Antiquated technology plagues the legal industry. While the legal market generates billions of dollars of revenue annually, the technology that supports it is decentralized, inconsistent and outdated. With the purpose of promoting modernization and illustrating how technology could improve existing inefficiencies in legal research, this thesis evaluates the creation and execution of an automated lawsuit classification system called the Civil Lawsuit Analysis System ("CLAS"). CLAS classifies the nature of a lawsuit based on information contained in a subset of attorney-authored court documents called "Memorandum of Points and Authorities" that are filed with trial courts during the process of the lawsuit. In light of the fact that no similar automated classification system exists today, developing the process by which CLAS operates proved to be a collaborative and time-intensive process that ultimately required specialized knowledge in both the computer science and the legal fields.

Human communication has features, such as syntax, unseen in any other form of animal communication. How did we come to use such a sophisticated form of communication? This paper addresses the issue of the origin of compositionality in languages using animat-based modeling. In the simulation, 100 software agents controlled by neural networks are given the ability to communicate with each other using sequences of signals. These agents---called animats---play communication games with each other, and if they are among the most effective communicators, they mate and produce offspring. Through evolution and learning, animats developed languages that exhibit some rudimentary compositionality in simple environment settings, but under more sophisticated environments, they struggled to establish effective protocols of communication. Results suggested that simple recurrent neural networks allowed animats to develop languages with rudimentary compositionality, and that imitation learning helped animats come to consensus on usage of signals.

Natural language is highly ambiguous, with the same word having different meanings depending on the context. While human readers often have no trouble interpreting the correct meaning, semantic ambiguity poses a significant problem for many natural language systems, such as those that translate text or perform machine reading. The task of identifying which meaning of a word is present in a given context is known as Word Sense Disambiguation (WSD), where a word's meanings are discretized into units referred to as senses. Because languages contain hundreds of thousand of unique words and each of those words can have multiple meanings, comprehensive sense-annotated corpora are often sparse, with only tens to low-hundreds of annotated examples of each word. As a result, creating high performance WSD systems requiring overcoming this data sparsity.

This thesis provides a three-fold approach to improving WSD performance in the face of data sparsity. First, we introduce two new algorithms that take the role of a lexicographer and automatically learn the senses of a word from example uses in a fully unsupervised way. We then demonstrate that these unsupervised systems can be combined with a limited amount of annotated data to create a semi-supervised WSD system that significantly outperforms a state-of-the-art supervised WSD system trained on the same data. Second, we propose a novel method for gathering high-quality sense annotations from large numbers of untrained, online workers, commonly referred to as crowdsourcing. Our method lowers the time and cost of building sense-annotated corpora, while maintaining as high a level of agreement between annotators, comparable with that of trained experts. Third, we analyze cases of ambiguity in sense annotations, when two annotators differ about which sense best describes the meaning of a particular usage of a word. To perform this analysis, we built the largest sense-annotated corpus where cases of semantic ambiguity are explicitly marked. Our analysis of this corpus revealed multiple causes for this ambiguity as well as how the ambiguity may be interpreted and resolved by natural language applications using ambiguous data. To complement this work on ambiguity, we have also introduced a new methodology for evaluating WSD systems that explicitly report ambiguous instances.