Question Detection Model for Call Transcripts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing (NLP) algorithms face challenges in accurately identifying questions in call transcripts, particularly in spoken conversations, due to their reliance on written text and inability to handle contextual errors, slang, and less structured sentences.
Innovation Solution
The use of multiple independent machine learning models, including inverse text normalization (ITN), term frequency inverse document frequency (TFIDF), sentence embeddings (such as BERT), and a support vector machine (SVM), to process and analyze sentences in call transcripts, generating confidence levels to determine if a sentence is a question.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing NLP algorithms are used to process call transcripts, then processing speed is maintained, but detection accuracy deteriorates due to inability to handle contextual errors, slang, and unstructured sentences
Solution Approach 1:
The patent combines multiple independent machine learning models (ITN, TFIDF, sentence embeddings, SVM) into an ensemble system. Each model processes the call transcript independently and their results are aggregated, allowing the system to leverage the strengths of each individual model while compensating for their individual weaknesses, thereby improving both accuracy and adaptability to spoken language variations
Solution Approach 2:
The system uses a composite approach by integrating different types of machine learning models with distinct computational characteristics. The ITN model handles normalization, TFIDF captures term importance, sentence embeddings provide semantic understanding, and SVM performs classification. This composite model architecture enables the system to handle diverse linguistic patterns in spoken conversations more effectively than single-model approaches
2Measurement precision
If multiple independent machine learning models are used to process sentences, then detection accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the question detection task into multiple independent sub-tasks, each handled by a specialized model. The ITN model segments the normalization task, TFIDF segments the feature extraction task, sentence embeddings segment the semantic understanding task, and SVM segments the classification task. This segmentation allows each component to be optimized independently while maintaining overall system manageability
Solution Approach 2:
The system introduces intermediary processing layers between the input transcript and final classification. The ITN model acts as an intermediary for text normalization, sentence embeddings serve as intermediaries for semantic representation, and the ensemble of models acts as intermediaries that transform raw text into structured features before SVM classification. These intermediaries simplify the overall processing by breaking down the complex transformation into manageable stages
Data Source
AI summary
Disclosed are some implementations of systems, apparatus, methods and computer program products for categorizing a sentence as a question. Rather than using a single model, several different models are leveraged to determine whether a sentence is a question. For example, the models can include an inverse text normalization (ITN) model, a sentence embeddings model, and a Term frequency inverse document frequency (TFIDF) model. The output of an ITN model is processed using a finite state transducer (FST) while the output of the sentence embeddings model and TFIDF model are processed using logistics regression (LR) models. A support vector machine (SVM) is then applied to the output of the FST and LR models to determine whether the sentence is a question.