Uphill Battles in Language Processing Scaling Early Achievements to Robust Methods NSF-supported workshop to be held in conjunction with EMNLP in Austin, Texas, with partial funding available for students and post-docs presenting posters at the workshop

Early researchers in Natural Language Processing had lofty goals, including getting computers to understand stories, engage in natural, cooperative dialogues with people, and translate text and speech fluently and accurately from one human language to another. While there were significant early achievements (including systems such as SHRDLU, LUNAR and COOP), the knowledge they were based on and the techniques they employed could not be scaled up for practical use.

While much of what early researchers set out to achieve has been either forgotten or side-lined in favor of what can be done by exploiting large data sets and processing power, its potential value has not gone away: There is much to be gained from recognizing not just what was said, but why; from identifying conclusions people naturally draw from both what has been said and hasn't; and from representing domains in a sufficiently rich way to reduce reliance on what is explicit in a text. As such, we believe there can be a broad and positive impact of reviving early aspirations in the current context of large data sets, "deep" and probabilistic methods, and (especially) methods that aim to combine the capabilities of logical and data-driven conclusions.

This workshop will remind the community of early goals and attempts to achieve them, the uphill battles that remain, and hopefully revive them in a context in which much more can be done. The workshop will be split into four sessions, covering different areas and challenges. Each session will start with four short (10 minute) presentations, two by established researchers who carried out early work in the area, and two by more junior researchers who are known for their work on specific problems in the area. Each session will have a moderator, who will then invite the speakers and audience to discuss open questions. The four sessions and the invited speakers are listed below.