Giannis Tzimas

We introduce the String Ensemble Chords Dataset (SECD), a large-scale controlled compositional audio dataset comprising 287,088 harmonic-interval and chord instances constructed through additive superposition of professionally recorded isolated string notes from the Philharmonia Orchestra into duo-, trio-, and quartet-like four-voice mixtures. Each mixture includes exact per-voice metadata for absolute pitch, dynamic marking, and playing technique, and the corpus is organised into six dataset groups covering harmonic intervals, triads, and seventh chords under loose and strict metadata-consistency conditions. To demonstrate dataset utility, we define four representative and reproducible reference benchmarks: ensemble size recognition, triad chord quality identification, per-instrument dynamics classification, and playing technique-family recognition. Baseline Audio Spectrogram Transformer (AST) models achieve test accuracies of 98.67%, 93.73%, 98.19%, and 99.39%, with corresponding macro-F1 scores of 98.64%, 93.73%, 98.01%, and 97.29%, under a complete-instance-disjoint, in-domain evaluation protocol. These results provide reproducible reference performance for the selected SECD tasks and demonstrate the corpus’s utility for controlled analysis of harmonic, timbral, dynamic, and textural attributes in classical string audio. The full SECD corpus is released through Zenodo as constructed audio mixtures with accompanying metadata, while the project GitHub repository provides the EXP1–EXP4 benchmark code, saved split definitions, and the mini-SECD demonstration package for lightweight reproducibility.

This paper addresses the challenge of deduplicating job postings in large, heterogeneous datasets by introducing an efficient, multi-stage methodology that combines embedding-based filtering with Large Language Model (LLM) validation. The proposed system begins with data preprocessing and semantic vectorization of key textual fields using a text embedding model. To reduce the computational cost of exhaustive pairwise comparisons, a clustering-based grouping mechanism is employed to restrict comparisons to semantically coherent clusters. Candidate duplicates are then validated using LLMs, which assess semantic equivalence across highlighted differences in job titles, descriptions, companies, and locations. The proposed system is evaluated on an augmented dataset of 50,000 job postings, producing 6669 candidate pairs for validation. Among the evaluated models, GPT-4o achieved the highest F1-score (95.10%), while the lightweight Phi-4 model demonstrated competitive performance (92.58%) with significantly lower computational cost. These findings demonstrate that the proposed hybrid framework achieves high semantic precision while remaining scalable for continuous large-scale deployment.

In this article, we present a tool for matching resumes to job posts and vice versa (job post to resumes). With minor modifications, it may also be adapted to other domains where text matching is necessary. This tool may help organizations save time during the hiring process, as well as assist applicants by allowing them to match their resumes to job posts they have selected. To achieve text matching without any model training (zero-shot matching), we constructed dynamic structured prompts that consisted of unstructured and semi-structured job posts and resumes based on specific criteria, and we utilized the Chain of Thought (CoT) technique on the Mistral model (open-mistral-7b). In response, the model generated structured (segmented) job posts and resumes. Then, the job posts and resumes were cleaned and preprocessed. We utilized state-of-the-art sentence similarity models hosted on Hugging face (nomic-embed-text-v1-5 and google-embedding-gemma-300m) through inference endpoints to create sentence embeddings for each resume and job post segment. We used the cosine similarity metric to determine the optimal matching, and the matching operation was applied to eleven different occupations. The results we achieved reached up to 87% accuracy for some of the occupations and underscore the potential of zero-shot techniques in text matching utilizing LLMs. The dataset we used was from indeed.com, and the Spring AI framework was used for the implementation of the tool.

This study presents a method for detecting and removing duplicate job postings in large datasets with emphasis on key attributes such as job title, location, company name, and job description. The approach begins with a preprocessing phase that standardizes text data—normalizing formats, removing special characters, and resolving lexical variations—to ensure consistency and compatibility. For deduplication, we utilize WordLlama, a fast and lightweight NLP toolkit optimized for fuzzy deduplication and similarity detection. Furthermore, we evaluate the performance of various Large Language Models (LLM) in identifying duplicates, measuring accuracy through precision and recall metrics. The objective is to determine which model best captures semantic similarities in job postings and achieves the highest deduplication accuracy. This comparison offers valuable insights into the effectiveness of LLMs for large-scale, text-based deduplication in the context of job postings.

Interval Time Series are present in everyday life. An example is the opening and closing value of some stock in a market for certain time intervals. The forecasting plays an essential role in many financial organisations. The development of new mathematical tools or improving the existing ones will lead to more accurate forecasting techniques. Interval Arithmetic is a mathematical field that uses intervals by nature and algorithms that use it are involved in the solution of Interval Time Series. Another classical algorithm is VAR models. In this paper, a method that comes from the insurance sector is used to forecast Brent Oil monthly values. The innovation of what we propose is that it converts the Interval Time Series system into a single time series and can propagate the results back to each time series of the system. This way the researcher works with only one time series instead of two (or more). The forecasting algorithm is a choice of the researcher, expediting the development of forecasting (even ARIMA can be applied). We demonstrate our methodology in forecasting the Brent Oil monthly prices by applying the ANFIS algorithm.

In the continuously changing labor market, understanding the dynamics of online job postings is crucial for economic and workforce development. With the increasing reliance on Online Job Portals, analyzing online job postings has become an essential tool for capturing real-time labor-market trends. This paper presents a comprehensive methodology for processing online job postings to generate labor-market intelligence. The proposed methodology encompasses data source selection, data extraction, cleansing, normalization, and deduplication procedures. The final step involves information extraction based on employer industry, occupation, workplace, skills, and required experience. We address the key challenges that emerge at each step and discuss how they can be resolved. Our methodology is applied to two use cases: the first focuses on the analysis of the Greek labor market in the tourism industry during the COVID-19 pandemic, revealing shifts in job demands, skill requirements, and employment types. In the second use case, a data-driven ontology is employed to extract skills from job postings using machine learning. The findings highlight that the proposed methodology, utilizing NLP and machine-learning techniques instead of LLMs, can be applied to different labor market-analysis use cases and offer valuable insights for businesses, job seekers, and policymakers.

In this article, we investigate the potential of synthetic resumes as a means for the rapid generation of training data and their effectiveness in data augmentation, especially in categories marked by sparse samples. The widespread implementation of machine learning algorithms in natural language processing (NLP) has notably streamlined the resume classification process, delivering time and cost efficiencies for hiring organizations. However, the performance of these algorithms depends on the abundance of training data. While selecting the right model architecture is essential, it is also crucial to ensure the availability of a robust, well-curated dataset. For many categories in the job market, data sparsity remains a challenge. To deal with this challenge, we employed the OpenAI API to generate both structured and unstructured resumes tailored to specific criteria. These synthetically generated resumes were cleaned, preprocessed and then utilized to train two distinct models: a transformer model (BERT) and a feedforward neural network (FFNN) that incorporated Universal Sentence Encoder 4 (USE4) embeddings. While both models were evaluated on the multiclass classification task of resumes, when trained on an augmented dataset containing 60 percent real data (from Indeed website) and 40 percent synthetic data from ChatGPT, the transformer model presented exceptional accuracy. The FFNN, albeit predictably, achieved lower accuracy. These findings highlight the value of augmented real-world data with ChatGPT-generated synthetic resumes, especially in the context of limited training data. The suitability of the BERT model for such classification tasks further reinforces this narrative.

This article presents a study on the multi-class classification of job postings using machine learning algorithms. With the growth of online job platforms, there has been an influx of labor market data. Machine learning, particularly NLP, is increasingly used to analyze and classify job postings. However, the effectiveness of these algorithms largely hinges on the quality and volume of the training data. In our study, we propose a multi-class classification methodology for job postings, drawing on AI models such as text-davinci-003 and the quantized versions of Falcon 7b (Falcon), Wizardlm 7B (Wizardlm), and Vicuna 7B (Vicuna) to generate synthetic datasets. These synthetic data are employed in two use-case scenarios: (a) exclusively as training datasets composed of synthetic job postings (situations where no real data is available) and (b) as an augmentation method to bolster underrepresented job title categories. To evaluate our proposed method, we relied on two well-established approaches: the feedforward neural network (FFNN) and the BERT model. Both the use cases and training methods were assessed against a genuine job posting dataset to gauge classification accuracy. Our experiments substantiated the benefits of using synthetic data to enhance job posting classification. In the first scenario, the models’ performance matched, and occasionally exceeded, that of the real data. In the second scenario, the augmented classes consistently outperformed in most instances. This research confirms that AI-generated datasets can enhance the efficacy of NLP algorithms, especially in the domain of multi-class classification job postings. While data augmentation can boost model generalization, its impact varies. It is especially beneficial for simpler models like FNN. BERT, due to its context-aware architecture, also benefits from augmentation but sees limited improvement. Selecting the right type and amount of augmentation is essential.