Section: New Results

Natural language processing in multimedia data

Text detection in videos

Participants : Khaoula Elagouni, Pascale Sébillot.

Texts embedded in multimedia documents often provide high level semantic clues that can be used in several applications or services. We thus aim at designing efficient Optical Character Recognition (OCR) systems able to recognize these texts. During the last three years, we have proposed three novel approaches, robust to text variability (different fonts, colors, sizes, etc.) and acquisition conditions (complex background, non-uniform lighting, low resolution, etc.). The first approach relies on a segmentation step and computes nonlinear separations between characters well adapted to the local morphology of images. The two other ones, called segmentation-free approaches, avoid the segmentation step by integrating a multi-scale scanning scheme: The first one relies on a graph model, while the second one uses a particular connectionist recurrent model able to handle spatial constraints between characters. In 2013, a precise evaluation and comparison between these approaches was conducted and published in [16] .

Combining lexical cohesion and disruption for topic segmentation

Participants : Guillaume Gravier, Pascale Sébillot, Anca-Roxana Simon.

Topic segmentation classically relies on one of two criteria, either finding areas with coherent vocabulary use or detecting discontinuities. We proposed a segmentation criterion combining both lexical cohesion and disruption, enabling a trade-off between the two [58] . We provide the mathematical formulation of the criterion and an efficient graph based decoding algorithm for topic segmentation. Experimental results on standard textual data sets and on a more challenging corpus of automatically transcribed broadcast news shows demonstrate the benefit of such a combination. Gains were observed in all conditions, with segments of either regular or varying length and abrupt or smooth topic shifts. Long segments benefit more than short segments.However the algorithm has proven robust on automatic transcripts with short segments and limited vocabulary reoccurrences.

Information extraction and text mining

Participants : Vincent Claveau, Marie Béatrice Arnulphy.

Following the work initiated in the previous period, we have kept on working on relation extraction. During this year, we have proposed a new prototype that still relies on a supervised machine learning approach but we now rely on the sequence built from the shortest syntactic path between the entities, as it is done in many studies. These paths of lemmas are then used in a kNN whose similarity score is based on language modeling techniques. Based on this new prototype, we have participated to several tracks of the BioNLP challenges concerning the automatic extraction of relations in a specialized corpus. Results obtained with this simple and non-domain specific technique were relatively good, with a second and fourth ranks among the participants for the two tasks concerned [26] .

We also pursued previous work on supervised techniques for entity extraction and classification. Instead of working on complex machine learning approaches, we rather use simple methods but the focus is set on clever similarity computing between training examples and candidates for which we make the most of existing information retrieval techniques. Our approach has been evaluated through our participation to BioNLP-ST13 competition, where it has been ranked first [26] .

We have also proposed unsupervised techniques for knowledge discovery, more precisely, to bring out coherent groups of entities. Existing techniques are usually based on clustering; the challenge is then to define a notion of similarity between the relevant entities. In this work, we have proposed to divert conditional random fields (CRF) in order to calculate indirectly the similarities among text sequences. Our approach consists in generating artificial labeling problems on the data to be processed to reveal regularities in the labeling of the entities. The good results obtained shows the validity of our approach [27] and opens many research avenues for other knowledge discovery tasks.

Unsupervised approaches to fine-grained morphological analysis

Participants : Vincent Claveau, Ewa Kijak.

Following the work initiated in the previous years, we have kept on studying fine-grained morphological analysis for biomedical information retrieval. In the biomedical field, the key to access information is the use of specialized terms (like photochemotherapy). These complex morphological structures may prevent a user querying for gastrodynia to retrieve texts containing stomachalgia. The original unsupervised technique proposed in 2012 has been further developed and tested. In particular, during this year, we have shown that it largely outperforms state-of-the-art tools (e.g., Morfessor and Derif) for morphological segmentation tasks. It also offers indirect morpho-lexical resources that are more reliable than hand-coded ones used in most state-of-the-art tools [11] .

Tree-structured named entities recognition

Participants : Christian Raymond, Davy Weissenbacher.

Many natural language processing tasks needs the production of tree-structured outputs, like syntactic parsing, named entities recognition or language understanding. Currently, only machine learning based systems are robust enough to process the raw and noisy automatic transcribed speech while no machine learning paradigm are able to learn directly the tree structure in a reasonable time. In this work, we studied a solution to tackle the problem of predicting tree structured named entities from speech contents. We investigate a fast and robust decomposition strategy that was implemented and ranked best at the ETAPE NER evaluation campaign with results far better than those of the other participant systems [54] .

Fast machine learning algorithm for efficient combination of various features

Participant : Christian Raymond.

Currently, in the field of natural language processing the machine learning algorithm "boosting over decision stumps" is often designed as the best off-the-shell classifier. It's actually widely used for his abilities to work on relatively big dataset, to operate intrinsically feature selection and to produce very good decision rules. We investigated a slight modification of this algorithm where the decision stumps are replaced by bonsai trees. Bonsai trees are small decision trees (with low depth) that can capture some structure in the data that decision stumps can not. This modification allows the boosting algorithm to exhibits better (or in the worst case similar) performances with a lower number of iteration the original algorithm needs. Thus allows in some cases a big improvement in term of performance for a lower cost in term of learning time. An application on image processing (typed/hand classification) exhibited interesting results in  [94]