Model evaluation in computational historical linguistics

Gerhard Jäger

doi:10.1515/tl-2019-0020

Abstract

Abstract This is a reply to the comments by Hammarström et al. (This volume) and List (This volume) on the target article Computational Historical Linguistics (This volume). There I proposed several methodological principles for research in Computational Historical Linguistics pertaining to suitable techniques for model fitting and model evaluation. Hammarström et al. debate the usefulness of these principles, and List proposes a novel evaluation measure specifically aimed at the task of proto-form reconstruction. This reply will focus on the role of model evaluation in our field.

Highlights

A major motivation for writing the target article was to initiate a debate within the research community about standards of model fitting and model evaluation in Computational Historical Linguistics (CHL)
Suffice it to say that we are still quite far from a full understanding of the causal factors involved in the data generating process in phylogenetic inference, and even further from adequately capturing these factors in an explanatory probabilistic model
One of the central points I wanted to make in the target article was that evaluating computational and statistical models in historical linguists should be a central concern in our research practice

Summary

Introduction

A major motivation for writing the target article was to initiate a debate within the research community about standards of model fitting and model evaluation in Computational Historical Linguistics (CHL). The empirically observed value clearly falls outside the distribution of possible values predicted by the model being used for phylogenetic inference (see Figure 1) This result emphasizes Hammarström et al.’s point that the pattern one word per concept that holds for the training data is not predicted to hold by the statistical model used. Suffice it to say that we are still quite far from a full understanding of the causal factors involved in the data generating process in phylogenetic inference (and other aspects of CHL), and even further from adequately capturing these factors in an explanatory probabilistic model As long as this is the case, predictive checks are an important tool to test in what respects our preliminary models perform well, and in what respects they do not. The method used in the pilot study identifies the likeliest proto-sound in each column of the progressive alignment instead

Reply to List

Conclusion