Early trials of this open-source, distributed system, developed under the IST project PENG, have only recently completed. But coordinator Gabriella Pasi says the participants were impressed with the results.
"Selected Swiss journalists and students assessed the performance of the system's various modules," she says. "For example they checked the effectiveness and accuracy of the information filtering, comparing the results with those from existing systems. They then looked at the integrated system and praised its user-friendliness."
Pasi adds that they liked the system's ability to find relevant information â measured in terms of recall (the proportion of retrieved and relevant documents compared to all documents in the collection) and precision (the ratio of retrieved and relevant documents to all the documents retrieved). A more detailed system trial is due for completion in November 2006.
More than just 'push'
The project originated in research carried out by several partners on information retrieval and filtering. Pasi notes that, "Our project proposal predated the launch of present news-aggregation services, which focus on 'pushing' out information based on user needs." The PENG system, by contrast, offers two distinct techniques: information filtering (push) and information retrieval (pull).
Current news-aggregation systems work very much like internet search engines, pushing out information based on certain user criteria. If users require further filtering, they must create a profile for themselves â which can result in the generation of somewhat limited lists. This process works well for journalists receiving information from online news agencies that produce a continuous news stream; but they still face the problem of selecting the most relevant news.
The PENG system enables users to go much further. By personalising filters, they can pick up targeted information from agencies and combine this with data retrieved from the web or specialised archives. They can also place constraints on the content they seek â such as the media category or trustworthiness of sources â to generate highly specific information. The system then calls on various modules to edit and summarise all this information automatically, before presenting it as the user wishes.
Innovative fuzzy algorithm
Pasi highlights the system's ability to learn user preferences over time. It can also deal with human vagueness or imprecision, such as in the filtering or interaction with the software.
The partners have also developed a new filtering algorithm. Based on categories, it can cluster news from agencies into thematic cluster groups such as sports or politics, for creating data subsets based on common characteristics (e.g. people with a certain hair colour). After these subsets are defined, the system can describe each group (e.g. this is the group with black hair).
"Of the two possible approaches to data clustering," says Pasi, "we chose 'unsupervised' because this approach does not force us to select a priori categories." She adds that the PENG system can display audiovisual content, but works mainly with textual information.
PENG was completed in August 2006. Though the complete system exists only as a prototype, project partner ATOS Origin is examining the possibility of using certain modules in standalone applications. The company is also interested in marketing the project's clustering algorithm, which could be used not only for filtering news but also for image gathering or e-commerce applications.
Professor Gabriella Pasi
Consiglio Nazionale Delle Ricerche ITC-CNR
Via Bassini N. 15
Tel: +39 02 2369 9489
Source: IST Results Portal