From user-generated text to insight context-aware measurement of social impacts and interactions using natural language processing

"Recent improvements in information and communication technologies have contributed to an increasingly globalized and connected world. The digital data that are created as the result of people's online activities and interactions consist of different types of personal and social information that can be used to extract and understand people's implicit or explicit beliefs, ideas, and biases. This thesis leverages methods and theories from natural language processing and social sciences to study and analyze the manifestations of various attributes and signals, namely social impacts, personal values, and moral traits, in user-generated texts. This work provides a comprehensive understanding of people's viewpoints, social values, and interactions and makes the following contributions. First, we present a study that combines review mining and impact assessment to provide an extensive discussion on different types of impact that information products, namely documentary films, can have on people. We first establish a novel impact taxonomy and demonstrate that, with a rigorous analysis of user-generated texts and a theoretically grounded codebook, classification schema, and prediction model, we can detect multiple types of (self-reported) impact in texts and show that people's language can help in gaining insights about their opinions, socio-cultural information, and emotional states. Furthermore, the results of our analyses show that documentary films can shift peoples' perceptions and cognitions regarding different societal issues, e.g., climate change, and using a combination of informative features (linguistic, syntactic, and psychological), we can predict impact in sentences with high accuracy. Second, we investigate the relationship between principles of human morality and the expression of stances in user-generated text data, namely tweets. More specifically, we first introduce and expand the Moral Foundations Dictionary and operationalize moral values to enhance the measurement of social effects. In addition, we provide detailed explanation on how morality and stance are associated in user-generated texts. Through extensive analysis, we show that discussions related to various social issues have distinctive moral and lexical profiles, and leveraging moral values as an additional feature can lead to measurable improvements in prediction accuracy of stance analysis. Third, we utilize the representation of emotional and moral states in texts to study people's interactions in two different social networks. Moreover, we first expand the analysis of structural balance to include direction and multi-level balance assessment (triads, subgroups, and the whole network). Our results show that analyzing different levels of networks and using various linguistic cues can grant a more inclusive view of people and the stability of their interactions; we found that, unlike sentiments, moral statuses in discussions stay balanced throughout the networks even in the presence of tension. Overall, this thesis aims to contribute to the emerging field of ""social"" NLP and broadens the scope of research in it by (1) utilizing a combination of novel taxonomies, datasets, and tools to examine user-generated texts and (2) providing more comprehensive insights about human language, cultures, and experiences."

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Information Sciences

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Rezapour, Rezvaneh

Contributors dc:contributor

• Diesner, Jana
• Underwood, Ted
• Girju, Roxana
• Karahalios, Karrie

• Information sciences

Statement dc:rights

• Copyright 2021 Rezvaneh Rezapour

Language dc:language

en, eng

Handle dc:identifier

http://hdl.handle.net/2142/113860

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F113860