UC San Diego

Text classification plays a fundamental role in transforming unstructured text data to structured knowledge. State-of-the-art text classification techniques rely on heavy domain-specific annotations to build massive machine(deep) learning models. Although these deep learning models exhibit superior performance, the lack of training data and expensive human effort in the manual annotation is a key bottleneck that forbids them from being adopted in many practical scenarios. To address this bottleneck, our research exploits the data and develops a family of data-driven text classification frameworks with minimal supervision, for e.g. class names, a few label-indicative seed words per class.

The massive volume of text data and complexity of natural language pose significant challenges to categorizing the text corpus without human annotations. For instance, the user- provided seed words can have multiple interpretations depending on the context, and their respective user-intended interpretation has to be identified for accurate classification. Moreover, metadata information like author, year, and location is widely available in addition to the text data, and it could serve as a strong, complementary source of supervision. However, leveraging metadata is challenging because (1) metadata is multi-typed, therefore it requires systematic modeling of different types and their combinations, (2) metadata is noisy, some metadata entities (e.g., authors, venues) are more compelling label indicators than others. And also, the label set is typically assumed to be fixed in traditional text classification problems. However, in many real-world applications, new classes especially more fine-grained ones will be introduced as the data volume increases. The goal of our research is to create general data-driven methods that transform real-world text data into structured categories of human knowledge with minimal human effort.

This thesis outlines a family of weakly supervised text classification approaches, which upon combining can automatically categorize huge text corpus into coarse and fine-grained classes, with just label hierarchy and a few label-indicative seed words as supervision. Specifically, it first leverages contextualized representations of word occurrences and seed word information to automatically differentiate multiple interpretations of a seed word, and thus result- ing in contextualized weak supervision. Then, to leverage metadata, it organizes the text data and metadata together into a text-rich network and adopt network motifs to capture appropriate combinations of metadata. Finally, we introduce a new problem called coarse-to-fine grained classification, which aims to perform fine-grained classification on coarsely annotated data. Instead of asking for new fine-grained human annotations, we opt to leverage label surface names as the only human guidance and weave in rich pre-trained generative language models into the iterative weak supervision strategy. We have performed extensive experiments on real-world datasets from different domains. The results demonstrate significant advantages of using contextualized weak supervision and leveraging metadata, and superior performance over baselines.