Section: New Results

Speech-to-Speech Translation and Langage Modeling

Participants : Kamel Smaïli, David Langlois, Sylvain Raybaud, Motaz Saad, Denis Jouvet, Cyrine Nasri.

machine translation, statistical models

Sylvain Raybaud has just defended his thesis untitled “De l'utilisation de mesures de confiance en traduction automatique : évaluation, post-édition et application à la traduction de la parole.”. His contributions are the following: study and evaluation of confidence measures for Machine Translation, an original algorithm to automatically build an artificial corpus with errors for training the confidence measures, development of an entire speech-to-text translation system.

In the scope of Confidence Measures, we participated to the World Machine Translation evaluation campaign (WMT2012 http://www.statmt.org/wmt12/quality-estimation-task.html ). More precisely, we proposed a Quality Estimation system to the Quality Estimation shared task. The goal was to predict the quality of translations generated by an automatic system. Each translated sentence is given a score between 1 and 5. The score is obtained using several numerical or boolean features calculated according to the source and target sentences. We perform a linear regression of the feature space against scores in the range [1:5]. To this end, we use a Support Vector Machine. We experiment with two kernels: linear and radial basis function. In our system we use the features from the shared task baseline system and our own features (based on the work from the Sylvain Raybaud's thesis). This leads to 66 features. To deal with this large number of features, we propose an in-house feature selection algorithm. Our system came 5th among 19 systems. This work was publish in [24] . In the continuation of this research, we contributed to the development of a Quality Estimation tool (quest: https://github.com/lspecia/quest ). For that, David Langlois was invited by Lucia Specia at University of Sheffield, Computer Sciences department, Natural Language Processing group. We added our own features into quest. This tool is dedicated to be available for the research community.

Another objective of our research work, with the Cyrine Nasri's Phd thesis, is to retrieve bilingual phrases for machine translation. As in fact, current statistical machine translation systems usually build an initial word-to-word alignment before learning phrase translation pairs. This operation needs many matching between different single words of both considered languages. We propose a new approach for phrase-based machine translation which does not need any word alignments. It is based on inter-lingual triggers determined by Multivariate Mutual Information. This algorithm segments sentences into phrases and finds their alignments simultaneously. Inspite of the youth of this method, experiments showed that the results are competitive but needs some more efforts in order to overcome the one of state-of-the-art methods.

Another aspect of the research of the group is to work on under resourced language related to Arabic. In fact, in several countries through the Arabic world, only few people speak the modern standard Arabic language. People speak something which is inspired from Arabic but could be very different from the modern standard Arabic. This one is reserved for the official broadcast news, official discourses and so on. The study of dialect is more difficult than any other natural language because it should be noted that this language is not written. A preliminary work has been done knowing that our final objective is to propose a machine translation between the different Arabic dialects and modern standrad Arabic. This issue is very difficult and challenging because no corpus does exist, vernaculars are different even within the same country, etc.

Last, Motaz Saad has started his thesis in November 2011. His objective is to work on opinion analysis in multilingual documents from internet. During this year, he retrieved comparable corpus from the web, and proposed a method to align these corpora at document level. He proposed algorithms to measure the degree of comparability between documents. He submitted his work to the International Conference on Corpus Linguistics (CICL2013).

In the framework of the ETAPE evaluation campaign a new machine learning based process was developed to select the most relevant lexicon to be used for the transcription of the speech data (radio and TV shows). The approach relies on a neural network trained to distinguish between words that are relevant for the task and those that are not. After training, the neural network (NN) is applied to each possible word (extracted from a very large text corpus). Then the words that have the largest NN output score are selected for creating the speech recognition lexicon. Such an approach can handle counts of occurences of the words in various data subsets, as well as other complementary informations, and thus offer more perspectives than the traditional unigram-based selection procedures.