Learning natural language interfaces with neural models

Abstract

Language is the primary and most natural means of communication for humans. The learning curve of interacting with various devices and services (e.g., digital assistants, and smart appliances) would be greatly reduced if we could talk to machines using human language. However, in most cases computers can only interpret and execute formal languages. In this thesis, we focus on using neural models to build natural language interfaces which learn to map naturally worded expressions onto machineinterpretable representations. The task is challenging due to (1) structural mismatches between natural language and formal language, (2) the well-formedness of output representations, (3) lack of uncertainty information and interpretability, and (4) the model coverage for language variations. In this thesis, we develop several flexible neural architectures to address these challenges. We propose a model based on attention-enhanced encoder-decoder neural networks for natural language interfaces. Beyond sequence modeling, we propose a tree decoder to utilize the compositional nature and well-formedness of meaning representations, which recursively generates hierarchical structures in a top-down manner. To model meaning at different levels of granularity, we present a structure-aware neural architecture which decodes semantic representations following a coarse-to-fine procedure. The proposed neural models remain difficult to interpret, acting in most cases as a black box. We explore ways to estimate and interpret the model’s confidence in its predictions, which we argue can provide users with immediate and meaningful feedback regarding uncertain outputs. We estimate confidence scores that indicate whether model predictions are likely to be correct. Moreover, we identify which parts of the input contribute to uncertain predictions allowing users to interpret their model. Model coverage is one of the major reasons resulting in uncertainty of natural language interfaces. Therefore, we develop a general framework to handle the many different ways natural language expresses the same information need. We leverage external resources to generate felicitous paraphrases for the input, and then feed them to a neural paraphrase scoring model which assigns higher weights to linguistic expressions most likely to yield correct answers. The model components are trained end-to-end using supervision signals provided by the target task. Experimental results show that the proposed neural models can be easily ported across tasks. Moreover, the robustness of natural language interfaces can be enhanced by considering the output well-formedness, confidence modeling, and improving model coverage.