I believe that machine translation (MT) research provides a likely practical route towards eventually developing major components of truly intelligent large-scale AI systems. My concrete plans for the next few years include (1) further developing rapid-deployment MT technology and implementing systems for new languages, (2) investigating other applications of the Multi-Engine MT (MEMT) architecture, (3) continuing to work on Translingual Information Retrieval, and (4) continuing to play a major role in the LTI's educational activities, especially in advising students and teaching.
I believe that my position in the systems scientist track at the LTI provides the most likely route towards eventually developing major components of truly intelligent and useful large-scale AI systems. Since this will probably be several decades in the making, it is important to have interesting and useful intermediate goals to work towards in the meantime. Machine translation provides just the right sort of technically challenging, economically important domain for work on natural language processing, with subproblems ranging in difficulty from those that already have commercially viable solutions to those requiring essentially all of AI to be solved. Thus while intelligence involves much more than just language use, truly human-level language use will require most of what we think of as intelligence.
The central theme of my recent research in MT has been the Multi-Engine MT (MEMT) architecture: applying several MT engines using different technologies to the same text, and selecting the best combined results of the different engines as the system output; in many MEMT systems, the other alternatives are retained to provide the user with a menu of alternative translations. MEMT began within the Pangloss project (see below) primarily as a means for achieving ``assimilation'' MT: translating documents coming into an organization from outside, and thus lacking the domain and language restrictions currently necessary for fully-automatic high-quality translation.
Near the end of the Pangloss project, we began thinking about means for achieving ``rapid-deployment MT'', by which we mean producing initial versions of MT systems for new languages at usable levels of quality in weeks or months, rather than years. We realized that MEMT provided a means for achieving this, by allowing us to combine simple, very-rapidly-deployable MT engines with better but slower-to-deploy MT engines, producing graceful improvement over time; this led to the Diplomat project. About the same time, the SRI Cambridge group in England adopted MEMT for their speech translation system, to allow ``anytime'' translation: in a real-time MT system, the results of quicker engines are used if the system must give an answer before better-but-slower engines have returned results. Within the LTI, MEMT has been used similarly in prototype real-time closed-caption translation systems. At some point during these developments, I recognized a theoretical basis for thinking about MEMT: MEMT attempts to exploit the differences between different MT technologies, to combine their strengths and ameliorate their weaknesses. As described above, this can be used to permit non-domain-restricted MT, to facilitate rapid-deployment MT, or to ensure real-time MT.
It has been gratifying to see the MEMT approach achieve wide-spread use in MT for assimilation and speech; in addition to the SRI Cambridge speech and LTI closed-captioning systems, it is being used in LTI's NICE project for the Organization of American States (OAS), and in the LTI's new Nespole and LingWear projects. Diplomat has also been used by the Informedia project for experiments in Translingual IR. Besides these practical successes, we carried out an evaluation in 1998 that demonstrated that the output set of translations from the MEMT Croatian/English system in Diplomat (see below) were significantly better than any one of the component engines. This was presented at AMTA-98, the main MT conference (Hogan and Frederking 1998).
In addition to exploring these particular directions further, I am interested in seeing whether this concept can be extended in other directions. My involvement in our TLIR work (see below) evolved as an outgrowth of my assimilation MT work. Similarly, my assimilation MT work has led to my involvement in several nascent projects related to translating television closed-captions, as mentioned in passing above.
Pangloss: As mentioned above, the Pangloss project (co-directed at CMU with Jaime Carbonell and Lori Levin, and joint with New Mexico State University and the University of Southern California) developed the first Multi-Engine MT (MEMT) system, for human-assisted Spanish-to-English translation of newspaper articles. The Pangloss MEMT system combined the results of Example-Based MT, Transfer-Based MT, and Knowledge-Based MT (KBMT). I co-invented the MEMT architecture (with Sergei Nirenburg) during this project. My main role in the project (as ``System Integrator'') was to make sure that the various parts of the system, developed at different sites, worked together. This involved high-level work participating in designs to make sure they would be ``integratable'' when finished, and lower-level work, programming and supervising programmers to achieve actual integration of the software.
Diplomat: The Diplomat project (co-directed with Jaime Carbonell and Alex Rudnicky) had as its goal to develop and demonstrate techniques for ``rapid-deployment'' speech-to-speech MT. This filled a government need arising from the rapidly changing worldwide political and economic situations we now face. We worked with Croatian, Korean, Haitian Creole, and Spanish, and carried out initial investigations of Arabic. The MT portion of the system was based on the MEMT architecture developed under Pangloss. In addition to building demonstration systems, we gave some careful thought to the remaining underlying research issues, beginning (on the MT side) with the (semi-)automatic learning of morphological rules from examples; in this respect, I feel that I helped define the topic of Christopher Hogan's LTI PhD thesis topic. In addition to rapid deployment, Diplomat was a demonstration vehicle for wearable computing; so we worked closely both with CMU's wearable computers group and a commercial wearable computer company, Teltronics. Diplomat was initially funded after a crash demonstration for Croatian that I spear-headed.
TransLingual IR (TLIR): In a somewhat different vein, I played a significant role (along with Jaime Carbonell, Yiming Yang, and Ralf Brown) in the LTI's initial work in the area of information retrieval across language boundaries; e.g., using a query in English to retrieve relevant documents in other languages. Our comparative study of different MT-based and IR-based methods resulted in a Distinguished Paper award at IJCAI-97, and several speaking invitations. My resulting trip to the Systems Engineering Research Institue (SERI) in South Korea raised the possibility of some joint work in this area, with SERI initially agreeing to become LTI affiliates, until an ill-timed crash in the South Korean currency prohibited foreign research contracts.
The Tongues project is the most direct result of the Diplomat project. The main disappointment of Diplomat was our inability to convince anyone in the DoD to actually test it in the field, although we gave many well-received demos, and even sent a (text-only) version of our system to Haiti under Army Research Lab (ARL) auspices. Tongues is a Lockheed Martin-led effort, under US Army funding, to build a prototype for a real field-deployed system based on Diplomat technology. Its initial users will be Army chaplains, who must carry out civil affairs and enemy POW (EPW) interactions as part of their duties. I am leading the LTI's development of the MT and speech synthesis components, under contract to Lockheed Martin, with a major demo for the Army in December 2000. A successful demo could lead to fast-track deployment by the US Army.
KDI-UA/Tracat: I have also played a small role in the NSF-funded KDI Universal Access project, primarily in co-supervising (with Eric Nyberg) the development of a prototype object-oriented GUI for integrated Translingual IR, gisting-MT, and summarization, implemented by several students as Independent Studies. This work was also funded by my part of the DARPA TIDES-funded Tracat project extension (led by Bill Scherlis), which was to be bridge funding to a TIDES project which went unfunded in the end.
Nespole/LingWear: I believe that my part in these two projects resulted from the successful demonstrations of the Diplomat project. My role in Nespole (a joint NSF/European Union project) is to test the MEMT architecture in the German/English component of a multi-lingual speech-to-speech translation system, in the domain of travel agency conversations. As for LingWear, it was funded under DARPA's TIDES program, one of whose goals is to further explore the rapid-deployment style of MT pioneered by Diplomat. LingWear will be translating Mandarin text into English, and bidirectionally translating Korean/English speech-to-speech.
MuchMore: I am currently a co-PI on MuchMore, a joint NSF/European Union project that is just beginning (in June 2000). In this project, we will be using MeSH, an internationally-standardized taxonomy of medical categories, as a bridge between queries in one language and medical document archives in a different language. We will be working primarily with German and English (as well as French and Chinese). We will use the taxonomy by statistically classifying each document and query into multiple MeSH categories, and then using the match between the categories of the query and the documents to rank the documents for retrieval.
I developed significant parts of two central LTI courses, Algorithms for Natural Language Processing (11-711) and Machine Translation (11-731), five lectures each, and teach these each year. Teaching more often than I had in the past has also been a worthwhile experience, although the current level is probably the most I can support as research faculty. As mentioned in my CV, I served on Stephen Beale's LTI PhD thesis committee, and am currently serving on Christopher Hogan's LTI PhD thesis committee, after helping to initially define his thesis topic. I also feel that I have contributed in some measure to Chris's development into a top-notch researcher, as well as providing an important educational experience to the several MS and undergraduate students who have worked on my projects (also those of my staff programmers who have gone on to graduate school). I look forward to continuing all of these educational activities.