Communications

Inferring new knowledge from data requires interpreting these data using complex mathematical models taking into account contextual information and background knowledge.  The goal of research in knowledge processing is thus to find the best possible models for performing a given "intelligent" task.  The quality of a model can be evaluated by measuring how it agrees with representative sample data sets for the task under study. Data is also important to train the models through automatic learning procedures. For the sake of reproducibility of experiments, the data used to test a model should be made publicly and easily available.  Furthermore, to avoid any bias in comparing different models despite their learning ability, they should be tested before any experimental result is released, implying that all tests occur almost simultaneously. This leads to a specific organization, often called an evaluation campaign, where several research teams harmonize their needs and coordinate their efforts around a synchronized common test. Producing the needed data and organizing evaluation campaigns thus constitute critical resources for the success of the associated researches and must be granted much attention.  Since they are not pure research activities, they should be carried out by dedicated teams, working at the service of the research community at large with a strong public support.  Such resources are becoming a strategic component of any ambitious research effort in the field, and offer a natural means to coordinate work at the international level.

This talk will highlight some open challenges in automatic processing of spoken language.  Automatic speech processing has witnessed substantial advances over the last decade with growing interest in transcription systems for automatic structuring of audio supporting a variety of applications such as information archival and retrieval, media monitoring, automatic subtitling, question-answering, speech translation and speech analytics. Much of the information on the web is not in a textual format, and therefore will escape detection and categorization via text-based methods. Today's systems rely on large quantities of audio and textual data for model training, which often entail high development costs. Therefore research is needed on developing generic recognition models and automatic learning from unannotated data.  Important future research will address keeping language models up-to-date, automatic topic identification, and enriched transcriptions providing annotations for speaker turns, language, acoustic conditions, etc.  While the performance of speech recognition technology has improved dramatically for a number of 'dominant' languages, technologies for language and speech processing are available only for a small proportion of the world's languages. A major challenge is to have a wider coverage of languages, allowing all citizens to interact in their own native language. With more and more connectivity and mobility, a related challenge is dealing with language (code) switching.

Unsupervised, fully automated understanding of web pages designed for humans seemed an insurmountable challenge just a few years ago. In this talk, we will summarize resent advances in data extraction based on the combination of noisy, but large-scale entity recognition with high-level ontology constraints that have brought unsupervised data extraction at high accuracy within our grasp. In the DIADEM ERC project, this trend is being used to develop the first integrated system and methodology to design extraction ontologies that enable fully automated, high accuracy data extraction in a given domain. The talk is concluded by an outlook into how the increasing amounts of available structured knowledge can contribute in the future to bootstrap the extraction of further knowledge from unstructured sources.