This thesis focuses on improving the pre-training of natural language models using unsupervised raw data to make them more efficient and aligned with downstream applications. In the first part, we introduce three alternative pre-training objectives to BERT's Masked Language Modeling (MLM), namely Random Token Substitution (RTS), Cluster-based Random Token Substitution (C-RTS), and Swapped Language Modeling (SLM). These objectives involve token swapping instead of masking, with RTS and C-RTS aiming to predict token originality and SLM predicting the original token values. Results show that RTS and C-RTS require less pre-training time while maintaining performance comparable to MLM. Surprisingly, SLM outperforms MLM on certain tasks despite using the same computational budget. In the second part, we proposes self-supervised pre-training tasks that align structurally with downstream applications, reducing the need for labeled data. We use large corpora like Wikipedia and CC-News to train models to recognize if text spans originate from the same paragraph or document in several ways. By doing continuous pre-training, starting from existing models like RoBERTa, ELECTRA, DeBERTa, BART, and T5, we demonstrate significant performance improvements in tasks like Fact Verification, Answer Sentence Selection, and Summarization. These improvements are especially pronounced when limited annotation data is available. The proposed objectives also achieve state-of-the-art results on various benchmark datasets, including FEVER (dev set), ASNQ, WikiQA, and TREC-QA, as well as enhancing the quality of summaries. Importantly, these techniques can be easily integrated with other methods without altering the internal structure of Transformer models, making them versatile for various NLP applications.

Many important classification problems, such as object classification, speech recognition, and machine translation, have been tackled by the supervised learning paradigm in the past, where training corpora of parallel input-output pairs are required with high cost. To remove the need for the parallel training corpora has practical significance for real-world applications, and it is one of the main goals of unsupervised learning. Recently, encouraging progress in unsupervised learning for solving such classification problems has been made and the nature of the challenges has been clarified. In this article, we review this progress and disseminate a class of promising new methods to facilitate understanding the methods for machine learning researchers. In particular, we emphasize the key information that enables the success of unsupervised learning - the sequential statistics as the distributional prior in the labels. Exploitation of such sequential statistics makes it possible to estimate parameters of classifiers without the need of paired input-output data. In this paper, we first introduce the concept of Caesar Cipher and its decryption, which motivated the construction of the novel loss function for unsupervised learning we use throughout the paper. Then we use a simple but representative binary classification task as an example to derive and describe the unsupervised learning algorithm in a step-by-step, easy-to-understand fashion. We include two cases, one with Bigram language model as the sequential statistics for use in unsupervised parameter estimation, and another with a simpler Unigram language model. For both cases, detailed derivation steps for the learning algorithm are included. Further, a summary table compares computational steps of the two cases in executing the unsupervised learning algorithm for learning binary classifiers.