This dissertation introduces the Expressive-Story Translator (EST) content planner and Fabula Tales sentence planner in a storytelling natural language generation framework. Both planners operate in a domain independent manner, abstractly modeling a variety of stories regardless of story vocabulary. The EST captures story semantics from a narrative representation and constructs text plans to maintain semantic content through rhetorical relations. Content planning is performed using these relations to enhance narrative effects, such as modeling emotions and temporal reordering. The EST transforms the story into semantic-syntactic structures interpreted by the parameterizable sentence planner, Fabula Tales. The semantic-syntactic integration allows Fabula Tales to employ narrative sentence planning devices to change narrator point of view, insert direct speech acts, and supplement character voice using operations for lexical selection, aggregation, and pragmatic marker insertion. The frameworks are evaluated using traditional machine translation metrics, narrative metrics, and overgenerate and rank to holistically test the effectiveness of each generated retelling. This work shows how different framings affect reader perception of stories and its characters, and uses statistical analysis of reader feedback to build story models tailored for specific narration preferences.

Conversational AI has seen tremendous progress in recent years, achieving near-human or even surpassing human performance in certain well-defined tasks, including speech recognition and question answering. Yet it tends to struggle with tasks which are less constrained, in particular those that involve producing human language. Current approaches to natural language generation (NLG) in dialogue systems still heavily rely on techniques that lack scalability and transferability to different domains, despite the general embrace of more robust methods by the NLG community, in particular deep learning (neural) models. These methods rely on large amounts of annotated data, yet they tend to produce generic, robotic, and boring responses that lack most of the human language nuances that make conversation creative and varied.

While the naturalness of the generated language is an important factor affecting the perceived quality of a dialogue system, semantic accuracy is also extremely important. If a system is not semantically accurate, it may provide the user with incorrect information or contradict its earlier responses. In this thesis, we focus on the task of generating an utterance from a structured meaning representation (MR). To support our work, we create and release a new parallel corpus with more varied dialogue acts and more conversational utterances than previous MR-to-text corpora. We explore different ways of promoting output diversity in neural data-to-text generation while ensuring high semantic accuracy by developing new methods to help deep learning NLG models produce diverse utterances that are faithful to their MRs. This is an important step toward making conversational AI more reliable and pleasant to interact with.

We first observe in our initial experiments that NLG models have the ability to produce more diverse and natural-sounding texts when explicitly prompted to, however, this diversity comes at the expense of semantic accuracy. This leads us to develop a set of methods for automatically assessing and enforcing semantic accuracy in the generated utterances. We introduce a general tool to find a semantic alignment between an utterance and the corresponding input, which can be used for automatically evaluating the accuracy of generated utterances and ranking a pool of candidate utterances a model produces. We also propose a novel semantically attention-guided decoding method for neural encoder-decoder models, which utilizes the models' own knowledge acquired from training in a way that enables them to track semantic accuracy during inference and rerank generated utterance candidates accordingly. We show on multiple datasets that both of these methods have an ability to dramatically reduce semantic errors in model outputs, while maintaining their overall quality and fluency.

We then systematically explore Monte-Carlo Tree Search (MCTS) as a way to simultaneously optimize both semantic accuracy and stylistic diversity during inference. To guide the MCTS, we propose a new referenceless automatic metric for utterance evaluation. Our results show that, using this novel method, we can successfully increase diversity while maintaining, or even improving, semantic accuracy.

Conversation is an essential component of social behavior, one of the primary means by which humans express emotions, moods, attitudes, and personality. Conversation is also critical to storytelling, where key information is often revealed by what a character says, how s/he says it, and how s/he reacts to what other characters say.

Interactive narrative systems (INS) are a type of playable media whose applications range from simple entertainment to systems for learning, training, and decision making. Many forms of INS involve interactions with virtual human characters. Thus a key technical capability for such systems is the ability to support natural conversational interaction. While most INS use hand-crafted character dialogue to produce high quality utterances, they suffer from problems of portability and scalability, or what has been called the authoring bottleneck. We believe Natural Language Generation (NLG) is part of the solution to alleviate such burden from authors by automatically generating character dialogue.

Here we focus on the issue of character voice. One way to produce believable, dramatic dialogue is to build stylistic models with linguistic features related to NLG decisions. Film/television dialogue are exemplars of many different linguistic styles that were designed to express dramatic characters. Thus we construct a corpus of film/television character dialogue from screenplays and transcripts publicly available from websites such as the Internet Movie Script Database. We apply content analysis and language modeling techniques to extract relevant linguistic features to build character-based stylistic models. We also apply machine learning techniques to discriminate characters base on available metadata such as genre, year, and director.

This thesis consists of two parts. The first part involves building a basic character model with film dialogue, and then applying the model to an existing expressive NLG engine to generate different character voices. We then evaluate the generation experiment with a perceptual study, which suggests several natural extensions.

The second part involves building a more refined model with television dialogue in order to explore a broader range of stylistic features that can be used to express dramatic characters. We test the model-fit of character models in two ways: 1) ranking experiments to pick out corresponding character’s utterances from a pool of mixed, original characters utterances, and 2) a second generation experiment to test user perceptions of characters.

Stories play an important role in human perception of the world and therefore the computational analysis of narrative structure is a key area in natural language processing. The focus of this thesis is to develop and evaluate computational models for two main elements of the narrative structure: Events and Desires. Our work first aims to test a theory that proposes a linear structure of narratives and identifies different parts of a story based on their function. Unlike most of the previous work that use the news articles or other simpler and more conventional genres, we use a corpus of personal stories from social media that have a wider range of topical content and variations of discourse relations.

We present an unsupervised method for modeling narrative events, focusing on specific event relations based on the Penn Discourse Treebank’s definition of contingency. We use a weakly supervised approach to extract the key events from stories and create a topic-sorted corpus of personal narratives using a bootstrapping method. We additionally propose new evaluation methods for testing the contingent event pairs. Our results show that most of the relations we learn from blog stories are not found in the existing event collections.

In our final contribution, we develop supervised methods for modeling the protagonist’s goals and their outcome in personal narratives, as a sub-problem of modeling affects. Our studies show that both prior and post context are useful for modeling desire fulfillment. In addition, we show that exploiting narrative structure is helpful, both directly in terms of the utility of discourse relation features and indirectly by using a sequential model. We further examine our analysis of the human desires by identifying and studying the expressions of unfulfilled goals.

Neural models for generating text from structured representations of meaning have recently gained popularity in the natural language generation (NLG) community. Instead of using a traditional NLG pipeline involving separate modules for sentence planning and surface realization, neural models combine these steps into a single end-to-end framework. This new paradigm allows for low effort data-driven generation, but makes it very unclear how to control model output and produce the required semantics with the desired syntactic or stylistic constructions for a given application.

This thesis takes on the task of learning to control neural natural language generation systems, with the goal of producing natural language outputs that are both semantically correct and stylistically varied. We tackle three critical bottlenecks of neural NLG: how to introduce a mechanism to produce style with neural generators, how to systematically acquire massive amounts of data required for training them, and how to jointly control semantic and stylistic choices, to allow for more diverse model outputs. We address the style bottleneck by experimenting with different methods for supervision in neural models with a synthetic dataset that we build (PersonageNLG), showing that we can produce diverse sentence planning operations in our model outputs. We address the data bottleneck by using freely available review data to create a massive, highly descriptive, and stylistically diverse corpus for training neural generators (YelpNLG), instead of relying on crowdsourcing. We address the control bottleneck by constructing stylistically rich meaning representations derived from review text based on parse information and freely-available ontologies, providing different forms of supervision to our neural models, and allowing us to produce outputs exhibiting a rich array of stylistic variation from semantically-grounded inputs.

We show that by controlling the nature of our input data and how it is represented in our models, we can control a model's ability to produce a required style without sacrificing its ability to produce fluent outputs that express the required content. Our data and experiments introduce novel methods for producing stylistic variation within a neural natural language generation pipeline, and are generalizable to new domains and style choices.

Human conversants in dialog adjust their behavior to their conversational partner in many ways. In terms of language, they adapt to their partners both lexically and syntactically, by using the same referring expressions or sentence structure. In terms of gesture, they mimic their partners' gestural form, frequency, expanse and speed. However, adaptation is not about simply copying dialog partners' words, syntax, and gestures. The process of adaptation to the partner takes place under other special constraints, e.g. providing coherent and informative turns in conversation, expressing one's own personality, or achieving other social and interpersonal goals. How do speakers adapt to one another and at the same time achieve their own conversational goals?

In this thesis, we first carry out an exploratory study to show that adapting to different linguistic features results in different style perceptions: adapting to hedges increase perceptions of friendliness, while adapting to syntactic structures increases perceptions of naturalness. On the basis of these results, we propose an adaptation measure that allows us to capture and model adaptation behaviors that may orient to different levels or types of linguistic representations, such as lexical, syntactical, or stylistic variations in the way speakers talk. We build and compare linguistic adaptation models using our measure with four human dialog corpora, and with different feature sets to represent different levels of linguistic representations.

We also explore gestural adaptation on both particular gesture forms and gesture style. We set up our experiment in the form of two virtual gents co-telling a story. We first verify that we can express the personality of virtual agents through varying gesture parameters such as speed, height, and expanse. We then show that human subjects prefer adaptive to nonadaptive virtual agents, where the adaptive virtual agent adapts the gesture form and personality of their dialog partner.

Depression and loneliness can be a serious social problem that affects people's physical and mental well-being. Our long-term goal is to develop a chatbot for well-being, WellBe, that can listen to people talk about their common daily events, acknowledge their feelings, and make responses to improve their well-being. We posit that developing WellBe requires: (1) the ability to classify user utterances to recognize the affective state of the user and the activity type that the user is talking about; and (2) effective dialogue strategies that condition on the classifier output and utilize theories of well-being. We assembled three corpora of user utterances representing both user affects and activities, and conducted a set of experiments within and across datasets. Our results show that a fine-tuned BERT model achieves F1 measures as high as .88 for user affect classification and .92 for user activity classification, and that it generalizes well across the different corpora. We then design rule-based response strategies that utilize these affect and activity classifiers, and conduct a human evaluation where we compare them with a baseline strategy that does not have access to the classifier output. The results show that responses that rely on the affect and activity classifiers are more engaging than the baseline, and that different response strategies are preferred depending on whether user affect is negative or positive.

Argumentation is an interactive process and frequently occurs in conversations. Social networks and online debates provide a two-way communication platform with a huge amount of opinion and argument rich information. However, the massive amount of information available today is overwhelming to our brain and its ability to efficiently absorb and comprehend. Identifying and integrating arguments across these discussions requires computational methods that can facilitate users to systematically search, analyze and summarize arguments as well as reason about the relationship among arguments. We develop techniques to recognize the specific arguments and counter arguments people tend to advance and group them across discussions as Facets. This entails two sub-tasks: (i) Extract and summarize important arguments, and (ii) Discover similar repeated argument aspects that appear across multiple dialogues on a topic. We present a systematic approach to leverage pyramid based summarization framework to identify central propositions as those arguments that people find most salient, and then rank and select arguments in social media dialogue, which is a novel method for ranking arguments in conversational data. Our results show that adding contextual knowledge from a dialogue improves argument extraction. We introduce a new task of Argument Facet Similarity (AFS), where we develop a new corpus aimed at identifying various facets across opinionated dialogue. A graded argument similarity model was defined that takes as input two sentential arguments and returns a scalar value that categorizes their similarity (AFS). The prediction output obtained from the proposed model improves the results obtained from previous work that defines similarity of short texts as Semantic Textual Similarity (STS).

Research on open-domain social dialogue systems has exploded over the last few years with the advent of large language models (LLMs) that can chat about any topic. Unlike traditional dialogue systems, open-domain dialogue systems cannot assume any specific information need or domain restrictions - the only inherent goal is to converse socially. While modern systems have access to more information and better tools, foundational components of natural human-human conversation remain elusive, i.e., intimacy and agency. In this thesis, we hypothesize that personalization is pivotal in fostering this genuine connection between users and open-domain dialogue systems.

Our first hypothesis is that personalizing the conversation to specific user interests will build a sense of understanding, rapport, and agency. To investigate this, we heuristically combine the results of an extensive natural language understanding pipeline with handcrafted rules to build a user modeling mechanism; this user model then personalizes the experience through response adaptation and topic-promotion strategies, resulting in a statistically significant positive impact on perceived conversation quality and length when evaluated at scale with a testbed open-domain dialogue system, that real Amazon Echo users access. Analyzing the user models unveils nuanced insights into user preferences, emphasizing a desire for more personalized experiences and receptiveness toward personal questions. This leads to our second hypothesis - asking appropriate personalized follow-up questions (PQs) helps to create a more engaged user experience that increases user satisfaction. Our initial test of this hypothesis uses a crowdsourced corpus of PQs (Would You Rather and Hypothetical) in the testbed system's dialogue policy. Our evaluation of the policy shows that it results in extended topical depth, leading to statistically significant longer, more highly rated conversations.

However, crowdsourcing PQs for every user interest does not scale. Question Generation tasks generally focus on factual questions from textual excerpts. Instead, we create a specialized training dataset of PQs more suitable for the novel task of Personal Question Generation. We first identify over 400 common user interests by sampling ~39,000 user models collected during user interactions with our testbed system. Then, we translate these into prompts and use the LLM GPT-3.5 to generate ~19,000 PQs and associated system answers. Evaluating the impact of this pre-generated data when used in our testbed system's dialogue policy results in statistically significant positive effects on perceived conversation quality. Statistically significant results also suggest that deep, user-centric PQs are the most effective means of increasing intimacy and engagement.

We then use the corpus of ~19,000 PQs to fine-tune a RedPajama 3B prompt-based PQ generator, which further shows the positive impact of producing highly tailored questions when evaluated in our testbed system. To evaluate our hypothesis independently from our testbed system, we synthetically generate a corpus of 2,000 long synthetic social dialogues that strongly aim to resemble real user conversations. We use these social dialogues to compare our fine-tuned PQ generator against 5 other state-of-the-art LLMs. Positive results affirm the importance of PQs in social conversation while also validating our model as a strong baseline for the task of Personalized Question Generation.

Traditional statistical natural language generation (NLG) systems require substantial hand-engineering: many of the components, such as content planners, sentence planners and surface realizers must all be designed and created manually and updated when a new type of utterance is required. Neural natural language generation (NNLG) models, on the other hand, learn to generate text through processing massive amounts of data in end-to-end encoder-decoder frameworks, where syntactic properties are learned automatically by the model. While the learning components of NNLG models are mostly accomplished automatically, they do require, however, that the training data be collected and labeled, which can often be a laborious process. The advantages of not needing handcrafted templates and syntax-to-semantics dictionaries may thereby be offset by the need to retrain neural models with new data as new domains are added, effectively replacing a "knowledge bottleneck" with a "data bottleneck".

To overcome the data bottleneck, we experiment with methods to leverage existing datasets to allow our NNLG models to generalize to novel meaning representations and sentence planning operations. We explore the generation of artificial data and mixing data from different sources as a way to augment existing training data available to the NNLG. Given our different methods for augmentation, we evaluate whether NNLGs can learn syntactic and semantic generalizations.

In this thesis, we developed methods to enable NNLG models to perform three types of generalization. First, we generalized multiple stylistic features to a single supervision token which represented the personality formed by the stylistic features. We then generated output with two different personality supervision tokens to create novel personalities not present in the training data. Second, we performed generalization on sentence planning operations. Sentence planning is the module of NLG models that affects the way individual propositions are combined into sentences, which usually has an effect on the final style of the realization. We generalized specific sentence planning operations, such as sentence scoping and contrastive discourse structuring, with values and attribute combinations beyond what was seen in the original training data. Finally, we combined training data from different sources to produce outputs for meaning representations that blend the ontologies from both sources. In all three experiments, we investigated different representations and architectures to enable models to generalize.

Our contributions include the development of methods that enable NNLG models to generalize and thereby expand beyond their original training data. This is an important extension of results to date and a significant step in making NNLG models more useful at low data cost. We also generated and released multiple datasets which we used to train and test our models.