Romaric Besançon

Generative Event Schema Induction with Entity Disambiguation

This paper presents a generative model to event schema induction. Previous methods in the literature only use head words to represent entities. However, elements other than head words contain useful information. For instance, an armed man is more discriminative than man. Our model takes into account this information and precisely represents it using probabilistic topic distributions. We illustrate that such information plays an important role in parameter estimation. Mostly, it makes topic distributions more coherent and more discriminative. Experimental results on benchmark dataset empirically confirm this enhancement.

Filtering and clustering relations for unsupervised information extraction in open domain

Information Extraction has recently been extended to new areas by loosening the constraints on the strict definition of the extracted information and allowing to design more open information extraction systems. In this new domain of unsupervised information extraction, we focus on the task of extracting and characterizing a priori unknown relations between a given set of entity types. One of the challenges of this task is to deal with the large amount of candidate relations when extracting them from a large corpus. We propose in this paper an approach for the filtering of such candidate relations based on heuristics and machine learning models. More precisely, we show that the best model for achieving this task is a Conditional Random Field model according to evaluations performed on a manually annotated corpus of about one thousand relations. We also tackle the problem of identifying semantically similar relations by clustering large sets of them. Such clustering is achieved by combining a classical clustering algorithm and a method for the efficient identification of highly similar relation pairs. Finally, we evaluate the impact of our filtering of relations on this semantic clustering with both internal measures and external measures. Results show that the filtering procedure doubles the recall of the clustering while keeping the same precision.

Cross-media feedback strategies: merging text and image information to improve image retrieval

The CEA-LIST/LIC2M develops both cross-language information retrieval systems and content-based image retrieval systems. The ad hoc and medical tasks of the ImageCLEF campaign offered us the opportunity to perform some experiments on merging the results of the two systems. The results obtained show that the performance of each system highly depends on the corpus and the task: feedback strategies can improve the results, but the parameters used are to be tuned according to the confidence of each system on the task and corpus: for the ad hoc task, text retrieval performs good whereas results of image retrieval are poor. On the other hand, for the medical task, the image retrieval performs better, and text retrieval can improve overall results only with reinforcement strategies.

Concept-based searching and merging for Multilingual information retrieval: First experiments at CLEF 2003

This article presents the LIC2M’s crosslingual retrieval system which participated in the Small Multilingual Track of CLEF 2003. This system is based on a deep linguistic analysis of documents and queries that aims at categorizing them in terms of concepts and implements an original search algorithm inherited from the SPIRIT (EMIR) system that takes into account this categorization.

Using temporal cues for segmenting texts into events

One of the early application of Information Extraction, motivated by the needs for intelligence tools, is the detection of events in news articles. But this detection may be difficult when news articles mention several occurrences of events of the same kind, which is often done for comparison purposes. We propose in this article new approaches to segment the text of news articles in units relative to only one event, in order to help the identification of relevant information associated with the main event of the news. We present two approaches that use statistical machine learning models (HMM and CRF) exploiting temporal information extracted from the texts as a basis for this segmentation. The evaluation of these approaches in the domain of seismic events show that with a robust and generic approach, we can achieve results at least as good as results obtained with a specialized heuristic approach.

Overview of CLEF 2008 INFILE pilot track

The INFILE campaign was run for the first time as a pilot track in CLEF 2008. Its purpose was the evaluation of cross-language adaptive filtering systems. It used a corpus of 300,000 newswires from Agence France Presse (AFP) in three languages: Arabic, English and French, and a set of 50 topics in general and specific domain (scientific and technological information). Due to delays in the organization of the task, the campaign only had 3 submissions (from one participant) which are presented in this article.

Information filtering evaluation: overview of CLEF 2009 INFILE track

The INFILE@CLEF 2009 is the second edition of a track on the evaluation of cross-language adaptive filtering systems. It uses the same corpus as the 2008 track, composed of 300,000 newswires from Agence France Presse (AFP) in three languages: Arabic, English and French, and a set of 50 topics in general and specific domains (scientific and technological information). In 2009, we proposed two tasks : a batch filtering task and an interactive task to test adaptive methods. Results for the two tasks are presented in this paper.

Data fusion of retrieval results from different media: experiments at ImageCLEF 2005

The CEA-LIST/LIC2M develops both multilingual text retrieval systems and content-based image indexing and retrieval systems. These systems are developed independently. The merging of the results of the two systems is one of the important research interests in our lab. We tested several simple merging techniques in the ImageCLEF 2005 campaign. The analysis of our results show that improved performance can be obtained by appropriately merging the two media. However, an a-priori tuning of the merging parameters is difficult because the performance of each system highly depends on the corpus and queries.

General learning approach for event extraction: Case of management change event

Starting from an ontology of a targeted financial domain corresponding to transaction, performance and management change news, relevant segments of text containing at least a domain keyword are extracted. The linguistic pattern of each segment is automatically generated to serve initially as a learning model. Each pattern is composed of named entities, keywords and articulation words. Some generic named entities like organizations, persons, locations, dates and grammatical annotations are generated by an automatic tool. During the learning step, each relevant segment is manually annotated with respect to the targeted entities (roles) structuring an event of the ontology. Information extraction is processed by associating a role with a specific entity. By alignment of generic entities to specific entities, some strings of a text are automatically annotated. An original learning approach is presented. Experiments with the management change event showed how recognition rates are improved by using different generalization tools.

Combining Word and Entity Embeddings for Entity Linking

The correct identification of the link between an entity mention in a text and a known entity in a large knowledge base is important in information retrieval or information extraction. The general approach for this task is to generate, for a given mention, a set of candidate entities from the base and, in a second step, determine which is the best one. This paper proposes a novel method for the second step which is based on the joint learning of embeddings for the words in the text and the entities in the knowledge base. By learning these embeddings in the same space we arrive at a more conceptually grounded model that can be used for candidate selection based on the surrounding context. The relative improvement of this approach is experimentally validated on a recent benchmark corpus from the TAC-EDL 2015 evaluation campaign.