| Advanced Research |
Fluential's R&D team conducts the latest cutting-edge research in the field of computational linguistics and integrates new technologies into its products. The fundamental idea behind the research is to enable us to develop accurate speech-to-speech translation systems rapidly and with minimal human effort. Our research focus is concentrated on a number of fields of natural language processing, such as information retrieval, speech recognition, and machine translation.
1. Unified Recognition/Translation SystemFluential is developing a state-of-the-art system, which integrates the process of speech recognition with the process of machine translation. These two processes are generally separated, resulting in sub-optimal performance when the two are combined in series. By integrating the entire recognition lattice into our state-of-the-art machine translation system, the unified system is able to achieve superior translation, which is both quick and robust to errors in speech recognition that are common in noisy environments.
2. Information Structuring and RetrievalFluential is developing methods for information structuring and retrieval, which aids in the rapid design and development of speech translation systems. We research methods to organize unstructured textual corpora of arbitrary size into a groups and hierarchies of sub-groups. Using learning algorithms, such as Kernel methods, we are able to extract task-specific information using rich feature sets and classify new text into set hierarchy. These methods enable recognition systems to be both scalable and robust to speech recognition errors.
3. Mining the WebWe research algorithms to mine the vast resources of the web for textual data that is relevant to various speech translation tasks. We have developed algorithms that are able to iteratively acquire similar textual information with minimal seed training data. This provides further training corpora for both speech recognition and machine translation models.
4. Synergy of Human and Statistical EffortsIn our goal to provide the highest quality translation using a limited amount of speech data, we continue to develop tools that allow people with only a facility in the second language to design recognition and translation grammars semi-automatically. Our algorithms are able to integrate human knowledge where the statistical system is not able to learn, given the limited resources available to it. |
|
 |
|
