Beyond SensEval - Determining Interlingua Utility for MT Call for Papers Workshop Description: While it is agreed that interlingual transfer is the ultimate goal in Machine Translation (MT), much work still needs to be done to build interlingual representations for MT systems. It is difficult to determine if a representation is a good one and, failing a gold standard, a useful one. Evaluation of interlingual representations involves several levels of measurement. The representation can be measured in ontological terms and through coverage, depth, complexity and resulting graph structure. The representation and accompanying tools can be measured through the ability to analyze data into the representation consistently, through evaluating inter-annotator agreement. The representation can be measured through the application of the resulting structure to a task, in this case MT. Here, a given text is first analyzed into an interlingual (IL) representation. Then, data is generated from the IL representation, such as generating sentence output that can be compared with the original text. Each of these evaluation strategies is complex as each involves more than one source of variation. In this workshop, we explore the problem of evaluating interlingual representations in the MT context. For the morning portion of the workshop, we invite submissions related to the problem of evaluating interlingual representations and the resulting text. For the afternoon session, we encourage participation in the task presented next. The Task At the Fifth Interlingua Workshop, held in October 2002, the focus was on inter-coder reliability in coding thematic roles. Participants were provided with a dependency structure for each of 11 sentences. Each word was then to be assigned a thematic role from a list of thematic roles previously provided and defined by the workshop organizers. At the Sixth Interlingua Workshop, held in October 2003, the participants marked up and compared events, objects, and states in a multilingual corpus of a UNESCO Courier article in fifteen languages (plus English). Although participants will be invited to write a short paper for the workshop, the primary aim is to determine an upper limit on the validity of an Interlingua for translation purposes. This year's task will involve an exercise of Manual Interlingual Translation. There are two phases to the task: Task A(nalysis) and Task G(eneration). For Task A, each participant is to provide four items: (1) a foreign language text, (2) one or more English translations, (3) an interlingual representation of the foreign language text, and (4) a description of the Interlingua used. The document of interest should not be more than 300 words (English translation words that is). Participants who do not have access to parallel text for the language of their interest should contact Nizar Habash (habash@cs.columbia.edu) to help locate such text. In Task G, participants will receive the Interlingua and Interlingua description submitted by other participants. The result of Task G is an English translation created from the Interlingua. Participants will provide a (joint) written report for the workshop on the process and results of their analysis and generation. These reports will be presented during the morning session of the workshop. The afternoon will be devoted to a general discussion of the task and examination of Interlingua utility, Manual Translation Quality (ala some automatic metric such as Bleu), cross-linguistic variation, and variation across multiple English versions of the same text. Pairs of participants who score the best Manual Translation Quality will receive a valuable prize and the admiration and envy of their colleagues. Submission Guidelines: For the paper-only portion of the workshop, participants should send it in Word or PDF format via email by Friday July 23, 2004 to Nizar Habash (habash@cs.columbia.edu). Include contact info for authors, title, abstract, and full text of 4-6 pages. A workshop URL will be created for the dissemination of ongoing information. Accepted workshop papers will be published by AMTA, and authors will be asked to follow AAAI formatting instructions for their final copy. These instructions can be found at http://www.aaai.org/Publications/Templates/aaai.pdf and a template can be downloaded from http://www.aaai.org/Publications/Templates/Author-kit.zip. But note that the initial submission need not conform to these guidelines. Proposed Schedule: 11 Jun 2004: Call for participation / papers released 09 Jul 2004: Intent to participate in task due 23 Jul 2004: Non-task papers due 02 Aug 2004: Results of Task A due 16 Aug 2004: Results of Task G due / notification of accepted papers 23 Aug 2004: Camera ready papers due 28 Sep 2004: Workshop Acceptance Criteria: Participants will be invited by the committee, which will base its decisions on the originality of the work and relevance to the goal of addressing issues common to both research communities. Workshop Organizers: Dr. Nizar Habash, Computer Science Department, Columbia University. habash@cs.columbia.edu (Chair) Dr. Bonnie Dorr, Computer Science Department, University of Maryland College Park. bonnie@umiacs.umd.edu Dr. Eduard Hovy, Director of the Natural Language Group, Information Sciences Institute, University of Southern California. hovy@isi.edu Florence Reeder, MITRE. freeder@mitre.org