Section: New Results

Machine translation and language modeling

Participants : Kamel Smaïli, David Langlois, Denis Jouvet, Emmanuel Vincent, Motaz Saad, Cyrine Nasri.

machine translation, statistical models

Language modeling

Vocabulary selection

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 (text tokens 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 occurrences of the words in various data subsets, as well as other complementary information, and thus offer more perspectives than the traditional unigram-based selection procedures [50] .

Music language modeling

Similarly to speech, music involves several levels of information, from the acoustic signal up to cognitive quantities such as composer style or key, through mid-level quantities such as a musical score or a sequence of chords. The dependencies between mid-level and lower- or higher-level information can be represented through acoustic models and language models, respectively. We pursued our pioneering work on music language modeling, with a particular focus on log-linear interpolation of multiple conditional distributions. We applied it to the joint modeling of “horizontal” (sequential) and “vertical” (simultaneous) dependencies between notes for polyphonic pitch estimation [26] and to the joint modeling of melody, key and chords for automatic melody harmonization [25] . We also proposed a new Bayesian n-gram topic modeling and estimation technique, which we applied to genre-dependent modeling of chord sequences and to music genre classification [74] .

Quality estimation of machine translation

In the scope of Confidence Measures, we participated to the World Machine Translation evaluation campaign for the second year (WMT2013 http://www.statmt.org/wmt13/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 0 and 1. The score is obtained by using several numerical or boolean features calculated according to the source and target sentences. We performed a linear regression of the feature space against scores in the range [0 ;1], to this end, we use a Support Vector Machine with 66 features. In this new participation, we proposed to increase the size of the training corpus. For that, we decided to use the post-edited and reference corpora in the training step after assigning a score to each sentence of these corpora. Then, we tune these scores on a development corpus. This leads to an improvement of 10.5% on the development corpus, in terms of Mean Average Error (average difference between reference and predicted scores), but achieves only a slight improvement on the test corpus. This work has been published in [51] .

Comparable corpora and multilingual sentiment analysis

In the PhD Thesis of Motaz Saad, we work on collecting comparable corpora. For that purpose we presented a method which extracts and aligns comparable corpora at the article level from Wikipedia encyclopedia based on interlanguage links. To evaluate the closeness of corpora we proposed several comparability measures. Our evaluations show that the proposed comparability measures are able to capture the comparability degree of any comparable corpora [60] . We go further on the comparability of multilingual corpora by studying their comparability in terms of sentiment. The final objective is to propose a multilingual press review concerning a given topic. This review should use several multilingual resources (electronic newspapers), and should class resources according to the including sentiments (fear, joy...about the subject), polarity (against or not to the subject)...This conducts to study opinions across different languages by comparing the underlying messages written by different people having different opinions. We propose "Sentiment based Comparability Measures" to compare opinions in multilingual comparable articles without translating source/target into the same language [27] .

Machine translation of arabic dialect

The translation of Arabic dialect constitutes a real challenge since it is an under-resourced language. In fact, Modern Standard Arabic is as any other evoluated language, it means it could be processed by the available tools but unfortunately in Arabic countries people speak an Arabic language which is inspired from the standard one but is different. Our objective is then to propose a speech to speech system converting modern standard Arabic to Algerian dialect. After collecting corpus, we decided to propose a method allowing to diacritize dialects in order to be able in the following to develop an acoustic model. For that, we considered the issue of diacritization as a machine translation issue, and we have developed a statistical machine translation which learns to transform an undiacritized corpus into a diacritized one [44] .