The development of a digital story-retell elicitation and analysis tool through citizen science data collection, software development and machine learning

BACKGROUND: In order to leverage the potential benefits of technology to speech and language therapy language assessment processes, large samples of naturalistic language data must be collected and analysed. These samples enable the development and testing of novel software applications with data relevant to their intended clinical application. However, the collection and analysis of such data can be costly and time-consuming. This paper describes the development of a novel application designed to elicit and analyse young children’s story retell narratives to provide metrics regarding the child’s use of grammatical structures (micro-structure) and story grammar (macro-structure elements). Key aspects for development were (1) methods to collect story retells, ensure accurate transcription and segmentation of utterances; (2) testing the reliability of the application to analyse micro-structure elements in children’s story retells and (3) development of an algorithm to analyse narrative macro-structure elements. METHODS: A co-design process was used to design an app which would be used to gather story retell samples from children using mobile technology. A citizen science approach using mainstream marketing via online channels, the media and billboard ads was used to encourage participation from children across the United Kingdom. A stratified sampling framework was used to ensure a representative sample was obtained across age, gender and five bands of socio-economic disadvantage using partial postcodes and the relevant indices of deprivation. Trained Research Associates (RA) completed transcription and micro and macro-structure analysis of the language samples. Methods to improve transcriptions produced by automated speech recognition were developed to enable reliable analysis. RA micro-structure analyses were compared to those generated by the digital application to test its reliability using intra-class correlation (ICC). RA macro-structure analyses were used to train an algorithm to produce macro-structure metrics. Finally, results from the macro-structure algorithm were compared against a subset of RA macro-structure analyses not used in training to test its reliability using ICC. RESULTS: A total of 4,517 profiles were made in the app used in data collection and from these participants a final set of 599 were drawn which fulfilled the stratified sampling criteria. The story retells ranged from 35.66 s to 251.4 s in length and had word counts ranging from 37 to 496, with a mean of 148.29 words. ICC between the RA and application micro-structure analyses ranged from 0.213 to 1.0 with 41 out of a total of 44 comparisons reaching ‘good’ (0.70–0.90) or ‘excellent’ (>0.90) levels of reliability. ICC between the RA and application macro-structure features were completed for 85 samples not used in training the algorithm. ICC ranged from 0.5577 to 0.939 with 5 out of 7 metrics being ‘good’ or better. CONCLUSION: Work to date has demonstrated the potential of semi-automated transcription and linguistic analyses to provide reliable, detailed and informative narrative language analysis for young children and for the use of citizen science based approaches using mobile technologies to collect representative and informative research data. Clinical evaluation of this new app is ongoing, so we do not yet have data documenting its developmental or clinical sensitivity and specificity.