Statistical language model trained with semantic variants

  • US 7,912,702 B2
  • Filed: 10/31/2007
  • Issued: 03/22/2011
  • Est. Priority Date: 11/12/1999
  • Status: Active Grant
  • ×
    • Pin Icon | RPX Insight
    • Pin
First Claim
Patent Images

1. A method of generating a statistical language model (SLM) grammar for a task domain which includes semantically variant words and phrases, the method comprising the steps of:

  • (a) providing a set of content words which can be associated with user questions in the task domain; and

    using a computer system;

    (b) determining semantic variants for each word in said set of content words;

    wherein said semantic variants include at least synonyms;

    (c) forming a semantic set of questions related to said user questions based on said synonyms;

    (d) performing semantic decoding on said semantic set of questions, to identify a disambiguated set of questions; and

    (e) configuring n-gram probabilities for words and phrases in said SLM grammar based on said set of disambiguated questions;

    wherein said SLM grammar is configured to recognize semantic variants of questions posed to a natural language speech recognition engine.

View all claims