Neural Network Entity Extraction from Conversation Transcripts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches to labeling text sequences struggle with accuracy when the label space is large, such as in clinical conversations where many symptoms can have multiple statuses, leading to data sparsity and poor performance with limited training data.
Innovation Solution
The system employs a neural network-based approach that first identifies relevant portions of the input sequence and their context, then generates entity names and statuses in parallel, rather than directly predicting entity-name-status pairs, allowing for accurate predictions even with large label spaces and limited training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional tagging models directly identify which text tokens should be assigned which labels, then the approach is simple to implement, but accuracy deteriorates when the label space is large
Solution Approach 1:
The patent segments the labeling task into two independent sub-tasks: (1) identifying which text tokens reference entities, and (2) determining the status of each referenced entity. This segmentation transforms a single large-label-space prediction problem into two smaller prediction problems, improving accuracy while maintaining implementation simplicity through modular neural network components.
2Device complexity
If conventional models make one prediction over the entire cross-product space, then the model structure is simple, but data sparsity increases when training data is limited
Solution Approach 1:
The model segments the prediction process into two stages: first predicting entity references, then predicting entity statuses. This reduces data sparsity by distributing training examples across two smaller label spaces rather than one large cross-product space, improving reliability when training data is limited while keeping the model structure manageable through sequential processing.
Solution Approach 2:
The patent transforms the prediction problem from a two-dimensional cross-product space (entity × status) into two separate one-dimensional prediction spaces. This dimensional transformation reduces the complexity of the label space from |entities|×|statuses| to |entities|+|statuses|, mitigating data sparsity issues with limited training data.
3Adaptability or versatility
If the label space is large (cross-product of multiple individual label spaces), then more comprehensive labeling is achieved, but conventional approaches show poor accuracy
Solution Approach 1:
The patent maintains comprehensive labeling coverage by independently predicting both entity references and entity statuses, then combining them. This segmentation approach achieves the same comprehensive label space coverage as conventional methods but improves accuracy by training on two smaller, less sparse label spaces rather than one large cross-product space.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for extracting entities from conversation transcript data. One of the methods includes obtaining a conversation transcript sequence, processing the conversation transcript sequence using a span detection neural network configured to generate a set of text token spans; and for each text token span: processing a span representation using an entity name neural network to generate an entity name probability distribution over a set of entity names, each probability in the entity name probability distribution representing a likelihood that a corresponding entity name is a name of the entity referenced by the text token span; and processing the span representation using an entity status neural network to generate an entity status probability distribution over a set of entity statuses.