This thesis delves into a fortiori arguments in deductive reasoning, underscoring their relevance in various domains such as law, philosophy, and artificial intelligence. The research is centred on employing GPT-3.5-turbo to automate the analysis of these arguments, with a focus on understanding intricate reasoning processes, generating clear and coherent explanations, and creating novel arguments. The methodology encompasses a series of tasks including detailed reasoning, interpretation, and the augmentation of a fortiori arguments. It involves meticulously identifying these arguments in diverse contexts, differentiating comparative elements, and categorizing them based on their logical structure. Extensive experiments reveals the challenges encountered by GPT-3.5-turbo in accurately detecting and classifying a fortiori arguments. Nevertheless, the model demonstrates a performance that rivals specialized models, particularly in extracting key components and interpreting underlying properties. The integration of external information into the model's processing significantly elevates the quality of the generated explanations. Additionally, the model exhibits a noteworthy capability in augmenting arguments, thus contributing to the enrichment of the data set. Despite facing certain limitations, this thesis makes significant contributions to the fields of artificial intelligence and logical reasoning. It introduces novel methodologies, establishes a rigorous evaluation framework, and provides deep insights that set the stage for future advancements in automated logical reasoning. The findings and methodologies presented herein not only underscore the potential of AI in complex reasoning tasks but also highlight areas for future research and development.

Large Language Models (LLMs) and chatbots show significant promise in streamlining the legal intake process. This advancement can greatly reduce the workload and costs for legal aid organizations, improving availability while making legal assistance more accessible to a broader audience. However, a key challenge with current LLMs is their tendency to overconfidently deliver an immediate 'best guess' to a client's question based on the output distribution learned over the training data. This approach often overlooks the client's actual intentions or the specifics of their legal situation. As a result, clients may not realize the importance of providing essential additional context or expressing their underlying intentions, which are crucial for their legal cases. Traditionally, logic based decision trees have been used to automate intake for specific access to justice issues, such as immigration and eviction. But those solutions lack scalability. We demonstrate a proof-of-concept using LLMs to elicit and infer clients' underlying intentions and specific legal circumstances through free-form, language-based interactions. We also propose future research directions to use supervised fine-tuning or offline reinforcement learning to automatically incorporate intention and context elicitation in chatbots without explicit prompting.

In response to legislation mandating companies to honor the \textit{right to be forgotten} by erasing user data, it has become imperative to enable data removal in Vertical Federated Learning (VFL) where multiple parties provide private features for model training. In VFL, data removal, i.e., \textit{machine unlearning}, often requires removing specific features across all samples under privacy guarentee in federated learning. To address this challenge, we propose \methname, a novel Gradient Boosting Decision Tree (GBDT) framework that effectively enables both \textit{instance unlearning} and \textit{feature unlearning} without the need for retraining from scratch. Leveraging a robust GBDT structure, we enable effective data deletion while reducing degradation of model performance. Extensive experimental results on popular datasets demonstrate that our method achieves superior model utility and forgetfulness compared to \textit{state-of-the-art} methods. To our best knowledge, this is the first work that investigates machine unlearning in VFL scenarios.

Aiming to train and deploy predictive models, organizations collect large amounts of detailed client data, risking the exposure of private information in the event of a breach. To mitigate this, policymakers increasingly demand compliance with the data minimization (DM) principle, restricting data collection to only that data which is relevant and necessary for the task. Despite regulatory pressure, the problem of deploying machine learning models that obey DM has so far received little attention. In this work, we address this challenge in a comprehensive manner. We propose a novel vertical DM (vDM) workflow based on data generalization, which by design ensures that no full-resolution client data is collected during training and deployment of models, benefiting client privacy by reducing the attack surface in case of a breach. We formalize and study the corresponding problem of finding generalizations that both maximize data utility and minimize empirical privacy risk, which we quantify by introducing a diverse set of policy-aligned adversarial scenarios. Finally, we propose a range of baseline vDM algorithms, as well as Privacy-aware Tree (PAT), an especially effective vDM algorithm that outperforms all baselines across several settings. We plan to release our code as a publicly available library, helping advance the standardization of DM for machine learning. Overall, we believe our work can help lay the foundation for further exploration and adoption of DM principles in real-world applications.

We introduce the Song Describer dataset (SDD), a new crowdsourced corpus of high-quality audio-caption pairs, designed for the evaluation of music-and-language models. The dataset consists of 1.1k human-written natural language descriptions of 706 music recordings, all publicly accessible and released under Creative Common licenses. To showcase the use of our dataset, we benchmark popular models on three key music-and-language tasks (music captioning, text-to-music generation and music-language retrieval). Our experiments highlight the importance of cross-dataset evaluation and offer insights into how researchers can use SDD to gain a broader understanding of model performance.

With the proliferation of social media, accurate detection of hate speech has become critical to ensure safety online. To combat nuanced forms of hate speech, it is important to identify and thoroughly explain hate speech to help users understand its harmful effects. Recent benchmarks have attempted to tackle this issue by training generative models on free-text annotations of implications in hateful text. However, we find significant reasoning gaps in the existing annotations schemes, which may hinder the supervision of detection models. In this paper, we introduce a hate speech detection framework, HARE, which harnesses the reasoning capabilities of large language models (LLMs) to fill these gaps in explanations of hate speech, thus enabling effective supervision of detection models. Experiments on SBIC and Implicit Hate benchmarks show that our method, using model-generated data, consistently outperforms baselines, using existing free-text human annotations. Analysis demonstrates that our method enhances the explanation quality of trained models and improves generalization to unseen datasets. Our code is available at https://github.com/joonkeekim/hare-hate-speech.git.

Explainable AI (XAI) aims to provide insight into opaque model reasoning to humans and as such is an interdisciplinary field by nature. In this paper, we interviewed 10 practitioners to understand the possible usability of training data attribution (TDA) explanations and to explore the design space of such an approach. We confirmed that training data quality is often the most important factor for high model performance in practice and model developers mainly rely on their own experience to curate data. End-users expect explanations to enhance their interaction with the model and do not necessarily prioritise but are open to training data as a means of explanation. Within our participants, we found that TDA explanations are not well-known and therefore not used. We urge the community to focus on the utility of TDA techniques from the human-machine collaboration perspective and broaden the TDA evaluation to reflect common use cases in practice.

The increased use of AI systems is associated with multi-faceted societal, environmental, and economic consequences. These include non-transparent decision-making processes, discrimination, increasing inequalities, rising energy consumption and greenhouse gas emissions in AI model development and application, and an increasing concentration of economic power. By considering the multi-dimensionality of sustainability, this paper takes steps towards substantiating the call for an overarching perspective on "sustainable AI". It presents the SCAIS Framework (Sustainability Criteria and Indicators for Artificial Intelligence Systems) which contains a set 19 sustainability criteria for sustainable AI and 67 indicators that is based on the results of a critical review and expert workshops. This interdisciplinary approach contributes a unique holistic perspective to facilitate and structure the discourse on sustainable AI. Further, it provides a concrete framework that lays the foundation for developing standards and tools to support the conscious development and application of AI systems.

Effective and rapid decision-making from randomized controlled trials (RCTs) requires unbiased and precise treatment effect inferences. Two strategies to address this requirement are to adjust for covariates that are highly correlated with the outcome, and to leverage historical control information via Bayes' theorem. We propose a new Bayesian prognostic covariate adjustment methodology, referred to as Bayesian PROCOVA, that combines these two strategies. Covariate adjustment in Bayesian PROCOVA is based on generative artificial intelligence (AI) algorithms that construct a digital twin generator (DTG) for RCT participants. The DTG is trained on historical control data and yields a digital twin (DT) probability distribution for each RCT participant's outcome under the control treatment. The expectation of the DT distribution, referred to as the prognostic score, defines the covariate for adjustment. Historical control information is leveraged via an additive mixture prior with two components: an informative prior probability distribution specified based on historical control data, and a weakly informative prior distribution. The mixture weight determines the extent to which posterior inferences are drawn from the informative component, versus the weakly informative component. This weight has a prior distribution as well, and so the entire additive mixture prior is completely pre-specifiable without involving any RCT information. We establish an efficient Gibbs algorithm for sampling from the posterior distribution, and derive closed-form expressions for the posterior mean and variance of the treatment effect parameter conditional on the weight, in Bayesian PROCOVA. We evaluate efficiency gains of Bayesian PROCOVA via its bias control and variance reduction compared to frequentist PROCOVA in simulation studies that encompass different discrepancies. These gains translate to smaller RCTs.

As the ongoing conflict between Israel and Hamas and its devastating effects play out in real time on social media, users are continuing to criticise tech firms for what they say is unfair content censorship – pulling into sharp focus longstanding concerns about the opaque algorithms that shape our online worlds. From the early days of the conflict, social media users have expressed outrage at allegedly uneven censorship of pro-Palestinian content on platforms like Instagram and Facebook. Meta has denied intentionally suppressing the content, saying that with more posts going up about the conflict, "content that doesn't violate our policies may be removed in error". But a third-party investigation (commissioned by Meta last year and conducted by the independent consultancy Business for Social Responsibility) had previously determined Meta had violated Palestinian human rights by censoring content related to Israel's attacks on Gaza in 2021, and incidents in recent weeks have revealed further issues with Meta's algorithmic moderation. Instagram's automated translation feature mistakenly added the word "terrorist" to Palestinian profiles and WhatsApp, also owned by Meta, created auto-generated illustrations of gun-wielding children when prompted with the word "Palestine".