Semantic Characteristics of Schizophrenic Speech

Kfir Bar, Vered Zilberstein, Ido Ziv, Heli Baram, Nachum Dershowitz, Samuel Itzikowitz, Eiran Vadim HarelNAACL 2019

TL;DR

This paper explores using computational linguistics to automatically detect semantic disturbances in the Hebrew speech of schizophrenia patients. By analyzing topic shifts (derailment) and the semantic similarity of modifiers, they show that patients exhibit more disorganized speech than controls, leading to a classifier that can detect schizophrenia with 81.5% accuracy.

What Problem Does This Paper Address?

Diagnosing schizophrenia often relies on subjective assessments of thought disorders, which manifest as disorganized speech. While previous research has explored linguistic markers of these disorders, particularly in English, there's a need for more objective, computational methods, especially for languages like Hebrew, which has complex morphology. This paper aims to bridge this gap by analyzing semantic characteristics of transcribed Hebrew speech from schizophrenia patients and controls to build an automated detection system.

Methods

The researchers employed two main experimental approaches:

  • Measuring Derailment: They calculated a "derailment score" for each speech response by measuring the semantic similarity between adjacent words. They used word embeddings (specifically, fastText pretrained vectors for Hebrew) to represent words as vectors and cosine similarity to quantify their semantic relatedness. Crucially, they found that focusing only on content words (nouns, verbs, adjectives, adverbs) provided a more robust measure of derailment, distinguishing patients from controls better than using all words.
  • Measuring Incoherence (Modifier Usage): They investigated how patients and controls use adjectives and adverbs (modifiers) to describe nouns and verbs. Using a dependency parser (YAP) for Modern Hebrew, they identified noun-adjective and verb-adverb pairs. They then compared the modifiers used by participants to those found in external Hebrew corpora. The hypothesis is that controls will use modifiers that are semantically closer to typical modifiers for a given noun or verb, indicating more coherent speech.

Datasets

  • Interview Transcripts: The primary dataset consists of transcribed interviews from 51 men (aged 19-63), divided into two groups:
  • Patients: 24 inpatients at Beer Yaakov Mental Health Center diagnosed with schizophrenia.
  • Controls: 27 individuals recruited via social media advertisement.
  • Approximately two-thirds of the patients were in schizophrenia remission at the time of the interview.
  • External Hebrew Corpora: For the modifier analysis, several external Hebrew corpora were used to establish typical modifier usage. These included medical websites ("Doctors," "Infomed," "To Be Healthy") and news archives ("HaAretz").

Key Insights and Ideas

  • Content Words Matter Most for Derailment: Focusing on semantic similarity of content words (nouns, verbs, adjectives, adverbs) rather than all words significantly improves the ability to detect derailment, suggesting that functional words and fillers can obscure the underlying semantic disorganization.
  • Semantic Similarity of Modifiers as a Coherence Measure: The way patients use adjectives and adverbs to describe nouns and verbs differs from controls. Controls tend to use modifiers that are more semantically aligned with typical usage, suggesting a more coherent descriptive style.
  • Hebrew Morphology Presents Unique Challenges: The highly inflected nature of Hebrew and the absence of written vowels require specialized tools (like YAP) and careful handling of word forms (using lemmas for modifier analysis) to accurately capture semantic relationships.
  • Speech is a Richer Data Source than Social Media: The authors argue that speech from psychiatric interviews is a more reliable data source for detecting thought disorders than social media posts, likely due to the controlled environment and direct interaction.

Key Results

  • Derailment: When using only content words, patients showed significantly lower derailment scores than controls, indicating more topic shifts. This finding was consistent across different window sizes (k) for calculating similarity.
  • Modifier Usage: Controls generally had significantly higher scores for both adjective and adverb similarity, suggesting they use modifiers more in line with typical usage compared to patients.
  • Classification: A classifier trained on these semantic features (derailment scores and modifier coherence scores) achieved 81.5% accuracy in distinguishing between patients and controls. The most important features for classification were the two derailment scores.

Limitations and Future Work

  • Sample Size: The authors acknowledge that the number of participants used for training the classifiers might be considered relatively small.
  • Intellectual Differences: The study did not measure participants' IQ, and some differences in intellect could potentially influence the results.
  • Education Levels: There was a notable difference in educational background between the groups, with a higher proportion of controls having post-high-school education.
  • Future Work: The authors plan to collect more speech samples, explore additional semantic and grammatical textual characteristics, and investigate the role of modifiers further. They also aim to address limitations like the influence of intellect and education in future studies.

Summary updated

BibTeX
@inproceedings{bar2019semantic,
  author = {Kfir Bar and Vered Zilberstein and Ido Ziv and Heli Baram and Nachum Dershowitz and Samuel Itzikowitz and Eiran Vadim Harel},
  title = {{Semantic Characteristics of Schizophrenic Speech}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure