[Paper Review] Enriching Complex Networks with Word Embeddings for Detecting Mild Cognitive Impairment from Speech Transcripts

Leandro Borges dos Santos, Edilson A. Corrêa+4 more|arXiv (Cornell University)|Jan 1, 2017

Topic Modeling29 references9 citations

TL;DR

This paper proposes a novel method, Complex Networks Enriched with Word Embeddings (CNE), to detect Mild Cognitive Impairment (MCI) from speech transcripts by modeling transcripts as complex networks and integrating word embeddings for improved semantic representation. CNE outperformed traditional Bag-of-Words and linguistic feature approaches on two datasets (DementiaBank and Cinderella), demonstrating higher accuracy and promise for large-scale MCI screening.

ABSTRACT

Mild Cognitive Impairment (MCI) is a mental disorder difficult to diagnose.Linguistic features, mainly from parsers, have been used to detect MCI, but this is not suitable for large-scale assessments.MCI disfluencies produce nongrammatical speech that requires manual or high precision automatic correction of transcripts.In this paper, we modeled transcripts into complex networks and enriched them with word embedding (CNE) to better represent short texts produced in neuropsychological assessments.The network measurements were applied with well-known classifiers to automatically identify MCI in transcripts, in a binary classification task.A comparison was made with the performance of traditional approaches using Bag of Words (BoW) and linguistic features for three datasets: DementiaBank in English, and Cinderella and Arizona-Battery in Portuguese.Overall, CNE provided higher accuracy than using only complex networks, while Support Vector Machine was superior to other classifiers.CNE provided the highest accuracies for Dementia-Bank and Cinderella, but BoW was more efficient for the Arizona-Battery dataset probably owing to its short narratives.The approach using linguistic features yielded higher accuracy if the transcriptions of the Cinderella dataset were manually revised.Taken together, the results indicate that complex networks enriched with embedding is promising for detecting MCI in large-scale assessments.1 talkbank.org/DementiaBank/

Motivation & Objective

• To address the challenge of large-scale, automated detection of Mild Cognitive Impairment (MCI) from speech transcripts.
• To overcome limitations of traditional linguistic parsing and Bag-of-Words (BoW) models in capturing semantic nuances in disfluent, non-grammatical speech.
• To improve representation of short, impaired speech transcripts by integrating word embeddings into complex network structures.
• To evaluate the performance of CNE against established methods (BoW and linguistic features) across multiple multilingual datasets.

Proposed method

• Transcripts were modeled as complex networks by treating words as nodes and word co-occurrences as edges.
• Word embeddings (e.g., Word2Vec or similar) were used to enrich network nodes with dense semantic representations.
• Network measurements (e.g., degree centrality, clustering coefficient, betweenness) were extracted as feature vectors.
• These features were fed into standard classifiers, with Support Vector Machine (SVM) showing superior performance.
• The approach was evaluated on three datasets: DementiaBank (English), Cinderella (Portuguese), and Arizona-Battery (Portuguese).
• Performance was compared against BoW and linguistic feature baselines, with manual revision applied to the Cinderella dataset for linguistic features.

Experimental results

Research questions

• RQ1 Can enriching complex networks with word embeddings improve the detection of Mild Cognitive Impairment in speech transcripts compared to traditional methods?
• RQ2 How does the CNE approach perform relative to Bag-of-Words and linguistic feature-based models across different linguistic and clinical datasets?
• RQ3 Does manual correction of transcripts enhance the performance of linguistic feature-based models in MCI detection?
• RQ4 Is the CNE approach robust across diverse datasets with varying transcript lengths and linguistic characteristics?

Key findings

• CNE achieved the highest accuracy on the DementiaBank (English) and Cinderella (Portuguese) datasets compared to BoW and linguistic features.
• For the Arizona-Battery dataset, which contains shorter narratives, BoW outperformed CNE, suggesting context-dependent effectiveness.
• Support Vector Machine (SVM) was the most effective classifier among those tested, outperforming other models in all datasets.
• Manual revision of transcripts significantly improved the accuracy of linguistic feature-based models on the Cinderella dataset.
• The integration of word embeddings into complex networks enhanced semantic representation, particularly for disfluent and non-grammatical speech typical in MCI.
• Overall, CNE demonstrates strong potential for scalable, automated MCI detection in clinical and large-scale settings.