Academic and Professional Information

Computer Science Department
Stony Brook University Stony Brook, NY 11794-4400
631-632-8654
svetastenchikova at gmail dot com

My research interests are in the areas of task-oriented dialog systems and automatic question answering.

Dialog systems are a rich application area for artificial intelligence and machine learning research as well as a challenging engineering task. In my work I apply knowledge from psychology and linguistics to improve usability of dialog systems. I am interested in research on adaptive dialog systems that over time learn from the user and improve their performance. In my research I look at directive and responsive adaptation. Directive adaptation occurs when a system guides a user into using particular words or grammar that can be better recognized by a system. For example, a system may use a particular form in its prompts: "from X to Y" to to prime a user and increase the chances of them using the same form in an utterances. Responsive adaptation is practiced in a system that changes its behavior adjusting to a user. Adjusting acoustic of a speech recognizer to a particular speaker, a a language model to a dialog state, language generation component to better serve a particular user are the examples of Responsive adaptation.

I am also interested in interactive question answering application of Natural Language Processing. Information extraction plays an important role in modern society with the increasing amount of information available on the web both published by accredited newspapers and independently created. Question answering is an interface on the information extraction that allows users find information using natural language queries. Question answering uses and benefits from improvement of natural language components: taggers, parsers, chunkers, ontologies, etc. I designed and developed a question answering system which I use to evaluate natural language tools and components. I also work on creating interactive speech interface for question answering.

Refereed Conference and Workshop Papers

Predicting Concept Types in User Corrections in Dialog In review

Analysis of Non-Understandings in the Communicator Corpus In review

Exact Phrases in Information Retrieval for Question Answering S. Stoyanchev, Y. C. Song, and W. Lahti In Proceedings of Information Retrieval for Question Answering workshop, Manchester, UK, August 2008 slides

Name-Aware Speech Recognition for Interactive Question Answering S. Stoyanchev, D. Hakkani-Tur, and G. Tur In the proceedings of ICASSP-2008, IEEE International Conference on Acoustics, Speech, and Signal Processing, Las Vegas, NV, April 2008 poster

Measuring Adaptation Between Dialogs S. Stoyanchev, A. Stent SIGDIAL-2007, 8th SIGdial Workshop on Discourse and Dialogue, Antwerp, Belgium, September 1-2, 2007. slides

RavenCalendar: A Multimodal Dialog System for Managing a Personal Calendar S. Stenchikova, B. Mucha, S. Hoffman, A. Stent NAACL HLT Demonstration Program, pages 15-16, Rochester, New York, USA, April 2007

Dialog Systems for Surveys: the Rate-a-Course System A. Stent, S. Stenchikova, and M.Marge Proceedings of the 1st IEEE/ACL Workshop on Spoken Language Technology. SLT 2006.

QASR: Question Answering Using Semantic Roles for Speech Interface S. Stenchikova, D. Hakkani-Tur, and G. Tur Proceedings of ICSLP-Interspeech 2006, Pittsburgh, PA

QASR: Spoken Question Answering Using Semantic Role Labeling. S. Stenchikova, D. Hakkani-Tur, G. Tur ASRU-2005, 9th biannual IEEE workshop on Automatic Speech Recognition and Understanding, Cancun, Mexico, December, 2005 (Demonstration)

Unpublished

Thesis Proposal. Svetlana Stoyanchev Defended February 2008. slides

Semantics and Question Answering: an approach to “why” questions. Svetlana Stenchikova, December 2006.

• I collaborate with a dialogs research group at Carnegie Melon University. I use the transcribed and annotated speech data from Let’s Go! dialog system, a real live dialog system that provides bus information to Pittsburgh residents. In my research on responsive adaptation I predict whether a user utterance contains a concept (place, time, or bus route) using prosodic features of an utterance as well as output from a speech recognizer using a generic language model. Prediction of a concept allows an adjustment of a language model for optimal recognition of an utterance. In my work on responsive adaptation I focus on user’s adaptation to the system in the forms of time concept which can be rendered as five pm, five o’clock, five. I adjust both system language understanding capability as well as the forms of concept generated by the system and observe how the interplay between the generation and understanding affects the user. In particular, I am interested in the question: What has a stronger affect on a user, system’s misunderstandings or priming in system prompts? You can find more information on these projects in my thesis proposal .
• Adaptation in dialogs. Calendar system. I have designed and built a personal calendar dialog system with voice interface. The dialog system uses RavenClaw/Olympus architecture developed at CMU. The calendar application serves as a platform for psycholinguistic experiments. I am currently designing user studies to investigate lexical and syntactic adaptation in task-oriented human-computer dialogs.
• Rate-a-course survey system. We are currently looking at adaptation and user modeling in the context of the Stony Brook Rate-A-Course dialog system, which gathers information from students about different aspects of their courses, such as instructor, assignments, text book, etc. At the moment, we are looking at the impact of two types of adaptation on system performance and survey effectiveness: system vs. user initiative; and lexical adaptation. First, we are looking at whether having the system follow the user's goals affects the type and amount of feedback the user provides. In our experiment, the system may have the user talk about all course topics in a specific order, or it may let the user choose topic order or even which topics to discuss.

• I have designed and implemented an open-domain question answering system. The system uses AQUAINT corpus indexed by Lucene as a data source. We use named entity tagging capability of Lydia system. We evaluate information retrieval (IR) performance on Web dataset and on AQUAINT dataset. We use linguistic processing to analyze questions and extract noun, verb and prepositional phrases. The extracted phrases are passed as exact phrases to a search query for Web and Lucene search interfaces. In the future work we would like to use other capabilities of Lydia system to improve IR of question answering. One such capability is expansion of named entities using synonym sets generated by the Lydia system during corpus pre-processing. So far, the tests did not yield positive results because of the noise in automatically generated synonym sets.
• Interactive Question Answering. We have added a voice interface to the system using SRI's Dynaspeak speech recognizer. We investigate how adapting speech recognition language model to the topic of the question can improve speech recognition. In this work we try to improve question recognition by allowing dialog interaction for question specification. This approach deals with frequent misrecognitions of names and unique words.

• Question Answering with semantic roles. Built question answering system. Applied Semantic Role Labeling to the Question Answering Task in order to improve the performance of the QA system and to ensure that the answers more concise and grammatically correct Mentors: Gokhan Tur and Dilek Tur.
• Misunderstandings in the human-computer dialogs. We analyze the corpora of the COMMUNICATOR dialogs by hand-annotating the misunderstanding segments. The annotated segments are used for finding the segments in other dialogs. This turns out to be not so easy and we are trying different techniques: heuristics, machine learning (Boosting, CRF), frequences of the n-grams in errors and outside. Analyzing the annotated miscommunication segments helps us identify successful dialog strategies.
• Implementation and enhencement of a probabilistic English Language Parser as a part of the Oak system. Advisor professor Sekine. The parser is based on the Collins' Lexicalized probabilistic parser (described in Collins' PhD dissertation) with the addition of the improvement using coordination structure information described in: Sadao Kurohashi and Makoto Nagao, A Syntactic Analysis Method of Long Japanese Sentences based on the Detection of Conjunctive Structures, Journal of Computational Linguistics, Vol.20, No.4, pp.507-534 (1994.10). The research purpose of the project is to see if the parsing functionality may be improved using the coordinate structures in the language.
• Analysis of charismatic speach project. Supervised by professor Julia Hirshberg, Columbia University.
• Worked on SMoTE project (Statistical Models of Translation Equivalency) with a group of researchers headed by professor Dan Melamed Designed and implemented a dynamic programming algorithm for a bootstrapping part of a Machine Translation system. Bootstrapping produces a parse tree for one of the languages in the input bitext (a pair of mutual translation texts) based on a parse structure of another language and on word links between the two parts of the bitext.