ITIsmail Hakki Toroslu

Ismail Hakki Toroslu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Ismail Hakki Toroslu has a total of 73 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    OpenAI API compatible AI Inference Service support in HPC environment

    by Adam Matuš, Tomáš Martinovič, Arif Görkem Özer, Jakub Konvička, Firat Cekinel, Pinar Karagoz, Ismail Hakki Toroslu, Jakub Krejčí, Jan Martinovič

    Driven by the rise of Artificial Intelligence (AI) and Large Language Models (LLMs), the demand for high-density GPU resources has escalated significantly. High-Performance Computing (HPC) centers possess the necessary hardware, yet their conventional infrastructure and software ecosystems make hosting user-friendly AI services highly complex. This paper presents an innovative inference service designed specifically for HPC environments. By integrating batch scheduling, strategic project pre-allocations, and the High-End Application Environment (HEAppE) middleware, the service exposes a seamless, cloud-like Application Programming Interface (API) for LLMs, which can also be extended for broader AI inference tasks including agentic AI. To support common generative use-cases, such as text-to-text or image-to-text tasks, the service is designed to be fully compatible with the industry-standard OpenAI API. We evaluate the performance of this solution using standardized benchmarks against a bare-metal baseline to demonstrate its minimal orchestration overhead. This service has been developed within the scope of the Horizon Europe project EXA4MIND.

  • Open Access

    Querying Beyond Keywords: Translating Natural Language to Elasticsearch DSL With Vector Search Support

    by Aysel Yagmur Ozdemir, Pinar Karagoz, Ismail Hakki Toroslu

    Despite Elasticsearch’s popularity in modern search pipelines, translating natural language queries (NLQs) into its domain-specific language (DSL) remains challenging, especially with the integration of semantic vector search. Existing benchmarks primarily focus on keyword-based query translation, overlooking the increasingly essential hybrid scenarios where keyword filters are combined with vector-based similarity retrieval. This paper introduces the first benchmark designed explicitly for evaluating the NLQ-to-Elasticsearch DSL translation in hybrid search contexts. Leveraging a carefully curated subset of the WikiSQL dataset enriched with related Wikipedia text embeddings, our benchmark supports structured queries integrated with dense semantic similarity conditions. Through this approach, we provide a structured and controlled evaluation framework capable of assessing not only the syntactic accuracy of generated Elasticsearch DSL queries but also the semantic reasoning capabilities of Large Language Models (LLMs). Our experiments measure the zero-shot and few-shot performance of state-of-the-art LLMs on hybrid query generation tasks, illuminating their strengths and limitations in capturing complex user intents across keyword-based and vector similarity dimensions. In addition to prompting-based evaluation, we investigate the impact of parameter-efficient fine-tuning on a representative open-source model to assess its potential for narrowing the performance gap with proprietary systems. Two central conclusions emerge from our analysis. First, GPT models generally outperform open-source systems in absolute accuracy and reliability under pure prompting settings. Second, certain open-source models, particularly Qwen2.5-Coder variants, demonstrate substantial few-shot gains, and with parameter-efficient fine-tuning, can significantly reduce this gap, matching or even surpassing GPT-5 under zero-shot prompting on specific query categories. Overall, the proposed benchmark highlights both the current limitations and the untapped potential of open models in hybrid NLQ-to-DSL translation, advancing research toward more expressive and semantically grounded natural-language search interfaces.

  • EmbMerge: A Transformer-Based Method for Fusing CDR Lists

    by Mehmet Erdeniz Aydoğdu, Yağmur Duru Tüfekçioğlu, Ismail Sengor Altingovde, Pinar Karagoz, Ismail Hakki Toroslu

  • Open Access

    terazi: AI Fairness Tool for Doubly Imbalanced Data

    by Asli Umay Ozturk, Yigit Sever, Ata Yalcin, Viktoria Pauw, Stephan Hachinger, Ismail Hakki Toroslu, Pinar Karagoz


  • Open Access

    Fair for a few: Improving Fairness in Doubly Imbalanced Datasets

    by Ata Yalcin, Asli Umay Ozturk, Yigit Sever, Viktoria Pauw, Stephan Hachinger, Ismail Hakki Toroslu, Pinar Karagoz

    Fairness has been identified as an important aspect of Machine Learning and Artificial Intelligence solutions for decision making. Recent literature offers a variety of approaches for debiasing, however many of them fall short when the data collection is imbalanced. In this paper, we focus on a particular case, fairness in doubly imbalanced datasets, such that the data collection is imbalanced both for the label and the groups in the sensitive attribute. Firstly, we present an exploratory analysis to illustrate limitations in debiasing on a doubly imbalanced dataset. Then, a multi-criteria based solution is proposed for finding the most suitable sampling and distribution for label and sensitive attribute, in terms of fairness and classification accuracy


  • Open Access

    DocSpider: a dataset of cross-domain natural language querying for MongoDB

    by Arif Görkem Özer, Recep Firat Cekinel, Ismail Hakki Toroslu, Pinar Karagoz

    Natural language querying allows users to formulate questions in a natural language without requiring specific knowledge of the database query language. Large language models have been very successful in addressing the text-to-SQL problem, which is about translating given questions in textual form into SQL statements. Document-oriented NoSQL databases are gaining popularity in the era of big data due to their ability to handle vast amounts of semi-structured data and provide advanced querying functionalities. However, studies on text-to-NoSQL systems, particularly on systems targeting document databases, are very scarce. In this study, we utilize large language models to create a cross-domain natural language to document database query dataset, DocSpider, leveraging the well-known text-to-SQL challenge dataset Spider. As a document database, we use MongoDB. Furthermore, we conduct experiments to assess the effectiveness of the DocSpider dataset to fine-tune a text-to-NoSQL model against a cross-language transfer learning approach, SQL-to-NoSQL, and zero-shot instruction prompting. The experimental results reveal a significant improvement in the execution accuracy of fine-tuned language models when utilizing the DocSpider dataset.


  • Masking the Bias: From Echo Chambers to Large Scale Aspect-Based Sentiment Analysis

    by Yeonjung Lee, Yusuf Mücahit Çetinkaya, Emre Külah, Hakkı Toroslu, Hasan Davulcu


  • Open Access

    NeSHFS: Neighborhood Search with Heuristic-based Feature Selection for Click-Through Rate Prediction

    by Dogukan Aksu, Ismail Hakki Toroslu, Hasan Davulcu

    Click-through-rate (CTR) prediction plays an important role in online advertising and ad recommender systems. In the past decade, maximizing CTR has been the main focus of model development and solution creation. Therefore, researchers and practitioners have proposed various models and solutions to enhance the effectiveness of CTR prediction. Most of the existing literature focuses on capturing either implicit or explicit feature interactions. Although implicit interactions are successfully captured in some studies, explicit interactions present a challenge for achieving high CTR by extracting both low-order and high-order feature interactions. Unnecessary and irrelevant features may cause high computational time and low prediction performance. Furthermore, certain features may perform well with specific predictive models while underperforming with others. Also, feature distribution may fluctuate due to traffic variations. Most importantly, in live production environments, resources are limited, and the time for inference is just as crucial as training time. Because of all these reasons, feature selection is one of the most important factors in enhancing CTR prediction model performance. Simple filter-based feature selection algorithms do not perform well and they are not sufficient. An effective and efficient feature selection algorithm is needed to consistently filter the most useful features during live CTR prediction process. In this paper, we propose a heuristic algorithm named Neighborhood Search with Heuristic-based Feature Selection (NeSHFS) to enhance CTR prediction performance while reducing dimensionality and training time costs. We conduct comprehensive experiments on three public datasets to validate the efficiency and effectiveness of our proposed solution.

  • Toward a Programmable Humanizing Artificial Intelligence Through Scalable Stance-Directed Architecture

    by Yusuf Mücahit Çetinkaya, Yeonjung Lee, Emre Külah, İsmail Hakkı Toroslu, Michael A. Cowan, Hasan Davulcu

    The rise of harmful online content underscores the urgent need for AI systems to effectively detect, filter those, and foster safer and healthier communication. This article introduces a novel approach to mitigate toxic content generation propensities of Large Language Models (LLMs) by fine-tuning them with a programmable stance-directed focus on core human values and common good. We propose a streamlined keyword coding and processing pipeline to generate weakly labeled data to train AI models that can avoid toxicity and champion civil discourse. We also developed a toxicity classifier and an Aspect-based Sentiment Analysis (ABSA) model to assess and control the effectiveness of a humanizing AI model. We evaluate the proposed pipeline using a contentious real-world Twitter dataset on U.S. race relations. Our approach successfully curbs the toxic content generation propensity of an unrestricted LLM by a significant 85%.