Research

  • Checking and validation of WATT translations
    Develop a verification and validation system to determine the effort required by professional translators to finalize a translation of an automated "first estimate". The system will diagnose the effectiveness of the translation software used to produce a warning bulletin and will apply to the RALI WATT system, the current system used by the Translation Bureau and other automatic translation tools such as Google Translate. The prototype will show the source bulletin, its "first estimate" automatic translation and the portion that it thinks it requires special attention. He will also consider identifying potential problems in the source text.

  • WATT adaptation for alerts sent on social networks
    Adapt WATT for translating weather reports broadcasted via social media (ie Twitter). This new approach will not necessarily focus on the literal translation of a message, but rather on its meaning, and, if necessary, may also use two tweets to ensure translation. There are some issues related to translating formatted social media messages, such as their length, the presence of idiosyncratic forms (such as ubiquitous hashtags), and content. RALI has conducted experiments to translate government text feeds on Twitter and will adapt this work to EC messages. 
     
  • WATT improvement
    Continue improvements to RALI's WATT machine translation technology. The current WATT system is based on a static mixture of translation memory and a motor based on statistical sentences powered by manual lexicons (geographical names, etc.). Having a diagnostic tool would be helpful to improve the current WATT system. Adapting to social media translations would provide another way to adapt the capabilities of the WATT system. Analyzes and updates of the WATT system from the above objectives will be the subject of the third objective over time.
     
  • Selective and customized environmental information display
    Mohamed Mouine, in his thesis Ph.D., was interested in the problem of generating a personalized visualization meeting the needs and preferences of a user. He developed MeteoVis, a weather report generator. Since there is little information on the user's profile, he relied on similar users to calculate his needs and preferences. Unsupervised learning was used to group similar users together by calculating a rate. similarity of user profiles in the same cluster to weight needs and preferences. With the help of independent users, evaluation and comparison experiments of our tool were conducted to compare them with the one currently used by Environment Canada. The results of this evaluation show that the visualizations generated by MeteoVis are better than the current bulletins prepared by EC.
     
  • Machine translation of weather forecasts

    There are two issues regarding Machine Translation in this project: translating weather alerts and providing weather forecasts in Inuktitut.

    In order to translate weather bulletins and alerts from English to French and vice-versa, we developed a phrase-based, statistical machine translation. The latter relies heavily on example translations in order to train efficient both language and translation statistical models. We therefore concentrated our efforts on the creation of a training corpus, built from various sources of human-produced, bilingual weather alerts and reports. We successfully designed a system, which receives raw weather updates from Environment Canada (EC) then extracts all relevant sentences and aligns them with each other when they are a translation of one another. 

    The second issue is far more challenging since we lack training material dedicated to translating weather material from, say English to Inuktitut. Furthermore, Inuktitut is known as being an agglutinative language, which means that the notion of word on which most statistical translation systems are built is a delicate one to handle. First, we addressed the problem of learning automatically in an unsupervised manner the morphological structure of a language. Discovering the morphological structure of a language is precisely the purpose of the Morpho Challenge workshop, which is part of the EU Network of Excellence PASCAL Challenge Program. Therefore, we decided to participate for the first time in this event. Jean-François Lavallée (a master student at RALI) devised a new system based on analogical learning. Its performance for Finnish, English, German and Turkish ranged among the top systems submitted at the shared task. Indeed, the system he developed ranked first among those who participated for the first time to this task, and very close from the best system which is a combination of existing current state-of-the-art systems.
     

  • Data‐mining of relevant information patterns
    We worked on setting up the base ingredient necessary prior to applying any data-mining algorithm: the data collection. We first acquired a detailed understanding of all the information contained in the Meteocode XML data files made available to us. These contain short and medium term predictions for many environmental variables (temperature, precipitation, atmospheric pressure, wind speed and direction, etc...) on a grid of geographical locations across Canada, and are updated every 6 to 12 hours. We set up a relational database and developed software that regularly fetches these Meteocode files, parses them, and populates the database with the extracted information. This system will thus collect the historical prediction data that will be needed to train statistical models.

    To allow exploration of this wealth of data, we set up an experimental internal web page. Through it we can now easily display the evolution of any of the predicted environmental variables across time at geographical locations of our choosing, both in table and graphical format. This is but a starting point for the future exploration of ways of presenting relevant meteorological information to specialized audiences, with a level of details that is not currently available on Environment Canada's site for the general public. More importantly, with this data infrastructure in place, it will now be straightforward to pull datasets suitable for researching and training novel statistical learning algorithms as planned.