UC Santa Barbara

Automating user tasks with natural language utterances, such as answering questions

over Wikipedia or booking flight tickets on the Web, is a key component in designing

intelligent systems. Natural language is usually preferred as a unified interface for these

systems and requires no domain expertise for users; however, understanding wide range of

diverse inputs and resolving errors that occur during this process are still open challenges

and the topics of this thesis.

Traditional machine learning systems for natural language interfaces usually require

large-scale labeled datasets with handcrafted rules to train and evaluate the performances

of the respective models. Firstly, the handcrafted design constrains the scope to a limited

set of domains and prevents the adaptation to new tasks. Additionally, large-scale labeled

data collection is generally domain dependent, costly, and time consuming. These systems

further assume that the underlying database, such as Freebase, is accessible and can be

queried indefinitely which is prohibitive when learning from constrained user interfaces,

such as Web pages. Last but not least, current systems focus on training offline in a

closed loop where users are excluded from the system inference process. They lack the

capabilities to continuously learn from users.

In this thesis, we address the drawbacks of the existing systems and propose data

efficient and user-centric solutions. We classify the natural language inference problem

based on two different perspectives: Accessiblity of the system functions – unconstrained

or constrained user interfaces, and nature of user involvement during inference – non-interactive or interactive user interfaces. We first develop neural network based systems

for non-interactive and unconstrained users interfaces with different data types (i.e. structured and unstructured). The system is trained to learn a continuous representation of

user utterance, generate and rank candidate answers from underlying database using

this representation. We augment these systems with an extractive candidate refinement

framework by integrating task-oriented human-machine dialogues. Our system is able to

understand, point, and refine the error in candidates by asking users validation questions

and offering alternatives. We also address the limitations of unconstrained user interfaces and propose reinforcement learning methods to develop policies that are capable of

learning from more constrained web interfaces. The policies are trained on a variety of

web pages, such as flight booking and social media interaction, with task-based reward

signals and no human supervision. We test the performance of our models with simulated as well as real users. Empirical results show that the proposed models are able to

learn from limited supervised data and have successful dialogues with users. We observe

improvements in answer prediction accuracy, task success rate, and real user ratings.