Unsupervised structure induction and multimodal grounding

Abstract

Structured representations build upon symbolic abstraction (e.g., words in natural language and visual concepts in natural images), offer a principled way of encoding our perceptions about the physical world, and enable the human-like generalization of machine learning systems. The predominant paradigm for learning structured representations of the observed data has been supervised learning, but it is limited in several respects. First, supervised learning is challenging given the scarcity of labeled data. Second, conventional approaches to structured prediction have been relying on a single modality (e.g., either images or text), ignoring the learning cues that may have been specified in and can be readily obtained from other modalities of data. In this thesis, we investigate unsupervised approaches to structure induction in a multimodal setting.

Unsupervised learning is inherently difficult in general, let alone inducing complex and discrete structures from data without direct supervision. By considering the multimodal setting, we leverage the alignments between different data modalities (e.g., text, audio, and images) to facilitate the learning of structure-induction models, e.g., knowing that the individual words in ``a white pigeon'' always appear with the same visual object, a language parser is likely to treat them as a whole (i.e., phrase). The multimodal learning setting is practically viable because multimodal alignments are generally abundant. For example, they can be found in online posts such as news and tweets that usually contain images and associated text, and in (YouTube) videos, where audio, scripts, and scenes are synchronized and grounded in each other.

We develop structure-induction models, which are capable of exploiting bimodal image-text alignments, for two modalities: (1) for natural language, we consider unsupervised syntactic parsing with phrase-structure grammars and regularize the parser by using visual image groundings; and (2) for visual images, we induce scene graph representations by mapping arguments and predicates in the text to their visual counterparts (i.e., visual objects and relations among them) in an unsupervised manner. While useful, crossmodal alignments are not always abundantly available on the web, e.g., the alignments between non-speech audio and text. We tackle the challenge by sharing the visual modality between image-text alignment and image-audio alignment; images function as a pivot and connect audio and text. The contributions of this thesis span from model development to data collection. We demonstrated the feasibility of applying multimodal learning techniques to unsupervised structure induction and multimodal alignment collection. Our work opens up new avenues for multimodal and unsupervised structured representation learning.