T-FREX: A Transformer-based Feature Extraction Method from Mobile App Reviews - Abstract - Transformer-based feature extraction from mobile app reviews using LLMs

Oct 28, 2026

Mobile app reviews serve as large-scale software knowledge sources, enabling maintenance, evolution, and feedback analysis. Effective feature extraction—identifying app functionalities or characteristics—is essential for assessing feature acceptance, discovering feature requests, and prioritizing development. Traditional syntactic pattern methods are often context-agnostic and hard to replicate. T-FREX proposes an automatic Transformer-based approach that builds a ground-truth dataset, fine-tunes LLMs for token classification, evaluates against truth and baselines, and validates new feature quality via human assessment.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Introduction

Background and Motivation

Challenges in Feature Mining

Transformer-Based LLM Approach

Method: T-FREX

Feature Extraction as NER

Dataset Creation and Fine-Tuning

Evaluation Strategy

T-FREX: A Transformer-based Feature Extraction Method from Mobile App Reviews Quim Motger 1 , Alessio Miaschi2 , Felice Dell’Orletta2 , Xavier Franch 1 , Jordi Marco 1 1Universitat Politècnica de Catalunya, Barcelona {quim.motger,xavier.franch, [jordi.marco}@upc.edu](jordi.marco}@upc.edu) 2Institute for Computational Linguistics "A. Zampolli" (CNR-ILC), ItaliaNLP Lab, Pisa {alessio.miaschi, [felice.dellorletta}@ilc.cnr.it](felice.dellorletta}@ilc.cnr.it) Abstract—Mobile app reviews are a large-scale data source for software-related knowledge generation activities, including software maintenance, evolution and feedback analysis. Effective extraction of features (i.e., functionalities or characteristics) from these reviews is key to support analysis on the acceptance of these features, identification of relevant new feature requests and prioritization of feature development, among others. Traditional methods focus on syntactic pattern-based approaches, typically context-agnostic, evaluated on a closed set of apps, difficult to replicate and limited to a reduced set and domain of apps. Meanwhile, the pervasiveness of Large Language Models (LLMs) based on the Transformer architecture in software engineering tasks lays the groundwork for empirical evaluation of the performance of these models to support feature extraction. In this study, we present T-FREX, a Transformer-based, fully automatic approach for mobile app review feature extraction. First, we collect a set of ground truth features from users in a real crowdsourced software recommendation platform and transfer them automatically into a dataset of app reviews. Then, we use this newly created dataset to fine-tune multiple LLMs on a named entity recognition task under different data configurations. We assess the performance of T-FREX with respect to this ground truth, and we complement our analysis by comparing T-FREX with a baseline method from the field. Finally, we assess the quality of new features predicted by T-FREX through an external human evaluation. Results show that T-FREX outperforms on average the traditional syntacticbased method, especially when discovering new features from a domain for which the model has been fine-tuned. Index Terms—feature extraction, mobile apps, reviews, token classification, named entity recognition, large language models I. INTRODUCTION Mobile app repositories provide valuable access to timely large-scale datasets of software-related information [1] . These repositories include heterogeneous, multiple-purpose platforms, from app stores to sideloading repositories and search engines [2] . One of the most popular contributions across these platforms is the publication of app reviews, in which users express multiple facets such as personal opinions or experiences, bug reports, inquiries or requests [3] . This information is relevant to multiple software engineering processes, including requirements elicitation and prioritization, release planning, validation analysis and software evolution [3]–[7] . App features are considered a core descriptor for understanding and categorizing app reviews [8]–[10] . In this context, a feature is considered as a distinct function or capability within a mobile application serving a particular purpose or need [11] . App feature extraction supports featurerelated knowledge generation, in which mobile app developers can potentially rely on to improve user experience, enhance app functionality, identify user preferences, and make datadriven decisions for app development strategies [3],[12],[13] . Consequently, mining large amounts of app reviews to extract app features has become a relevant task. Nevertheless, mining features from app reviews presents particular challenges. It requires the daily analysis of thousands of short documents, each with a limited length, composed of an average of a few dozen words per review [14] . Beyond measurable characteristics, user-generated documents tend t

FAQs

What problem does T-FREX address?

T-FREX addresses automatic extraction of app features from mobile app reviews by re-framing feature extraction as a named entity recognition-style token classification task.

How is training data created for T-FREX?

Ground truth features are collected from users in a real crowdsourced software recommendation platform and transferred into a dataset of app reviews for model fine-tuning.

How is the quality of predicted new features validated?

Predicted new features are assessed through an external human evaluation and compared against the ground truth and a baseline method from the field.

Oct 28, 2026

Upload by Quinn Holloway