Cross-linguistic Data-driven Measurement of Child Language Development

In order to solve the puzzle of child language acquisition, researchers have developed a number of metrics. All of these metrics involve examining a transcript of utterances, and producing a score that corresponds to language development. The most commonly used metrics rely on surface-level features, which are used due to computational ease and availability. An alternative approach is to, given a transcript, produce a number corresponding to the age of the child examined in the transcript. Using this approach, we can track a child’s linguistic development based on age—a number easily understood without intimate knowledge of the process. This age prediction task has been approached from a Machine Learning perspective, involving the extraction of features from transcripts and using these to predict age [SS12, LS14]. This research has shown that simple features correlate well with age; however, it has only been done for English-speaking children. In this paper, we explore this age prediction approach for language development to children speaking other languages. Using transcripts of children speaking Spanish, Japanese, and Hebrew from the CHILDES database, we examine the age prediction task using similar simple syntactic feature templates as those that were used in previous research in English. We find that approaches using only syntactic features perform at comparable levels to those using more languagespecific, content-based features. Our best results show the ability to predict a child’s age based on syntactic features from transcripts within two months of their actual age, with strong correlations between predicted and actual age for the data set. Additionally, we compare performance to contentbased features and find no significant improvement over syntactic features. We suggest future experiments in order to determine the best feature sets for a given language, and call for increased data collection in this area. With the increased availability of child language transcripts, such a cross-linguistic data-driven approach has the ability to influence, motivate, and assist the research area of child language development.