Recently, significant progress has been made in 2D human pose estimation. While some research has focused on enhancing the accuracy of keypoint detection, others have aimed at reducing model size. However, most models excel in either one aspect or the other, but rarely both simultaneously. In this paper, we address the challenge of balancing accuracy and inference speed. Inspired by large-kernel convolutions and attention mechanisms, we introduce LKConvPose, a hybrid CNN architecture that achieves high keypoint detection accuracy with low computational cost. Specifically, LKConvPose-S attains 76.5 AP on the COCO validation dataset using only 4 GFLOPs, making it the most efficient model at its scale.

Graphical user interface (GUI) prototyping helps to clarify requirements and keep stakeholders engaged in software development. While contemporary approaches retrieve GUIs relevant to a user’s query, little support exists for the actual reuse, i.e., for using an existing GUI to create a new one. To shorten the gap, we investigate GUI-centered reuse via one of the latest artificial intelligence (AI) techniques—vision-language models (VLMs). We report an empirical study involving 73 university students working on ten GUI reuse tasks. Each task is associated with different reuse directions recommended by VLMs and by a natural language (NL) method. In addition, a focused GUI element is provided to offer a starting point for making the actual changes. Our results show that VLMs significantly outperform the NL method in making reuse recommendations, but surprisingly, the focused GUI elements are not consistently modified during reuse. With the assessments made by four experienced designers, we further offer insights into the creativity of human-reuse and AI-reuse results.

Context: As mobile applications (apps) widely spread throughout our society and daily life, various personal information is constantly demanded by apps in exchange for more intelligent and customized functionality. An increasing number of users are voicing their privacy concerns through app reviews on app stores. Objective: The main challenge of effectively mining privacy concerns from user reviews lies in that reviews expressing privacy concerns are overridden by a large number of reviews expressing more generic themes and noisy content. In this work, we propose a novel automated approach to overcome that challenge. Method: Our approach first employs information retrieval and document embeddings to extract candidate privacy reviews in an unsupervised manner , which are further labeled to prepare the annotation dataset. Then, supervised classifiers are trained to automatically identify privacy reviews. Finally, an interpretable topic mining algorithm is designed to detect privacy concern topics contained in the privacy reviews. Results: Experimental results show that the best performing document embedding achieves an average precision of 96.80% in the top 100 retrieved candidate privacy reviews, outperforming the taxonomy-based baseline, which achieves 73.87%. All trained privacy review classifiers achieve an F 1 score above 91%, surpassing the keyword-matching baseline by as much as 7.5% and the large language model baseline by up to 2.74%. For detecting privacy concern topics from privacy reviews, our proposed algorithm achieves both better topic coherence and topic diversity than three strong topic modeling baselines, including LDA . Conclusion: Empirical evaluation results demonstrate the effectiveness of our approach in identifying privacy reviews and detecting user privacy concerns in app reviews.

Creativity focuses on the generation of novel and useful ideas. In this paper, we propose an approach to automatically generating creative requirements candidates via the adversarial examples resulted from applying small changes (perturbations) to the original requirements descriptions. We present an architecture where the perturbator and the classifier positively influence each other. Meanwhile, we ensure that each adversarial example is uniquely traceable to an existing feature of the software, instrumenting explainability. Our experimental evaluation of six datasets shows that around 20% adversarial shift rate is achievable. In addition, a human subject study demonstrates our results are more clear, novel, and useful than the requirements candidates outputted from a state-of-the-art machine learning method. To connect the creative requirements closer with software development, we collaborate with a software development team and show how our results can support behavior-driven development for a web app built by the team.

Requirements terminology defines and unifies key specialized and/or technical concepts of the software system, which is significant for understanding the application domain in requirements engineering (RE). However, manual terminology extraction from natural language requirements is laborious and expensive, especially with large scale requirements specifications. In this paper, we aim to employ natural language processing (NLP) techniques and machine learning (ML) algorithms to automatically extract and rank the requirements terms to support high-level feature modeling. To this end, we propose an automatic framework composed of noun phrase identification technique for requirements terms extraction and TextRank combined with semantic similarity for terms ranking. The final ranked terms are organized as a hierarchy, which can be used to help name elements when performing feature modeling. In the quantitative evaluation, our extraction method performs better than three baseline methods in recall with comparable precision. Moreover, our adapted TextRank algorithm can rank more relevant terms at the top positions in terms of average precision compared with most baselines. An illustrative example on the smart home domain further shows the usefulness of our framework in aiding elements naming during feature modeling. The research results suggest that proper adoption and adaption of NLP and ML techniques according to the characteristics of specific RE task could provide automation support for problem domain understanding.

In previous similarity-based WSD systems, studies have allocated much effort on learning comprehensive sense embeddings using contextual representations and knowledge sources.However, the context embedding of an ambiguous word is learned using only the sentence where the word appears, neglecting its global context.In this paper, we investigate the contribution of both word-level and senselevel global context of an ambiguous word for disambiguation.Experiments have shown that the Context-Oriented Embedding (COE) can enhance a similarity-based system's performance on WSD by relatively large margins, achieving state-of-the-art on all-words WSD benchmarks in knowledge-based category.

This paper aims at demonstrating non-functional requirements analysis requirements analysis normally used supervised methods which need a lot of manual annotation work. Using unsupervised approaches to classify non-functional requirements can save a lot of labor and time, but the accuracy of the existing approaches is relatively low. In order to solve the dilemma, we propose a new clustering approach in this paper. The approach is an improved version of the previous aspect segmentation approach, but differs in terms of classification strategy, the representation of the review sentences, and the strategy for selecting new keywords. Experiments are conducted and compared on a software reviews dataset. Results show an improved performance of the new approach.