Call Transcript Classification Using Embeddings and LLM Prompts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for classifying fraudulent calls in telecommunications networks are defeatable by telephone number spoofing and constructing a comprehensive database of suspicious numbers is difficult, requiring significant computational resources and storage.
Innovation Solution
A method using a large-language model (LLM) is employed to classify calls by computing an embedding vector of a call transcript, searching for similar examples in a database, and prompting the LLM with a limited number of relevant examples to maintain accuracy while reducing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a database of suspicious telephone numbers is constructed to classify fraudulent calls, then classification accuracy is improved, but device complexity and storage requirements increase significantly
Solution Approach 1:
The patent replaces the traditional mechanical database lookup system with an AI-based semantic analysis system. Instead of matching calls against a static database of suspicious numbers, the system uses natural language processing to analyze call transcripts and determine fraud likelihood, thereby reducing database complexity while maintaining or improving classification accuracy
Solution Approach 2:
The patent changes the classification parameters from static telephone number matching to dynamic transcript-based semantic analysis. By analyzing the content and context of call transcripts rather than relying on pre-defined number lists, the system adapts to new fraud patterns without requiring database updates, reducing both storage and complexity
2Measurement precision
If thousands of training examples are used to finetune a machine-learning model for call classification, then classification accuracy is improved, but processing requirements and storage needs increase
Solution Approach 1:
The patent applies partial action by using a pre-trained large language model that has already learned general language patterns, requiring only minimal domain-specific adaptation rather than extensive training from scratch. This reduces processing energy while maintaining the ability to accurately classify calls through the model's inherent language understanding capabilities
Solution Approach 2:
The patent utilizes a large language model that has been pre-trained on vast amounts of general text data before being applied to call classification. This preliminary training phase, which occurs beforehand, enables the model to perform classification tasks with minimal additional processing, reducing real-time energy consumption while preserving high accuracy
3Measurement precision
If multiple embedding vectors are stored in the database for each example text, then classification accuracy is improved, but storage requirements increase
Solution Approach 1:
The patent extracts only the essential semantic representation (single embedding vector) from each example text rather than storing multiple variations or redundant representations. This extraction of core semantic meaning maintains classification accuracy by capturing the essential fraud indicators while minimizing storage requirements through efficient vector representation
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
A computer-implemented method for classifying a call in a telecommunications network, comprising: accessing a transcript of the call, computing an embedding vector of the transcript; searching in a database of embedding vectors, each representing an example text associated with a known classification, for a first embedding vector with a defined degree of similarity to the embedding vector of the transcript; and constructing a first prompt comprising: an instruction directed to a large-language model to classify the call using a first example; the transcript; and the first example, the first example being: an example text represented by the first embedding vector and the associated known classification. The method further comprises prompting the large-language model with the first prompt; receiving a response to the first prompt from the large-language model, comprising a classification of the call; and, in response to the classification meeting a criterion, initiating an action at the telecommunications network.