PortDial Highlights

In PortDial industry and academia collaborated seamlessly towards transferring state-of-the-art grammar enhancement and localization technologies in a commercial system. Working with real data and real services was essential in guiding our solution that has been fully integrated in the PortDial commercial platform.
  1. Human-in-the-loop paradigm: Instead of following a fully automated approach for grammar induction (as it is typical in the research lab) we have opted for a human-in-the-loop approach were the grammar developer iteratively post-edits and corrects the grammar fragments suggested by the system. We have shown that by using user/developer feedback in the grammar induction process performance (precision of induced rules) increases threefold, adding significant value to the grammar development cycle.
  2. Knowledge-based (top-down) approach using lexicalized ontologies is a viable approach for grammar induction. Good performance is achieved when combined with the bottom-up (data-driven) approach.
  3. Web data harvesting is an good alternative to using service data. Excellent grammar induction performance is achieved when appropriate web querying and filtering algorithms are used. The method is appropriate also for language porting (localization), where a seed corpus is translated with an off-the-shelf  system and used to harvest web data for grammar induction.
Overall, the PortDial platform achieved 60-70% accuracy in (non-terminal) rule induction and reduced the grammar development/tuning cycle by approximately 60% exceeding expectations at project start.
Comments