Work

Learning Natural Language Interfaces using Deep Neural Networks

Year: 2019
Type: dissertation
Abstract: Automating user tasks with natural language utterances, such as answering questionsover Wikipedia or booking flight tickets on the Web, is a key component in designingintelligent systems. Natural language is usually preferred as a unified interface for thesesystems and requires no domain expertise for users; however, understanding wide range ofdiverse inputs and resolving errors that occur during this process are still open challengesand the topics of this thesis.Traditional machine learning systems for natural language interfaces usually requirelarge-scale labeled datasets with handcrafted rules to train and evaluate the performancesof the respective models. Firstly, the handcrafted design constrains the scope to a limitedset of domains and prevents the adaptation to new tasks. Additionally, large-scale labeleddata collection is generally domain dependent, costly, and time consuming. These systemsfurther assume that the underlying database, such as Freebase, is accessible and can bequeried 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 aclosed loop where users are excluded from the system inference process. They lack thecapabilities to continuously learn from users.In this thesis, we address the drawbacks of the existing systems and propose dataefficient and user-centric solutions. We classify the natural language inference problembased on two different perspectives: Accessiblity of the system functions – unconstrainedor constrained user interfaces, and nature of user involvement during inference – non-interactive or interactive user interfaces. We first develop neural network based systemsfor non-interactive and unconstrained users interfaces with different data types (i.e. structured and unstructured). The system is trained to learn a continuous representation ofuser utterance, generate and rank candidate answers from underlying database usingthis representation. We augment these systems with an extractive candidate refinementframework by integrating task-oriented human-machine dialogues. Our system is able tounderstand, point, and refine the error in candidates by asking users validation questionsand offering alternatives. We also address the limitations of unconstrained user interfaces and propose reinforcement learning methods to develop policies that are capable oflearning from more constrained web interfaces. The policies are trained on a variety ofweb pages, such as flight booking and social media interaction, with task-based rewardsignals 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 tolearn from limited supervised data and have successful dialogues with users. We observeimprovements in answer prediction accuracy, task success rate, and real user ratings.
Language: English
Open Access status: green
OpenAlex ID: W2995612754
Cites: 102
Cited by:
Related to: 10