| Abstract | Statistical models try to generalize knowledge starting from the frequency of probabilistic events in the data. If more data is available, events are more often observed and models are more efficient. Natural Language Processing approaches based on those models are therefore dependant on the quantity and availability of these resources. Thus, there is a permanent need for generating and updating the learning data.
This dependency touches Statistical Machine Translation, which requires multilingual resources. This thesis refers to four articles tackling two tasks that contribute significantly to this dependency: the Bilingual Documents Alignment (BDA) and the Bilingual Lexicons Induction (BLI).
The first publication describes the system submitted for the BDA shared task of the WMT16 conference. Developed on a search engine, our system indexes bilingual web sites and tries to identify the English-French pages linked by translation. The alignment is realized using a "bag of words" representation and a bilingual lexicon. The tool we have developed allowed us to evaluate more than 1,000 configurations and identify one yielding decent performances on this particular task.
The three other articles are concerned with the BLI task. The first one focuses on the so-called standard approach, and proposes a breadth parameter exploration in the Semantic Web context. The second article compares the standard approach with more recent techniques based on interlingual representation of words, or the so-called embeddings, issued from neural networks. The last contribution reports the enhanced global performances on the task, combining the outputs of the two studied approaches through supervised reclassification.
|
