Sentiment Analysis for Multi-Turn Dialogue Understanding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing models for chat dialogue modeling are inadequate in multi-turn conversation scenarios, suffering from passive behavior and incoherent responses, and rely on explicit human annotations and knowledge graphs that are difficult to obtain for real-world scenarios.
Innovation Solution
The method involves extracting internal knowledge from dialogues to understand fine-grained sentiment information, adapting aspect-based sentiment analysis to conversational scenarios, and using a processor-based system to extract sentiment expressions, generate polarity values, and determine target mentions based on sentence embeddings and contextual representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If knowledge graphs are used to enhance dialogue understanding, then the quality of dialogue responses improves, but the complexity of obtaining and processing knowledge graphs increases
Solution Approach 1:
The patent extracts sentiment information and entity mentions directly from dialogue text without requiring external knowledge graphs. The system identifies sentiment expressions, determines their polarity, and links them to target entities through processing only the dialogue content itself, eliminating the need for complex KG construction and entity linking
Solution Approach 2:
The system performs self-service by automatically extracting structured sentiment knowledge from unstructured dialogue text using neural network models. The sentiment analysis module independently identifies sentiment expressions, determines polarity, and associates them with entities without requiring external human annotations or pre-built knowledge bases
2Measurement precision
If human annotations are used to create benchmark datasets, then the accuracy of sentiment analysis improves, but the time and cost of data preparation increases
Solution Approach 1:
The system automatically extracts sentiment information, polarity, and entity mentions from dialogue text using trained neural network models without requiring manual human annotation. The sentiment analysis module processes dialogue data independently, identifying sentiment expressions and their targets through automated NLP techniques
Solution Approach 2:
The patent uses pre-trained neural network models for sentiment analysis that have been trained on general sentiment data. These models can be applied directly to dialogue datasets without requiring re-annotation, allowing the system to leverage preliminary training to achieve accurate sentiment extraction in domain-specific conversations
3Measurement precision
If entity linking is performed as a pre-requisite step, then the accuracy of knowledge graph utilization improves, but the overall processing time and potential error propagation increases
Solution Approach 1:
The patent removes the entity linking step entirely from the processing pipeline. Instead of linking entities to knowledge graphs, the system directly extracts entity mentions from dialogue text and associates them with sentiment expressions through neural network-based mention extraction, eliminating the bottleneck of entity linking
Solution Approach 2:
The system segments the sentiment analysis task into independent components: sentiment expression identification, polarity determination, and target entity extraction. Each component is processed separately through dedicated neural network modules, allowing parallel processing and eliminating the sequential dependency of entity linking
Data Source
AI summary
There is included a method and apparatus for sentiment analysis for multi-turn conversations comprising computer code for obtaining input dialogues; extracting sentiment expressions based on sentence embeddings corresponding to the input dialogues; generating polarity values based on the sentence embeddings corresponding to the input dialogues; and determining a target mention associated with at least one of the sentiment expressions based on the sentiment expressions and the sentence embeddings, wherein the determining of the target mention includes generating rich contextual representations based on the sentence embeddings and the sentiment expressions; and determining the target mention based on calculated boundaries, wherein the calculated boundaries are generated using the rich contextual representations.