Made of Steel? Learning Plausible Materials for Components in the Vehicle Repair Domain

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1420–1435 May 2-6, 2023 ©2023 Association for Computational Linguistics Made of Steel? Learning Plausible Materials for Components in the Vehicle Repair Domain Annerose Eichel, Helena Schlipf, Sabine Schulte im Walde Institute for Natural Language Processing, University of Stuttgart {annerose.eichel,schulte}@ims.uni-stuttgart.de, helena.schlipf@gmail.com Abstract We propose a novel approach to learn domain- specific plausible materials for components in the vehicle repair domain by probing Pre- trained Language Models (PLMs) in a cloze task style setting to overcome the lack of an- notated datasets. We devise a new method to aggregate salient predictions from a set of cloze query templates and show that domain- adaptation using either a small, high-quality or a customized Wikipedia corpus boosts per- formance. When exploring resource-lean al- ternatives, we find a distilled PLM clearly out- performing a classic pattern-based algorithm. Further, given that 98% of our domain-specific components are multiword expressions, we suc- cessfully exploit the compositionality assump- tion as a way to address data sparsity. 1 Introduction Connecting a symptom to an underlying cause is a crucial building block for natural language un- derstanding across domains. For example, as il- lustrated in Fig. 1, a standard approach that hu- man mechanics in vehicle repair shops apply when tracing the cause of a symptom, is to exploit the link between a vehicle component and the com- ponent’s materials. This study tackles the task of automatically learning plausible materials for ve- hicle components. The information is crucial in vehicle repair shops, as mechanics are faced with constantly growing vehicle complexity, making it hard to manually identify the cause of a malfunc- tion. While domain-specific information on vehicle components is often available, plausible domain- specific material information typically needs to be gathered from external data sources. For example, from “Brake disk are usually manufactured from gray cast iron”, one may retrieve the information that brake disks consist of iron (and possibly further materials not mentioned here). We propose a novel approach to leverage state- of-the-art pretrained language models (PLMs) for Figure 1: Conceptual overview for learning domain- specific plausible materials (e.g., aluminium) for vehicle components (e.g., engine valves rocker arm). the task of learning plausible materials in highly domain-specific contexts, such as vehicle compo- nents in the vehicle repair domain. Our approach can also be applied to similarly specific domains, e.g., to learn which materials or substances are plausibly used in products in the health domain, textile industry, etc. As we tackle a task from ma- terial science, we make a contribution to a field that is not only very challenging, but heavily under- investigated from an NLP perspective. We focus on the PLM RoBERTa (Liu et al., 2019) and develop an approach that successfully learns plausible ma- terials in English for components in the vehicle re- pair domain. We further exploit domain-adaptation variants and show (i) that harnessing a small high- quality domain-specific corpus boosts the perfor- mance over an out-of-the-box vanilla RoBERTa, and …

Content truncated to 500 words