CRII: RI: Learning Structured Prediction Models with Auxiliary Supervision

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Many machine learning problems involve making joint predictions over a set or a sequence of mutually dependent outputs. As an example, consider recognizing a handwritten word, where each character must be recognized in order for the word to be understood. It is important to consider the correlations between the predictions of adjacent characters to aid the individual predictions of characters. Structured prediction models are proposed to solve problems of this type. They have been shown to be successful in many real-world applications, including speech recognition, natural language understanding, and object detection in images. Despite its success, training a structured prediction model requires an extensive collection of training data. However, obtaining human annotations with complex structures is costly. For example, it takes a professionally trained linguist several minutes to label a syntactic parse tree for a single sentence, making it expensive to obtain high-quality annotations. The objective of this research is to develop methods that utilize learning signals that do not directly aim to achieve the target tasks. The outcome of this project will create a fundamental shift in the applicability of structured prediction models, enabling applications in which complex decisions are required and annotated data are expensive to acquire. This will bring new collaboration opportunities with other areas, including education, healthcare, and social and behavioral sciences. The goals of this project are to study algorithms for training structured prediction models from heterogeneous learning signals, design automatic algorithms for mining useful information from massive columns of structured and unstructured data, and apply the proposed techniques in real-world applications. The proposed algorithms will be evaluated on a broad range of natural language processing applications, including the algebra word problem, co-reference resolution, and grammatical error correction. The results of the project will be disseminated by publishing papers, releasing open-source software and data sets, organizing workshops and tutorials, and creating new courses on natural language processing and machine learning.