In recent years, the field of Natural Language Processing has seen significant revolution by the introduction of Transformers, a stack of multiple layers of attention and non-linearity, capable of performing almost any task and the backbone for large foundation models.
In contrast to traditional NLP pipelines, this architecture is able to learn the features required to perform a specific task without any assumptions about the structure of language. The only remaining hard-coded aspect is the way the text input is fed to these models. This preprocessing step, known as the tokenization step, divides the input into chunks which could be as fine-grained as bytes or as coarse-grained as words. The most popular approach is to use subwords, such as Byte Pair Encodings or word pieces which lie between characters and words. However, it has been shown that hard-coding the input representations, has its own drawbacks. In particular, it would lead to sub-optimal performance in downstream tasks. In addition, to perform different tasks we need different levels of representations. In this thesis, we define and address the novel task of inducing units from text in an unsupervised manner. This work is a step towards completely end-to-end models which can decide which level of representation is the most suitable for them to perform a specific task.

Our contributions are two-fold:
First, we design models which are able to induce units without supervision at different levels. And second, since the task is novel, we need novel evaluations to show its effectiveness. Therefore, for every model we develop, we design and/or gather the set of tasks which evaluate and interpret the performance of our models.

In the first chapter, we design a model to induce morpheme-like units from a sequence of characters. We adapt a method from object discovery in vision, called Slot Attention for our purpose. We propose to evaluate this model by introducing bi-directional probing evaluation. In the second chapter, we design a model which induces word-like units from a sequence of characters by integrating non-parametric variational information bottleneck in the last layers of a transformer encoder. In the next chapter, we move to the multi-modal domain and starting from subwords, we design a model which induces phrases from image captions by aligning them to the objects in the image. Lastly, we explore a task-driven approach towards inducing entities.