RALI has undertaken research in Information Retrieval (IR) for a number of years. The research covers various topics, such as the following ones:
- IR models
- Cross-Language IR and IR in different languages
- Web data mining for IR
- Past projects on IR applications
The current methods of IR or search engines often seem to be based solely on intuition or heuristics. Are there theories for IR? The answer is yes.
The undergoing research at RALI on IR models aims to develop methods capable of better satisfying user’s information needs. This can only be done through a better understanding of the user’s query, the document contents and connections between them, as well as the context of search. Different types of information and resources are used: the general linguistic knowledge, statistics on word and phrase usages in queries and documents, query logs, the search history, etc. Currently we are trying to implement this idea by extending approaches based on statistical language models.
How can one find documents written in Chinese with a query in English or French (when relevant documents exist only in Chinese)? The key is (query) translation. However, machine translation is neither the only possible, nor necessarily the best solution for this translation task. At RALI, we use statistical translation models which are constructed automatically from a large set of parallel or comparable texts. This method turns out to be of low cost (no manual intervention is needed) and of high quality (it can outperform the best MT system existing currently).
In parallel, we also develop specific methods for different languages such as French, Chinese and Arabic.
Web Data mining for IR
Much useful information exists on the Web for different purposes. Our research aims to develop automatic ways to collect it. The PTMiner system is developed to automatically identify parallel texts on the Web, in order to construct statistical translation models. This system has been able to mine a large number of parallel texts for English-French, English-Chinese, English-Italian, and English-German.
Web documents also contain many translations of technical terms and named entities, which are often not included in bilingual dictionaries, but can be automatically mined using patterns. In addition, we also work on approaches to mine relations between terms in the same language from large sets of documents, queries and query logs.
Search for medical documents has specific requirements, namely in evidence-based medicine: user’s information needs are usually specified on several aspects – Patient, Intervention, Comparison and Outcome (PICO). A medical intervention plays different roles when it is specified in Intervention or in Comparison. A finer matching requires us to automatically tag the role of medical concepts in documents and queries and to extend the traditional IR methods for such a structured query. We have constructed a test collection – CLIREC from PubMed and Cochrane reviews for PICO-based medical search, and this collection is publically available. This project is in collaboration with McGill University and University of British Columbia.
ERIC (Specialized Search Engine for the Construction sector)
ERIC is a research project in collaboration with the Faculté de l’aménagement and EBSI. It is supported by Bell-BUL and NSERC between 2001-2004. The goal is to identify precise answers to specialized questions asked by participants of the construction sector (architects, entrepreneurs, etc.). For example, one may ask what is the norm for foundation on frozen soil?
This project is supported by Nstein technologies and NSERC. It aims to automatically identify business opportunities (such as call for tenders) on the Internet for SMEs. Our research in this project covers the following aspects:
- automatically collecting calls for tenders on the Web
- matching calls for tenders and the need of an enterprise
- classifying calls for tenders
- searching calls for tenders in different languages
In this project supported by the CITÉ (Interdisciplinary Center on Emerging Technologies), we explore the possibility to find images from an image or textual query. We try to extract relationships between visual characteristics of a group of images (e.g. shape) and keywords (e.g. boat), and to use the relationships to bridge images with texts. The extraction of cross-media relationships is performed on a set of images annotated or accompanied by text. The experiments in ImageCLEF tests showed good performance and high potential of this method.