Dialogue Logicality Profiling for Explainable Alzheimer's Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models for Alzheimer's Disease detection focus on local low-level features within utterances, limiting the ability to generate global high-level representations of dialogue sessions, which biases predictive models and reduces explainability.
Innovation Solution
A method using a large language model to generate a linguistic deficit profile, followed by an embedding model to create vectors, and a pre-trained language model to concatenate these vectors for determining logicality in dialogue sentences, enhancing the accuracy of downstream models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If local low-level features (pauses, punctuation) are used for feature extraction, then the processing is simple and fast, but the modeling ability for patient-level detection is constrained and explainability is reduced
Solution Approach 1:
The patent segments the dialogue session into multiple utterances and further into linguistic units (words, phrases, sentences). This hierarchical segmentation allows the system to analyze both local low-level features (pauses, punctuation within utterances) and global high-level features (narrative structure, logical coherence across the entire session), resolving the contradiction between processing simplicity and detection accuracy
Solution Approach 2:
The patent introduces a new dimension of analysis by generating linguistic deficit profiles that capture global session-level narrative patterns beyond traditional local features. This adds a temporal and structural dimension to the feature space, enabling the model to consider both immediate utterance characteristics and overall dialogue coherence, thereby improving detection accuracy while maintaining processing efficiency
2Ease of manufacture
If only local low-level features within utterances are analyzed, then the feature extraction is straightforward, but the ability to generate global high-level representations of dialogue sessions is limited
Solution Approach 1:
The patent implements a nested structure where linguistic units (words, phrases) are nested within utterances, which are nested within the overall dialogue session. This nested architecture allows the system to maintain simple local feature extraction at each level while progressively building up to complex global representations, achieving both ease of implementation and comprehensive analysis capability
Solution Approach 2:
The patent introduces linguistic deficit profiles as an intermediary representation that bridges local utterance features and global session-level analysis. These profiles serve as a mediator that aggregates local features into meaningful global patterns, enabling the system to generate comprehensive high-level representations without overwhelming computational complexity
3Object-affected harmful factors
If transcript-based detection methods are used, then less sensitive data is involved compared to speech, but the investigation of linguistic deficits is limited without session-level understanding
Solution Approach 1:
The patent performs preliminary processing of speech transcripts into structured linguistic deficit profiles before downstream analysis. This preliminary action captures session-level narrative patterns and logical coherence in advance, ensuring that no linguistic deficit information is lost while working with text-based data that is less sensitive than raw speech
Data Source
AI summary
A method for determining logicality of dialogue sentences is provided. This method is performed by a processor, and includes the following steps: executing a large language model to generate a linguistic deficit profile according to a dialogue text and a prompt text, executing an embedding model to generate a first vector according to the linguistic deficit profile, executing a pre-trained language model to generate a plurality of second vectors according to the dialogue text, executing the pre-trained language model to concatenate the first vector with each second vector, and executing the pre-trained language model to generate a logicality determination result according to each second vector concatenated with the first vector.