Shareable filler model for grammar authoring

• US 7,865,357 B2
• Filed: 03/14/2006
• Issued: 01/04/2011
• Est. Priority Date: 03/14/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Wang, Ye-Yi, Ju, Yun-Cheng, Yu, Dong, Acero, Alejandro
• Primary Examiner(s): Smits; Talivaldis I
• Assistant Examiner(s): PULLIAS, JESSE SCOTT
• Application Number:
  US11/375,488
• Publication Number:
  US 20070219793A1
• Time in Patent Office: 1,757 Days
• Field of Search: None
• US Class Current: 704/9
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A method of forming a shareable filler model (shareable model for garbage words) from a word n-gram model is provided. The word n-gram model is converted into a probabilistic context free grammar (PCFG). The PCFG is modified into a substantially application-independent PCFG, which constitutes the shareable filler model.

Claims

1. A computer-implementable method comprising:

providing a generic word n-gram model;

converting, using a processor of a computer, the word n-gram model into a probabilistic context free grammar (PCFG);

modifying, using the processor of the computer, the PCFG into a substantially application-independent PCFG, which constitutes a generic shareable filler model, and

recognizing speech signals with the help of the shareable filler model using the processor of the computer,

wherein modifying the PCFG into a substantially application-independent PCFG comprises removing links from a start node of the PCFG to words in the PCFG to eliminate all links between the start node of the PCFG to words in the PCFG and only maintaining at least one link between the start node and at least one backoff node, the backoff node being utilized to estimate probabilities associated with the start node and the end node.

2. The method of claim 1 wherein modifying the PCFG into a substantially application-independent PCFG further comprises removing links from the words in the PCFG to an end node of the PCFG and only maintaining at least one link between the end node and the at least one backoff node.

3. The method of claim 2 wherein modifying the PCFG into a substantially application-independent PCFG further comprises updating backoff weights in the PCFG to compensate for the removed links from the start node to the words and from the words to the end node.

4. A language model comprising the shareable filler model formed in the method of claim 1.

5. A speech recognition system comprising the language model of claim 4.

6. A grammar comprising a pre-amble, a post-amble and a slot between the pre-amble and the post-amble, wherein each of the pre-amble and the post-amble comprises the shareable filler model formed in the method of claim 1.

7. A grammar authoring architecture comprising:

a memory storing at least one slot that carries semantic information and at least one filler in series with the at least one slot,

a processor executing a model in which the at least one filler is modeled with a PCFG that is generated from a generic word n-gram, the processor outputting a speech recognition result, and

wherein the PCFG is substantially application-independent and comprises internal words that are disconnected from its start node and its end node by elimination of all links between the start node of the PCFG and the internal words and elimination of all links between the end node of the PCFG and the internal words, and

wherein only at least one link between the start node and at least one backoff node and only at least one link between the end node and the least one backoff node are maintained, and

wherein the backoff node is utilized to estimate probabilities associated with the start node and the end node.

8. The architecture of claim 7 and further comprising context cues integrated with the at least one filler.

9. The architecture of claim 7 wherein the substantially application-independent PCFG further comprises backoff weights that are adjusted to compensate for the disconnected internal words.

10. The architecture of claim 9 and further comprising context cues integrated with the at least one filler.

11. A computer-implementable method comprising:

providing a shareable filler model;

interpolating, using a processor of a computer, the shareable filler model with a training set; and

recognizing speech signals with the help of the shareable filler model interpolated with the training set using the processor of the computer,

wherein the shareable filler model is a word n-gram based shareable filler model, and

wherein the word n-gram based shareable filler model comprises a substantially application-independent PCFG that is generated from a word n-gram, and

wherein the substantially application-independent PCFG comprises internal words that are disconnected from its start node and its end node by elimination of all links between the start node of the PCFG and the internal words and elimination of all links between the end node of the PCFG and the internal words, and

wherein only at least one link between the start node and at least one backoff node and only at least one link between the end node and the least one backoff node are maintained, and

wherein the backoff node is utilized to estimate probabilities associated with the start node and the end node.

12. A speech recognition system comprising a language model formed in the method of claim 11.