () [8,9], whch n the authors' opnon are the bass of the patent. The method of patent trend detecton conssts of the followng stages: ) dentfcaton of key s for each tme nterval; ) determnaton of trend s; 3) search of patents that contan the trend s; 4) cluster formaton for the patents that contan trend nformaton. One of the man tasks of the developed method of patent trends searchng s to present the trends n the form of nformatve, user-frendly terms that reflect the core of trends. Key s are very sutable for such terms. 3. Extracton of the s from the patent texts To extract the techncal functons n presentaton from patent texts, one used the prevously developed procedures [0] for segmentaton of complex sentences of patent texts, morphologcal and semantc analyss wth the constructon of dependency trees and buldng the deep-syntactc structures based on the Meanng-Text Theory for reduced Stanford dependences. Wth the correct work of the semantc analyzer, the root node of the deep-syntactc structure must be a verb (Fgure ) (t s extracted as an Acton ), then ts chld nodes are extracted wth actant relatons I ( Subect ) and II ( Obect ), and then for each node Subect, Obect and Acton the chld nodes are extracted wth attrbutve relaton accordng to the Meanng-Text Theory. The result of extractng the from the sentence The super-capactor electrode further comprsng a slane couplng agent s: acton: comprses (root node - comprses, attrbutve relaton further ); subect: super-capactor electrode (I actant relaton - electrode, attrbutve relaton supercapactor ); obect: slane couplng agent (II actant relaton - agent, attrbutve relaton slane, couplng ). Fgure. The example of the techncal functon To ncrease the nformaton content of the Subect-Acton-Obect structures, several s are combned nto the common structure accordng to the developed algorthm of groupng (comparson) s []. 4. Identfcaton of trend s To search for the patent trends, a method of searchng for trend s has been developed. The trend mples a textual descrpton of a techncal functon for whch the frequency of ts occurrence n patent texts ncreases wth a tme. However, there are random non-nformatve s, whch are characterzed by the mss of ther occurrence n the prevous tme perod and the begnnng of ther use

4 Internatonal Conference Informaton Technologes n Busness and Industry 08 IOP Publshng IOP Conf. Seres: Journal of Physcs: Conf. Seres (08) do :0.088/ /05/3/03065 n the current tme perod. In order to flter out such s, let us use the hypothess that n a short perod of tme the frequency of use of can not hardly change (decrease). Tr (vec vec + = (, TF ),(, TF ),...(, TF )} (), vec ) = (,Ch(TF +, TF )),...(,Ch(TF +, TF ))} k k k + TF TF + + Ch(TF, TF ) =, Ch(TF, TF ) LT TF + Kr where vec s the vector of s occurrences n the patents that are granted for the tme nterval ; TF - frequency of the -th n the patent texts; Tr(vec +, vec ) the trend vector (trend) of occurrence of all s between two neghborng tme ntervals; + Ch(TF, TF ) - the functon that determnes the trend (change of the occurrence of the -th ) between two neghborng tme nterval; Kr - the coeffcent of decreasng the mportance of the "rare" ; L T - the lmt of the mnmum allowable trend change for the tme nterval. Coeffcent Kr s meant for excludng from the lst of the trend s, whch rarely occur n the patent texts. Wth an nsgnfcant change of the absolute frequency of occurrence of such s, there s a sgnfcant ncrease of ther relatve change Ch between two neghborng tme ntervals. The experments carred out showed that the optmum value s Kr = 50. The L T parameter s desgned to regulate the number of fndng trend s. Its value s nversely proportonal to the number of s that can be consdered as trend and drectly proportonal to ther average sgnfcance as trends. The set of trend patents for a certan perod of tme s defned as follows: T =,,... } () N k k T : Ch(TF +, TF ) L T where T the set of the trend ; N - the number of the trend. 5. Search of patents that contan trend s The next step s to search for patents contanng trend s: T PAT = PAT,PAT,...PAT } (3) where T PAT set of trend patents N PAT - the number of trend patents. The authors have mplemented two approaches to the search for trend patents: The search for the trend s carred out only n the patent clams. The man dea of ths approach s that all s found n the clams are more sgnfcant than the s contaned n the other part of the patent text. PAT NPAT PAT T : Clam, T (4) where Clam patent clam PAT. The search for the trend s carred out n all parts of the patent. The man dea of ths approach s that trends can also occur n patents, the man descrpton (clam) of whch may 3

5 Internatonal Conference Informaton Technologes n Busness and Industry 08 IOP Publshng IOP Conf. Seres: Journal of Physcs: Conf. Seres (08) do :0.088/ /05/3/03065 not be relevant to the subect of the trend. Thus, the descrpton of the trend can be contaned n any part of the patent text (for example n the feld Desc full patent descrpton). PAT T : Clam, Abst, Desc }, T (5) PAT where Abst the feld Abstract of the PAT patent; Desc the feld Descrpton of the PAT patent. 6. Cluster formaton for patents contanng trend nformaton To realze the groupng of trend patents nto clusters, t s necessary to mplement the followng stages of analyss: buldng a term-document matrx; clusterng based on the Latent Drchlet Allocaton (LDA) [,3] model (Fgure ) and usng the constructed model to obtan the dstrbuton of vectors by the clusters (unnamed topcs). Fgure. Dstrbuton of vectors from unnamed topcs A term-document matrx s a table, whose rows are documents and columns are terms. Ths matrx descrbes the frequency of occurrence of each term n all documents. Terms are trend s, the documents trend patents. To construct the LDA model, a TF-based matrx should be generated. TF (term frequency) s the frequency of occurrence of the trend n the texts of trend patents. To ncrease the nformaton content of the trend s, several smlar s are oned nto one unque. Frst, t s necessary to obtan a thesaurus of all trend s of the patent database voc = t,t,t 3,..,t n },, where t s the next unque trend. Then each -th row of the term-document matrx wll be the resultng thesaurus, and each column wll represent the number of occurrences of the unque trend n the -th patent. The framework Spark and ts MLlb lbrary for machne learnng allow obtanng the thesaurus of all the terms of a set of documents and buldng a term-document matrx. Based on the LDA model, t s possble to obtan synchronous clusterng of trend patents and trend s on the same set of clusters (topcs). As a result, a soft cauterzaton s constructed - a trend patent may refer to several topcs wth varyng degrees of ownershp: _Cl = }, Pat_Cl = PAT } (6) where =,..,cl ndex of the cluster; cl the predetermned number of clusters; _Cl - cluster of the trend s; t T t-th trend n cluster _Cl ; Pat_Cl -th cluster of the trend patents; PAT k T PAT k-th trend patent n cluster Pat_Cl. t k 4

6 Internatonal Conference Informaton Technologes n Busness and Industry 08 IOP Publshng IOP Conf. Seres: Journal of Physcs: Conf. Seres (08) do :0.088/ /05/3/03065 Further, let us rank the obtaned topcs by the number of trend patents and correlate trend patents and ther trend. 7. Conclusons The developed method of dentfyng technology changes (detecton of patent trends) s focused on usng as trend terms not ndvdual ungrams or n-grams (usually for exstng methods and systems), but structured descrptons of techncal functons n the form (Subect, Acton, Obect). A trend s a textual descrpton of a techncal functon, for whch the frequency of ts occurrence n patent texts ncreases wth tme. For the flterng of random non-nformatve s, the hypothess s made that n a short tme perod, the frequency of use cannot change sharply (decrease). Two approaches to the search for trend patents have been mplemented: searchng only n the patent clam and searchng n full text of patent (trends can occur n patents, the clams of whch may not match wth the trendng topc). For the groupng of trend patents nto clusters, the followng stages of analyss are realzed: constructon of the term-document matrx and clusterng based on the LDA model. The termdocument matrx contans the frequences of occurrence of all trend s n all trend patents. Based on the LDA model, t s possble to obtan a synchronous clusterng of trend patents and trend s on the same set of clusters. The authors rank the obtaned topcs by the number of trend patents and correlate wth trend patents and ther trend 8. Acknowledgments The reported study was funded by the Russan Foundaton for Basc Research (research proect No A, research proect No A). References [] Fretsch R, Neuhäusler P, Rothengatter O 0 Patent applcatons - structures, trends and recent developments Studen zum deutschen Innovatonssystem 8 68 [] Bronwyn H H, Jaffe A B, Tratenberg M 00 NBER Workng Paper [3] Yoon B, Park Y 004 The Journal of Hgh Technology Management Research (5) [4] Kontostaths A, Galtsky L, et al 004 Survey of Text Mnng 85-4 [5] Park H, Yoon J, Km K 0 Identfyng patent nfrngement usng based semantc technologcal smlartes. Scentometrcs [6] Yufeng D U, Duo J I, Lxue J, et al. 06 Patent Smlarty Measure Based on Structure Chnese Sentence and Clause for Text Informaton Processng 30() [7] Cho S, Park H, Kang D, Lee J Y, Km K 0 An -based text mnng approach to buldng a technology tree for technology plannng Expert Systems wth Applcatons [8] Korobkn D M, Fomenkov S A, Kolesnkov S G, Golovanchkov A B 06 Techncal Functon Dscovery n Patent Databases for Generatng Innovatve Solutons Mult Conference on Computer Scence and Informaton Systems 06 ed P Kommers [9] Fnogeev A G, Parygn D S, Fnogeev A A 07 The convergence computng model for bg sensor data mnng and knowledge dscovery Human-centrc Computng and Informaton Scences 7 () [0] Korobkn D M, Fomenkov S A, Kravets A G, Kolesnkov S G 07 Methods of Statstcal and Semantc Patent Analyss CIT&DS 07 ed A Kravets, M Shcherbakov, et al (Germany, Sprnger) 48-6 [] Korobkn D M, Fomenkov S A, Kravets A G, Kolesnkov S G 07 Pror art canddate search on base of statstcal and semantc patent analyss Mult Conference on Computer Scence and Informaton Systems 07 ed Y Xao, A P Abraham

7 Internatonal Conference Informaton Technologes n Busness and Industry 08 IOP Publshng IOP Conf. Seres: Journal of Physcs: Conf. Seres (08) do :0.088/ /05/3/03065 [] Korobkn D M, Fomenkov S A, Kravets A G, Golovanchkov A B 06 Patent data analyss system for nformaton extracton tasks 3th Internatonal Conference on Appled Computng (AC) 06 ed H Weghorn 5-9 [3] Sadovnkova N, Parygn D, Kalnkna M, Sanzhapov D, Treu N N 07 Models and methods for the urban transt system research CIT&DS