Karim Benzineb’s research while affiliated with Shanghai University of International Business and Economics and other places

Patent classifications are built to set up some order in the growing number and diversity of inventions, and to facilitate patent information searches. The need to automate classification tasks appeared when the growth in the number of patent applications and of classification categories accelerated in a dramatic way. Automated patent classification systems use various elements of patents’ content, which they sort out to find the information most typical of each category. Several algorithms from the field of Artificial Intelligence may be used to perform this task, each of them having its own strengths and weaknesses. Their accuracy is generally evaluated by statistical means. Automated patent classification systems may be used for various purposes, from keeping a classification well organized and consistent, to facilitating some specialized tasks such as prior art search. However, many challenges remain in the years to come to build systems which are more accurate and allow classifying documents in more languages.

In this task 2,000 patents in three languages (English, French and German) were to be classified among approximately 600 categories. We used a classifier based on neural networks of the Winnow type. This classi- fier is already used for similar tasks in professional applications. We test- ed three different approaches to improve the classification accuracy: the first one aimed at solving the issue of poorly-documented categories, the second one was meant to enrich the overall training corpus and the third one was based on the processing of the corpus' collocations. Although we ranked first in this competition, none of the three approaches mentioned above provided for a clear improvement in classification accuracy; our re- sults were essentially due to the implementation of the classification algo- rithm itself.

In this experiment led at the University of Geneva (UniGE), we evaluated several similarity measures as well as the relevance of using automated classification to filter out search results. The patent field is particularly well suited to classification-based filtering because each patent is already classified. Our results show that such a filtering approach does not improve searching performances, but it does not have a negative impact on recall either. This last observation allows considering classification as a possible tool to reduce the search space without reducing the quality of search results.

Several studies have tried to improve retrieval performances based on automatic Word Sense Disambiguation techniques. So far, most attempts have failed. We try, through this paper, to give a deep analysis of the reasons behind these failures. During our participation at the Robust WSD task at CLEF 2008, we performed experiments on monolingual (English) and bilingual (Spanish to English) collections. Our official results and a deep analysis are described below, along with our conclusions and perspectives.

This task was meant to compare the results of two d ifferent retrieval techniques: the first one was ba sed on the words found in documents and query texts; the secon d one was based on the senses (concepts) obtained b y disambiguating the words in documents and queries. The underlying goal was to come up with a more prec ise knowledge about the possible improvements brought b y word sense disambiguation (WSD) in the informatio n retrieval process. The proposed task structure was interesting in that it drew up a clear separation b etween the actors (humans or computers): those who provide the corpus, those who disambiguate it, and those who q uery it. Thus it was possible to test the universality and t he interoperability of the methods and algorithms i nvolved.

Nous présentons ici un moteur d'indexation ayant fait l'objet d'un transfert technologique entre l'université et l'industrie. Ce moteur est actuellement intégré dans des applications utilisées par les organisations internationales. Les corpus indexés sont volumineux et multilingues. En partant des spécificités du cahier des charges, nous examinons les choix d'architecture et de technologies effectués pour répondre aux contraintes de performance et de volumétrie. L'utilisation optimale des ressources de mémoire, de calcul et de stockage est discutée. Le séquençage et la parallélisation des processus sont examinés. ABSTRACT. We present here an indexing engine which is covered by a technology transfer agreement between the University and the private sector. This engine is currently included in various applications used by international organizations. The document collections which are indexed are large and multilingual. The particular elements of the technical specifications are the starting point of our analysis; then we look at the design and technology choices made to meet the performance and volume constraints. The optimal use of memory, calculation and storage resources is discussed. The serialization and parallelization of processes are analyzed. MOTS-CLÉS : indexation, document, performance, architecture.