Language model adaptation for specific texts

• US 9,256,596 B2
• Filed: 06/18/2014
• Issued: 02/09/2016
• Est. Priority Date: 06/18/2014
• Status: Active Grant

• Current Assignee: Nice Ltd
• Original Assignee: Nice Systems Limited (Nice Ltd)
• Inventors: Nissan, Maor, Artzi, Shimrit, Bretter, Ronny
• Primary Examiner(s): Poon, King
• Assistant Examiner(s): Siddo, Ibrahim
• Application Number:
  US14/307,520
• Publication Number:
  US 20150370784A1
• Time in Patent Office: 601 Days
• US Class Current: 1/1
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/247 Thesauruses; Synonyms
• CPC Class Codes: G06F 40/268 Morphological analysis
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/42 Data-driven translation
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models

Abstract

A computerized method for adapting a baseline language model, comprising obtaining a textual corpus of documents that comprise textual expressions, incorporating in the baseline language model textual expressions from documents which are determined as relevant to a provided target text based on a plurality of different relevancy determinations between the documents and the provided target text, thereby adapting the baseline language model to form an adapted language model for recognizing terms of a context of the provided target text, wherein the method is automatically performed on an at least one computerized apparatus configured to perform the method.

Claims

1. A computerized method for adapting a baseline language model, comprising:

obtaining a textual corpus of documents that comprise textual expressions;

incorporating in the baseline language model textual expressions from documents which are determined as relevant to a provided target text based on a plurality of different relevancy determinations between the documents and the provided target text, thereby adapting the baseline language model to form an adapted language model for recognizing terms of a context of the provided target text,

wherein a first relevancy determination comprises a determination by sufficiently small evaluated perplexities of the baseline language model with respect to the target text, and

wherein a second relevancy determination comprises a determination of matches between stems of words in the textual expressions in the documents and stems of words of the target text, and

wherein a third relevancy determination comprises a determination of matches between words in the textual expressions of the documents with words of the target text that for the matching have been converted to synonyms thereof according to a preset dictionary of synonyms, and

wherein a fourth relevancy determination comprises a determination of semantic similarities of words in the textual expressions of the documents and words of the target text based on semantic distance and reduction thereof, and

wherein the method is automatically performed on an at least one computerized apparatus configured to perform the method.

2. The method according to claim 1, wherein the textual expressions from the documents which are determined as relevant to the provided target text are filtered by removing redundant duplicates.

3. The method according to claim 1, wherein said incorporating in the baseline language is carried out by an interpolation directed to diminish a perplexity of the adapted language model with respect to a provided preferred text.

4. The method according to claim 1, wherein said incorporating in the baseline language model is carried out by way of an intermediary language model formed with the textual expressions from the documents which are determined as relevant to the provided target text.

5. The method according to claim 1, wherein stems of words in the documents of the textual corpus are indexed for expediting references thereof, thus expediting performances of the plurality of different relevancy determinations.

6. The method according to claim 1, further comprising evaluating the performance of the adapted language model in recognizing speech related to the provided target text relative to the respective performance of the baseline language model, consequently determining which of the cited language models is more suitable for decoding speech related to the provided target text.