Abstract

Recent technological advancements in computerized text and speech analysis as well as machine learning methods have sparked a growing body of research investigating the algorithmic recognition of affect from the ubiquitous digital traces of natural language data and corresponding affect-linked language variations. Also, commercial interest to leverage these new data using AI for affect inferences is on the rise. However, due to the challenges associated with collecting data on subjective affect experience and corresponding language samples, previous research studies and commercial products have mostly relied on data sets from labelled text or enacted speech and, thereby, are focused on affect expression. This work leverages new smartphone-based data collection methods to collect self-reports on in-situ subjective affect experience and corresponding language samples in the wild to investigate between-person differences and within-person fluctuations in affect experience. The present dissertation aims to achieve three goals: (1) to investigate if between-person differences and within-person fluctuations in subjective affect experience are associated with and predictable from cues in spoken and written natural language, (2) to identify specific language characteristics, such as the use of specific word categories or voice parameters, that are associated with and predictive of affect experience, and (3) to analyze the influence of the context of language production on the associations and predictions of affect experience from natural language. This work is comprised of two empirical studies that analyze self-reports on subjective affect experience and natural language data collected with smartphones. Study 1 investigates predictions of between-person differences and within-person fluctuations in subjective momentary affect experience in more than 23000 speech samples from over 1000 participants in two data sets from Germany and the United States. In contrast to voice acoustics, which contain limited predictive information for affective arousal, state-of-the-art word embeddings yield significant above-chance predictions for affective arousal and valence. Moreover, interpretable machine learning methods are used to identify those voice features (i.e., loudness and spectral features) that are most predictive of affect experience. Finally, the work suggests that affect predictions from voice cues from semi-structured free speech are superior to those from read-out predefined sentences and that the emotional sentiment of the spoken content has no effect on affect predictions from voice cues. Study 2 analyzes patterns in written language data logged through smartphones' keyboards to investigate how between-person differences and within-person fluctuations in affect experience manifest in and are predictable from logged text data across different time frames and communication contexts. From a data set of more than 10 million typed words, features regarding typing dynamics, word use based on word dictionaries, and emoji and emoticon use are computed. From the data, distinct affect-linked language variations across communication contexts (private messaging versus public posting) and time frames (trait, weekly, daily, momentary) are identified (e.g., the use 1st person singular). Predictions of affect experience from machine learning algorithms, however, are not significantly better than chance. Results of this study highlight the challenges of using occurrence-counts, such as word dictionaries, for the assessment of subjective affect experience. By leveraging novel smartphone-based experience sampling and on-device language data collection in everyday life, the present dissertation shows how characteristics of spoken and written language are associated with and predictive of subjective affect experience. Thereby, this work highlights the utility of smartphones for investigating subjective affect experience in natural language in the wild, overcoming the caveats of prior research methods. Prediction results, however, challenge the optimistic prediction performances reported in prior works on the recognition of affect expression experience. Using statistical methods from the areas of description, prediction, and explanation, the present dissertation also reveals specific affect-linked language characteristics. Finally, results underline the relevance of the context of language production on language characteristics and corresponding affect predictions. The promising applications and potential future directions of this technology come with multiple challenges with regard to the conceptualization of affect, interdisciplinarity, ethics, and data privacy and security. If these challenges can be overcome, natural language analysis based on data collected with smartphones represents a promising tool to monitor affective well-being and to advance the affective sciences.