Mucahid Kutlu

Md. Arid Hasan, Maram Hasanain, Fatema Ahmad et al. · utoronto

Natural Question Answering (QA) datasets play a crucial role in evaluating the capabilities of large language models (LLMs), ensuring their effectiveness in real-world applications. Despite the numerous QA datasets that have been developed and some work has been done in parallel, there is a notable lack of a framework and large scale region-specific datasets queried by native users in their own languages. This gap hinders the effective benchmarking and the development of fine-tuned models for regional and cultural specificities. In this study, we propose a scalable, language-independent framework, NativQA, to seamlessly construct culturally and regionally aligned QA datasets in native languages, for LLM evaluation and tuning. We demonstrate the efficacy of the proposed framework by designing a multilingual natural QA dataset, MultiNativQA, consisting of ~64k manually annotated QA pairs in seven languages, ranging from high to extremely low resource, based on queries from native speakers from 9 regions covering 18 topics. We benchmark open- and closed-source LLMs with the MultiNativQA dataset. We made the MultiNativQA dataset(https://huggingface.co/datasets/QCRI/MultiNativQA), and other experimental scripts(https://gitlab.com/nativqa/multinativqa) publicly available for the community.

Mohamed Bayan Kmainasi, Mucahid Kutlu, Ali Ezzat Shahroor et al.

Hateful memes often require compositional multimodal reasoning: the image and text may appear benign in isolation, yet their interaction conveys harmful intent. Although thinking-based multimodal large language models (MLLMs) have recently advanced vision-language understanding, their capabilities remain underexplored for hateful meme analysis. We propose a reinforcement learning based post-training framework that improves reasoning in thinking-based MLLMs through task-specific rewards and a novel Group Relative Policy Optimization (GRPO) objective. Specifically, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful meme understanding, (ii) extend an existing hateful meme dataset by generating weakly or pseudo-supervised chain-of-thought rationales via distillation, and (iii) introduce a GRPO-based objective that jointly optimizes meme classification and explanation quality to encourage fine-grained, step-by-step reasoning. Experiments on the Hateful Memes benchmark show that our approach achieves state-of-the-art performance, improving accuracy and F1 by approximately 1 percent and explanation quality by approximately 3 percent. We will publicly release our code, dataset extensions, and evaluation resources to support reproducibility.

Dilara Dogan, Bahadir Altun, Muhammed Said Zengin et al.

The recent advances in natural language processing have yielded many exciting developments in text analysis and language understanding models; however, these models can also be used to track people, bringing severe privacy concerns. In this work, we investigate what individuals can do to avoid being detected by those models while using social media platforms. We ground our investigation in two exposure-risky tasks, stance detection and geotagging. We explore a variety of simple techniques for modifying text, such as inserting typos in salient words, paraphrasing, and adding dummy social media posts. Our experiments show that the performance of BERT-based models fined tuned for stance detection decreases significantly due to typos, but it is not affected by paraphrasing. Moreover, we find that typos have minimal impact on state-of-the-art geotagging models due to their increased reliance on social networks; however, we show that users can deceive those models by interacting with different users, reducing their performance by almost 50%.

Kaan Efe Keleş, Ömer Kaan Gürbüz, Mucahid Kutlu

Potential harms of Large Language Models such as mass misinformation and plagiarism can be partially mitigated if there exists a reliable way to detect machine generated text. In this paper, we propose a new watermarking method to detect machine-generated texts. Our method embeds a unique pattern within the generated text, ensuring that while the content remains coherent and natural to human readers, it carries distinct markers that can be identified algorithmically. Specifically, we intervene with the token sampling process in a way which enables us to trace back our token choices during the detection phase. We show how watermarking affects textual quality and compare our proposed method with a state-of-the-art watermarking method in terms of robustness and detectability. Through extensive experiments, we demonstrate the effectiveness of our watermarking scheme in distinguishing between watermarked and non-watermarked text, achieving high detection rates while maintaining textual quality.

Salah Feras Alali, Mohammad Nashat Maasfeh, Mucahid Kutlu et al.

With the incredible advancements in Large Language Models (LLMs), many people have started using them to satisfy their information needs. However, utilizing LLMs might be problematic for political issues where disagreement is common and model outputs may reflect training-data biases or deliberate alignment choices. To better characterize such behavior, we assess the stances of nine LLMs on 24 politically sensitive issues using five prompting techniques. We find that models often adopt opposing stances on several issues; some positions are malleable under prompting, while others remain stable. Among the models examined, Grok-3-mini is the most persistent, whereas Mistral-7B is the least. For issues involving countries with different languages, models tend to support the side whose language is used in the prompt. Notably, no prompting technique alters model stances on the Qatar blockade or the oppression of Palestinians. We hope these findings raise user awareness when seeking political guidance from LLMs and encourage developers to address these concerns.

Md Mustafizur Rahman, Mucahid Kutlu, Tamer Elsayed et al.

To create a new IR test collection at low cost, it is valuable to carefully select which documents merit human relevance judgments. Shared task campaigns such as NIST TREC pool document rankings from many participating systems (and often interactive runs as well) in order to identify the most likely relevant documents for human judging. However, if one's primary goal is merely to build a test collection, it would be useful to be able to do so without needing to run an entire shared task. Toward this end, we investigate multiple active learning strategies which, without reliance on system rankings: 1) select which documents human assessors should judge; and 2) automatically classify the relevance of additional unjudged documents. To assess our approach, we report experiments on five TREC collections with varying scarcity of relevant documents. We report labeling accuracy achieved, as well as rank correlation when evaluating participant systems based upon these labels vs.\ full pool judgments. Results show the effectiveness of our approach, and we further analyze how varying relevance scarcity across collections impacts our findings. To support reproducibility and follow-on work, we have shared our code online: https://github.com/mdmustafizurrahman/ICTIR_AL_TestCollection_2020/.

Shammur Absar Chowdhury, Hind Almerekhi, Mucahid Kutlu et al.

This paper presents a comprehensive overview of the first edition of the Academic Essay Authenticity Challenge, organized as part of the GenAI Content Detection shared tasks collocated with COLING 2025. This challenge focuses on detecting machine-generated vs. human-authored essays for academic purposes. The task is defined as follows: "Given an essay, identify whether it is generated by a machine or authored by a human.'' The challenge involves two languages: English and Arabic. During the evaluation phase, 25 teams submitted systems for English and 21 teams for Arabic, reflecting substantial interest in the task. Finally, seven teams submitted system description papers. The majority of submissions utilized fine-tuned transformer-based models, with one team employing Large Language Models (LLMs) such as Llama 2 and Llama 3. This paper outlines the task formulation, details the dataset construction process, and explains the evaluation framework. Additionally, we present a summary of the approaches adopted by participating teams. Nearly all submitted systems outperformed the n-gram-based baseline, with the top-performing systems achieving F1 scores exceeding 0.98 for both languages, indicating significant progress in the detection of machine-generated text.

Firoj Alam, Md Arid Hasan, Sahinur Rahman Laskar et al. · utoronto

The rapid advancement of large language models (LLMs) has raised concerns about cultural bias, fairness, and their applicability in diverse linguistic and underrepresented regional contexts. To enhance and benchmark the capabilities of LLMs, there is a need to develop large-scale resources focused on multilingual, local, and cultural contexts. In this study, we propose the NativQA framework, which can seamlessly construct large-scale, culturally and regionally aligned QA datasets in native languages. The framework utilizes user-defined seed queries and leverages search engines to collect location-specific, everyday information. It has been evaluated across 39 locations in 24 countries and in 7 languages -- ranging from extremely low-resource to high-resource languages -- resulting in over 300K Question-Answer (QA) pairs. The developed resources can be used for LLM benchmarking and further fine-tuning. The framework has been made publicly available for the community (https://gitlab.com/nativqa/nativqa-framework).

Basak Demirok, Mucahid Kutlu

While large language models provide significant convenience for software development, they can lead to ethical issues in job interviews and student assignments. Therefore, determining whether a piece of code is written by a human or generated by an artificial intelligence (AI) model is a critical issue. In this study, we present AIGCodeSet, which consists of 2.828 AI-generated and 4.755 human-written Python codes, created using CodeLlama 34B, Codestral 22B, and Gemini 1.5 Flash. In addition, we share the results of our experiments conducted with baseline detection methods. Our experiments show that a Bayesian classifier outperforms the other models.

Tarık Saraç, Selin Mergen, Mucahid Kutlu

In this paper, we present our work developed for the scientific web discourse detection task (Task 4a) of CheckThat! 2025. We propose a novel council debate method that simulates structured academic discussions among multiple large language models (LLMs) to identify whether a given tweet contains (i) a scientific claim, (ii) a reference to a scientific study, or (iii) mentions of scientific entities. We explore three debating methods: i) single debate, where two LLMs argue for opposing positions while a third acts as a judge; ii) team debate, in which multiple models collaborate within each side of the debate; and iii) council debate, where multiple expert models deliberate together to reach a consensus, moderated by a chairperson model. We choose council debate as our primary model as it outperforms others in the development test set. Although our proposed method did not rank highly for identifying scientific claims (8th out of 10) or mentions of scientific entities (9th out of 10), it ranked first in detecting references to scientific studies.

Togay Yazar, Mucahid Kutlu, İsa Kerem Bayırlı

Over the past century, the Turkish language has undergone substantial changes, primarily driven by governmental interventions. In this work, our goal is to investigate the evolution of the Turkish language since the establishment of Türkiye in 1923. Thus, we first introduce Turkronicles which is a diachronic corpus for Turkish derived from the Official Gazette of Türkiye. Turkronicles contains 45,375 documents, detailing governmental actions, making it a pivotal resource for analyzing the linguistic evolution influenced by the state policies. In addition, we expand an existing diachronic Turkish corpus which consists of the records of the Grand National Assembly of Türkiye by covering additional years. Next, combining these two diachronic corpora, we seek answers for two main research questions: How have the Turkish vocabulary and the writing conventions changed since the 1920s? Our analysis reveals that the vocabularies of two different time periods diverge more as the time between them increases, and newly coined Turkish words take the place of their old counterparts. We also observe changes in writing conventions. In particular, the use of circumflex noticeably decreases and words ending with the letters "-b" and "-d" are successively replaced with "-p" and "-t" letters, respectively. Overall, this study quantitatively highlights the dramatic changes in Turkish from various aspects of the language in a diachronic perspective.

Basak Demirok, Mucahid Kutlu, Selin Mergen

As large language models (LLMs) rapidly advance, their role in code generation has expanded significantly. While this offers streamlined development, it also creates concerns in areas like education and job interviews. Consequently, developing robust systems to detect AI-generated code is imperative to maintain academic integrity and ensure fairness in hiring processes. In this study, we introduce MultiAIGCD, a dataset for AI-generated code detection for Python, Java, and Go. From the CodeNet dataset's problem definitions and human-authored codes, we generate several code samples in Java, Python, and Go with six different LLMs and three different prompts. This generation process covered three key usage scenarios: (i) generating code from problem descriptions, (ii) fixing runtime errors in human-written code, and (iii) correcting incorrect outputs. Overall, MultiAIGCD consists of 121,271 AI-generated and 32,148 human-written code snippets. We also benchmark three state-of-the-art AI-generated code detection models and assess their performance in various test scenarios such as cross-model and cross-language. We share our dataset and codes to support research in this field.

Umur Togay Yazar, Mucahid Kutlu

Dynamic structure of languages poses significant challenges in applying natural language processing models on historical texts, causing decreased performance in various downstream tasks. Turkish is a prominent example of rapid linguistic transformation due to the language reform in the 20th century. In this paper, we propose two methods for detecting synonyms used in different time periods, focusing on Turkish. In our first method, we use Orthogonal Procrustes method to align the embedding spaces created using documents written in the corresponding time periods. In our second method, we extend the first one by incorporating Spearman's correlation between frequencies of words throughout the years. In our experiments, we show that our proposed methods outperform the baseline method. Furthermore, we observe that the efficacy of our methods remains consistent when the target time period shifts from the 1960s to the 1980s. However, their performance slightly decreases for subsequent time periods.

Preslav Nakov, Giovanni Da San Martino, Tamer Elsayed et al.

We describe the fourth edition of the CheckThat! Lab, part of the 2021 Conference and Labs of the Evaluation Forum (CLEF). The lab evaluates technology supporting tasks related to factuality, and covers Arabic, Bulgarian, English, Spanish, and Turkish. Task 1 asks to predict which posts in a Twitter stream are worth fact-checking, focusing on COVID-19 and politics (in all five languages). Task 2 asks to determine whether a claim in a tweet can be verified using a set of previously fact-checked claims (in Arabic and English). Task 3 asks to predict the veracity of a news article and its topical domain (in English). The evaluation is based on mean average precision or precision at rank k for the ranking tasks, and macro-F1 for the classification tasks. This was the most popular CLEF-2021 lab in terms of team registrations: 132 teams. Nearly one-third of them participated: 15, 5, and 25 teams submitted official runs for tasks 1, 2, and 3, respectively.

Md Mustafizur Rahman, Dinesh Balakrishnan, Dhiraj Murthy et al.

Building a benchmark dataset for hate speech detection presents various challenges. Firstly, because hate speech is relatively rare, random sampling of tweets to annotate is very inefficient in finding hate speech. To address this, prior datasets often include only tweets matching known "hate words". However, restricting data to a pre-defined vocabulary may exclude portions of the real-world phenomenon we seek to model. A second challenge is that definitions of hate speech tend to be highly varying and subjective. Annotators having diverse prior notions of hate speech may not only disagree with one another but also struggle to conform to specified labeling guidelines. Our key insight is that the rarity and subjectivity of hate speech are akin to that of relevance in information retrieval (IR). This connection suggests that well-established methodologies for creating IR test collections can be usefully applied to create better benchmark datasets for hate speech. To intelligently and efficiently select which tweets to annotate, we apply standard IR techniques of {\em pooling} and {\em active learning}. To improve both consistency and value of annotations, we apply {\em task decomposition} and {\em annotator rationale} techniques. We share a new benchmark dataset for hate speech detection on Twitter that provides broader coverage of hate than prior datasets. We also show a dramatic drop in accuracy of existing detection models when tested on these broader forms of hate. Annotator rationales we collect not only justify labeling decisions but also enable future work opportunities for dual-supervision and/or explanation generation in modeling. Further details of our approach can be found in the supplementary materials.

Md Mustafizur Rahman, Mucahid Kutlu, Matthew Lease

Research community evaluations in information retrieval, such as NIST's Text REtrieval Conference (TREC), build reusable test collections by pooling document rankings submitted by many teams. Naturally, the quality of the resulting test collection thus greatly depends on the number of participating teams and the quality of their submitted runs. In this work, we investigate: i) how the number of participants, coupled with other factors, affects the quality of a test collection; and ii) whether the quality of a test collection can be inferred prior to collecting relevance judgments from human assessors. Experiments conducted on six TREC collections illustrate how the number of teams interacts with various other factors to influence the resulting quality of test collections. We also show that the reusability of a test collection can be predicted with high accuracy when the same document collection is used for successive years in an evaluation campaign, as is common in TREC.

Ammar Rashed, Mucahid Kutlu, Kareem Darwish et al.

On June 24, 2018, Turkey conducted a highly consequential election in which the Turkish people elected their president and parliament in the first election under a new presidential system. During the election period, the Turkish people extensively shared their political opinions on Twitter. One aspect of polarization among the electorate was support for or opposition to the reelection of Recep Tayyip Erdoğan. In this paper, we present an unsupervised method for target-specific stance detection in a polarized setting, specifically Turkish politics, achieving 90% precision in identifying user stances, while maintaining more than 80% recall. The method involves representing users in an embedding space using Google's Convolutional Neural Network (CNN) based multilingual universal sentence encoder. The representations are then projected onto a lower dimensional space in a manner that reflects similarities and are consequently clustered. We show the effectiveness of our method in properly clustering users of divergent groups across multiple targets that include political figures, different groups, and parties. We perform our analysis on a large dataset of 108M Turkish election-related tweets along with the timeline tweets of 168k Turkish users, who authored 213M tweets. Given the resultant user stances, we are able to observe correlations between topics and compute topic polarization.

Yavuz Selim Kartal, Busra Guvenen, Mucahid Kutlu

The massive amount of misinformation spreading on the Internet on a daily basis has enormous negative impacts on societies. Therefore, we need automated systems helping fact-checkers in the combat against misinformation. In this paper, we propose a model prioritizing the claims based on their check-worthiness. We use BERT model with additional features including domain-specific controversial topics, word embeddings, and others. In our experiments, we show that our proposed model outperforms all state-of-the-art models in both test collections of CLEF Check That! Lab in 2018 and 2019. We also conduct a qualitative analysis to shed light-detecting check-worthy claims. We suggest requesting rationales behind judgments are needed to understand subjective nature of the task and problematic labels.

Mucahid Kutlu, Tyler McDonnell, Aashish Sheshadri et al.

Crowdsourcing offers an affordable and scalable means to collect relevance judgments for IR test collections. However, crowd assessors may show higher variance in judgment quality than trusted assessors. In this paper, we investigate how to effectively utilize both groups of assessors in partnership. We specifically investigate how agreement in judging is correlated with three factors: relevance category, document rankings, and topical variance. Based on this, we then propose two collaborative judging methods in which a portion of the document-topic pairs are assessed by in-house judges while the rest are assessed by crowd-workers. Experiments conducted on two TREC collections show encouraging results when we distribute work intelligently between our two groups of assessors.

Mucahid Kutlu, Vivek Khetan, Matthew Lease

Because researchers typically do not have the time or space to present more than a few evaluation metrics in any published study, it can be difficult to assess relative effectiveness of prior methods for unreported metrics when baselining a new method or conducting a systematic meta-review. While sharing of study data would help alleviate this, recent attempts to encourage consistent sharing have been largely unsuccessful. Instead, we propose to enable relative comparisons with prior work across arbitrary metrics by predicting unreported metrics given one or more reported metrics. In addition, we further investigate prediction of high-cost evaluation measures using low-cost measures as a potential strategy for reducing evaluation cost. We begin by assessing the correlation between 23 IR metrics using 8 TREC test collections. Measuring prediction error wrt. R-square and Kendall's tau, we show that accurate prediction of MAP, P@10, and RBP can be achieved using only 2-3 other metrics. With regard to lowering evaluation cost, we show that RBP(p=0.95) can be predicted with high accuracy using measures with only evaluation depth of 30. Taken together, our findings provide a valuable proof-of-concept which we expect to spur follow-on work by others in proposing more sophisticated models for metric prediction.

Maram Hasanain, Reem Suwaileh, Tamer Elsayed et al.

This article introduces a new language-independent approach for creating a large-scale high-quality test collection of tweets that supports multiple information retrieval (IR) tasks without running a shared-task campaign. The adopted approach (demonstrated over Arabic tweets) designs the collection around significant (i.e., popular) events, which enables the development of topics that represent frequent information needs of Twitter users for which rich content exists. That inherently facilitates the support of multiple tasks that generally revolve around events, namely event detection, ad-hoc search, timeline generation, and real-time summarization. The key highlights of the approach include diversifying the judgment pool via interactive search and multiple manually-crafted queries per topic, collecting high-quality annotations via crowd-workers for relevancy and in-house annotators for novelty, filtering out low-agreement topics and inaccessible tweets, and providing multiple subsets of the collection for better availability. Applying our methodology on Arabic tweets resulted in EveTAR , the first freely-available tweet test collection for multiple IR tasks. EveTAR includes a crawl of 355M Arabic tweets and covers 50 significant events for which about 62K tweets were judged with substantial average inter-annotator agreement (Kappa value of 0.71). We demonstrate the usability of EveTAR by evaluating existing algorithms in the respective tasks. Results indicate that the new collection can support reliable ranking of IR systems that is comparable to similar TREC collections, while providing strong baseline results for future studies over Arabic tweets.

Mucahid Kutlu, Tamer Elsayed, Matthew Lease

While test collections provide the cornerstone for Cranfield-based evaluation of information retrieval (IR) systems, it has become practically infeasible to rely on traditional pooling techniques to construct test collections at the scale of today's massive document collections. In this paper, we propose a new intelligent topic selection method which reduces the number of search topics needed for reliable IR evaluation. To rigorously assess our method, we integrate previously disparate lines of research on intelligent topic selection and deep vs. shallow judging. While prior work on intelligent topic selection has never been evaluated against shallow judging baselines, prior work on deep vs. shallow judging has largely argued for shallowed judging, but assuming random topic selection. We argue that for evaluating any topic selection method, ultimately one must ask whether it is actually useful to select topics, or should one simply perform shallow judging over many topics? In seeking a rigorous answer to this over-arching question, we conduct a comprehensive investigation over a set of relevant factors never previously studied together 1) topic selection method 2) the effect of topic familiarity on human judging speed and 3) how different topic generation processes impact (i) budget utilization and (ii) the resultant quality of judgments. Experiments on NIST TREC Robust 2003 and Robust 2004 test collections show that not only can we reliably evaluate IR systems with fewer topics, but also that 1) when topics are intelligently selected, deep judging is often more cost-effective than shallow judging in evaluation reliability and 2) topic familiarity and topic generation costs greatly impact the evaluation cost vs. reliability trade-off. Our findings challenge conventional wisdom in showing that deep judging is often preferable to shallow judging when topics are selected intelligently.