Topic Modeling Using Speaker-Specific Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current topic modeling techniques, such as the latent Dirichlet allocation (LDA) model, face challenges in accurately identifying topics in verbal communications due to equal treatment of words from different speakers, regional and personal variations in language, and the assumption of uniform topic distribution, leading to misidentification and inefficiencies in modeling common versus rare topics.
Innovation Solution
An improved topic modeling approach that segregates words spoken by different parties and uses an inferred prior distribution to model the probability of topics and words, incorporating the Pitman-Yor process for more accurate language modeling, and dynamically builds models based on user input to account for regional and personal variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If the LDA model treats all words equally regardless of speaker, then the model is simpler to implement, but topic modeling accuracy deteriorates due to skewing from agent or caller language patterns
Solution Approach 1:
The patent segments the corpus into separate speaker-specific corpora (agent corpus and caller corpus) and trains separate language models for each speaker type. This segmentation allows the system to capture distinct language patterns of different speakers while maintaining model simplicity through modular architecture.
Solution Approach 2:
The patent applies local quality by creating speaker-specific language models that are tailored to each speaker type's unique language patterns. Instead of using a single global model, the system adapts the language model to local characteristics of each speaker group, improving accuracy without significantly increasing complexity.
2Device complexity
If the LDA model assumes uniform topic distribution, then the model is computationally simpler, but it fails to accurately represent the likelihood of different topics occurring in real communications
Solution Approach 1:
The patent changes the parameter of topic distribution from uniform to non-uniform by introducing topic probabilities that reflect real-world occurrence patterns. The system learns topic distributions from data that capture the actual likelihood of different topics, making the model more realistic while managing complexity through efficient probability estimation.
3Quantity of substance
If the LDA model uses a single language model for all speakers, then the model is more compact, but it cannot account for regional and personal variations in speech patterns
Solution Approach 1:
The patent segments the language modeling into speaker-specific models rather than using a single monolithic model. This segmentation enables the system to handle regional and personal language variations by capturing speaker-specific patterns while keeping each individual model relatively compact.
Solution Approach 2:
The patent creates a universal framework that can handle multiple speaker types and language variations through a consistent segmentation approach. The same segmentation methodology applies to different speaker groups, making the system versatile in handling various language patterns while maintaining structural consistency.
Data Source
AI summary
Systems and methods for improving call topic models are described herein. In an embodiment a server computer receives call transcript data comprising an electronic digital representation of a verbal transcription of a call between a first person of a first person type and a second person of a second person type. The server computer splits the call transcript data into first person type data comprising words spoken by the first person in the call and second person type data comprising words spoken by the second person type in the call. The server computer uses a stored topic model to determine a topic of the call, the topic model simultaneously modeling the first person type data as a function of a first probability distribution of words used by the first person type for one or more topics and the second person type data as a function of a second probability distribution of words used by the second person type for the one or more topics, both the first probability distribution of words and the second probability distribution of words being modeled as a function of a third probability distribution of words for the one or more topics. The server computer then stores a data record identifying the topic of the call and/or stores data identifying the topic of the call with the call transcripts.