Deep Wide Neural Network Sentiment Prediction for Long Text Threads
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sentiment analysis techniques are ineffective for analyzing longer text sequences, such as customer support tickets, which often lack obvious affect words and contain ambiguous or neutral sentiment, leading to unreliable sentiment classification and skewed results.
Innovation Solution
The implementation of a deep and wide neural network-based model using transfer learning, which incorporates both deep and wide features, including raw text sequences and domain features, to determine case-level and thread-level sentiment scores, improving sentiment prediction accuracy through machine-learned models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing sentiment analysis techniques are applied to longer text sequences, then the analysis can be performed on more diverse text types, but the reliability of sentiment classification deteriorates due to lack of obvious affect words and ambiguous sentiment
Solution Approach 1:
The patent segments long text sequences into multiple shorter segments or sentences, analyzes sentiment for each segment individually, and then aggregates the results. This allows the model to handle longer texts by breaking them into manageable units that can be processed effectively, maintaining reliability while increasing adaptability to various text lengths
Solution Approach 2:
The patent introduces attention mechanisms that add a new dimension to the analysis by weighting different parts of the text differently. Instead of treating all text equally, the model learns to focus on sentiment-relevant portions while downweighting neutral or ambiguous sections, thereby maintaining classification reliability across diverse text lengths
2Measurement precision
If traditional sentiment analysis models are used, then the model complexity remains low, but the sentiment prediction accuracy deteriorates for long and unbalanced text threads
Solution Approach 1:
The patent employs deep neural networks with multiple layers and parameters that automatically learn from data. By increasing model capacity through additional parameters and layers, the system achieves higher sentiment prediction accuracy for complex text patterns while using regularization techniques to prevent overfitting
Solution Approach 2:
The patent introduces embedding layers as intermediaries that transform raw text into dense vector representations. These embeddings serve as a bridge between raw input and classification, capturing semantic relationships and enabling more accurate sentiment prediction without requiring excessive model complexity
3Measurement precision
If domain-specific features are incorporated into the model, then the sentiment analysis becomes more accurate for customer support contexts, but the model complexity increases
Solution Approach 1:
The patent merges domain-specific features (such as customer support metadata, ticket categories, and agent information) with the neural network model. By integrating these features into the overall architecture, the model captures domain-specific patterns while maintaining a unified structure that manages complexity through systematic feature integration
Data Source
AI summary
Embodiments of the disclosed technologies use machine learning to produce thread level classification data and case level classification data.