The patent system is designed to encourage disclosure of new technologies and novel ideas by granting exclusive rights on the use of inventions to their inventors, for a limited period of time. Before a patent can be granted, patent oces around the world perform thorough searches to ensure that no previous similar disclosures were made. In the intellectual property terminology, such kind of searches are called prior art searches. In some industries, the number of granted patents a company owns has a high impact on the market value of the company. This underlines the importance of well-performed prior art searches. Together with the TrecChem track [5], also organized by our institution, the ClefIp eort comes to complete the work that is being done in the series of Ntcir workshops (see for example [ 4]). The rst ClefIp track ran within Clef 20091. The purpose of the track was twofold: to encourage and facilitate research in the area of patent retrieval by providing a large clean data set for experimentation; to create a large test collection of patents in the three main European languages for the evaluation of crosslingual information access. The ClefIp data set includes documents published by the European Patent Oce (Epo) which contain a mixture of English, German and French content. The track focused on the task of prior art search. In 2010 and 2011, the ClefIp track was organized as a benchmarking activity (lab) in the Clef conference. In these years, the main goal of the ClefIp eort remained the same to foster research in the patent retrieval area, and provide a large clean data set. To this end, the number of tasks in the track was increased and the data set was enlarged. Recognizing the importance of patent classications in the daily activity of an intellectual property professional, in 2010 the ClefIp benchmarking activity included a patent classication task. The participants were asked to classify 1 http://www.clef-campaign.org
given patent application documents according to the International Patent
Classication ( Ipc) system [
This year (2011), in addition to the (now classic) prior art search task and the patent classication task two patent image related tasks were oered. Often, patent applications contain images that clarify details about the invention they describe. Images in patents may be drawn by hand, by computer, or both, may contain text, and are generally black-and-white (i.e. not even monochrome). Depending on the technological area of a patent, images may be technical drawings of a mechanical component, or an electric component, owcharts if the patent describes, for example, a workow, chemical structures, tables, etc. When a patent expert browses through a list of search results given by a search engine, he or she can very quickly dismiss irrelevant patents to the patent application by just glancing at the images in the retrieved patents. The number of documents to be looked at in more detail is thus greatly reduced. With the Image-based Document Retrieval and Image-based Classication tasks in ClefIp we try to make this aspect of an IP expert’s daily work familiar to the research communities 12 international teams have participated this year, we present here an overview their work and research results. The paper is structured as follows: Section 2 describes the test collection used this year, section 3 presents the participating teams and gives an overview of the methods the teams involved. In the same section we also present the main measurements done in this track. 2 2.1
The ClefIp collection contains patents, physically stored as a collection of
Xml les encoding patent documents. A patent document may be an application
document, a search report, or a granted patent document. Each patent document
is assigned a kind code, which appears as a sux to the patent identier (e.g.
EPnnnnnnn-A1, WO-nnnnnnnnnn-A2). In the case of the Epo, patent application
documents that include a search report carry the code A1, patent application
search reports carry the code A3, granted patent documents carry the code B1,
etc.2 For a description of key terms and steps in a patent’s lifecycle see [
An important tool in organizing the large amount of patent data which patent
oces regulate is the classication system . A patent classication system ‘sorts’
the patents according to the technical area they belong to, and it is a basis for a
quick investigation of the state of the art in a eld [
Kind codes used by the Wipo are listed at http://www.wipo.int/patentscope/en/ wo_publication_information/kind_codes.html
The 2011 ClefIp data collection is based on the 2010 data, and is extracted from the Marec3 data corpus. The ClefIp collection contains mainly patent documents published by the Epo.
Two important additions were made to the 2010 collection corpus. The rst
one was to include in the distributed corpus certain patents published by the
World Intellectual Patent Organization ( Wipo). A high percentage of the Epo
patents contained in the ClefIp corpus are patent applications internationally
led under the Patent Cooperation Treaty ( Pct [
In number of documents, adding the Wipo patent documents to the collection corpus increased with 1.2 million patent documents, to a nal number of approximately 3.5 million Xml documents, referring to approximately 1.5 million patents. The images corresponding to the 47 thousands Xml documents in the three Ipc subclasses in the ImgPac task occupy 5.4 Gb for 290880 ti les.
The same as in the previous years, the test collection corpus was delivered to the participants as is, without merging the documents related to the same patent into one document. Each patent is identied by a unique patent number a string identifying the publishing oce (EP for the Epo and WO for the Wipo) followed by a series of digits. Corresponding to each patent is a directory containing the Xml documents representing the patent documents related to that patent. For the EP patents, the layout is 00000n/nn/nn/nn/*.xml, where the sequence of digits in the directory layout corresponds to the one in the patent number. For the WO patents, the layout is 00nnnn/nn/nn/nn/*.xml, where the rst 4 digits (after ‘00’) represent the document publication year.
For example, to patent EP 0981201 corresponds the directory containing les EP-0981201-A2.xml, EP-0981201-A3.xml, and EP-0981201-B1.xml: > pwd EP/000000/98/12/01 > ls EP-0981201-A2.xml EP-0981201-A3.xml EP-0981201-B1.xml To patent WO 1994030029 corresponds the directory containing the le WO1994030029-A1.xml: > pwd WO/001994/03/00/29 > ls WO-1994030029-A1.xml 3 The Marec data corpus is a collection of over 19 million patent documents, in Xml format, made available by the Irf for research purposes.
The patent image les are stored as tif les in one separate folder, the correspondence between the image le and its Xml le is established with a small set of rules.
All textual documents in the ClefIp collection contain the following main Xml elds: bibliographic data, abstract, description, and claims. Not all documents actually have content in these elds. The content of the various Xml elds can be English, German, or French. Some elds may occur more than once, each time with a dierent language. The Xml patent le have also a document language (English, German or French), this not excluding that its subelds occur with a dierent language attribute than the document language. For example, granted EP patent documents (EP-nnnnnnn-Bn.xml) must contain claims in three languages (English, German and French). 2.2
Tasks and Topics 5 tasks were proposed to the participants: Prior Art Candidates Search ( Pac), Patent Classication ( Cls1), Rened Patent Classication ( Cls2) Patent Imagebased Retrieval (ImgPac ), and Patent Image-based Classication ( ImgCls ). The topics for each of the tasks were chosen from the same topic pool we have used in 2010. We will detail each of the proposed tasks in the following. Prior Art Candidates Search. The rst task in this track ( Pac) consisted in nding patent documents in the target collection that may invalidate a given patent application. The participants were provided with one set of 3973 topics. The topics were formulated as ‘Find all patents in the collection that potentially invalidate patent application EP-nnnnnnn-An.’, where the Xml le storing the patent application document EP-nnnnnnn-An (which we call the topic le or topic patent ) was given in an attached archive. A third of the topic les’ document language was English, another third was German, and the last third was French. The task did not restrict the language used for retrieving the documents, but participants were encouraged to use the multilingual characteristic of the collection (namely, that claims in granted patent documents are provided in three languages). A small set of 300 training topics was also provided, and participants were allowed to use the ClefIp 2010 topics sets in their system training.
Patent Classication. The second task in the ClefIp track (Cls1) required to classify a given patent document according to the Ipc system. the topics were formulated as ‘Classify patent document EP-nnnnnnn-An according to the Ipc system.’, where the Xml le storing the patent application document EPnnnnnnn-An was given in an attachment. The Ipc system is hierarchically organized into sections, classes, subclasses, groups and subgroups. The classication was to be given at the subclass level. The set of classication topics contained 3000 patent documents, a dierent set than the one used in the Pac task. Again, the task didn’t restrict the language used for classication, but the topic language was English for one third of the topics, German for the next third, and French for the last third of the topics. Participants could use the ClefIp collection corpus as a training set.
Rened Patent Classication. This task is, at least in formulation, very similar to the Cls1 task. It required the participants to classify given patent documents according to the Ipc system. The subclass was given, the participants had to return the group/subgroup classication levels.
Patent Image-based Prior Art Search. This task (ImgPac ) was introduced as a pilot task. It has the same aim as the Pac task, except that the images and Xml les corresponding to the patents were available and were to be used. Only the three Ipc subclasses listed above were used, as for these classes, patent searchers often rely on visual comparison of images in the patents to nd relevant prior art. The queries consisted of the text and complete set of images of 211 patents, with the topics formulated in the same way as for the Pac task. Patent Image-based Classication. The aim of the image classication task (ImgCls ) was to automatically classify patent images based on visual content. For the ImgCls task, only images extracted from patents, not full patents, were provided. Participants to this task did not need the full ClefIp corpus. The classication was into nine classes: drawing, chemical structure, program listing, gene sequence, ow chart, graph, mathematics, table, and symbol. Training data with between 300 and 6,000 training images for each of these classes was provided, and only these data were to be used to train image type classication techniques. The task was to train a classier using the provided training data, and test the resulting classier on a set of 1,000 patent images. 2.3
Relevance Assessments
The relevance assessments used to evaluate the Pac and ImgPac submissions
were obtained automatically from the patent citations stored in the collection
documents. Since the average number of citations per patent in the ClefIp
collection is low (below 4), we have looked for methods to extend the set of
relevant documents per topic. For this we used an extended list of citations,
where to the patents listed in the patent’s search report (the direct citations), we
added also the patent citations listed in the family members of the topic patent,
as well as the family members of the cited patents. For detailed explanations of
the citation extraction procedure, we point the reader to the overview article [
The relevance assessments used to evaluate the Cls1 and Cls2 submissions were also obtained automatically from the documents that originated the Cls topics. We have extracted the Ipc codes, restricted to the subclass level, and group/subgroup level respectively, from the patent documents.
The relevance assessments for the ImgCls topics were done manually by us.
With the exception of the ImgCls task, a submission consisted of a single text
le with at most 1,000 answers per topic, in the standard format used for the
Trec submissions. The submissions to the ImgCls task consisted of a single
text le with one entry per topic, each entry containing the topic id and nine
space separated values one for each class used in the classication. 12
participating groups have submitted a total of 77 runs, (unequally) distributed over the
ve proposed task. Table 1 shows the list of participating groups, marking the
tasks where runs were submitted. The submissions were uploaded to the Direct
system [
P C C Im Im DE x DE x RU x IN x AT NL x AT AT x CH x NL x x x KR x x x FR This section is based on the descriptions provided by the participants. We present here which Xml elds were used in document processing, what kind of pre and postprocessing was done, the retrieval and ranking system that was used to 4 The Direct system is currently developed under the Promise project obtain the results, crosslanguage techniques involved, as well as any other relevant details. ⋆ The hildesheim and chemnitz groups collaborated in try to identify how patent IR is aected by the query length and the use of linguistics in the retrieval process. Their method was developed on top of the Xtrieval framework, which provides a common interface to several search engines. In preprocessing, they used a specic stopword list, especially created for the patent domain. Subsequently, each language was indexed separately. In addition to the text, the Ipc classes were also added to the index. Like other groups this year, they extracted dierent types of phrases in order to improve the precision of the results. However, their method was not purely statistical, but also used a rule based dependency parser.
The search process considered three types of queries: term-based, phrasesbased and a combination of the two. They also compared very long queries with short (and precise) queries. Their results would indicate that using long queries is better than using small precise ones. It is arguable however, how indicative of the content of the given patent application a short query can be. Furthermore, they show that using linguistic phrases did not increase the eectiveness of the retrieval system. ⋆ At the time of writing this text, we have no information on how the experiments submitted by the hprussia participant were obtained. ⋆ The team from Hyderabad combines three methods in order to retrieve and rank prior art. At the base, their method relies on the Lemur toolkit and the translation of queries into English. First, they apply a key phrase extraction method in order to create queries from the topic patent documents. Then, they identify references to other patens within the text of the document and use them in two ways: rst, to create a document vector based on the Ipc codes assigned to the current patent application and to all the referenced patents. Second, to add them directly to the result list.
Without using this citation list, the best results appear to be those using both the Ipc information and the text search. When using citations, the Ipcinformation seems to reduce the quality of the results. ⋆ The joanneum group participated in the ImgCls task. They use features such as Local Binary Patterns (LBP), MPEG-7 features as well as OCR’d text extracted from the images. Support Vector Machines are used for the classication. Runs either use a single type of feature, or combine the classication results of dierent types of features using late fusion. The best run is the one using the LBP only. ⋆ The lugano and tuwien-2 teams took up on the Pac task by focusing on generating patent summaries that improve the query formulation. Topic patent documents are summarized by the PatTextTiling technique, which is an adaptation of the TextTiling summarization algorithm. The topics’ Ipc codes are used to dene a relevance set (corpus documents that share Ipc codes with the topic) which is used to get closer to a better relevance model. Query models can be built summary-based or description-based (from the description document sections). After removing stop-words and stemming, the patent document corpus was indexed using Terrier. The ranking model used is BM25. The experiments are ltered by the topic’s Ipc code and patent citations extracted from the topic’s text are added to the results. Only English topics were considered, for this reason the plot values for the German and French languages are missing in Figure 3. ⋆ In its approach to tackle the Patent Classication tasks, the nijgmenen team use the Linguistic Classication System ( Lcs) to implement three classiers: Naive Bayes, Winnow, and SVM lihgt. Various experimental settings were compared to gradually improve classication results. Among such settings are the use of dierent document sections, of patent citations, dierent document representations, and dierent training data sets. Parameter tuning during training also contributed to better classication results. In the training phase only English abstracts and descriptions were used. As metadata the Ipc codes, and the applicant, inventor, and address elds were extracted. It turned out that applicant, inventor and address information did not contribute much to the classication results. To the abstracts and (rst 400 words of) the description thus extracted the Aegir dependency parser was applied, its results being added to the document representations. Patent citations in the topic les were used to rerank the classication results.
For the Prior Art Candidates search task, the nijgmenen group teamed up with the spinque group, focusing on using bag-of-words approach enriched with syntacticsemantic information. Only the English content of the titles, abstracts, claims and descriptions was used. (For this reason, Figure 7 doesn’t have plot entries for the nijgmenen experiments.) In a separate processing step, the Ipc information and the rst 400 words of the description were extracted. The extracted English content was cleaned up by removing image references and claim headers, sentenced, and parsed with Aegir. Both the corpus and the topics were processed in this way. Selecting the query terms was done based on their relevance to Ipc classes, relevance computed with the ( Lcs) software. Finally, the retrieval was done using with the Spinque framework, which allows the denition of search strategies via a graphical user interface. ⋆ The wisenut group participated both in the Prior Art Candidate search and Classication tasks, although the Pac participation was due to implementing a KNN-like classication using Pac search results. This solution was chosen based on the experience that training and classifying documents suer from having few training documents, not enough memory for the (usually) large classication model, and from using too much processing time. To obtain the Pac search results, the Xml les were processed in the following way: weighted keywords were selected out of the title, abstract, description, and claims eld. Then, after Pos tagging, functional words, stop words and noncontent words are removed and cooccurrence terms are added. Terms and text content in German and French were translated into English using the MyMemory translation service 5. The system used to index, search and classify is based on Lucene, to which a simple Java-based interface was implemented. The interface included also an application for classication that applied a improved weighting scheme in the KNN classication to obtain the Cls1 and Cls2 results. ⋆ The xerox-sas group participated in both the ImgPac and ImgCls tasks. Images for both tasks are represented using Fisher vectors. For the ImgCls task, linear classiers are used. The best results are obtained for the run in which the images were articially rotated and added to the training set, to take into account that the images are sometimes rotated. For the ImgPac task, different strategies are investigated to compare one set of images to another (as patents consisting of a group of images, not single images, are to be retrieved). Runs are also created in which the predicted ImgCls image classes are used. For the retrieval of the patent text, dierent sections of the patent are weighted dierently. Similarities are also calculated based on shared Ipc categories and similarities of the patent citation graph, with late fusion used to combine the similarities. A weighted late fusion strategy was again used to combine the text and image ranking, with the text rankings weighted higher than the image rankings. While the visual retrieval performs poorly, when combined with text retrieval it outperforms the text-only retrieval. 3.2
Evaluation Results We have evaluated the submitted experiments using the most common metrics in IR. Before we ran the evaluation software, some simple cleanup of the data was done. This included replacing whitespace sequences with only one blank space, ltering out experiment entries which did not belong to a given topic patent documents.
For each submitted Pac experiment we computed the following measures: For each submitted Cls experiment we computed the following measures:
All computations were done using the trec_eval 9.0 software provided by Nist. Figures 1 through 8 show some of the calculated measures. Detailed values for each of the mentioned measures were sent to the lab participants and are soon to be published into a technical report. 5 MyMemory, http://mymemory.translated.net 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 map
P P at 100
The gures below use a shortened version of the original run names, whose length would have made the pictures less legible. The mapping between the original run name and the shortened versions is shown in the appendix. Note that, although run names might be the same, the experiment les are dierent between tasks. For each submitted ImgCls experiment, we computed Equal .y5H .3hC .h2C .u1L .Lu2 .Lu3 .Lu4 .h7C .h1C i.3N i.1N .4hC .y3H i.7W i.6W i.5W i.8W i.4W i.3W i.2W i.1W .y4H .y6H .y2H i.2N .y1H .h6C i.4N .h5C .1pH Error Rate (EER) and Area Under Curve (AUC) of a ROC curve, and True Positive Rate (TPR) per class averaged over all classes, as well as confusion matrices. This was done using a custom-written script running in Octave 6. The results of all runs are summarized in Table 2. 4
We have given an account of the benchmarking activities done in the frame of the ClefIp 2011. Compared to the last year, collaborations between research groups has intensied. Another positive observation is that participants are drawing on the research results obtain in the previous years to improve their retrieval methods. The coagulation of research groups leads to a consolidation of the methods used for patent retrieval and allows them to reach a maturity
Rat100 all en de fr 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0
C H W W W W W W W W H H H H C C H
Fig. 3. Map measures per languages for the Pac runs Pat1 Pat5 Rat5 F1at5 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
0 0.9 0.8 0.7 0.6
0.25
all en de fr 0.35 0.3 0.25 0.2 0.15 0.1 0.05 joanneum.alphacentauri 0.15 0.91 0.66 joanneum.arcturus 0.24 0.81 0.50 joanneum.betelgeuse 0.18 0.90 0.62 joanneum.canopus 0.16 0.91 0.65 joanneum.procyon 0.37 0.67 0.27 joanneum.rigel 0.16 0.90 0.63 joanneum.sirius 0.16 0.91 0.64 joanneum.vega 0.32 0.72 0.28 xerox-sas.RUNORH 0.06 0.98 0.85 xerox-sas.RUNORH_ROTRAIN 0.04 0.99 0.91 xerox-sas.FV_ORH_SP 0.08 0.92 0.85 xerox-sas.MEAN_ALL 0.08 0.91 0.85
Table 2. Summary of the ImgCls run results. level which would make them a candidate for commercial exploitation. However, it also reduces the diversity of the submissions. The number of participating groups was lower this year and at the workshop we need to explore the ways in which the evaluation of patent retrieval tools needs to go ahead.
One such was was identied as of last year. Image retrieval is extremely important for many technologies patented. However, participation in the image tasks was low in the rst. The ImgPac pilot task is challenging due to the multimodal nature of the retrieval task, the large amount of data and the full patent retrieval (containing a set of images) as opposed to single image retrieval. It is hoped that this task can lead to more collaboration between image and text retrieval groups in the next years. The ImgCls task was planned so as to have a lower threshold of entry for groups with image classication expertise. While 6 groups registered to obtain the data, only 2 participated.
Acknowledgements This work was partly supported by the EU Network of Excellence PROMISE 7 (FP7-258191) and the Austrian Research Promotion Agency (FFG) FIT-IT project IMPEx 8 (No. 825846).
Original id NIJMEGEN.RUN_ADMWCIT NIJMEGEN.RUN_ADMWOWCIT NIJMEGEN.RUN_ADMWOW NIJMEGEN.RUN_ADMW NIJMEGEN.RUN_ADWOWCIT NIJMEGEN.RUN_ADWOW NIJMEGEN.RUN_ADWTCIT NIJMEGEN.RUN_ADWT WISENUT.WISENUT_R1_BASE WISENUT.WISENUT_R2_BASE_10 WISENUT.WISENUT_R3_BASE_20 WISENUT.WISENUT_R4_BASE_30 WISENUT.WISENUT_R5_CO WISENUT.WISENUT_R6_CO_10 WISENUT.WISENUT_R7_CO_20 WISENUT.WISENUT_R8_CO_30
id NIJMEGEN.RUN_WINNOW_WORDS Ni.1 WISENUT.WISENUT_R1_BASE Wi.1 WISENUT.WISENUT_R2_BASE_10 Wi.2 WISENUT.WISENUT_R3_BASE_20 Wi.3 WISENUT.WISENUT_R4_BASE_30 Wi.4 WISENUT.WISENUT_R5_CO Wi.5 WISENUT.WISENUT_R6_CO_10 Wi.6 WISENUT.WISENUT_R7_CO_20 Wi.7 WISENUT.WISENUT_R8_CO_30 Wi.8
id XEROX-SAS.3MAX3MEAN Xe.1 XEROX-SAS.3MAX3MEAN_LATEMONO Xe.2 XEROX-SAS.3MAX3MEAN_MT_CIT Xe.3 XEROX-SAS.3MAX_LATEMONO Xe.4 XEROX-SAS.3MAX_MT Xe.5 XEROX-SAS.FVORH_3MAX Xe.6 XEROX-SAS.FVORH_3MAX3MEAN Xe.7 XEROX-SAS.MAXMEANMODAD Xe.8 XEROX-SAS.MAXMEANMODAD_MT Xe.9 Table 6. Original and short run XEROX-SAS.MAX_MT_CIT Xe.10 names for the Pac task