Method and apparatus for detection of sentiment in automated transcriptions

• US 8,412,530 B2
• Filed: 02/21/2010
• Issued: 04/02/2013
• Est. Priority Date: 02/21/2010
• Status: Active Grant

• Current Assignee: Nice Ltd
• Original Assignee: Nice Systems Limited (Nice Ltd)
• Inventors: Pereg, Oren, Ashkenazi, Tzach
• Primary Examiner(s): Smits, Talivaldis Ivars
• Application Number:
  US12/709,482
• Publication Number:
  US 20110208522A1
• Time in Patent Office: 1,136 Days
• Field of Search: None
• US Class Current: 704/270
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence

Abstract

A method for automatically detecting sentiments in an audio signal of an interaction held in a call center, including, receiving the audio signal from a logging and capturing unit. Performing audio analysis on the audio signal to obtain text spoken within the interaction. Segmenting the text into context units according to acoustic information acquired from the audio signal to identify units of speech bound by non-speech segments, wherein each context unit includes one or more words. Extracting a sentiment candidate context unit from the context units using a phonetic based search. Extracting linguistic features from the text of the sentiment candidate context unit and acoustic features from a segment of the audio signal associated with the sentiment candidate context unit. Determining in accordance with the linguistic features and acoustic features whether the sentiment candidate context unit is valid or erroneous, and determining sentiment polarity and intensity.

Claims

1. A method for automatically detecting sentiments in an audio signal of an interaction held in a call center, comprising:

receiving the audio signal from a logging and capturing unit associated with the call center;

performing audio analysis on the audio signal to obtain text spoken within the interaction;

segmenting the text into context units according to acoustic information acquired from the audio signal to identify units of speech bound by non-speech segments;

wherein each context unit includes one or more words;

extracting a sentiment candidate context unit from the context units using a phonetic based search;

extracting linguistic features from the text of the sentiment candidate context unit and acoustic features from a segment of the audio signal associated with the sentiment candidate context unit;

determining in accordance with the linguistic features and acoustic features whether the sentiment candidate context unit is valid or erroneous; and

determining sentiment polarity and intensity for the sentiment candidate context unit.

2. The method of claim 1 further comprising determining a total interaction score indicating sentiment polarity and intensity of the interaction.

3. The method of claim 2 wherein determining the total interaction score is performed using an interaction sentiment model.

4. The method of claim 3 further comprising a training step for generating the interaction sentiment model.

5. The method of claim 4 wherein the training step comprises:

receiving a training interaction;

performing audio analysis on the training interaction for obtaining;

segmenting the text into context units;

determining sentiment context units from the context units;

extracting features related to the interaction;

receiving tagging data related to the interaction; and

generating the interaction sentiment model using the features and the tagging data.

6. The method of claim 1 wherein the audio analysis comprises activating at least one engine selected from the group consisting of:

a speech to text engine;

a word spotting engine;

an emotion detection engine; and

a talkover engine.

7. The method of claim 1 wherein determining whether the candidate context unit is valid or erroneous is performed using a validity model.

8. The method of claim 5 further comprising a training step for generating the validity model.

9. The method of claim 8 wherein the training step comprises:

receiving a training interaction;

performing audio analysis on the training interaction for obtaining text;

segmenting the text into context units;

extracting a training sentiment candidate context unit from the context units;

extracting features from the training sentiment candidate context unit or from the audio signal;

receiving tagging data for the training sentiment candidate context unit; and

generating a model using the training sentiment candidate context unit and the tagging data.

10. The method of claim 1 wherein determining the sentiment polarity and intensity is performed using a sentiment polarity and intensity model.

11. The method of claim 10 further comprising a training step for generating the polarity or intensity model.

12. The method of claim 11 wherein the training step comprises:

receiving a training interaction;

performing audio analysis on the training interaction for obtaining text;

segmenting the text into context units;

extracting a training sentiment candidate context unit from the context units;

extracting features from the training sentiment candidate context unit and from the audio signal;

receiving tagging data for the training sentiment candidate context unit; and

generating the polarity or intensity model using the training sentiment candidate context unit and the tagging data.

13. The method of claim 1 wherein the feature is selected from the group consisting of:

a term in the sentiment candidate context unit;

a feature associated with the term;

number of adjectives in the sentiment candidate context unit;

number of adverbs in the sentiment candidate context unit;

whether the sentiment candidate context unit contains a modal or a pronoun;

speech rate in the sentiment candidate context unit;

silence average duration in the sentiment candidate context unit;

duration of the sentiment candidate context unit;

main speaker in the sentiment candidate context unit;

number of positive or negative dictionary terms in the sentiment candidate context unit;

whether the term is an exact match with a dictionary term or not;

a score indicating a level of match between the term and the dictionary term;

score based on terms found in the sentiment candidate context unit;

interaction duration;

repetitions within the interaction of a term related to a word in the sentiment candidate context unit;

part of speech of a word within the sentiment candidate context unit;

a score obtained from a second audio analysis engine operated on the interaction or part thereof;

a feature of a context unit neighboring the sentiment candidate context unit;

a position of a context unit; and

a relative position of a context unit.

14. The method of claim 1 further comprising a step of further analysis, selected from the group consisting of:

topic extraction, link analysis, and root cause analysis.

20. A non-transitory computer readable storage medium containing a set of instructions for a general purpose computer, the set of instructions comprising instructions to perform the method of claim 1.

15. An apparatus for automatically detecting sentiments in an audio signal of an interaction held in a call center, comprising:

an interaction receiving component arranged to receive an audio signal representing the interaction;

an extraction component for obtaining text spoken within the interaction;

a context unit segmentation component arranged to segment the audio signal into context units according to acoustic information acquired from the audio signal to identify units of speech bound by non-speech segments;

wherein each context unit includes one or more words;

a sentiment candidate identification component for identifying a sentiment candidate from the context units using a phonetic based search;

a feature extraction component arranged to extract linguistic features from the text of the sentiment candidates and acoustic features from a segment of the audio signal associated with the sentiment candidates; and

a classification component arranged to apply a model to the linguistic features and acoustic features to determine whether the audio signal contains sentiment, or sentiment polarity and intensity of the sentiment within the audio signal.

16. The apparatus of claim 15 further comprising a natural language processing component arranged to perform linguistic analysis on text extracted by the extraction component.

17. The apparatus of claim 15 wherein the extraction component comprises at least one engine selected from the group consisting of:

a speech to text engine, a word spotting engine;

an emotion detection engine; and

a talkover engine.

18. The apparatus of claim 15 further comprising a model training component arranged to train the model.

19. The apparatus of claim 15 wherein the feature extraction component is arranged to extract at least one feature selected from the group consisting of:

a term;

a feature associated with the term;

number of adjectives in a context unit;

number of adverbs in a context unit;

whether the context unit contains a modal or a pronoun;

speech rate in context unit;

silence average duration in the context unit;

duration of the context unit;

main speaker in the context unit;

number of positive or negative dictionary terms in the context unit;

whether the term is an exact match with a dictionary term or not;

a score indicating a level of match between the term and the dictionary term;

score based on terms found in the context unit;

interaction duration;

repetitions within the interaction of a term related to a word in the context unit;

part of speech of a word within the context unit;

a score obtained from a second audio analysis engine operated on the interaction or part thereof;

a feature of a second context unit neighboring the context unit;

a position of a context unit; and

a relative position of a context unit.