arXiv subjects

Gianluca Demartini

Publications and source records attributed to Gianluca Demartini.

At least 37 records · Page 2Linked to original sources

arXiv2025

LLM-based Semantic Augmentation for Harmful Content Detection

Recent advances in large language models (LLMs) have demonstrated strong performance on simple text classification tasks, frequently under zero-shot settings. However, their efficacy declines when tackling complex social media challenges such as propaganda detection, hateful meme classification, and toxicity identification. Much of the existing work has focused on using LLMs to generate synthetic training data, overlooking the potential of LLM-based text preprocessing and semantic augmentation. In this paper, we introduce an approach that prompts LLMs to clean noisy text and provide context-rich explanations, thereby enhancing training sets without substantial increases in data volume. We systematically evaluate on the SemEval 2024 multi-label Persuasive Meme dataset and further validate on the Google Jigsaw toxic comments and Facebook hateful memes datasets to assess generalizability. Our results reveal that zero-shot LLM classification underperforms on these high-context tasks compared to supervised models. In contrast, integrating LLM-based semantic augmentation yields performance on par with approaches that rely on human-annotated data, at a fraction of the cost. These findings underscore the importance of strategically incorporating LLMs into machine learning (ML) pipeline for social media classification tasks, offering broad implications for combating harmful content online.

cs.CL↗

arXiv2025

Leveraging Semantic Type Dependencies for Clinical Named Entity Recognition

Previous work on clinical relation extraction from free-text sentences leveraged information about semantic types from clinical knowledge bases as a part of entity representations. In this paper, we exploit additional evidence by also making use of domain-specific semantic type dependencies. We encode the relation between a span of tokens matching a Unified Medical Language System (UMLS) concept and other tokens in the sentence. We implement our method and compare against different named entity recognition (NER) architectures (i.e., BiLSTM-CRF and BiLSTM-GCN-CRF) using different pre-trained clinical embeddings (i.e., BERT, BioBERT, UMLSBert). Our experimental results on clinical datasets show that in some cases NER effectiveness can be significantly improved by making use of domain-specific semantic type dependencies. Our work is also the first study generating a matrix encoding to make use of more than three dependencies in one pass for the NER task.

cs.CL↗

arXiv2025

Mapping and Influencing the Political Ideology of Large Language Models using Synthetic Personas

The analysis of political biases in large language models (LLMs) has primarily examined these systems as single entities with fixed viewpoints. While various methods exist for measuring such biases, the impact of persona-based prompting on LLMs' political orientation remains unexplored. In this work we leverage PersonaHub, a collection of synthetic persona descriptions, to map the political distribution of persona-based prompted LLMs using the Political Compass Test (PCT). We then examine whether these initial compass distributions can be manipulated through explicit ideological prompting towards diametrically opposed political orientations: right-authoritarian and left-libertarian. Our experiments reveal that synthetic personas predominantly cluster in the left-libertarian quadrant, with models demonstrating varying degrees of responsiveness when prompted with explicit ideological descriptors. While all models demonstrate significant shifts towards right-authoritarian positions, they exhibit more limited shifts towards left-libertarian positions, suggesting an asymmetric response to ideological manipulation that may reflect inherent biases in model training.

cs.CL↗

arXiv2025

The Impact of Persona-based Political Perspectives on Hateful Content Detection

While pretraining language models with politically diverse content has been shown to improve downstream task fairness, such approaches require significant computational resources often inaccessible to many researchers and organizations. Recent work has established that persona-based prompting can introduce political diversity in model outputs without additional training. However, it remains unclear whether such prompting strategies can achieve results comparable to political pretraining for downstream tasks. We investigate this question using persona-based prompting strategies in multimodal hate-speech detection tasks, specifically focusing on hate speech in memes. Our analysis reveals that when mapping personas onto a political compass and measuring persona agreement, inherent political positioning has surprisingly little correlation with classification decisions. Notably, this lack of correlation persists even when personas are explicitly injected with stronger ideological descriptors. Our findings suggest that while LLMs can exhibit political biases in their responses to direct political questions, these biases may have less impact on practical classification tasks than previously assumed. This raises important questions about the necessity of computationally expensive political pretraining for achieving fair performance in downstream tasks.

cs.CL↗

arXiv2025

Are Large Language Models Good Data Preprocessors?

High-quality textual training data is essential for the success of multimodal data processing tasks, yet outputs from image captioning models like BLIP and GIT often contain errors and anomalies that are difficult to rectify using rule-based methods. While recent work addressing this issue has predominantly focused on using GPT models for data preprocessing on relatively simple public datasets, there is a need to explore a broader range of Large Language Models (LLMs) and tackle more challenging and diverse datasets. In this study, we investigate the use of multiple LLMs, including LLaMA 3.1 70B, GPT-4 Turbo, and Sonnet 3.5 v2, to refine and clean the textual outputs of BLIP and GIT. We assess the impact of LLM-assisted data cleaning by comparing downstream-task (SemEval 2024 Subtask "Multilabel Persuasion Detection in Memes") models trained on cleaned versus non-cleaned data. While our experimental results show improvements when using LLM-cleaned captions, statistical tests reveal that most of these improvements are not significant. This suggests that while LLMs have the potential to enhance data cleaning and repairing, their effectiveness may be limited depending on the context they are applied to, the complexity of the task, and the level of noise in the text. Our findings highlight the need for further research into the capabilities and limitations of LLMs in data preprocessing pipelines, especially when dealing with challenging datasets, contributing empirical evidence to the ongoing discussion about integrating LLMs into data preprocessing pipelines.

cs.CL↗

arXiv2025

Plan-Then-Execute: An Empirical Study of User Trust and Team Performance When Using LLM Agents As A Daily Assistant

Since the explosion in popularity of ChatGPT, large language models (LLMs) have continued to impact our everyday lives. Equipped with external tools that are designed for a specific purpose (e.g., for flight booking or an alarm clock), LLM agents exercise an increasing capability to assist humans in their daily work. Although LLM agents have shown a promising blueprint as daily assistants, there is a limited understanding of how they can provide daily assistance based on planning and sequential decision making capabilities. We draw inspiration from recent work that has highlighted the value of 'LLM-modulo' setups in conjunction with humans-in-the-loop for planning tasks. We conducted an empirical study (N = 248) of LLM agents as daily assistants in six commonly occurring tasks with different levels of risk typically associated with them (e.g., flight ticket booking and credit card payments). To ensure user agency and control over the LLM agent, we adopted LLM agents in a plan-then-execute manner, wherein the agents conducted step-wise planning and step-by-step execution in a simulation environment. We analyzed how user involvement at each stage affects their trust and collaborative team performance. Our findings demonstrate that LLM agents can be a double-edged sword -- (1) they can work well when a high-quality plan and necessary user involvement in execution are available, and (2) users can easily mistrust the LLM agents with plans that seem plausible. We synthesized key insights for using LLM agents as daily assistants to calibrate user trust and achieve better overall task outcomes. Our work has important implications for the future design of daily assistants and human-AI collaboration with LLM agents.

cs.HC↗

arXiv2025

Exploring Wikipedia Gender Diversity Over Time $\unicode{x2013}$ The Wikipedia Gender Dashboard (WGD)

The Wikipedia editors' community has been actively pursuing the intent of achieving gender equality. To that end, it is important to explore the historical evolution of underlying gender disparities in Wikipedia articles. This paper presents the Wikipedia Gender Dashboard (WGD), a tool designed to enable the interaction with gender distribution data, including the average age in every subclass of individuals (i.e. Astronauts, Politicians, etc.) over the years. Wikipedia APIs, DBpedia, and Wikidata endpoints were used to query the data to ensure persistent data collection. The WGD was then created with Microsoft Power BI before being embedded on a public website. The analysis of the data available in the WGD found that female articles only represent around 17% of English Wikipedia, but it has been growing steadily over the last 20 years. Meanwhile, the average age across genders decreased over time. WGD also shows that most subclasses of `Person' are male-dominated. Wikipedia editors can make use of WGD to locate areas with marginalized genders in Wikipedia, and increase their efforts to produce more content providing coverage for those genders to achieve better gender equality in Wikipedia.

cs.CY↗

arXiv2024

Enhancing Media Literacy: The Effectiveness of (Human) Annotations and Bias Visualizations on Bias Detection

Marking biased texts is a practical approach to increase media bias awareness among news consumers. However, little is known about the generalizability of such awareness to new topics or unmarked news articles, and the role of machine-generated bias labels in enhancing awareness remains unclear. This study tests how news consumers may be trained and pre-bunked to detect media bias with bias labels obtained from different sources (Human or AI) and in various manifestations. We conducted two experiments with 470 and 846 participants, exposing them to various bias-labeling conditions. We subsequently tested how much bias they could identify in unlabeled news materials on new topics. The results show that both Human (t(467) = 4.55, p < .001, d = 0.42) and AI labels (t(467) = 2.49, p = .039, d = 0.23) increased correct detection compared to the control group. Human labels demonstrate larger effect sizes and higher statistical significance. The control group (t(467) = 4.51, p < .001, d = 0.21) also improves performance through mere exposure to study materials. We also find that participants trained with marked biased phrases detected bias most reliably (F(834,1) = 44.00, p < .001, η2part = 0.048). Our experimental framework provides theoretical implications for systematically assessing the generalizability of learning effects in identifying media bias. These findings also provide practical implications for developing news-reading platforms that offer bias indicators and designing media literacy curricula to enhance media bias awareness.

cs.HC↗

arXiv2024

Optimizing LLMs with Direct Preferences: A Data Efficiency Perspective

Aligning the output of Large Language Models (LLMs) with human preferences (e.g., by means of reinforcement learning with human feedback, or RLHF) is essential for ensuring their effectiveness in real-world scenarios. Despite significant advancements in LLM alignment techniques, the impact of different type of preference data on model performance has yet to be systematically explored. In this study, we investigate the scalability, data efficiency, and effectiveness of Direct Preference Optimization (DPO) in fine-tuning pre-trained LLMs, aiming to reduce their dependency on extensive amounts of preference data, which is expensive to collect. We (1) systematically compare the performance of models fine-tuned with varying percentages of a combined preference judgement dataset to define the improvement curve of DPO and assess its effectiveness in data-constrained environments; and (2) provide insights for the development of an optimal approach for selective preference data usage. Our study reveals that increasing the amount of data used for training generally enhances and stabilizes model performance. Moreover, the use of a combination of diverse datasets significantly improves model effectiveness. Furthermore, when models are trained separately using different types of prompts, models trained with conversational prompts outperformed those trained with question answering prompts.

cs.AI↗

arXiv2024

Personas with Attitudes: Controlling LLMs for Diverse Data Annotation

We present a novel approach for enhancing diversity and control in data annotation tasks by personalizing large language models (LLMs). We investigate the impact of injecting diverse persona descriptions into LLM prompts across two studies, exploring whether personas increase annotation diversity and whether the impacts of individual personas on the resulting annotations are consistent and controllable. Our results show that persona-prompted LLMs produce more diverse annotations than LLMs prompted without personas and that these effects are both controllable and repeatable, making our approach a suitable tool for improving data annotation in subjective NLP tasks like toxicity detection.

cs.CL↗

arXiv2024

Fairness without Sensitive Attributes via Knowledge Sharing

While model fairness improvement has been explored previously, existing methods invariably rely on adjusting explicit sensitive attribute values in order to improve model fairness in downstream tasks. However, we observe a trend in which sensitive demographic information becomes inaccessible as public concerns around data privacy grow. In this paper, we propose a confidence-based hierarchical classifier structure called "Reckoner" for reliable fair model learning under the assumption of missing sensitive attributes. We first present results showing that if the dataset contains biased labels or other hidden biases, classifiers significantly increase the bias gap across different demographic groups in the subset with higher prediction confidence. Inspired by these findings, we devised a dual-model system in which a version of the model initialised with a high-confidence data subset learns from a version of the model initialised with a low-confidence data subset, enabling it to avoid biased predictions. Our experimental results show that Reckoner consistently outperforms state-of-the-art baselines in COMPAS dataset and New Adult dataset, considering both accuracy and fairness metrics.

cs.LG↗

arXiv2024

Hate Speech Detection with Generalizable Target-aware Fairness

To counter the side effect brought by the proliferation of social media platforms, hate speech detection (HSD) plays a vital role in halting the dissemination of toxic online posts at an early stage. However, given the ubiquitous topical communities on social media, a trained HSD classifier easily becomes biased towards specific targeted groups (e.g., female and black people), where a high rate of false positive/negative results can significantly impair public trust in the fairness of content moderation mechanisms, and eventually harm the diversity of online society. Although existing fairness-aware HSD methods can smooth out some discrepancies across targeted groups, they are mostly specific to a narrow selection of targets that are assumed to be known and fixed. This inevitably prevents those methods from generalizing to real-world use cases where new targeted groups constantly emerge over time. To tackle this defect, we propose Generalizable target-aware Fairness (GetFair), a new method for fairly classifying each post that contains diverse and even unseen targets during inference. To remove the HSD classifier's spurious dependence on target-related features, GetFair trains a series of filter functions in an adversarial pipeline, so as to deceive the discriminator that recovers the targeted group from filtered post embeddings. To maintain scalability and generalizability, we innovatively parameterize all filter functions via a hypernetwork that is regularized by the semantic affinity among targets. Taking a target's pretrained word embedding as input, the hypernetwork generates the weights used by each target-specific filter on-the-fly without storing dedicated filter parameters. Finally, comparative experiments on two HSD datasets have shown advantageous performance of GetFair on out-of-sample targets.

cs.CL↗

arXiv2024

Estimating Gender Completeness in Wikipedia

Gender imbalance in Wikipedia content is a known challenge which the editor community is actively addressing. The aim of this paper is to provide the Wikipedia community with instruments to estimate the magnitude of the problem for different entity types (also known as classes) in Wikipedia. To this end, we apply class completeness estimation methods based on the gender attribute. Our results show not only which gender for different sub-classes of Person is more prevalent in Wikipedia, but also an idea of how complete the coverage is for difference genders and sub-classes of Person.

cs.CY↗

arXiv2024

Identification of Regulatory Requirements Relevant to Business Processes: A Comparative Study on Generative AI, Embedding-based Ranking, Crowd and Expert-driven Methods

Organizations face the challenge of ensuring compliance with an increasing amount of requirements from various regulatory documents. Which requirements are relevant depends on aspects such as the geographic location of the organization, its domain, size, and business processes. Considering these contextual factors, as a first step, relevant documents (e.g., laws, regulations, directives, policies) are identified, followed by a more detailed analysis of which parts of the identified documents are relevant for which step of a given business process. Nowadays the identification of regulatory requirements relevant to business processes is mostly done manually by domain and legal experts, posing a tremendous effort on them, especially for a large number of regulatory documents which might frequently change. Hence, this work examines how legal and domain experts can be assisted in the assessment of relevant requirements. For this, we compare an embedding-based NLP ranking method, a generative AI method using GPT-4, and a crowdsourced method with the purely manual method of creating relevancy labels by experts. The proposed methods are evaluated based on two case studies: an Australian insurance case created with domain experts and a global banking use case, adapted from SAP Signavio's workflow example of an international guideline. A gold standard is created for both BPMN2.0 processes and matched to real-world textual requirements from multiple regulatory documents. The evaluation and discussion provide insights into strengths and weaknesses of each method regarding applicability, automation, transparency, and reproducibility and provide guidelines on which method combinations will maximize benefits for given characteristics such as process usage, impact, and dynamics of an application scenario.

cs.CL↗

arXiv2023

Perspectives on Large Language Models for Relevance Judgment

When asked, large language models (LLMs) like ChatGPT claim that they can assist with relevance judgments but it is not clear whether automated judgments can reliably be used in evaluations of retrieval systems. In this perspectives paper, we discuss possible ways for LLMs to support relevance judgments along with concerns and issues that arise. We devise a human--machine collaboration spectrum that allows to categorize different relevance judgment strategies, based on how much humans rely on machines. For the extreme point of "fully automated judgments", we further include a pilot experiment on whether LLM-based relevance judgments correlate with judgments from trained human assessors. We conclude the paper by providing opposing perspectives for and against the use of~LLMs for automatic relevance judgments, and a compromise perspective, informed by our analyses of the literature, our preliminary experimental evidence, and our experience as IR researchers.

cs.IR↗

arXiv2023

Data Bias Management

Due to the widespread use of data-powered systems in our everyday lives, concepts like bias and fairness gained significant attention among researchers and practitioners, in both industry and academia. Such issues typically emerge from the data, which comes with varying levels of quality, used to train supervised machine learning systems. With the commercialization and deployment of such systems that are sometimes delegated to make life-changing decisions, significant efforts are being made towards the identification and removal of possible sources of data bias that may resurface to the final end user or in the decisions being made. In this paper, we present research results that show how bias in data affects end users, where bias is originated, and provide a viewpoint about what we should do about it. We argue that data bias is not something that should necessarily be removed in all cases, and that research attention should instead shift from bias removal towards the identification, measurement, indexing, surfacing, and adapting for bias, which we name bias management.

cs.LG↗

arXiv2023

On the Impact of Data Quality on Image Classification Fairness

With the proliferation of algorithmic decision-making, increased scrutiny has been placed on these systems. This paper explores the relationship between the quality of the training data and the overall fairness of the models trained with such data in the context of supervised classification. We measure key fairness metrics across a range of algorithms over multiple image classification datasets that have a varying level of noise in both the labels and the training data itself. We describe noise in the labels as inaccuracies in the labelling of the data in the training set and noise in the data as distortions in the data, also in the training set. By adding noise to the original datasets, we can explore the relationship between the quality of the training data and the fairness of the output of the models trained on that data.

cs.CV↗

arXiv2022

Crowdsourced Fact-Checking at Twitter: How Does the Crowd Compare With Experts?

Fact-checking is one of the effective solutions in fighting online misinformation. However, traditional fact-checking is a process requiring scarce expert human resources, and thus does not scale well on social media because of the continuous flow of new content to be checked. Methods based on crowdsourcing have been proposed to tackle this challenge, as they can scale with a smaller cost, but, while they have shown to be feasible, have always been studied in controlled environments. In this work, we study the first large-scale effort of crowdsourced fact-checking deployed in practice, started by Twitter with the Birdwatch program. Our analysis shows that crowdsourcing may be an effective fact-checking strategy in some settings, even comparable to results obtained by human experts, but does not lead to consistent, actionable results in others. We processed 11.9k tweets verified by the Birdwatch program and report empirical evidence of i) differences in how the crowd and experts select content to be fact-checked, ii) how the crowd and the experts retrieve different resources to fact-check, and iii) the edge the crowd shows in fact-checking scalability and efficiency as compared to expert checkers.

cs.IR↗