Relation classification with limited supervision

Large reams of unstructured data, for instance in form textual document collections containing entities and relations, exist in many domains. The process of deriving valuable domain insights and intelligence from such documents collections usually involves the extraction of information such as the relations between the entities in such collections. Relation classification is the task of detecting relations between entities. Supervised machine learning models, which have become the tool of choice for relation classification, require substantial quantities of annotated data for each relation in order to perform optimally. For many domains, such quantities of annotated data for relations may not be readily available, and manually curating such annotations may not be practical due to time and cost constraints. In this work, we develop both model-specific and model-agnostic approaches for relation classification with limited supervision. We start by proposing an approach for learning embeddings for contextual surface patterns, which are the set of surface patterns associated with entity pairs across a text corpus, to provide additional supervision signals for relation classification with limited supervision. We find that this approach improves classification performance on relations with limited supervision instances. However, this initial approach assumes the availability of at least one annotated instance per relation during training. In order to address this limitation, we propose an approach which formulates the task of relation classification as that of textual entailment. This reformulation allows us to use the textual descriptions of relations to classify their instances. It also allows us to utilize existing textual entailment datasets and models to classify relations with zero supervision instances. The two methods proposed previously rely on the use of specific model architectures for relation classification. Since a wide variety of models have been proposed for relation classification in the literature, a more general approach is thus desirable. We subsequently propose our first model-agnostic meta-learning algorithm for relation classification with limited supervision. This algorithm is applicable to any gradient-optimized relation classification model. We show that the proposed approach improves the predictive performance of two existing relation classification models when supervision for relations is limited. Next, because all the approaches we have proposed so far assume the availability of all supervision needed for classifying relations prior to model training, they are unable to handle the case when new supervision for relations becomes available after training. Such new supervision may need to be incorporated into the model to enable it classify new relations or to improve its performance on existing relations. Our last approach addresses this short-coming. We propose a model-agnostic algorithm which enables relation classification models to learn continually from new supervision as it becomes available, while doing so in a data-efficient manner and without forgetting knowledge of previous relations.