Language models (LMs) trained on vast quantities of text data can acquire sophisticated skills such as generating summaries, answering questions or generating code. However, they also manifest behaviors that violate human preferences, e.g., they can generate offensive content, falsehoods or perpetuate social biases. In this thesis, I explore several approaches to aligning LMs with human preferences. First, I argue that aligning LMs can be seen as Bayesian inference: conditioning a prior (base, pretrained LM) on evidence about human preferences (Chapter 2). Conditioning on human preferences can be implemented in numerous ways. In Chapter 3, I investigate the relation between two approaches to finetuning pretrained LMs using feedback given by a scoring function: reinforcement learning from human feedback (RLHF) and distribution matching. I show that RLHF can be seen as a special case of distribution matching but distributional matching is strictly more general. In chapter 4, I show how to extend the distribution matching to conditional language models. Finally, in chapter 5 I explore a different root: conditioning an LM on human preferences already during pretraining. I show that involving human feedback from the very start tends to be more effective than using it only during supervised finetuning. Overall, these results highlight the room for alignment techniques different from and complementary to RLHF.

Multi-modal relation extraction (MMRE) is a challenging task that aims to identify relations between entities in text leveraging image information. Existing methods are limited by their neglect of the multiple entity pairs in one sentence sharing very similar contextual information (ie, the same text and image), resulting in increased difficulty in the MMRE task. To address this limitation, we propose the Variational Multi-Modal Hypergraph Attention Network (VM-HAN) for multi-modal relation extraction. Specifically, we first construct a multi-modal hypergraph for each sentence with the corresponding image, to establish different high-order intra-/inter-modal correlations for different entity pairs in each sentence. We further design the Variational Hypergraph Attention Networks (V-HAN) to obtain representational diversity among different entity pairs using Gaussian distribution and learn a better hypergraph structure via variational attention. VM-HAN achieves state-of-the-art performance on the multi-modal relation extraction task, outperforming existing methods in terms of accuracy and efficiency.

Large Language Models (LLMs) exhibit impressive zero/few-shot inference and generation quality for high-resource languages (HRLs). A few of them have been trained on low-resource languages (LRLs) and give decent performance. Owing to the prohibitive costs of training LLMs, they are usually used as a network service, with the client charged by the count of input and output tokens. The number of tokens strongly depends on the script and language, as well as the LLM's subword vocabulary. We show that LRLs are at a pricing disadvantage, because the well-known LLMs produce more tokens for LRLs than HRLs. This is because most currently popular LLMs are optimized for HRL vocabularies. Our objective is to level the playing field: reduce the cost of processing LRLs in contemporary LLMs while ensuring that predictive and generative qualities are not compromised. As means to reduce the number of tokens processed by the LLM, we consider code-mixing, translation, and transliteration of LRLs to HRLs. We perform an extensive study using the IndicXTREME classification and six generative tasks dataset, covering 15 Indic and 3 other languages, while using GPT-4 (one of the costliest LLM services released so far) as a commercial LLM. We observe and analyze interesting patterns involving token count, cost, and quality across a multitude of languages and tasks. We show that choosing the best policy to interact with the LLM can reduce cost by 90% while giving better or comparable performance compared to communicating with the LLM in the original LRL.

The prevalence of digital media and evolving sociopolitical dynamics have significantly amplified the dissemination of hateful content. Existing studies mainly focus on classifying texts into binary categories, often overlooking the continuous spectrum of offensiveness and hatefulness inherent in the text. In this research, we present an extensive benchmark dataset for Amharic, comprising 8,258 tweets annotated for three distinct tasks: category classification, identification of hate targets, and rating offensiveness and hatefulness intensities. Our study highlights that a considerable majority of tweets belong to the less offensive and less hate intensity levels, underscoring the need for early interventions by stakeholders. The prevalence of ethnic and political hatred targets, with significant overlaps in our dataset, emphasizes the complex relationships within Ethiopia's sociopolitical landscape. We build classification and regression models and investigate the efficacy of models in handling these tasks. Our results reveal that hate and offensive speech can not be addressed by a simplistic binary classification, instead manifesting as variables across a continuous range of values. The Afro-XLMR-large model exhibits the best performances achieving F1-scores of 75.30%, 70.59%, and 29.42% for the category, target, and regression tasks, respectively. The 80.22% correlation coefficient of the Afro-XLMR-large model indicates strong alignments.

Evaluation benchmarks play a pivotal role in the development of artificial intelligence (AI) systems. Traditionally, natural language processing (NLP) benchmarks have primarily focused on assessing specific and relatively straightforward abilities. However, the advent of large language models (LLMs), also known as foundation models, has brought about a paradigm shift. These powerful models have demonstrated a wide array of novel capabilities, prompting a redirection in the evaluation focus towards more general and intricate skills, such as comprehensive world knowledge and complex reasoning abilities. To align with the remarkable advancements in LLMs, new benchmarks have emerged to probe the diverse and multifaceted capabilities of these models. For instance, MMLU [8], HellaSwag [25], ARC [4], and TruthfulQA [10] are benchmark datasets that have garnered widespread recognition among researchers and are frequently employed on leaderboards to evaluate the performance of language models. However, these benchmarks are primarily tailored to the English language, resulting in a limited understanding of LLMs' capabilities in other languages, including Vietnamese. Despite the recent surge in powerful Vietnamese LLMs, such as Vistral-7B-Chat [12], PhoGPT-4B-Chat [13], and VinaLLaMA-7B-Chat [16], benchmarking these models on datasets translated from English to Vietnamese, even with perfect translations, cannot adequately assess the true quality of these language models concerning their knowledge about core interests of Vietnamese users.

Paraphrase generation is a pivotal task in natural language processing (NLP). Existing datasets in the domain lack syntactic and lexical diversity, resulting in paraphrases that closely resemble the source sentences. Moreover, these datasets often contain hate speech and noise, and may unintentionally include non-English language sentences. This research introduces ParaFusion, a large-scale, high-quality English paraphrase dataset developed using Large Language Models (LLM) to address these challenges. ParaFusion augments existing datasets with high-quality data, significantly enhancing both lexical and syntactic diversity while maintaining close semantic similarity. It also mitigates the presence of hate speech and reduces noise, ensuring a cleaner and more focused English dataset. Results show that ParaFusion offers at least a 25% improvement in both syntactic and lexical diversity, measured across several metrics for each data source. The paper also aims to set a gold standard for paraphrase evaluation as it contains one of the most comprehensive evaluation strategies to date. The results underscore the potential of ParaFusion as a valuable resource for improving NLP applications.

The staggering pace with which the capabilities of large language models (LLMs) are increasing, as measured by a range of commonly used natural language understanding (NLU) benchmarks, raises many questions regarding what "understanding" means for a language model and how it compares to human understanding. This is especially true since many LLMs are exclusively trained on text, casting doubt on whether their stellar benchmark performances are reflective of a true understanding of the problems represented by these benchmarks, or whether LLMs simply excel at uttering textual forms that correlate with what someone who understands the problem would say. In this philosophically inspired work, we aim to create some separation between form and meaning, with a series of tests that leverage the idea that world understanding should be consistent across presentational modes - inspired by Fregean senses - of the same meaning. Specifically, we focus on consistency across languages as well as paraphrases. Taking GPT-3.5 as our object of study, we evaluate multisense consistency across five different languages and various tasks. We start the evaluation in a controlled setting, asking the model for simple facts, and then proceed with an evaluation on four popular NLU benchmarks. We find that the model's multisense consistency is lacking and run several follow-up analyses to verify that this lack of consistency is due to a sense-dependent task understanding. We conclude that, in this aspect, the understanding of LLMs is still quite far from being consistent and human-like, and deliberate on how this impacts their utility in the context of learning about human language and understanding.

As AI systems enter into a growing number of societal domains, these systems increasingly shape and are shaped by user preferences, opinions, and behaviors. However, the design of AI systems rarely accounts for how AI and users shape one another. In this position paper, we argue for the development of formal interaction models which mathematically specify how AI and users shape one another. Formal interaction models can be leveraged to (1) specify interactions for implementation, (2) monitor interactions through empirical analysis, (3) anticipate societal impacts via counterfactual analysis, and (4) control societal impacts via interventions. The design space of formal interaction models is vast, and model design requires careful consideration of factors such as style, granularity, mathematical complexity, and measurability. Using content recommender systems as a case study, we critically examine the nascent literature of formal interaction models with respect to these use-cases and design axes. More broadly, we call for the community to leverage formal interaction models when designing, evaluating, or auditing any AI system which interacts with users.

Recent geospatial machine learning studies have shown that the results of model evaluation via cross-validation (CV) are strongly affected by the dissimilarity between the sample data and the prediction locations. In this paper, we propose a method to quantify such a dissimilarity in the interval 0 to 100%, and from the perspective of the data feature space. The proposed method is based on adversarial validation, which is an approach that can check whether sample data and prediction locations can be separated with a binary classifier. To study the effectiveness and generality of our method, we tested it on a series of experiments based on both synthetic and real datasets and with gradually increasing dissimilarities. Results show that the proposed method can successfully quantify dissimilarity across the entire range of values. Next to this, we studied how dissimilarity affects CV evaluations by comparing the results of random CV and of two spatial CV methods, namely block and spatial+ CV. Our results showed that CV evaluations follow similar patterns in all datasets and predictions: when dissimilarity is low (usually lower than 30%), random CV provides the most accurate evaluation results. As dissimilarity increases, spatial CV methods, especially spatial+ CV, become more and more accurate and even outperforming random CV. When dissimilarity is high (>=90%), no CV method provides accurate evaluations. These results show the importance of considering feature space dissimilarity when working with geospatial machine learning predictions, and can help researchers and practitioners to select more suitable CV methods for evaluating their predictions.

This paper presents a systematic literature review of published studies on AI-based automated speech therapy tools for persons with speech sound disorders (SSD). The COVID-19 pandemic has initiated the requirement for automated speech therapy tools for persons with SSD making speech therapy accessible and affordable. However, there are no guidelines for designing such automated tools and their required degree of automation compared to human experts. In this systematic review, we followed the PRISMA framework to address four research questions: 1) what types of SSD do AI-based automated speech therapy tools address, 2) what is the level of autonomy achieved by such tools, 3) what are the different modes of intervention, and 4) how effective are such tools in comparison with human experts. An extensive search was conducted on digital libraries to find research papers relevant to our study from 2007 to 2022. The results show that AI-based automated speech therapy tools for persons with SSD are increasingly gaining attention among researchers. Articulation disorders were the most frequently addressed SSD based on the reviewed papers. Further, our analysis shows that most researchers proposed fully automated tools without considering the role of other stakeholders. Our review indicates that mobile-based and gamified applications were the most frequent mode of intervention. The results further show that only a few studies compared the effectiveness of such tools compared to expert Speech-Language Pathologists (SLP). Our paper presents the state-of-the-art in the field, contributes significant insights based on the research questions, and provides suggestions for future research directions.