Semantic Coherence Dataset: Speech transcripts.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Speech Analysis

Background:

• Probabilistic measures like perplexity are increasingly used to analyze spoken language.
• Perplexity has shown promise in discriminating between healthy controls and individuals with Alzheimer Disease.
• Existing datasets may not fully support the nuanced investigation of semantic coherence in speech.

Purpose of the Study:

• To introduce a novel dataset for experimenting with semantic coherence metrics in spoken language.
• To evaluate the stability and utility of perplexity as a measure of semantic coherence.
• To facilitate research into both intra-subject and inter-subject semantic coherence.

Main Methods:

• Development of the Semantic Coherence Dataset comprising speech transcripts.
• Utilizing language models to compute perplexity scores on speech data.
• Analyzing perplexity scores for intra-subject coherence (single speaker, different speech types) and inter-subject coherence (multiple speakers).

Main Results:

• The Semantic Coherence Dataset provides extensive data (over 40 hours) for linguistic analysis.
• Perplexity demonstrates potential as a stable metric for assessing semantic coherence across different speech contexts.
• The dataset enables the study of how speaker-specific models perform on varied transcripts.

Conclusions:

• The Semantic Coherence Dataset is a valuable resource for advancing research in semantic coherence and language model evaluation.
• Perplexity shows promise as a robust tool for analyzing spoken language features and clinical applications.
• Further research can leverage this dataset to explore advanced linguistic features and develop new analytical measures.