Automating Intended Target Identification for Paraphasias in Discourse Using a Large Language Model.

PURPOSE: To date, there are no automated tools for the identification and fine-grained classification of paraphasias within discourse, the production of which is the hallmark characteristic of most people with aphasia (PWA). In this work, we fine-tune a large language model (LLM) to automatically predict paraphasia targets in Cinderella story retellings. METHOD: Data consisted of 332 Cinderella story retellings containing 2,489 paraphasias from PWA, for which research assistants identified their intended targets. We supplemented these training data with 256 sessions from control participants, to which we added 2,415 synthetic paraphasias. We conducted four experiments using different training data configurations to fine-tune the LLM to automatically "fill in the blank" of the paraphasia with a predicted target, given the context of the rest of the story retelling. We tested the experiments' predictions against our human-identified targets and stratified our results by ambiguity of the targets and clinical factors. RESULTS: The model trained on controls and PWA achieved 50.7% accuracy at exactly matching the human-identified target. Fine-tuning on PWA data, with or without controls, led to comparable performance. The model performed better on targets with less human ambiguity and on paraphasias from participants with fluent or less severe aphasia. CONCLUSIONS: We were able to automatically identify the intended target of paraphasias in discourse using just the surrounding language about half of the time. These findings take us a step closer to automatic aphasic discourse analysis. In future work, we will incorporate phonological information from the paraphasia to further improve predictive utility. SUPPLEMENTAL MATERIAL: https://doi.org/10.23641/asha.24463543.

Consistency and reliability of automated language measures across expressive language samples in autism.

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder with substantial clinical heterogeneity, especially in language and communication ability. There is a need for validated language outcome measures that show sensitivity to true change for this population. We used Natural Language Processing to analyze expressive language transcripts of 64 highly-verbal children and young adults (age: 6-23 years, mean 12.8 years; 78.1% male) with ASD to examine the validity across language sampling context and test-retest reliability of six previously validated Automated Language Measures (ALMs), including Mean Length of Utterance in Morphemes, Number of Distinct Word Roots, C-units per minute, unintelligible proportion, um rate, and repetition proportion. Three expressive language samples were collected at baseline and again 4 weeks later. These samples comprised interview tasks from the Autism Diagnostic Observation Schedule (ADOS-2) Modules 3 and 4, a conversation task, and a narration task. The influence of language sampling context on each ALM was estimated using either generalized linear mixed-effects models or generalized linear models, adjusted for age, sex, and IQ. The 4 weeks test-retest reliability was evaluated using Lin's Concordance Correlation Coefficient (CCC). The three different sampling contexts were associated with significantly (P < 0.001) different distributions for each ALM. With one exception (repetition proportion), ALMs also showed good test-retest reliability (median CCC: 0.73-0.88) when measured within the same context. Taken in conjunction with our previous work establishing their construct validity, this study demonstrates further critical psychometric properties of ALMs and their promising potential as language outcome measures for ASD research.

ParAlg: A Paraphasia Algorithm for Multinomial Classification of Picture Naming Errors.

PURPOSE: A preliminary version of a paraphasia classification algorithm (henceforth called ParAlg) has previously been shown to be a viable method for coding picture naming errors. The purpose of this study is to present an updated version of ParAlg, which uses multinomial classification, and comprehensively evaluate its performance when using two different forms of transcribed input. METHOD: A subset of 11,999 archival responses produced on the Philadelphia Naming Test were classified into six cardinal paraphasia types using ParAlg under two transcription configurations: (a) using phonemic transcriptions for responses exclusively (phonemic-only) and (b) using phonemic transcriptions for nonlexical responses and orthographic transcriptions for lexical responses (orthographic-lexical). Agreement was quantified by comparing ParAlg-generated paraphasia codes between configurations and relative to human-annotated codes using four metrics (positive predictive value, sensitivity, specificity, and F1 score). An item-level qualitative analysis of misclassifications under the best performing configuration was also completed to identify the source and nature of coding discrepancies. RESULTS: Agreement between ParAlg-generated and human-annotated codes was high, although the orthographic-lexical configuration outperformed phonemic-only (weighted-average F1 scores of .78 and .87, respectively). A qualitative analysis of the orthographic-lexical configuration revealed a mix of human- and ParAlg-related misclassifications, the former of which were related primarily to phonological similarity judgments whereas the latter were due to semantic similarity assignment. CONCLUSIONS: ParAlg is an accurate and efficient alternative to manual scoring of paraphasias, particularly when lexical responses are orthographically transcribed. With further development, it has the potential to be a useful software application for anomia assessment. SUPPLEMENTAL MATERIAL: https://doi.org/10.23641/asha.22087763.

Refining Semantic Similarity of Paraphasias Using a Contextual Language Model.

PURPOSE: ParAlg (Paraphasia Algorithms) is a software that automatically categorizes a person with aphasia's naming error (paraphasia) in relation to its intended target on a picture-naming test. These classifications (based on lexicality as well as semantic, phonological, and morphological similarity to the target) are important for characterizing an individual's word-finding deficits or anomia. In this study, we applied a modern language model called BERT (Bidirectional Encoder Representations from Transformers) as a semantic classifier and evaluated its performance against ParAlg's original word2vec model. METHOD: We used a set of 11,999 paraphasias produced during the Philadelphia Naming Test. We trained ParAlg with word2vec or BERT and compared their performance to humans. Finally, we evaluated BERT's performance in terms of word-sense selection and conducted an item-level discrepancy analysis to identify which aspects of semantic similarity are most challenging to classify. RESULTS: Compared with word2vec, BERT qualitatively reduced word-sense issues and quantitatively reduced semantic classification errors by almost half. A large percentage of errors were attributable to semantic ambiguity. Of the possible semantic similarity subtypes, responses that were associated with or category coordinates of the intended target were most likely to be misclassified by both models and humans alike. CONCLUSIONS: BERT outperforms word2vec as a semantic classifier, partially due to its superior handling of polysemy. This work is an important step for further establishing ParAlg as an accurate assessment tool.

"Um" and "Uh" Usage Patterns in Children with Autism: Associations with Measures of Structural and Pragmatic Language Ability.

Pragmatic language difficulties, including unusual filler usage, are common among children with Autism Spectrum Disorder (ASD). This study investigated "um" and "uh" usage in children with ASD and typically developing (TD) controls. We analyzed transcribed Autism Diagnostic Observation Schedule (ADOS) sessions for 182 children (117 ASD, 65 TD), aged 4 to 15. Although the groups did not differ in "uh" usage, the ASD group used fewer "ums" than the TD group. This held true after controlling for age, sex, and IQ. Within ASD, social affect and pragmatic language scores did not predict filler usage; however, structural language scores predicted "um" usage. Lower "um" rates among children with ASD may reflect problems with planning or production rather than pragmatic language.

Combining voice and language features improves automated autism detection.

Variability in expressive and receptive language, difficulty with pragmatic language, and prosodic difficulties are all features of autism spectrum disorder (ASD). Quantifying language and voice characteristics is an important step for measuring outcomes for autistic people, yet clinical measurement is cumbersome and costly. Using natural language processing (NLP) methods and a harmonic model of speech, we analyzed language transcripts and audio recordings to automatically classify individuals as ASD or non-ASD. One-hundred fifty-eight participants (88 ASD, 70 non-ASD) ages 7 to 17 were evaluated with the autism diagnostic observation schedule (ADOS-2), module 3. The ADOS-2 was transcribed following modified SALT guidelines. Seven automated language measures (ALMs) and 10 automated voice measures (AVMs) for each participant were generated from the transcripts and audio of one ADOS-2 task. The measures were analyzed using support vector machine (SVM; a binary classifier) and receiver operating characteristic (ROC). The AVM model resulted in an ROC area under the curve (AUC) of 0.7800, the ALM model an AUC of 0.8748, and the combined model a significantly improved AUC of 0.9205. The ALM model better detected ASD participants who were younger and had lower language skills and shorter activity time. ASD participants detected by the AVM model had better language profiles than those detected by the language model. In combination, automated measurement of language and voice characteristics successfully differentiated children with and without autism. This methodology could help design robust outcome measures for future research. LAY SUMMARY: People with autism often struggle with communication differences which traditional clinical measures and language tests cannot fully capture. Using language transcripts and audio recordings from 158 children ages 7 to 17, we showed that automated, objective language and voice measurements successfully predict the child's diagnosis. This methodology could help design improved outcome measures for research.

A Pseudo-Value Approach to Analyze the Semantic Similarity of the Speech of Children With and Without Autism Spectrum Disorder.

Conversational impairments are well known among people with autism spectrum disorder (ASD), but their measurement requires time-consuming manual annotation of language samples. Natural language processing (NLP) has shown promise in identifying semantic difficulties when compared to clinician-annotated reference transcripts. Our goal was to develop a novel measure of lexico-semantic similarity - based on recent work in natural language processing (NLP) and recent applications of pseudo-value analysis - which could be applied to transcripts of children's conversational language, without recourse to some ground-truth reference document. We hypothesized that: (a) semantic coherence, as measured by this method, would discriminate between children with and without ASD and (b) more variability would be found in the group with ASD. We used data from 70 4- to 8-year-old males with ASD (N = 38) or typically developing (TD; N = 32) enrolled in a language study. Participants were administered a battery of standardized diagnostic tests, including the Autism Diagnostic Observation Schedule (ADOS). ADOS was recorded and transcribed, and we analyzed children's language output during the conversation/interview ADOS tasks. Transcripts were converted to vectors via a word2vec model trained on the Google News Corpus. Pairwise similarity across all subjects and a sample grand mean were calculated. Using a leave-one-out algorithm, a pseudo-value, detailed below, representing each subject's contribution to the grand mean was generated. Means of pseudo-values were compared between the two groups. Analyses were co-varied for nonverbal IQ, mean length of utterance, and number of distinct word roots (NDR). Statistically significant differences were observed in means of pseudo-values between TD and ASD groups (p = 0.007). TD subjects had higher pseudo-value scores suggesting that similarity scores of TD subjects were more similar to the overall group mean. Variance of pseudo-values was greater in the ASD group. Nonverbal IQ, mean length of utterance, or NDR did not account for between group differences. The findings suggest that our pseudo-value-based method can be effectively used to identify specific semantic difficulties that characterize children with ASD without requiring a reference transcript.

Evaluating atypical language in autism using automated language measures.

Measurement of language atypicalities in Autism Spectrum Disorder (ASD) is cumbersome and costly. Better language outcome measures are needed. Using language transcripts, we generated Automated Language Measures (ALMs) and tested their validity. 169 participants (96 ASD, 28 TD, 45 ADHD) ages 7 to 17 were evaluated with the Autism Diagnostic Observation Schedule. Transcripts of one task were analyzed to generate seven ALMs: mean length of utterance in morphemes, number of different word roots (NDWR), um proportion, content maze proportion, unintelligible proportion, c-units per minute, and repetition proportion. With the exception of repetition proportion (p [Formula: see text]), nonparametric ANOVAs showed significant group differences (p[Formula: see text]). The TD and ADHD groups did not differ from each other in post-hoc analyses. With the exception of NDWR, the ASD group showed significantly (p[Formula: see text]) lower scores than both comparison groups. The ALMs were correlated with standardized clinical and language evaluations of ASD. In age- and IQ-adjusted logistic regression analyses, four ALMs significantly predicted ASD status with satisfactory accuracy (67.9-75.5%). When ALMs were combined together, accuracy improved to 82.4%. These ALMs offer a promising approach for generating novel outcome measures.

Epidemiological surveys of ASD: advances and remaining challenges.

Recent worldwide epidemiological surveys of autism conducted in 37 countries are reviewed; the median prevalence of autism is .97% in 26 high-income countries. Methodological advances and remaining challenges in designing and executing surveys are discussed, including the effects on prevalence of variable case definitions and nosography, of reliance on parental reports only, case ascertainment through mainstream school surveys, innovative approaches to screen school samples more efficiently, and consideration of age in interpreting surveys. Directions for the future of autism epidemiology are discussed, including the need to systematically examine cross-cultural variation in phenotypic expression and developing surveillance programs.