Minimax-Based Feature Selection for Low-Resource Language Classification Theoretical Research Paper with A Worked Numerical Illustration

Abstract

Character and word features are commonly large sets that are used in language classification systems. If the resources are restricted in languages that have limited amount of labeled data, however, a feature set that works well for the overall corpus may work poorly for a small language subset or an unfamiliar domain of text. A minimax approach is proposed in this paper to choose a small set of features for the identification of a target language. Large validation loss is computed for a group of languages and domains, as opposed to the average loss. The framework is based on a feature budget constraint, regularized classifier training and group-specific evaluation. An equivalent adversarial weighting interpretation is responsible for the method's attention to poorly served groups. A finite numerical illustration compares six candidate feature subsets and illustrates how a compromise between the average performance and worst group protection can be achieved. Also included is a practical selection procedure and an empirical evaluation protocol. The paper is methodological: There is no report of any corpus experiment, and all the numbers lost in the illustration are assumed ones. The framework presents a testable method for compact language classification in limited resources, and highlights the importance of the robustness of group estimates and the design of validation.