Researchers Tackle Mobile NLP

BY BARBARA ROSARIO AND TIM PAEK

The ACL 2008 Workshop on Mobile Language Processing took place on June 20 in Columbus, Ohio following ACL-08: HLT with an invited talk by Dr. Lisa Stifelman, Principal User Experience Manager at Tellme/Microsoft, seven oral paper presentations, a poster and a demo session and a panel discussion. The workshop attracted around 30 participants on the last day of a long conference.

In anticipation of new and exciting applications for natural and spoken language processing on mobile devices, this workshop provided a forum for discussing some of the challenges that are unique to this domain. For instance, processing and memory limitations on small devices need to be addressed. Several papers addressed this issue. In "Information extraction using finite state automata and syllable n-grams" Seon et al. proposed a modified HMM for information extraction in a mobile environment. This kind of model has the advantage of being compact. Huggins-Daines et al. proposed a simple entropy-based technique to improve the scalability of acoustic models in embedded systems; they showed a significant speed-up in recognition with a negligible increase in word error rate ("Mixture Pruning and Roughening for Scalable Acoustic Models.") Ganchev and Dredze in "Small Statistical Models by Random Feature Mixing" showed how it is possible to do efficient NLP learning by reducing the number of parameters on resource constrained devices with little loss in performance; and "A Wearable Headset Speech-to-Speech Translation System" by Krstovski et al. shrunk a speech translation system to fit into a wearable speech-to-speech translation system.

Some applications and practical considerations may require a client/server or distributed architecture: what are the implications for language processing systems in using such architectures? Homola ("Distributed Database for Mobile NLP Applications") proposed a distributed database for lexical transfer in machine translation. The database contains data shared among multiple devices and automatically synchronizes them.

The limitation of the input and output channels necessitates typing on increasingly smaller keyboards which can be quite difficult, and similarly reading on small displays is challenging. Speech and multimodal interfaces, language generation and dialog systems would provide a natural way to interact with mobile devices. A multimodal dialogue system for interacting with a home entertainment center via a mobile device was proposed by Gruenstein et al. in "A Multimodal Home entertainment Interface via a Mobile Device."

Furthermore, the growing market of cell phones in developing regions can be used for delivering applications in the areas of health, education and economic growth to rural communities. For the health domain, Nikolova and Ma in their paper "Assistive Mobile Communication Support" discussed the role of mobile technologies in a system for communication support for people with speech and language disabilities.

The panel discussion was interesting and lively, covering issues such as "killer apps," thin vs. thick clients, computational power, greatest bottlenecks to getting NLP technology onto mobile devices etc. Our illustrious panel session members were: Susan Boyce (Tellme), Ken Church (MSR), Michael Johnston (AT&T Labs Research), Mike Phillips (vlingo) and Noah Smith (CMU).

We believe that the issues raised by the papers in this Workshop represent just the tip of the iceberg, and we hope that by raising awareness of these issues, more research will be aimed at mobile language processing.