James K. Baker
Distinguished Career Professor
Center for Innovations in Speech and Language
Language Technologies Institute
School of Computer Science
Carnegie Mellon University
jkbaker at cs dot cmu dot edu
I was the founder, CEO and Chairman of Dragon Systems, Inc. When Dragon introduced Dragon NaturallySpeaking, the first general purpose automatic dictation system in 1997, it was the culmination of a 25-year quest. I am now engaged in a second quest, seeking to expand the range and scope of speech and language technology to many more applications and languages. This expansion of scope will also require fundamental breakthroughs in technology and the acquisition and creation of a large collection of language data and knowledge. It will require the collaborative effort of hundreds or thousands of participants.
Grand Challenge in Speech and Language Technology
- Universal Communicator: The Universal Communicator will aid communication across barriers of language, geography and culture. It should work for all people regardless of educational background and computer skills. It should require no user training and should work for both written and spoken communication.
- ALLOW (All Living Languages Of the World): The technology base for the Universal Communicator, the ALLOW program aims at developing speech recognition, speech synthesis, machine-facilitated translation and optical character recognition for all languages of the world (about 7,000 languages).
- Saving Endangered Languages: Over half the languages of the world are endangered and likely to become extinct. The tools being developed for the ALLOW program will enable indigenous populations to better utilize computer technology in their own language. With blogs, podcasts, social networks and their own Wikipedia, these communities can document and preserve their language and culture and educate their younger generations in their heritage.
Game-Changing Paradigm Shift in Automatic Learning
- Socratic Paradigm: The Socratic paradigm is based on core principles of Socrates' philosophy illustrated by these quotations:
- "The only thing I know is that I don't know anything" (Which according to the Oracle of Delphi made Socrates the wisest of the Greeks.)
- "An unexamined life is not worth living" (In choosing a sentence of death rather than exile.)
- That is, question everything, especially the assumptions in our models and in our learning methods. Know what we know and know what we don't know.
- Socratic Validation: A method for using unlabeled data for development testing. How can you test performance if you don't know the correct answer? The more strongly you believe that this is impossible, the more you should be interested in this technology. (Under the Socratic paradigm, you ask a different question.)
- Learning from Ignorance: A method for transfer learning within a single system by deliberately designing a version of the system with limited capabilities. This technique, also called self-reference transfer learning, also allows any conventional supervised learning algorithm to be used with unlabeled data.
- Learning in the Field: A widely deployed system acquires an hour of data for every hour each copy of the system is in use. This is mostly unlabeled data, but with the new techniques in the Socratic paradigm, the quantity of this data changes the game. Thousands of semi-autonomous Socratic agents can be running in the background, testing every assumption and every element of the system. With an unlimited amount of fresh data, there is no danger of tuning the system to a fixed set of training data or a fixed set of development test data. Fortunately, there is also no need for this enormous quantity of data to be labeled manually.