An Investigation of the Relationship Between Automated Machine Translation Evaluation Metrics and User Performance on an Information Extraction Task

Abstract

This dissertation applies nonparametric statistical techniques toMachine Translation (MT) Evaluation using data from a MT Evaluation experiment conducted through a joint Army Research Laboratory (ARL) and Center for the Advanced Study of Language (CASL) project. In particular, the relationship between human task performance on an information extraction task with translated documents and well-known automated translation evaluation metric scores for those documents is studied. Findings from a correlation analysis of the connection between autometrics and task-based metrics are presented and contrasted with current strategies for evaluating translations. A novel idea for assessing partial rank correlation within the presence of grouping factors is also introduced. Lastly, this dissertation presents a framework for task-based machine translation (MT) evaluation and predictive modeling of task responses that gives new information about the relative predictive strengths of the different autometrics (and re-coded variants of them) within the statistical Generalized Linear Models developed in analyses of the Information Extraction Task data. This work shows that current autometrics are inadequate with respect to the prediction of task performance but, near adequacy can be accomplished through the use of re-coded autometrics in a logistic regression setting. As a result, a class of automated metrics that are best suitable for predicting performance is established and suggestions are offered about how to utilize metrics to supplement expensive and time-consuming experiments with human participants. Now users can begin to tie the intrinsic automated metrics to the extrinsic metrics for task they perform. The bottom line is that there is a need to average away MT dependence (averaged metrics perform better in overall predictions than original autometrics). Moreover, combinations of recoded metrics performed better than any individual metric. Ultimately, MT evaluation methodology is extended to create new metrics specially relevant to task-based comparisons. A formal method to establish that differences among metrics as predictors are strong enough not to be due by chance remains as future work. Given the lack of connection in the field of MT Evaluation between task utility and the interpretation of automated evaluation metrics, as well as the absence of solid statistical reasoning in evaluating MT, there is a need to bring innovative and interdisciplinary analytical techniques to this problem. Because there are no papers in the MT evaluation literature that have done statistical modeling before or that have linked automated metrics with how well MT supports human tasks, this work is unique and has high potential for benefiting the Machine Translation research community.