Shwetali Shimangaud

Shwetali Shimangaud, Lola Burgueño, Rijul Saini et al.

Context: Having domain models derived from textual specifications has proven to be very useful in the early phases of software engineering. However, creating correct domain models and establishing clear links with the textual specification is a challenging task, especially for novice modelers. Objectives: We propose an approach for determining the alignment between a partial domain model and a textual specification. Methods: To this aim, we use Natural Language Processing techniques to pre-process the text, generate an artificial natural language specification for each model element, and then use an LLM to compare the generated description with matched sentences from the original specification. Ultimately, our algorithm classifies each model element as either aligned (i.e., correct), misaligned (i.e., incorrect), or unclassified (i.e., insufficient evidence). Furthermore, it outputs the related sentences from the textual specification that provide the evidence for the determined class. Results: We have evaluated our approach on a set of examples from the literature containing diverse domains, each consisting of a textual specification and a reference domain model, as well as on models containing modeling errors that were systematically derived from the correct models through mutation. Our results show that we are able to identify alignments and misalignments with a precision close to 1 and a recall of approximately 78%, with execution times ranging from 18 seconds to 1 minute per model element. Conclusion: Since our algorithm almost never classifies model elements incorrectly, and is able to classify over 3/4 of the model elements, it could be integrated into a modeling tool to provide positive feedback or generate warnings, or employed for offline validation and quality assessment.

Sowmya Vajjala, Shwetali Shimangaud

Large Language Models revolutionized NLP and showed dramatic performance improvements across several tasks. In this paper, we investigated the role of such language models in text classification and how they compare with other approaches relying on smaller pre-trained language models. Considering 32 datasets spanning 8 languages, we compared zero-shot classification, few-shot fine-tuning and synthetic data based classifiers with classifiers built using the complete human labeled dataset. Our results show that zero-shot approaches do well for sentiment classification, but are outperformed by other approaches for the rest of the tasks, and synthetic data sourced from multiple LLMs can build better classifiers than zero-shot open LLMs. We also see wide performance disparities across languages in all the classification scenarios. We expect that these findings would guide practitioners working on developing text classification systems across languages.