Abstracts - 2006
Spoken Language Translation in the Flight Domain
C. Wang & S. Seneff
For the past several years, we have been developing computer systems that can help a student of a foreign language acquire proficiency by allowing them to engage in spoken converrsation with the computer in the foreign language. In order for such a system to be effective, it must also be able to provide assistance at any time by acting as an always-present tutor. The main task of the software tutor is to provide translation assistance form the native language into the second language. In prior research, we have developed a system configuration for the weather domain, which can translate from English into Chinese both the user's queries and the system's responses . Since we already have a sophisticated mixed-initiative dialogue system for the flight domain that can communicate in English, we were motivated to convert this system into a conversational domain for language learning as well. The level of complexity of queries in the flight domain is substantially higher than that of queries typical of the weather domain, so translation of these queries, which must be of extremely high quality in order not to mislead the student, is a very challenging task.
In order to assure high quality of the translation, we have adopted a linguistically motivated approach which involves parsing the input string into a hierarchical meaning representation, which we call a "semantic frame," and then using formal generation rules to paraphrase the semantic frame into Chinese. The resulting Chinese string is then processed through the Chinese understanding system. If it fails to parse, a more robust but less precise back-off method is adopted, based on a simple [key: value] representation of the concepts encapsulated in an electronic form (e-form), derived automatically through simple generation rules from the semantic frame. If the string generated from the e-form fails to parse in the Chinese grammar, the system apologizes for begin unable to translate the user's query. This technique assumes that, if the user can accurately imitate the provided translations, the Chinese grammar will be able to process their query. There is of course a risk of rejecting a perfect translation due to gaps in coverage of the Chinese grammar, but over time these gaps will eventually be filled.
Table 1: Examples of English-to-English paraphrases that yield simpler sentence structure to improve translation feasibility.
We applied this technique to a set of some 30,000 flight domain English queries obtained from prior data collection efforts. We found that most of the challenging problems due to differences in the two languages could be handled through the use of existing mechanisms in our Genesis language generation system . A novelty of the approach we are using in the flight domain as compared with our previous research in the weather domain is a different strategy for handling the back-off generation from the e-form. In the weather domain, we adopted an example-based translation approach, which involved finding a matching template in a corpus derived from a large set of examples, and then replacing the values in the matching template's translation string with those appropriate for the original query. Since the flight domain is far more complex than the weather domain in terms of both the number of possible atributes and the complexity of the clause structure, there turns out to be a much greater sparse data problem as well as a more challenging task of variable substitution. We therefore decided, instead, to utilize our Genesis system to supply direct generation rules from the e-form into a simplified English paraphrase. We then apply the standard formal method to the simplified string. Examples of English queries alongside their simplified English paraphrases are shown in Table 1.
To evaluate our translation framework, we selected a set of 683 test utterances that had not been used for developing the translation rules. We excluded from an original pool of test data any meta-queries, such as "start over" or "good bye," which are commands directed to the system that do not need to be translated, and any utterances that were only one word long, since these were uninteresting. We also excluded utterances whose transcript failed to parse. Results are summarized in Table 2. Fewer than 2% of the recognizer outputs failed to parse, and all but 5% generated a translation string (see (3) in the table). A significant fraction of the translations were rejected due to a failure to parse in the Chinese grammar (4). For 80% of the failed utterances, we were able to generate a translation of the simplified English paraphrase, and 43% of these yielded a translation that was judged well-formed by the Chinese parsing grammar. Overall, 77.6% of the original utterances produced a parseable translation string.
Table 2: Evaluation results for spoken language translation of an unseen test corpus in the flight domain. All utterances had transcripts that could parse in the original English grammar.
This research is ongoing, and we expect to continue to improve all aspects of the system. We have begun to design a translation game, that we hope students would find entertaining, which can be played at a Web site or over the telephone. The student would be given a randomly generated English utterance, and would be tasked with providing a Chinese equivalent utterance, either by speaking or by typing in pinyin. The system would compare the utterance's meaning with one obtained from its own internal translation, and if the utterance is judged correct, would congratulate the student and provide a new utterance to translate. As the game progresses, the system would keep track of how many turns it took the student to successfully translate each utterance, and would gradually increase or decrease the difficulty level depending on the student's performance. Explicit user enrollment would allow the system to personalize the difficulty level to match the student's capabilities.
 C. Wang and S. Seneff. High-quality Speech Translation for Language Learning, InSTIL Symposium on Computer Assisted Language Learning, pp. 99--102, Venice, Italy, 2004.
 L. Baptist and S. Seneff. GENESIS-II: A Versatile System for Language Generation in Conversational System Applications. In Proceedings of ICSLP , pp. 271--274, Beijing, China, October 2000.