Skip to main content

Thank you for visiting You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

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 \(= .07\)), nonparametric ANOVAs showed significant group differences (p\(< 0.01\)). 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\(< 0.01\)) 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.


Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by impairments in communication and social interaction, and restricted and repetitive patterns of behavior1. Language differences are prevalent among children with ASD, with most showing some expressive and receptive or pragmatic language impairments2,3,4,5. Pragmatic, semantic, and syntactic language usage predicts social skills in children with ASD6 and pragmatic language impairments have been linked to an increased prevalence of anxiety disorders among children with ASD7. Conversational language ability is a significant predictor of job independence and friendship in adults with ASD, both of which are metrics for measuring quality of life8. The presence of communication difficulties is a hallmark of autism, and as such are important targets for evaluation and intervention.

Several standardized measures are commonly used for clinical assessment of language in autism, such as the Clinical Evaluation of Language Fundamentals (CELF) and the Peabody Picture Vocabulary Test (PPVT)9,10. While such clinical measures offer direct assessment of language ability, they require specially trained professionals for administration, are time- and money-intensive, result in inaccurate performance due to child stress or anxiety, and their limited scope provides only partial examination of linguistic ability11. In addition, standardized language testing generates speech samples in highly constrained contexts and lacks ecological validity. One alternative is the parent-reported Children’s Communication Checklist-2 (CCC–2), which is used to assess communication behaviors12 and has been shown to accurately identify pragmatic language difficulties in children with ASD13. However, more objective language measures are called for, as the fallibility of parent-reported measures is well documented14,15.

In contrast to conventional language assessments, measures of expressive language can be generated from natural language sampling, which more accurately reflects an individual’s true verbal communication ability in a real-world context. Barokova and Tager-Flusberg (2020) and others have called for the development of measures which address linguistic heterogeneity among individuals with ASD, specifically citing the utility of natural language samples for their ability to capture differences across a wide range of age and language levels. This methodology uses a spoken language sample to evaluate features like talkativeness, vocabulary, syntax, utterance planning, and articulation16,17,18. Many studies have investigated the validity of such measures to discriminate by age17,19 and clinical group20,21,22,23, and have shown convergent validity with standardized language tests16. Expressive language is already well established as a clinical marker of language ability24,25,26.

Many measures of expressive language have been evaluated for their ability to distinguish between ASD and a typically developing (TD) control group. However, because studies have generally not included a non-ASD clinical control condition we cannot know if results of ASD and TD comparisons are specific to ASD. For example, Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder that has a high co-occurrence with autism, with both groups showing difficulties with language and social communication27,28. Children with ADHD show communication and pragmatic differences similar to children with ASD, although with lesser magnitude29,30. As ADHD is part of the differential diagnosis of ASD and shows overlap in language domains, it represents a useful comparison group to test how language features in ASD differ from other neurodevelopmental disorders, increasing the ability to examine the specificity of language measures.

Natural Language Processing (NLP) is a broad field at the intersection of computer science, linguistics, and artificial intelligence that aims to analyze speech and language automatically through the use of computational methods. NLP has been used widely in studies examining language use in neurodevelopmental disorders, including autism. Applications range from distributional semantic models used to identify unexpected words in narrative retellings by children with autism31 to word alignment models that identify pragmatically inappropriate language in children with ASD compared with their language-matched peers32. Other applications include vector-space techniques to detect semantic differences in language of children with ASD and typically developing controls33,34. Natural Language Processing is useful for identifying and accurately quantifying characteristics of language in ASD that would be labor-intensive or infeasible to do by hand. As such, there is great potential for using computational methods to produce novel language measures in ASD research; many measures of expressive language can be calculated automatically with appropriate software.

This study was designed to address a gap in the research: namely the power of automatically calculated measures of expressive language to discriminate between ASD and two non-ASD control groups, and the correlation of these measures with common standardized tests. This is part of a larger research project examining language in already-diagnosed children with ASD. Our long-term objectives are to develop quantitative automated tools for outcome measures of language features associated with ASD, though we do not intend the measures presented here to be immediately applicable to screening or diagnosis. This study uses transcripts of a common instrument of autism diagnosis, the Autism Diagnostic Observation Schedule (ADOS) as an expressive language sample. We examined seven measures of expressive language, which were automatically computed for one ADOS activity using NLP methods. Hereafter these are called Automated Language Measures (ALMs) for their ability to be automatically calculated. Our specific goals were to: 1. examine language differences (measured by ALMs) between ASD and two non-ASD control groups (ADHD and TD); 2. analyze the convergent validity of these measures with standardized language measures; 3. investigate the discriminant validity of individual ALMs in classifying ASD status; and 4. examine if gains in discriminant validity could be obtained by combining all ALMs together.



Participants aged 7 to 17 years with either ASD, ADHD, or TD were recruited for an fMRI study by community outreach and referrals from Oregon Health & Science University’s specialty clinics. Data collection occurred from 2012 to 2018. Potential participants came in for a screening visit to determine if they qualified for the study. During this initial visit, informed written consent or assent was obtained from all participants and their parents. Parents also completed a Developmental and Medical History survey. All children in the ASD and ADHD groups were directly assessed by experienced child psychiatrists and clinical psychologists who confirmed their diagnosis based on DSM-IV-TR criteria35. The research diagnostic team reviewed results of the standardized diagnostic assessments (both videos and scored protocols) and used best estimate procedures. ASD was ruled out in the TD and ADHD groups based on the ADOS, clinical interview, and parent-completed autism questionnaires (see below). Exclusion criteria for all groups included the presence of seizure disorder, cerebral palsy, pediatric stroke, history of chemotherapy, sensorimotor handicaps, closed head injury, thyroid disorder, schizophrenia, bipolar disorder, current major depressive episode, fetal alcohol syndrome, Tourette’s disorder, severe vision impairments, Rett’s syndrome, current use of psychoactive medications, non-English speaker, or an IQ below 70.

Of 289 screened participants, 104 were ultimately excluded from this study for not meeting strict diagnostic criteria and another 10 for failing to complete the initial assessment procedures, leaving a sample of 175 subjects (102 ASD, 45 ADHD, 28 TD) included in the main neuroimaging study. Of these, we excluded a further six subjects with ASD due to an untranscribable ADOS session, either for poor sound quality or non-compliance, leaving a total sample of 169 children for this analysis. Sample characteristics for the participants are given in Table 1.


Autism diagnosis and severity

The ADOS36 is a semi-structured, standardized assessment in which a trained examiner engages participants in activities that are designed to elicit social and communication behaviors indicative of symptoms of ASD as defined in the DSM-IV-TR35. In this study, all participants were administered Module 3 of the ADOS-2, designed for children and adolescents with fluent speech. Module 3 comprises 14 tasks that are generally administered in sequence although the tester has some flexibility to change the task order if clinically indicated. All ADOS interviews were administered by research assistants or a senior clinical psychologist trained to research reliability level. All administrations were videotaped and later transcribed. The Social Affect (SA) score (10 items; range 0-20) and the Restricted and Repetitive Behavior (RRB) score (4 items; range 0-8) were used in these analyses37. Higher scores indicate more severe ASD symptoms.

Other tests administered were the Autism Diagnostic Interview–Revised (ADI-R), a semi-structured, standardized interview designed to examine three major developmental domains (language and communication, reciprocal social interaction, and restricted, repetitive, and stereotyped behaviors and interests)38, and the Social Responsiveness Scale (SRS), a parent-completed measure of autistic symptomatology and associated social impairment suitable for 4–18 year olds39. Only caregivers of the ASD group were interviewed with the ADI-R; interviews were administered by trained interviewers. Data were reviewed by the diagnostic team and integrated in the best estimate clinical procedures used to confirm diagnoses.

Intellectual level

Intellectual level of participants was estimated with a short form of the Wechsler Intelligence Scale for Children 4th Edition (WISC)40. Three subtests were administered: Information, Block Design, and Vocabulary, allowing a full scale IQ to be estimated from the sum of scaled scores of the three subtests according to the formula set out by Sattler and Dumont (2004)41.

Language assessment

Language characteristics and linguistic pragmatic abilities were assessed using the parent-completed Children’s Communication Checklist, second edition (CCC-2)12. The CCC-2 is a widely-used, 70-item standardized checklist of pragmatic and social communication behaviors applicable to children ages 4:0 to 16:11. Caregivers are asked to make a frequency judgment about how often behaviors occur on 4-point scale. The CCC-2 is divided into 10 subscales measuring: (A) speech, (B) syntax, (C) semantics, (D) coherence, (E) inappropriate initiation, (F) stereotyped language, (G) the use of context, (H) non-verbal communication, (I) social relationships, and (J) interests. The first four subscales (A–D) evaluate articulation and phonology, language structure, vocabulary, and discourse; four other subscales (E-H) evaluate pragmatic aspects of communication as well as stereotyped language with atypical or unusual expressions and use of nonverbal communication like facial expressions, bodily movements, and gestures. The last two subscales (I and J) measure behaviors characteristic of children with ASD. Each subscale raw score is converted to age-standardized scores (mean = 10; SD = 3). A General Communication Composite (GCC) is derived by summing scores A to H (mean = 100; SD = 15). A Structural Language scale score is derived by averaging scores A to D, and a Pragmatic Language scale score is obtained by averaging scores E to H. Lower scores are indicative of more problems.

Table 1 Sample characteristics.



All ADOS administrations were audio and video recorded. The audio was transcribed according to modified SALT guidelines (Systematic Analysis of Language Transcripts)42 by a team of trained research assistants who were blind to the participants’ diagnostic status and intellectual abilities. The ADOS activities Make-Believe and Joint Interactive Play, Description of a Picture, Telling a Story From a Book, Cartoons, Conversation and Reporting, Emotions Conversation, Social Difficulties and Annoyance Conversation, Friends Relationships and Marriage Conversation, and Loneliness Conversation were transcribed. Speech was split into communication units, or c-units, consisting of a main clause and any subordinate, modifying clauses, or of speech fragments that constitute a whole utterance such as responses to questions. Special attention was paid to notation of disfluencies, or mazes, and any unintelligible speech was marked. Any disagreements between transcribers were resolved through discussion with a clinician and a consensus judgment. Transcribers participated in biannual consistency checks to review protocol and ensure continued standardization. Lab transcription guidelines are available upon request from the first author.

The ADOS is a recommended source of natural language for measuring expressive language ability and has been used in previous studies26,43,44. We analyzed the transcript of one ADOS task: Friends, Relationships, and Marriage. The focus of this conversation is the participant’s understanding of the nature of personal relationships, on why someone would want to engage in such relationships, and what the participant’s role might be in those relationships. This conversational activity is administered in the second half of the ADOS, after participants have “warmed up” to the testing situation. In addition, this task consistently yielded the most utterances of all ADOS tasks, providing sufficient speech data for analysis. The mean number of participant utterances for the activity was 77.37 (73.1 ASD, 71.4 TD, 90.3 ADHD). The audio had a mean length of 6.27 minutes (6.47 ASD, 5.43 TD, 6.36 ADHD), a satisfactory length for language sampling analysis45. This activity also has high standardization of conversational questions, leading to good comparability between participants. The examiner uses pre-established interview questions that are open-ended and designed to facilitate the flow of conversation. Follow-up probes are used at the examiner’s discretion to maintain that flow.

Automated language measures (ALMs)

We chose to examine seven expressive language measures that have been explored in previous studies on neurodevelopmental disorders. Mean Length of Utterance in Morphemes (MLUM) and Number of Distinct Word Roots (NDWR) were calculated on all complete, fluent, and intelligible c-units following Gorman et al. (2015). Um proportion, a measure of uh and um usage, was calculated as the total number of ums over the total number of um + uh. Content maze proportion, a measure of disfluency, was calculated as the number of content mazes over the number of content mazes + the number of fillers. Following MacFarlane et al. (2017) we use the term content maze to refer to disfluencies which contain content words (as opposed to the fillers ah, uh, um, mm, hmm, like, well, you know, I mean). An example is below:

  • Content maze (My mom) My dad picked me up.

  • Filler maze I love (uh) pancakes.

All mazes and fillers were marked during the transcription process, not as a post-hoc coding. Unintelligible proportion was calculated as the number of partially or fully unintelligible c-units over the total number of c-units. C-units per minute (CPM) was calculated as the number of attempted c-units per minute. Repetition proportion is a measure of a child’s repetition of examiner speech, and is calculated as the number of child words that are repeated in a set of two or more from the examiner’s immediately previous turn, divided by the total number of child words. An example is below:

  • Examiner And the moon was coming up.

  • Child The moon was coming up then.

The child repeats four words from the examiner (“the moon was coming up”) out of five total words, so the repetition proportion is 4/5. A set of repeated words is counted as a repetition if two or more words are exactly identical to the examiner’s speech, or if there is an approximate match between the words. Following van Santen et al. (2013), in the case of an approximate match the following were allowed: you, I, me, we, us could be substituted with each other; him/her and he/she could be substituted with each other; the, a, an, is, are, am, ’m, ’s, is, and ’re could be deleted.

These seven ALMs and their calculation methods are summarized in Table 2.

Table 2 ALM calculation.

MLUM and NDWR were calculated using software written by Gorman et al. (2015). Um proportion, unintelligible proportion, and c-units per minute were all calculated using software written by the authors. Content maze proportion was calculated using software adapted from MacFarlane et al. (2017)23. Repetition proportion was calculated using software from van Santen et al. (2013)21. All software was written in Python 2.746.

Statistical analyses

Our first aim was to explore differences between the three diagnostic groups for the seven ALMs. Because um proportion, unintelligible proportion, and repetition proportion were not normally distributed, standard ANOVA assumptions were violated; we therefore compared groups using nonparametric Kruskal-Wallis one-way ANOVAs. Effect sizes were calculated with eta-squared (\(\eta ^{2}\)). Significant results (\(\hbox {p} < 0.05\)) were followed up with post-hoc contrasts using the Games-Howell test.

For the second aim of examining the convergent validity of these ALMs, we calculated the Spearman’s rank correlations between each ALM and key language-related clinical scores from the ADOS and the CCC-2.

Our third aim was to establish the discriminant validity of the ALMs by evaluating their ability to classify ASD status. As post-hoc contrasts indicated that the TD and ADHD groups were not significantly different from each other, these were combined into a non-ASD control group for further analyses. The ALMs had heterogeneous distributions with some approximating normal distributions (MLUM, NDWR, CPM, content maze proportion) and others with extreme skewness (um proportion, unintelligible proportion, repetition proportion). In order to create a common scale, the seven ALMs were recoded as ordinal variables using the observed distribution on the whole sample. Values falling between the 1st and 50th centiles were scored 1, those between the 50th and the 75th centiles were scored 2, values between the 75th and 90th centile were scored 3, and a score of 4 was ascribed to values above the 90th centile. A slight adjustment to this recoding had to be made for um proportion, for which more than 10% of participants had a value of 0. Accordingly, the score of 4 was shifted to the 86th centile to include all zeros; the other score centiles were left unchanged. Scores were reversed for MLUM, NDWR, um proportion, and CPM to account for the inverse nature of those ALMs: lower values on these measures is associated with higher impairment. After recoding, higher impairment is shown by increasing values (from 1 to 4) for all ALMs. Seven binary logistic regression models using ASD status (\(0 = \hbox {non}\)-ASD; \(1 = \hbox {ASD}\)) as a dependent variable and each recoded ALM as independent variables were calculated. Models were estimated with and without adjustment on full scale IQ and chronological age. Adjusted models showed consistent superiority in terms of overall significance and proportion correctly classified; accordingly, all models were subsequently adjusted on IQ and age. Their predictive value was evaluated with the Wald statistic. For each model, we also report the -2 log-likelihood, the Nagelkerke pseudo-R square, and the accuracy, alongside sensitivity and specificity. We first estimated a baseline model using only IQ and age as independent variables; models containing each single ALM were subsequently compared to this baseline model to gauge the increment in predictive performance attributable to each ALM.

To address our fourth aim, whether combining the ALMs improves discriminant validity, we estimated a final logistic regression model using all seven ALMs in addition to IQ and age as independent variables. We report the same statistics for this model as for the individual logistic regression models. Finally, we calculated Receiver Operator Characteristic (ROC) curves of the baseline model, the seven individual ALM models, and the combined model with all seven ALMs. For each model, we used the prediction probabilities produced from the logistic regression predictive model. We plotted these nine ROC curves in a graph and report the corresponding area under the curve (AUC).

Following Perneger (1998) and Rothman (2014), we did not use Bonferroni’s adjustment for multiple tests. Throughout, a p-value of \(< 0.05\) was retained as a level of statistical significance. All analyses were performed using R statistical computing software47.

Ethical approval

This study was approved by the Oregon Health & Science University Institutional Review Board, Number 00000531, and all research was performed in accordance with their relevant guidelines and regulations.


Aim 1: examine language differences between groups

The means and standard deviations of the ALMs for each diagnostic group are shown in Table 3. The relative frequency distribution for each ALM is shown by ASD status in Fig. 1. Nonparametric Kruskal-Wallis one-way ANOVAs showed significant group differences at \(p < .01\) for all ALMs except repetition proportion (\(p = 0.078\)). The largest effect sizes were found for content maze proportion and c-units per minute. MLUM, NDWR, um proportion, and unintelligible proportion each had moderate effect sizes. Repetition proportion had a small effect size. In post-hoc tests, we found significant group differences between the TD and ASD group for all ALMs except NDWR (\(p = 0.528\)). Significant group differences were found between the ADHD and ASD group for all ALMs. No significant differences were found between the TD and ADHD groups for any of these seven ALMs. Thus, for subsequent analyses we combined the TD and ADHD groups into a non-ASD control group.

Table 3 Diagnostic group differences for ALMs.

The direction of differences between diagnostic groups fall into two sets. The ASD group had significantly higher means of content maze proportion, unintelligible proportion, and repetition proportion (where higher values are more characteristic of ASD) and significantly lower means of MLUM, NDWR, um proportion, and CPM (where lower values are more characteristic of ASD). This pattern is consistent with the expectation for increased language atypicalities in the ASD group.

We retained the ALM repetition proportion in later analyses despite it falling short of statistical significance in the Kruskal-Wallis test for two reasons: 1. repetitive speech has a high theoretical significance in language characterization in ASD and 2. it was possible that it would become a significant predictor once IQ, age, and other ALM scores were covaried.

Figure 1
figure 1

Relative frequency distribution of each ALM by ASD status.

Aim 2: convergent validity of ALMs

In order to evaluate the convergent validity of these ALMs, we evaluated how well they correlated with standardized clinical measures. We calculated Spearman rank correlation coefficients between each of the seven ALMs and the child’s CCC-2 and ADOS scores. From the CCC-2 we used the General Communication Composite (GCC), Structural score, and Pragmatic score. From the ADOS, we used the total Social Affect (SA) score and total Restricted and Repetitive Behavior (RRB) score. These results are summarized in Table 4.

Table 4 Relationship of ALMs to clinical scores using Spearman correlations.

With the exception of repetition proportion, all ALMs were significantly correlated with all three scores from the CCC-2, with absolute values of correlation coefficients ranging from .23 to .47. The direction of correlations followed the same groupings seen previously: content maze proportion, unintelligible proportion, and repetition proportion were negatively correlated with the CCC-2 scores. MLUM, NDWR, um proportion, and c-units per minute were all positively correlated with the CCC-2 scores. This was the expected direction of correlation, since lower scores on the CCC-2 indicate language difficulties which are associated with ASD. Notably, content maze proportion had the strongest association with the three CCC-2 scores (absolute value range .44 to .47). Conversely, repetition proportion had small and insignificant correlations with CCC-2 scores.

All ALMs were significantly correlated with the ADOS total SA score, with absolute values of correlation coefficients ranging from .20 to .46. Five ALMs were significantly correlated with the ADOS total RRB score, with absolute values of correlation coefficients ranging from .22 to .46, while NDWR and repetition proportion were not significantly correlated with the RRB score. Overall, RRB score correlations were of smaller magnitude than those with the SA score. The direction of correlations followed the same groupings seen previously. This was expected since higher ADOS SA and RRB scores are associated with an ASD diagnosis.

Aim 3: discriminant validity of ALMs

We then examined the discriminant validity of these seven ALMs to determine their ability to distinguish diagnostic groups by using logistic regression models. We excluded 10 participants from this analysis (7 ASD, 3 non-ASD): six participants with undefined um proportion (for whom um and uh did not occur in the Friends and Marriage Conversation), and four participants who did not have WISC IQ results. These 10 participants were excluded from all models to ensure comparability, leaving a sample of 89 ASD and 70 non-ASD. Results of the logistic regression models are summarized in Table 5.

Model 0 is the baseline model, using IQ and age as independent variables. Models 1-7 are logistic regression models using a single recoded ordinal ALM as well as IQ and age as independent variables.

Table 5 Logistic regression models for ALMs adjusted by age and IQ.

The baseline model 0 classification accuracy was 62.89%. As shown by the Wald test statistic, this effect was largely driven by IQ. Age did not significantly contribute to the model; however, we adjusted all subsequent models on both IQ and age as they are developmentally relevant metrics. Models 1-7 had classification accuracy values ranging from 63.52% for NDWR to 75.47% for content maze proportion. The model for content maze proportion had the highest accuracy out of the seven models by over five percentage points.

Specificity (the true negative rate) was 71.91% at the baseline model, reflecting the strong effect of IQ. However, sensitivity (the true positive rate) was only 51.43%. The models for MLUM and NDWR were not significant (Wald test: \(p >.45\)) and resulted in very modest improvements in sensitivity while having lower specificity compared to the baseline model. By contrast, the models for um proportion, content maze proportion, unintelligible proportion and c-units per minute each had significant Wald test statistics (\(p < .01\)). Improvements in sensitivity were attained in each of these four models; however, only content maze proportion resulted in gains in specificity as well. The ALM repetition proportion fell just short of significance (Wald test: \(p = .0545\)). Remarkably, this model had one of the highest levels of specificity (75.28%) of all ALM models though it remained poor for sensitivity (58.57%).

Aim 4: combined power of ALMs to increase discrimination between groups

Lastly, we estimated a logistic regression model using all seven recoded ordinal ALMs as well as IQ and age as independent variables and ASD status as the dependent variable (see model 8 in Table 5). This model had the highest classification accuracy of all models, 82.39%. It also had the highest specificity and sensitivity of 82.86% and 82.02%, respectively. Only three ALMs reached significance in the Wald test statistic: content maze proportion, unintelligible proportion, and c-units per minute. This indicates that each of these three ALMs uniquely improved the prediction over and beyond the effect of the other six ALMs and the covariates of age and IQ. The ALMs which did not significantly contribute to their respective individual models—MLUM, NDWR, and repetition proportion—also did not significantly contribute to the combined model. Additionally, um proportion lost the significance it had in its own model when combined with the other ALMs.

We estimated ROC curves for all models in Table 5. In these calculations, we used the class prediction probabilities estimated from the logistic regression models as the predictors in the ROC curves. The resulting nine ROC curves are shown in Fig. 2, alongside their AUC values. The combined model (model 8) had the highest AUC of 0.9223, whereas the baseline model (model 0) had the lowest AUC of 0.6868.

Figure 2
figure 2

AUC is area under the curve. Baseline is modeled with only IQ and age as independent variables. All other models are adjusted on IQ and age.

ROC curve for logistic regression models, evaluated on class probabilities.


We calculated seven automated language measures from the transcripts of one task of the ADOS. Six measures showed significant differences between ASD and a neurodiverse control group. Consistent with the high prevalence of language atypicalities in autism, the ASD group had significantly lower means of MLUM, um proportion, and c-units per minute than the TD and ADHD groups, and significantly lower means for NDWR than the ADHD group. The ASD group had significantly higher means of content maze proportion, unintelligible proportion, and repetition proportion than the TD and ADHD groups. There were no statistical differences between the TD and ADHD groups for any ALM. Furthermore, ALMs that discriminated between ASD and TD also discriminated between ASD and ADHD, suggesting that these ALMs capture features of language specific to ASD, and not simply those common to all neurodevelopmental disorders. Content maze proportion and c-units per minute were the most discriminant measures as shown by larger effect sizes in Kruskal-Wallis ANOVAs while MLUM, which measures syntactic complexity, and NDWR, which measures vocabulary size, differed only modestly across groups. Like the other ALMs, repetition proportion showed means for the ASD group that were indicative of more atypicality than for the other two groups, but differences fell just short of statistical significance. One plausible explanation for this result is the reduced statistical power due to the low base rate of child-examiner repetitions in our experiment. It is possible that analyses of longer transcripts (such as several or all ADOS tasks) or of more naturalistic speech samples would improve the sensitivity of repetition proportion to ASD anomalies.

The convergent validity of the ALMs was implied by small but significant Spearman rank correlations between the ALM scores and the clinical measures CCC-2 and ADOS, confirming that these ALMs are capturing language differences commonly seen in clinical groups. All ALMs except repetition proportion were significantly correlated with the General Communication Composite (GCC), the Structural score, and the Pragmatic score of the CCC-2. All ALMs were significantly correlated with the ADOS Social Affect (SA) score, and all except NDWR and repetition proportion were correlated with the ADOS Restricted and Repetitive Behavior (RRB) score. The clinical measure most correlated with the ALMs was the ADOS SA score (absolute value range .26 to .46; median .43). The smaller magnitude of RRB score correlations (absolute value range .05 to .46; median .23) is consistent with the fact that the SA score is measuring socio-communicative behaviors whereas the RRB score is more driven by non-language features of autism such as sensory-motor mannerisms or cognitive-behavioral inflexibility. Overall, the magnitudes of Spearman correlations were modest; none of the absolute values were above 0.5, which is a common rule of thumb for moderate correlation. It is notable, however, that clinical scores were provided by highly trained professionals spending 45-60 minutes with a participant or by parents who have observed their child over time and across contexts. In that respect, it is remarkable that ALMs computed on only six minutes of language correlated with those conventional scores.

In our investigation of discriminant validity, four ALMs made significant contributions to their individual logistic regression models as independent variables and achieved higher accuracy and sensitivity than a baseline model. Repetition proportion followed the same pattern, but did not quite reach significance. All ALMs increased the sensitivity, or true positive rate, of their model, resulting in more participants with ASD being correctly classified as having ASD. Content maze proportion and repetition proportion were the only ALMs that resulted in higher specificity than the baseline model 0. Those two ALMs were thus better than IQ and age alone at correctly identifying participants without ASD, since the proportion of true negatives, or participants correctly classified as non-ASD, improved. Overall, content maze proportion achieved the best sensitivity and specificity, with an overall accuracy of over 75%. Thus, um proportion, content maze proportion, unintelligible proportion, and c-units per minute each showed discriminant validity by being able to distinguish children with ASD from typically developing children and children with ADHD. An anonymous reviewer asked if the group differences we report could be attributed to the proportionally larger number of females in the TD group as compared to the ASD and ADHD groups. To address this issue, we estimated new versions of these logistic regression models with child sex as a covariate. We found that while sex was a significant covariate in most of the models, it did not change the significance of the ALMs in any of the models.

Since some ALMs made improvements to specificity while others improved only sensitivity, we investigated the discriminant validity of using all the ALMs together to differentiate children with and without ASD. When the ALMs were combined in one logistic regression model (model 8), overall accuracy improved substantially (82.4%), with high levels of specificity (82.9%) and sensitivity (82.0%) as well. Content maze proportion, unintelligible proportion, and c-units per minute significantly improved prediction of ASD status over and beyond the effect of other ALMs and of covariates. Of the four ALMs that were contributory in their own models, one (um proportion) lost statistical significance when combined with the other ALMs. This very likely reflects the co-linearity between um proportion and content maze proportion in model 8 (Spearman: -.39) due to the fact that both ALMs are conceptually similar measures of disfluency.

NDWR and MLUM remained non-contributory in model 8, and may not be relevant to ASD, however these ALMs could still be useful in a less verbal sample. Repetition proportion did not reach statistical significance (\(p = .075\)); as explained above, this may simply reflect its low sensitivity when measuring an infrequent language characteristic. Further analyses will determine if other and longer language samples would improve its performance; likewise, this ALM might be more relevant when used with samples of younger or more developmentally delayed participants whose language may be more repetitious. Thus, performance of this ALM remains to be tested in different samples.

The AUC values for the logistic regression prediction probabilities followed a similar pattern as accuracy, as shown in Table 5. Model 8 had the highest AUC of 0.9223, followed by model 4 (content maze proportion) which had an AUC of 0.8149. Model 0 had the lowest AUC of 0.6868. Notably, all AUC values are higher than their corresponding accuracy values. One plausible contributing factor is that the accuracy represents the logistic regression results for a cut-off value of 0.5. However, that is not necessarily the optimal cut-off for the prediction probabilities. The ROC curve instead calculates the true positive rate and false positive rate across many thresholds, and can thus capture results not only at 0.5, but at the optimal cut-off and others in between. These AUC values alongside the other classification results again demonstrate the discriminant validity of several of the ALMs. Based on the findings of our study sample, content maze proportion, unintelligible proportion, and c-units per minute stand out as the best ALMs to use together to distinguish between ASD and a neurodiverse control group.

Content maze proportion, a measure of disfluency, was the highest performing ALM according to many different metrics. It significantly differentiated ASD and non-ASD in a Kruskal-Wallis ANOVA, and had the largest eta-squared effect size of 0.2430. It showed the highest correlations with measures from both the CCC-2 and the ADOS (absolute value range 0.44 to 0.47), a similarly strong correlation as found by Abbeduto et al. (2020) in a fragile X sample between their measure of disfluency (a proportion of the total number of c-units that include a maze or verbal disfluency) and the standardized measures CELF, VABS (Vineland Adaptive Behavior Scales), and GFTA (Goldman-Fristoe Test of Articulation). Of the seven individual logistic regression models, it had the best fit according to both -2 log likelihood and Nagelkerke \(R^2\), and the highest accuracy, specificity, and sensitivity. It also contributed significantly to model 8. To the authors’ knowledge, only two prior studies have examined content mazes23,48. The lack of published literature on content mazes combined with the high performance of content maze proportion in this study indicate that it should be explored further. This ALM shows particular promise as a measurable language differential between ASD and non-ASD groups.

This study has several limitations. These results are constrained by our sample, which has a large age range (7-17 years), relatively high IQ (average 99-111 across clinical groups), and unequal group sizes. While age and IQ were included as co-variates in the logistic regression models, using a more tightly controlled sample may produce more robust results. While these seven ALMs showed themselves to be useful measures of expressive language, they are still relatively exploratory measures and could be refined further. Alternative calculations of some ALMs such as finding mean length of utterance in words24,49 or including incomplete or disfluent utterances, could change the performance of MLU. Likewise, NDWR could be more robust if it were calculated as type-token ratio instead50, or if it were replaced with a count of the total number of words. Um proportion could be calculated using different weightings51. While high reliability with by-hand calculation has been established for MLUM and content maze proportion, it has not yet been verified for the other ALMs23,52. The ALMs unintelligibility proportion and repetition proportion are of limited utility in our study because a large number of participants were at the floor of these measures, making them difficult to accurately analyze compared with the other measures. The ALM um proportion is limited in scope because a small number of participants have an undefined result due to the nature of the calculation. The ordinal recoding for the logistic regression models has limitations: the recoding was done to calibrate all of the ALMs to the same scale, but the choices of cut-offs for recoding were somewhat arbitrary. Another limitation lies in manual transcription, which is time consuming and costly. However, progress in voice recognition technology should help to bypass this step soon.

There are many future directions for this work. The ADOS is used widely, making our methodology readily replicable and accessible to many labs. MacFarlane et al. (2017) found ADOS activity to be a robust predictor of disfluency use, and Abbeduto et al. (2020) showed higher correlations between expressive language measures and clinical assessments in a narration task than in a conversation task. This suggests that using a different ADOS task, such as one focused on narrating a picture, may be a more robust way to measure the utility of our ALMs; we plan to examine this possibility in future work. Additionally, while establishing the convergent and discriminant validity of the ALMs is a first step towards establishing them as language outcome measures, more work is needed. Future steps for this process would include evaluating their test-retest reliability, consistency across different expressive language samples, and responsivity to real change. Finally, ALMs should be tested as objective, quantifiable measures of atypical language in samples of younger or more language impaired participants where they might contribute to novel screening and diagnostic tools that harness new technologies.


We applied computational methodology to a clinical speech corpus and quantified expressive language through Automated Language Measures. We established moderate convergent validity and good discriminant validity for these ALMs, with content maze proportion and c-units per minute showing especially promising results in discriminating children with ASD. ALM calculations are blind to participant status unlike most clinical measures, such as the CCC-2 and the ADOS, which can be influenced by the parent or professional’s prior knowledge of the child. Our results confirm the potential of using Natural Language Processing for evaluating language samples in ASD research. Compared to conventional measures of language analysis, NLP has the strong advantage of being automated, reliable, fast, and applicable across various age groups and sampling contexts. We do not yet expect these measures to be used as an alternative method of language evaluation. Instead, we intend this work to be a proof of concept of using Automated Language Measures to learn more about language impairment in children with ASD, without depending upon extensive language testing or biased parent reports. We see great value in further development of Automated Language Measures to aid in characterizing language in ASD and evaluating outcomes.


  1. American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders: DSM-V (American Psychiatric Association, Washington, DC, 2013).

  2. Tager-Flusberg, H. & Caronna, E. Language disorders: Autism and other pervasive developmental disorders. Pediatric Clinics of North America 54, 469–481. (2007).

    Article  PubMed  Google Scholar 

  3. Loucas, T. et al. Autistic symptomatology and language ability in autism spectrum disorder and specific language impairment. J. Child Psychol. Psychiatry 49, 1184–1192. (2008).

    Article  PubMed  Google Scholar 

  4. Andrés-Roqueta, C. & Katsos, N. The contribution of grammar, vocabulary and theory of mind in pragmatic language competence in children with autistic spectrum disorders. Fronti. Psychol. 8, (2017).

  5. Wittke, K., Mastergeorge, A. M., Ozonoff, S., Rogers, S. J. & Naigles, L. R. Grammatical language impairment in autism spectrum disorder: Exploring language phenotypes beyond standardized testing. Fronti. Psychol. 8, (2017).

  6. Levinson, S., Eisenhower, A., Bush, H. H., Carter, A. S. & Blacher, J. Brief report: Predicting social skills from semantic, syntactic, and pragmatic language among young children with autism spectrum disorder. J. Autism Dev. Disord. (2020).

  7. Rodas, N. V., Eisenhower, A. & Blacher, J. Structural and Pragmatic Language in Children with ASD: Longitudinal Impact on Anxiety and Externalizing Behaviors. J. Autism Dev. Disord. 47, 3479–3488. (2017).

    Article  PubMed  Google Scholar 

  8. Friedman, L., Sterling, A., DaWalt, L. S. & Mailick, M. R. Conversational language is a predictor of vocational independence and friendships in adults with ASD. J. Autism Dev. Disord. 49, 4294–4305. (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  9. Semel, E. M., Wiig, E. H. & Secord, W. CELF 4: 4 Screening Test (Pearson, PsyhCorp., San Antonio, TX, 2004).

    Google Scholar 

  10. Dunn, L. M. & Dunn, D. M. PPVT-4: Peabody Picture Vocabulary Test (Pearson Assessments, San Antonio, 2007).

    Google Scholar 

  11. Barokova, M. & Tager-Flusberg, H. Commentary: Measuring language change through natural language samples. J. Autism Dev. Disord. 50, 2287–2306. (2020).

    Article  PubMed  Google Scholar 

  12. Bishop, D. V. M. The Children’s Communication Checklist: CCC-2 (ASHA, 2003).

  13. Joanne, Volden & Linda, Phillips. Measuring pragmatic language in speakers with autism spectrum disorders: Comparing the children’s communication checklist–2 and the test of pragmatic language. Am. J. Speech Lang. Pathol. 19, 204–212. (2010).

    Article  Google Scholar 

  14. Jeste, S. S. & Geschwind, D. H. Clinical trials for neurodevelopmental disorders: At a therapeutic frontier. Sci. Transl. Medi. 8, 321fs1. (2016).

    CAS  Article  PubMed  Google Scholar 

  15. Abbeduto, L., McDuffie, A., Thurman, A. J. & Kover, S. T. Chapter three-language development in individuals with intellectual and developmental disabilities: From phenotypes to treatments. In Hodapp, R. M. & Fidler, D. J. (eds.) International review of research in developmental disabilities, vol. 50 of Fifty Years of Research in Intellectual and Developmental Disabilities, 71–118, (Academic Press, 2016).

  16. Abbeduto, L. et al. Expressive language sampling as a source of outcome measures for treatment studies in fragile X syndrome: Feasibility, practice effects, test-retest reliability, and construct validity. Journal of Neurodevelopmental Disorders 12, 10. (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  17. Channell, M. M., Loveall, S. J., Conners, F. A., Harvey, D. J. & Abbeduto, L. Narrative language sampling in typical development: implications for clinical trials. Am. J. Speech Lang. Pathol. 27, 123–135. (2018).

    Article  PubMed  PubMed Central  Google Scholar 

  18. Kover, S. T. & Abbeduto, L. Expressive language in male adolescents with fragile X syndrome with and without comorbid autism. J. Intelle. Disabil. Res. 54, 246–265. (2010).

    CAS  Article  PubMed  Google Scholar 

  19. Tilstra, J. & McMaster, K. Productivity, fluency, and grammaticality measures from narratives: potential indicators of language proficiency? Commun. Disord. Quart. 29, 43–53. (2007).

    Article  Google Scholar 

  20. Kover, S. T., McDuffie, A., Abbeduto, L. & Ted, B. W. Effects of sampling context on spontaneous expressive language in males with fragile X syndrome or down syndrome. J. Speech Lang. Hear. Res.arch 55, 1022–1038. (2012).

    Article  PubMed  PubMed Central  Google Scholar 

  21. van Santen, J. P. H., Sproat, R. W. & Hill, A. P. Quantifying repetitive speech in autism spectrum disorders and language impairment. Autism Res. 6, (2013).

  22. Irvine, C. A., Eigsti, I.-M. & Fein, D. A. Uh, Um, and Autism: Filler disfluencies as pragmatic markers in adolescents with optimal outcomes from autism spectrum disorder. J. Autism Dev. Disord. 46, 1061–1070. (2016).

    Article  PubMed  PubMed Central  Google Scholar 

  23. MacFarlane, H. et al. Quantitative analysis of disfluency in children with autism spectrum disorder or language impairment. PLoS ONE 12, (2017).

  24. Brown, R. A First Language: The Early Stages. A First Language: The Early Stages (Harvard U. Press, Oxford, England, 1973).

  25. Condouris, K., Meyer, E. & Tager-Flusberg, H. The relationship between standardized measures of language and measures of spontaneous speech in children with autism. Am. J. Speech Lang. Pathol. 12, 349–358. (2003).

    Article  PubMed Central  Google Scholar 

  26. Helen, Tager-Flusberg. et al. Defining spoken language benchmarks and selecting measures of expressive language development for young children with autism spectrum disorders. J. Speech Lang. Hear. Res. 52, 643–652. (2009).

    Article  Google Scholar 

  27. Sokolova, E. et al. A causal and mediation analysis of the comorbidity between attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). J. Autism Dev. Disord. 47, 1595–1604. (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  28. May, T. et al. Trends in the overlap of autism spectrum disorder and attention deficit hyperactivity disorder: prevalence, clinical management, language and genetics. Curr. Dev. Disord. Reports 5, 49–57. (2018).

    Article  Google Scholar 

  29. Bruce, B., Thernlund, G. & Nettelbladt, U. ADHD and language impairment. Eur. Child Adolesc. Psychiatry 15, 52–60. (2006).

    Article  Google Scholar 

  30. Geurts, H. M. & Embrechts, M. Language profiles in ASD, SLI, and ADHD. J. Autism Dev. Disord. 38, 1931. (2008).

    Article  PubMed  Google Scholar 

  31. Rouhizadeh, M., Prud’hommeaux, E., Roark, B. & van Santen, J. Distributional semantic models for the evaluation of disordered language. Proceedings of the conference. Association for Computational Linguistics. North American Chapter. Meeting 2013, 709–714 (2013).

    PubMed  PubMed Central  Google Scholar 

  32. Prud’hommeaux, E. & Rouhizadeh, M. Automatic detection of pragmatic deficits in children with autism. The ... Workshop on Child, Computer and Interaction 2012, 1–6 (2012).

    PubMed  PubMed Central  Google Scholar 

  33. Mikolov, T., Chen, K., Corrado, G. & Dean, J. Efficient estimation of word representations in vector space. arXiv:1301.3781 [cs] (2013).

  34. Goodkind, A., Lee, M., Martin, G. E., Losh, M. & Bicknell, K. Detecting language impairments in autism: A computational analysis of semi-structured conversations with vector semantics. Proceedings of the Society for Computation in Linguistics 1, 12–22 (2018).

    Google Scholar 

  35. American Psychiatric Association. Diagnostic and statistical manual of mental disorders: DSM-IV-TR (American Psychiatric Association, Washington, DC, 2000).

  36. Lord, C. et al. The autism diagnostic observation schedule–generic: A standard measure of social and communication deficits associated with the spectrum of autism. J. Autism Dev. Disord. 30, 205–223. (2000).

    CAS  Article  PubMed  Google Scholar 

  37. Gotham, K., Pickles, A. & Lord, C. Standardizing ADOS scores for a measure of severity in autism spectrum disorders. J. Autism Dev. Disord. 39, 693–705 (2009).

    Article  Google Scholar 

  38. Rutter, M., Le Couteur, A. & Lord, C. Autism diagnostic interview-revised. Los Angeles, CA: Western Psychological Services 29, 30 (2003).

    Google Scholar 

  39. Constantino, J. N. & Gruber, C. P. Social responsive scale (SRS) manual (Western Psychological Services, Los Angeles, CA, 2005).

    Google Scholar 

  40. Wechsler, D. Wechsler intelligence scale for children–fourth edition (WISC-IV) 4th edn. (The Psychological Corporation, San Antonio, TX, 2003).

    Google Scholar 

  41. Sattler, J. M. & Dumont, R. Assessment of children: WISC-IV and WPPSI-III supplement. Jerome M. Sattler, Publisher. Inc: San Diego, CA (2004).

  42. Miller, J. & Iglesias, A. SALT: Systematic analysis of language transcripts [Research version] (SALT Software, Middleton, WI, 2012).

    Google Scholar 

  43. Kover, S. T., Davidson, M. M., Sindberg, H. A. & Weismer, S. E. Use of the ADOS for assessing spontaneous expressive language in young children with ASD: A comparison of sampling contexts. J. Speech Lang. Hear. Res. 57, 2221–2233. (2014).

    Article  PubMed  PubMed Central  Google Scholar 

  44. Suh, J. et al. Narrative performance of optimal outcome children and adolescents with a history of an autism spectrum disorder (ASD). J. Autism Dev. Disord. 44, 1681–1694 (2014).

    Article  Google Scholar 

  45. Heilmann, J., Nockerts, A. & Miller, J. Language sampling: Does the length of the transcript matter? Langu. Speech Heari. Servi. Schools 41, 393–404. (2010).

    Article  PubMed  Google Scholar 

  46. Van Rossum, G. & Drake Jr, F. L. Python reference manual (Centrum voor Wiskunde en Informatica Amsterdam, 1995).

  47. RCoreTeam. R: A Language and Environment for Statistical Computing (R Foundation for Statistical Computing, Vienna, Austria, 2017).

  48. Lake, J. K., Humphreys, K. R. & Cardy, S. Listener vs. speaker-oriented aspects of speech: Studying the disfluencies of individuals with autism spectrum disorders. Psychonomic Bulle. Rev. 18, 135–140 (2011).

    Article  Google Scholar 

  49. Rice, Mabel L. et al. Mean length of utterance levels in 6-month intervals for children 3 to 9 years with and without language impairments. J. Speech Lang. Hearing Res. 53, 333–349. (2010).

    Article  PubMed  PubMed Central  Google Scholar 

  50. Watkins, R. V., Kelly, D. J., Harbers, H. M. & Hollis, W. Measuring children’s lexical diversity: Differentiating typical and impaired language learners. J. Speech Lang. Heari. Res. 38, 1349–1355. (1995).

    CAS  Article  Google Scholar 

  51. McGregor, K. K. & Hadden, R. R. Brief report: Um fillers distinguish children with and without ASD. J. Autism Dev. Disord. 50, 1816–1821. (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  52. Gorman, K. et al. Automated morphological analysis of clinical language samples. In Proceedings of the 2nd Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality, 108–116, (Association for Computational Linguistics, Denver, Colorado, 2015).

Download references


The authors wish to thank Dr. Damien Fair’s neuroimaging team which collected the baseline clinical data as part of their NIH funded studies R01MH115357 and R01MH086654. We thank, in particular, Michaela Cordova, Beth Calamé, and Julia Painter, and the children and their families who participated in this study.


This work was supported by the National Institute on Deafness and Other Communication Disorders of the National Institutes of Health under award R01DC012033 (PI: Dr. E. Fombonne).

Author information

Authors and Affiliations



A.S. and H.M. designed the study and generated the ADOS transcripts. A.S., H.M., E.F. analyzed the data and wrote the manuscript. J.A. participated in the software design. G.L. assisted in the data analysis and edited the manuscript. S.B. and J.D. provided expertise for the methodology and analysis. All authors contributed to the manuscript and approved its content.

Corresponding author

Correspondence to Alexandra C. Salem.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher's note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Salem, A.C., MacFarlane, H., Adams, J.R. et al. Evaluating atypical language in autism using automated language measures. Sci Rep 11, 10968 (2021).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI:

Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.


Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing