RI: Medium: Learning Disentangled Representations for Text to Aid Interpretability and Transfer

Northeastern University, Boston MA

Investigators

Abstract

Machine learning methods for natural language processing power many technologies that we use on a day-to-day basis, such as spam filters and translation software. The models underlying these techniques have become increasingly sophisticated, yielding improved performance but also increasing complexity. In particular, "neural network" based approaches have re-emerged as the dominant class of machine learning models for language processing. These approaches often perform better than their non-neural counterparts, but also have key downsides. First, training these models requires human effort and time to generate a sufficiently large set of training data in the form of manually annotated text. Second, it is often not obvious whether a model trained on one dataset will generalize to another. Finally, it is hard to discern why such models make the specific predictions that they do, largely because predictions are made on the basis of learned representations of texts which do not naturally afford transparency. This project proposes technical innovations to address these interrelated issues using "disentanglement". The idea is to design models such that the learned representations used to make predictions have known meaning. This approach has the potential to enable re-use of models (increasing efficiency and reducing human costs), and aid interpretability, so that one can have a better idea of why a model made a given prediction. To realize the above goals of improved interpretability and transferability of models, this work will develop and evaluate new models that learn representations in which certain dimensions are imbued with explicit semantics. This is a departure from current approaches, which indiscriminately code all attributes into a single (entangled) representation. To achieve disentanglement, this project will explore deep generative models and sparse, gated neural encoders. These will use inductive biases and light supervision strategies that guide models toward disentangled representations. For example, models will be penalized if distances in learned embedding spaces do not reflect human judgments concerning the relative similarities of instances with respect to specific aspects of interest. In other cases, "weak" supervision (e.g., rules) may provide adequate guidance for disentanglement. Finally, "probing" tasks constitute a third supervision strategy to be explored: This will involve the use of auxiliary tasks to provide "supervision" that guides individual aspect-wise embeddings of input. The project will develop and evaluate such models for representative problems in natural language processing, specifically: classification, sequence tagging, and summarization. Models will be evaluated both for predictive performance (including their generalizability to new domains and the efficiency with which they do so), and the degree to which learned representations are disentangled and capture the intended aspects. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.