Towards Causality Extraction from Requirements

TL;DR

This paper introduces a new NLP approach using Tree Recursive Neural Networks to extract causal relations from requirement texts, supported by a large annotated dataset to facilitate training and benchmarking.

Contribution

It presents a novel TRNN-based architecture for causality extraction and provides a large annotated dataset for training and evaluating such models.

Findings

Developed a TRNN architecture for causality extraction from requirements.

Created a dataset with over 212,000 sentences for training and benchmarking.

Encouraged community contribution to improve causality annotation.

Abstract

System behavior is often based on causal relations between certain events (e.g. If event1, then event2). Consequently, those causal relations are also textually embedded in requirements. We want to extract this causal knowledge and utilize it to derive test cases automatically and to reason about dependencies between requirements. Existing NLP approaches fail to extract causality from natural language (NL) with reasonable performance. In this paper, we describe first steps towards building a new approach for causality extraction and contribute: (1) an NLP architecture based on Tree Recursive Neural Networks (TRNN) that we will train to identify causal relations in NL requirements and (2) an annotation scheme and a dataset that is suitable for training TRNNs. Our dataset contains 212,186 sentences from 463 publicly available requirement documents and is a first step towards a gold standard corpus for causality extraction. We encourage fellow researchers to contribute to our dataset and help us in finalizing the causality annotation process. Additionally, the dataset can also be annotated further to serve as a benchmark for other RE-relevant NLP tasks such as requirements classification.