System and method for automatic call segmentation at call center

• US 8,750,489 B2
• Filed: 10/23/2008
• Issued: 06/10/2014
• Est. Priority Date: 10/23/2008
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Park, Youngja
• Primary Examiner(s): ADDY, THJUAN KNOWLIN
• Application Number:
  US12/257,037
• Publication Number:
  US 20100104086A1
• Time in Patent Office: 2,056 Days
• Field of Search: 379/265.09, 379/88.01, 379/266.1, 379/243, 379/88.04, 379/72, 379/85, 379/71, 704/246, 704/245, 704/251
• US Class Current: 379/266.1
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: H04M 2201/40 using speech recognition speech recognition per se G10L15/00
• CPC Class Codes: H04M 2201/41 using speaker recognition speaker recognition per se G10L17/00
• CPC Class Codes: H04M 2203/357 Autocues for dialog assistance
• CPC Class Codes: H04M 3/42221 Conversation recording systems at the subscriber's set H04M1/656
• CPC Class Codes: H04M 3/51 Centralised call answering arrangements requiring operator intervention , e.g. call or contact centers for telemarketing

Abstract

A system and method for automatic call segmentation including steps and means for automatically detecting boundaries between utterances in the call transcripts; automatically classifying utterances into target call sections; automatically partitioning the call transcript into call segments; and outputting a segmented call transcript. A training method and apparatus for training the system to perform automatic call segmentation includes steps and means for providing at least one training transcript with annotated call sections; normalizing the at least one training transcript; and performing statistical analysis on the at least one training transcript.

Claims

1. A method for automatic segmentation of call transcripts into content-based call segments comprising the steps of:

automatically detecting boundaries between utterances in call transcripts;

automatically classifying utterances into target call sections based on contents of at least one of the utterances;

automatically partitioning the call transcript into content-based call segments; and

outputting a segmented call transcript.

2. The method of claim 1 wherein partitioning comprises merging utterances into call segments.

3. The method of claim 2 wherein said merging comprises at least one of merging adjacent utterances which are classified into same call sections and merging short call sections into previous call sections.

4. The method of claim 3 wherein said merging short call sections comprises the steps of:

comparing a length of each utterance to a predefined standard; and

identifying an utterance as short based on a predetermined relationship to the predefined standard.

5. The method of claim 4 wherein the predefined standard comprises at least one of a number of words and a time duration for the utterance.

6. The method of claim 1 further comprising normalizing speech from the transcript prior to detecting utterance boundaries.

7. The method of claim 6 wherein the normalizing comprises the steps of:

normalizing all filler words into a first artificial generic word;

normalizing numeric tokens and special expressions into a second artificial generic word; and

retrieving lemma words from training data and recognizing the words in the call transcript.

8. The method of claim 1 further comprising determining if the call transcript is an automatic speech recognition (ASR) transcript.

9. The method of claim 8 wherein, when the call transcript is determined to be an ASR transcript, performing additional steps of:

recognizing each utterance; and

identifying a speaker as one of a customer and an agent for each utterance.

10. The method of claim 9 further comprising automatically extracting speaker identifying keywords from speaker annotated text.

11. The method of claim 9 wherein said identifying a speaker comprises applying learned speaker statistics comprise at least one of:

p(agent|w) is the probability of the speaker being the agent when word w is given; and

p(customer|w) is the probability of the speaker being the customer when word w is given.

12. The method of claim 1 wherein automatically detecting utterance boundaries comprises steps of:

determining word boundaries; and

deciding if determined word boundaries are utterance boundaries based on analysis of call transcript sections as compared to learned call statistics in an utterance boundary detection model.

13. The method of claim 12 wherein said learned call statistics comprise at least one of:

P_L(w<sub>i−

1</sub>, w_i) is the probability of w<sub>i−

1</sub> and w_i being the last two words of an utterance;

P_L(w_i) is the probability of w_i being the last word of an utterance;

P_F(w_j) is the probability of w_j being the first word of an utterance;

P_F(w_i,w_j+1) is the probability of w_j and w_j+1 being the first two words of an utterance;

C(w_i,w_j) is the correlation coefficient of w_i and w_j appearing as a bigram within an utterance, wherein the bigram is computed by using Dice coefficient

14. The method of claim 1 wherein automatically classifying based on contents of at least one of the utterances comprises steps of:

detecting at least one predefined feature in the utterance; and

computing a most likely target call section type for the utterance.

15. The method of claim 14 wherein said detecting at least one predefined feature comprises detecting at least one of speaker identification, call section type of a previous utterance, position of the utterance in the call transcript, a number of domain-specific words in the utterance and words that appear multiple times in the utterance.

16. A system for automatic segmentation of call transcripts into content-based call segments comprising:

a segmentator component for automatically detecting boundaries between utterances in the call transcripts, for automatically classifying utterances into target call sections based on contents of at least one of the utterances, and for automatically partitioning the call transcript into call segments; and

an annotator component for annotating the call transcript and for outputting a segmented call transcript.

17. The system of claim 16 further comprising at least one database for storing at least a call segment classification model and an utterance boundary detection model.

18. The system of claim 16 further comprising an automatic speech recognition (ASR) engine.

19. The system of claim 16 further comprising a speaker identification component.

20. The system of claim 16 further comprising a normalization component for normalizing said call transcript.

21. A method for training a system to perform automatic segmentation of call transcripts comprising the steps of:

providing at least one training transcript with content-based annotated call sections;

normalizing the at least one training transcript; and

performing statistical analysis on the at least one training transcript to pre-define features found in call sections of the at least one training transcript and compute statistics for pre-defined features.

22. The method of claim 21 wherein said performing statistical analysis comprise computing statistics for a plurality of features, said features comprising at least one of speaker identification statistics based on occurrence of agent keywords and customer keywords, domain specificity statistics for occurrence of domain specific terms in said call sections, position statistics regarding occurrence of utterances in said call sections, and word appearance statistics regarding occurrence of words in said call sections.

23. Apparatus for training a system to perform automatic segmentation of call transcripts comprising:

a normalization component for obtaining at least one training transcript with content-based annotated call sections and normalizing the at least one training transcript; and

a statistical analysis component for performing statistical analysis on the at least one training transcript to pre-define features found it call sections of the at least one training transcript and to compute statistics for pre-defined features.

24. The apparatus of claim 23 further comprising at least one model generation component for generating at least one of a content-based call segment classification model and an utterance boundary detection model based on the statistical analysis.