Model-based Specification and Analysis of Natural Language Requirements in the Financial Domain

Software requirements form an important part of the software development process. In many software projects conducted by companies in the financial sector, analysts specify software requirements using a combination of models and natural language (NL). Neither models nor NL requirements provide a complete picture of the information in the software system, and NL is highly prone to quality issues, such as vagueness, ambiguity, and incompleteness. Poorly written requirements are difficult to communicate and reduce the opportunity to process requirements automatically, particularly the automation of tedious and error-prone tasks, such as deriving acceptance criteria (AC). AC are conditions that a system must meet to be consistent with its requirements and be accepted by its stakeholders. AC are derived by developers and testers from requirement models. To obtain a precise AC, it is necessary to reconcile the information content in NL requirements and the requirement models.

In collaboration with an industrial partner from the financial domain, we first systematically developed and evaluated a controlled natural language (CNL) named Rimay to help analysts write functional requirements. We then proposed an approach that detects common syntactic and semantic errors in NL requirements. Our approach suggests Rimay patterns to fix errors and convert NL requirements into Rimay requirements. Based on our results, we propose a semiautomated approach that reconciles the content in the NL requirements with that in the requirement models. Our approach helps modelers enrich their models with information extracted from NL requirements. Finally, an existing test-specification derivation technique was applied to the enriched model to generate AC.

The first contribution of this dissertation is a qualitative methodology that can be used to systematically define a CNL for specifying functional requirements. This methodology was used to create Rimay, a CNL grammar, to specify functional requirements. This CNL was derived after an extensive qualitative analysis of a large number of industrial requirements and by following a systematic process using lexical resources. An empirical evaluation of our CNL (Rimay) in a realistic setting through an industrial case study demonstrated that 88% of the requirements used in our empirical evaluation were successfully rephrased using Rimay.

The second contribution of this dissertation is an automated approach that detects syntactic and semantic errors in unstructured NL requirements. We refer to these errors as smells. To this end, we first proposed a set of 10 common smells found in the NL requirements of financial applications. We then derived a set of 10 Rimay patterns as a suggestion to fix the smells. Finally, we developed an automatic approach that analyzes the syntax and semantics of NL requirements to detect any present smells and then suggests a Rimay pattern to fix the smell. We evaluated our approach using an industrial case study that obtained promising results for detecting smells in NL requirements (precision 88%) and for suggesting Rimay patterns (precision 89%).

The last contribution of this dissertation was prompted by the observation that a reconciliation of the information content in the NL requirements and the associated models is necessary to obtain precise AC. To achieve this, we define a set of 13 information extraction rules that automatically extract AC-related information from NL requirements written in Rimay. Next, we propose a systematic method that generates recommendations for model enrichment based on the information extracted from the 13 extraction rules. Using a real case study from the financial domain, we evaluated the usefulness of the AC-related model enrichments recommended by our approach. The domain experts found that 89% of the recommended enrichments were relevant to AC, but absent from the original model (precision of 89%).