The advent and proliferation of web technologies and the boom of big data has given rise to new modes of social interactions, as well as a deluge of information across numerous domains, topics and languages. This unstructured human-generated data contains rich semantic and stylistic signals reflecting the latent intentions of people, and is remarkably valuable for knowledge discovery. In order to effectively understand and lend structure to such massive quantities of data, it is not enough to merely extract and analyze salient patterns from it. It is equally important to understand and model the functional intentions, behavioral characteristics, reactions and responses of the authors and/or consumers of the data in question. In this dissertation, we extract and study the functional meaning, intentions and knowledge patterns in modern digital content in disparate contexts, which is the focus of an area called latent pragmatic analysis. We then propose some interesting avenues of future research.
Our first contribution is the development of context-aware knowledge harvesting techniques to automatically organize unstructured information into easily accessible, hierarchical schemas, thus avoiding the need for manual or expert curation. We propose a technique, ETF, that learns a ranking model utilizing semantic and graph theoretic features to insert newly emerging conceptual information into large, existing general-purpose knowledge stores like DBPedia. Second, we develop a machine learning algorithm, BOLT-K, to automatically learn ontology hierarchies for emergent topics or sub-domains. It significantly reduces the need for human supervision by augmenting the limited labeled training data, and transferring knowledge from existing, functionally related schemas. Third, to ensure the presence of factually accurate information in our constructed knowledge stores, we propose an encoder-decoder framework called FACE-KEG. It predicts the veracity of input textual information via a graph transformer network classifier, and also generates human-comprehensible explanations via a recurrent neural network, to justify its predictions.
Orthogonal to our prior efforts of studying and structuring digital content from the web, our subsequent work transitions into modeling the behavioral aspects and intentions of human users or creators of digital content. As our fourth and fifth contributions, we discovered and categorized user intentions or intents (our proposed technique is called OPINE), and domains(our proposed technique is called ADVIN) from human user conversations and user interactions with virtual assistants. Our deep learning models are independent of the topic of the input text utterances, require minimal domain-specific labeled training data and can be employed for various downstream conversational and search applications. Our final contributions augment the textual semantics and language cues with emotional signals, social network topology and alternate modalities (e.g. video, audio), to investigate multiple facets of online user behavior towards social good applications. To this end, our sixth contribution analyzed the multimodal dependency patterns between digital multimedia content attributes and the responses and reactions invoked among their target users or viewers. Our seventh contribution is a scalable, unsupervised model to realize the credibility and reliability of online social network users, by quantifying appropriate socio-psychological elements such as the social influence exerted by users in a network, the underlying network structure and the affective valence expressed in user content. Our eighth contribution successfully correlates distinctive online social platform cues (e.g. user activity, network engagement, affective valence, linguistic patterns) with theoretical insights from offline medical and psychological findings, with emphasis on clinical depression.
Through these research efforts, we emphasize that developing computational models enriched by domain specific insights and contextual information that adapt to users' behavior, needs and operations under various settings is of immense value in several disciplines, such as information retrieval, conversational search, marketing and crisis response. The long-term research goal of this dissertation is to develop novel, intelligent, context-aware systems that effectively unify and structure heterogeneous data from trustworthy sources and respond to human intents and behavior patterns.