Automatic Analysis of Flaws in Pre-Trained NLP Models

Most tools for natural language processing today are based on machine learning and come with pre-trained models. In addition, third-parties provide pre-trained models for popular NLP tools. The predictive power and accuracy of these tools depends on the quality of these models. Downstream researchers often base their results on pre-trained models instead of training their own. Consequently, pre-trained models are an essential resource to our community. However, to be best of our knowledge, no systematic study of pre-trained models has been conducted so far. This paper reports on the analysis of 274 pre-models for six NLP tools and four potential causes of problems: encoding, tokenization, normalization and change over time. The analysis is implemented in the open source tool Model Investigator. Our work 1) allows model consumers to better assess whether a model is suitable for their task, 2) enables tool and model creators to sanity-check their models before distributing them, and 3) enables improvements in tool interoperability by performing automatic adjustments of normalization or other pre-processing based on the models used.