Luigi Libero Lucio Starace

Sergio Di Meglio, Aniello Somma, Luigi Libero Lucio Starace et al.

Online property booking platforms are widely used and rely heavily on consistent, up-to-date information about accommodation facilities, often sourced from third-party providers. However, these external data sources are frequently affected by incomplete or inconsistent details, which can frustrate users and result in a loss of market. In response to these challenges, we present an industrial case study involving the integration of Large Language Models (LLMs) into CALEIDOHOTELS, a property reservation platform developed by FERVENTO. We evaluate two well-known LLMs in this context: Mistral 7B, fine-tuned with QLoRA, and Mixtral 8x7B, utilized with a refined system prompt. Both models were assessed based on their ability to generate consistent and homogeneous descriptions while minimizing hallucinations. Mixtral 8x7B outperformed Mistral 7B in terms of completeness (99.6% vs. 93%), precision (98.8% vs. 96%), and hallucination rate (1.2% vs. 4%), producing shorter yet more concise content (249 vs. 277 words on average). However, this came at a significantly higher computational cost: 50GB VRAM and $1.61/hour versus 5GB and $0.16/hour for Mistral 7B. Our findings provide practical insights into the trade-offs between model quality and resource efficiency, offering guidance for deploying LLMs in production environments and demonstrating their effectiveness in enhancing the consistency and reliability of accommodation data.

Anna Corazza, Sergio Di Martino, Adriano Peron et al.

In the context of End-to-End testing of web applications, automated exploration techniques (a.k.a. crawling) are widely used to infer state-based models of the site under test. These models, in which states represent features of the web application and transitions represent reachability relationships, can be used for several model-based testing tasks, such as test case generation. However, current exploration techniques often lead to models containing many near-duplicate states, i.e., states representing slightly different pages that are in fact instances of the same feature. This has a negative impact on the subsequent model-based testing tasks, adversely affecting, for example, size, running time, and achieved coverage of generated test suites. As a web page can be naturally represented by its tree-structured DOM representation, we propose a novel near-duplicate detection technique to improve the model inference of web applications, based on Tree Kernel (TK) functions. TKs are a class of functions that compute similarity between tree-structured objects, largely investigated and successfully applied in the Natural Language Processing domain. To evaluate the capability of the proposed approach in detecting near-duplicate web pages, we conducted preliminary classification experiments on a freely-available massive dataset of about 100k manually annotated web page pairs. We compared the classification performance of the proposed approach with other state-of-the-art near-duplicate detection techniques. Preliminary results show that our approach performs better than state-of-the-art techniques in the near-duplicate detection classification task. These promising results show that TKs can be applied to near-duplicate detection in the context of web application model inference, and motivate further research in this direction.