Bayesian Topic Model for Semantic Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing techniques for automatic classification of written and verbal communications require expert design or large amounts of labeled training data, making them costly and often unsuitable for auditing fairness or correctness.
Innovation Solution
A computer-implemented method using a structured Bayesian belief network to model call transcript data, allowing for the generation of statistical topic models that can transform unstructured text into interpretable topic representations without the need for labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional classification techniques are used, then classification accuracy can be improved, but the cost of labeled training data and expert design increases significantly
Solution Approach 1:
The patent applies preliminary action by pre-seeding the topic model with keyword groups before training. This allows the model to start with prior knowledge about relevant topics, reducing the need for extensive labeled training data while maintaining classification accuracy. The keyword groups are prepared in advance and integrated into the model structure before the actual training process begins.
Solution Approach 2:
The patent uses topic models as an intermediary between raw text data and classification labels. Instead of directly training classifiers on labeled data, the system first generates topic representations from unlabeled data, then uses these topics as intermediate features for classification. This intermediary step reduces dependency on labeled data while preserving classification accuracy.
2Ease of manufacture
If conventional admixture models are used, then semantic structure can be imposed without labeled data, but human intervention is required to interpret topics
Solution Approach 1:
The patent applies preliminary action by pre-seeding the topic model with keyword groups that provide semantic meaning to topics. Instead of generating unnamed topics that require human interpretation, the model is prepared in advance with meaningful keyword associations, making the output directly interpretable without requiring subsequent human intervention to label or understand the topics.
3Ease of operation
If expert-designed classification models are used, then interpretability for auditing can be improved, but the complexity and cost of model curation increases
Solution Approach 1:
The patent applies self-service by enabling the topic model to automatically generate interpretable topic representations without requiring expert curation. The model uses pre-seeded keyword groups and unlabeled data to self-organize meaningful topics that are inherently interpretable, eliminating the need for expert-designed classification schemas while maintaining interpretability for auditing purposes.
Data Source
AI summary
A computer-implemented method executed using a computing device comprises digitally generating and storing a machine learning statistical topic model in computer memory, the topic model being programmed to model call transcript data representing words spoken on a call as a function of one or more topics of a set of topics, the set of topics being modeled to comprise a set of pre-seeded topics and a set of non-pre-seeded topics, and the one or more topics being modeled as a function of a probability distribution of topics; programmatically pre-seeding the topic model with a set of keyword groups, each keyword group associating a respective set of keywords with a topic of the set of pre-seeded topics; programmatically training the topic model using unlabeled training data; conjoining a classifier to the topic model to create a classifier model, the classifier defining a joint probability distribution over topic vectors and one or more observed labels; programmatically training the classifier model using labeled training data; receiving target call transcript data comprising an electronic digital representation of a verbal transcription of a target call; programmatically determining, using the classifier model, at least one of one or more topics of the target call or one or more classifications of the target call; digitally storing the target call transcript data with additional data indicating the determined one or more topics of the target call and/or the determined one or more classifications of the target call; accessing, in computer storage, a first digitally stored electronic document comprising a first text; receiving computer input specifying a search query comprising one or more search terms; processing the search query using the classifier model to output a query topic vector representing a thematic content of the search query; processing the first text using the classifier model to output and store in the computer memory a first plurality of topic vectors each representing a topic in the text; using the query topic vector and the first plurality of topic vectors, calculating a plurality of similarity values, each of the similarity values representing a similarity of the query topic vector to a particular topic vector among the first plurality of topic vectors; outputting a visual display that specifies one or more topic vectors among the first plurality of topic vectors having one or more corresponding similarity values that are greater than a specified threshold similarity value.