NLI Shared Task 2013: MQ Submission

Our submission for this NLI shared task used for the most part standard features found in recent work. Our focus was instead on two other aspects of our system: at a high level, on possible ways of constructing ensembles of multiple classifiers; and at a low level, on the granularity of part-of-speech tags used as features. We found that the choice of ensemble combination method did not lead to much difference in results, although exploiting the varying behaviours of linear versus logistic regression SVM classifiers could be promising in future work; but part-of-speech tagsets showed noticeable differences. We also note that the overall architecture, with its feature set and ensemble approach, had an accuracy of 83.1% on the test set when trained on both the training data and development data supplied, close to the best result of the task. This suggests that basically throwing together all the features of previous work will achieve roughly the state of the art.