Question detection model for call transcript

A system for classifying one or more sentences to be a question or not a question using a call transcript of an audio or video call, the system comprising:

• a database system implemented using a server system, the database system configurable to cause:
• performing automatic speech recognition on the audio or video call such that the call transcript of the audio of video call is generated, the call transcript including a first sentence and a second sentence;
• applying inverse text normalization (ITN) to the first sentence such that a first result is generated;
• generating a first confidence level indicating whether the first result represents a question using a finite state transducer (FST);
• applying term frequency inverse document frequency (TFIDF) to the first sentence such that a second result is generated;
• generating a second confidence level indicating whether the first sentence is a question by applying a logistic regression (LR) model to the second result;
• generating a third result by applying a model of sentence embedding to the first sentence;
• generating a third confidence level indicating whether the first sentence is a question by applying the LR model to the third result;
• generating a first support vector machine (SVM) value indicating whether the first sentence is a question by applying a SVM to a first plurality of confidence levels, the first plurality of confidence levels including the first confidence level, the second confidence level, and the third confidence level;
• training and updating, using at least one processor, the SVM according to the first result, the second result, the third result, and the first SVM value such that an updated SVM is generated;
• and
• after training and updating the SVM, processing a second sentence by performing steps, comprising:
• applying ITN to the second sentence of the call transcript such that a fourth result is generated;
• generating a fourth confidence level indicating whether the fourth result represents a question using the FST;
• applying TFIDF to the second sentence such that a fifth result is generated;
• generating a fifth confidence level indicating whether the second sentence is a question by applying the LR model to the fifth result;
• generating a sixth result by applying the model of sentence embedding to the second sentence;
• generating a sixth confidence level indicating whether the second sentence is a question by applying the LR model to the sixth result;
• and
• generating a second SVM value indicating whether the second sentence is a question by applying the updated SVM to a second plurality of confidence levels, the second plurality of confidence levels including the fourth confidence level, the fifth confidence level, and the sixth confidence level.

A computer program product comprising computer-readable program code capable of being executed by one or more processors when retrieved from a non-transitory computer-readable medium, the program code comprising computer-readable instructions configurable to cause classifying one or more sentences to be a question or not a question using a call transcript of an audio or video call by performing steps, comprising:

• performing automatic speech recognition on the audio or video call such that the call transcript of the audio of video call is generated, the call transcript including a first sentence and a second sentence;
• applying inverse text normalization (ITN) to the first sentence such that a first result is generated;
• generating a first confidence level indicating whether the first result represents a question using a finite state transducer (FST);
• applying term frequency inverse document frequency (TFIDF) to the first sentence such that a second result is generated;
• generating a second confidence level indicating whether the first sentence is a question by applying a logistic regression (LR) model to the second result;
• generating a third result by applying a model of sentence embedding to the first sentence;
• generating a third confidence level indicating whether the first sentence is a question by applying the LR model to the third result;
• generating a first support vector machine (SVM) value indicating whether the first sentence is a question by applying a SVM to a first plurality of confidence levels, the first plurality of confidence levels including the first confidence level, the second confidence level, and the third confidence level;
• training and updating, using at least one processor, the SVM according to the first result, the second result, the third result, and the first SVM value such that an updated SVM is generated;
• and
• after training and updating the SVM, processing a second sentence by performing steps, comprising:
• applying ITN to the second sentence of the call transcript such that a fourth result is generated;
• generating a fourth confidence level indicating whether the fourth result represents a question using the FST;
• applying TFIDF to the second sentence such that a fifth result is generated;
• generating a fifth confidence level indicating whether the second sentence is a question by applying the LR model to the fifth result;
• generating a sixth result by applying a model for sentence embedding to the second sentence such that a sixth result;
• generating a sixth confidence level indicating whether the second sentence is a question by applying the LR model to the sixth result;
• and
• generating a second SVM value indicating whether the second sentence is a question by applying the updated SVM to a second plurality of confidence levels, the second plurality of confidence levels including the fourth confidence level, the fifth confidence level, and the sixth confidence level.