NAMEAWARE SPEECH RECOGNITION FOR INTERACTIVE QUESTION ANSWERING - PowerPoint PPT Presentation

Title: NAMEAWARE SPEECH RECOGNITION FOR INTERACTIVE QUESTION ANSWERING


1
NAME-AWARE SPEECH RECOGNITION FOR INTERACTIVE
QUESTION ANSWERING Svetlana Stoyanchev 3, Gokhan
Tur2 , Dilek Hakkani T\ur1 1 International
Computer Science Institute (ICSI), Speech Group,
Berkeley, CA, USA2 SRI International, Speech
Technology and Research (STAR) Lab., Menlo Park,
CA, USA 3 State University of New York (SUNY),
Stony Brook, NY, USA svetastenchikova_at_gmail.com
gokhan_at_speech.sri.com dilek_at_icsi.berkeley.edu
INTRODUCTION
DATASET
APPROACH
Goal Improving speech recognition in a
voice-enabled question answering application
through interactivity

• Allow interactivity in specifying a named
  entity.
• Recognize a named entity using grammars and
  retrieve matching documents
• Build a question-specific language model

40 questions from TREC 2007 competition AQUAINT
corpus (3 Gigabyte document collection)
Question Answering QA is a natural language
interface for information retrieval TREC is a
yearly competition. Participants are given a
document set and a set of questions factoid
(who, what, when, where), list, or other (find
other relevant information on the given
topic).
EXPERIMENTS AND RESULTS
Evaluation 3 speakers read 40 questions
twice Set 3 40 questions with a named entity
(NE) Set 4 40 questions without a named entity
CONTROL FLOW

• System asks a user to specify a named entity
  (1). A named entity is recognized using grammar
  constructed from a database of named entities
  (2). Next, we extract documents matching the
  named entity (3). Using these documents we build
  a language model using named entity-specific
  language model (4). In parallel Build a language
  model using TREC questions (5). Merge the two
  Language Models (6)
• In the final step we recognize the question
  using the new language model (7).

On what date did Michael Brown resign as a head
of FEMA?
Identify that the target is DATE, identify a
named entities Michael Brown and FEMA
NE Michael Brown, FEMA Phrases head of
FEMA Verb resigned
Search can use web or a document collection
Find Candidate Sentences, identify DATEs
On September 12, 2005, in the wake of what was
widely believed to be feckless handling of the
aftermath of Hurricane Katrina and facing
allegations that he had falsified portions of his
résumé, Brown resigned, it is mentioned
earlier in the document that Michael Brown was a
head of FEMA After his September 12 resignation,
Brown continued working for FEMA
Cheating model (contains questions)
1
2
September 12, 2005 September 12
Candidate answers
Identify match between September12, 2005 and
September 12.
3

• Motivation
• The word error rates of the state-of-the-art
  open-domain speech recognition technology are
  around 25-30.
• Performance is known to be even lower for names
  and rare words.
• Named entities (NE) are strongly associated with
  the content words. For example for the target
  name Gordon Gekko, one question used in TREC 2004
  evaluations is In what film is Gordon Gekko the
  main character?, including non-function words
  related to the movie industry, such as film or
  character.
• The goal is to capture content words using the
  documents where the named entity appears
  frequently.

• Conclusions
• The question-specific model achieves 32.2
  reduction in word error rate from the baseline
  using the questions where pronominal references
  are resolved. (from 58.36 to 26.02)
• Used name specific language models in question
  answering task where target name in question is
  asked to the user beforehand.
• Used TREC benchmark questions on the AQUAINT
  corpus.
• Future Work
• Grounding process of the named entity
• ask a user for the type of the named entity (e.g.
  person, organization, location, movie) and
  associations
• build a focused grammar a user may be
  specifying
• Orhan Pamuk, a Turkish writer.

5
4
6
7
Related Work

• R. Iyer and M. Ostendorf, Modeling long distance
  dependencies
• in language Topic mixtures versus dynamic cache
  model IEEE Transactions on Speech and Audio
  Processing
• D. Gildea and T. Hofmann, Topic-based language
  models using EM in Proceedings of Eurospeech
• F. Bechet, G. Riccardi, and D. Hakkani-Tur
  Mining spoken dialog corpora for system
  evaluation and modeling