Natural Language Inference (NLI) research involves the development of models that can mimic human inference processes based on natural language and classify the inference relation between sentences. For example, given the premise that ``In 2019, the Raptors won their first Eastern Conference title, and the team's first NBA Finals", it follows that ``The Raptors beat another team in the 2019 NBA Finals". but it does not follow that ``The Golden State Warriors won the last game of the NBA Finals in 2019".The goal of NLI is to build machines that can take pairs of premise and hypothesis as input and correctly predict the inference relation between them, reverse-engineering the inference process of a human. NLI is a fundamental task with a simple and generic formalization such that NLI models can be practically useful in all kinds of NLP applications. In recent years, there has been emerging interest and research in data-driven natural language inference.This thesis starts with several key applications of data-driven NLI modules, including sentence-based NLI modeling, how to effectively use the NLI model as a key natural language understanding (NLU) module in both an automatic fact-checking system for claim verification and in an open-domain dialogue system for improving dialogue consistency. Empirical results not only demonstrate valuable use cases of NLI models in NLP applications but, more importantly, reveal the fact that the data is a key factor that contributes to the success of the usage of NLI models. That leads to the second part of this thesis, namely, adversarial NLI, a research endeavor that embodies a dynamic human-and-model-in-the-loop learning paradigm for NLI via competitive iterations between model training and crowd-sourcing to push the limit of NLU.