"What's my model inside of?": Exploring the role of environments for grounded natural language understanding

In contrast to classical cognitive science which studied brains in isolation, ecological approaches focused on the role of the body and environment in shaping cognition. Similarly, in this thesis we adopt an ecological approach to grounded natural language understanding (NLU) research. Grounded language understanding studies language understanding systems situated in the context of events, actions and precepts in naturalistic/simulated virtual environments. Where classic research tends to focus on designing new models and optimization methods while treating environments as given, we explore the potential of environment design for improving data collection and model development. We developed novel training and annotation approaches for procedural text understanding based on text-based game environments. We also drew upon embodied cognitive linguistics literature to propose a roadmap for grounded NLP research, and to inform the development of a new benchmark for measuring the progress of large language models on challenging commonsense reasoning tasks. We leveraged the richer supervision provided by text-based game environments to develop Breakpoint Transformers, a novel approach to modeling intermediate semantic information in long narrative or procedural texts. Finally, we integrated theories on the role of environments in collective human intelligence to propose a design for AI-augmented "social thinking environments" for knowledge workers like scientists.

BRAIn: Bayesian Reward-conditioned Amortized Inference for natural language generation from feedback

Following the success of Proximal Policy Optimization (PPO) for Reinforcement Learning from Human Feedback (RLHF), new techniques such as Sequence Likelihood Calibration (SLiC) and Direct Policy Optimization (DPO) have been proposed that are offline in nature and use rewards in an indirect manner. These techniques, in particular DPO, have recently become the tools of choice for LLM alignment due to their scalability and performance. However, they leave behind important features of the PPO approach. Methods such as SLiC or RRHF make use of the Reward Model (RM) only for ranking/preference, losing fine-grained information and ignoring the parametric form of the RM (eg., Bradley-Terry, Plackett-Luce), while methods such as DPO do not use even a separate reward model. In this work, we propose a novel approach, named BRAIn, that re-introduces the RM as part of a distribution matching approach.BRAIn considers the LLM distribution conditioned on the assumption of output goodness and applies Bayes theorem to derive an intractable posterior distribution where the RM is explicitly represented. BRAIn then distills this posterior into an amortized inference network through self-normalized importance sampling, leading to a scalable offline algorithm that significantly outperforms prior art in summarization and AntropicHH tasks. BRAIn also has interesting connections to PPO and DPO for specific RM choices.

Can Large Language Models Learn Independent Causal Mechanisms?

Despite impressive performance on language modelling and complex reasoning tasks, Large Language Models (LLMs) fall short on the same tasks in uncommon settings or with distribution shifts, exhibiting some lack of generalisation ability. This issue has usually been alleviated by feeding more training data into the LLM. However, this method is brittle, as the scope of tasks may not be readily predictable or may evolve, and updating the model with new data generally requires extensive additional training. By contrast, systems, such as causal models, that learn abstract variables and causal relationships can demonstrate increased robustness against changes in the distribution. One reason for this success is the existence and use of Independent Causal Mechanisms (ICMs) representing high-level concepts that only sparsely interact. In this work, we apply two concepts from causality to learn ICMs within LLMs. We develop a new LLM architecture composed of multiple sparsely interacting language modelling modules. We introduce a routing scheme to induce specialisation of the network into domain-specific modules. We also present a Mutual Information minimisation objective that trains a separate module to learn abstraction and domain-invariant mechanisms. We show that such causal constraints can improve out-of-distribution performance on abstract and causal reasoning tasks.

Active Learning for Graphs with Noisy Structures

Graph Neural Networks (GNNs) have seen significant success in tasks such as node classification, largely contingent upon the availability of sufficient labeled nodes. Yet, the excessive cost of labeling large-scale graphs led to a focus on active learning on graphs, which aims for effective data selection to maximize downstream model performance. Notably, most existing methods assume reliable graph topology, while real-world scenarios often present noisy graphs. Given this, designing a successful active learning framework for noisy graphs is highly needed but challenging, as selecting data for labeling and obtaining a clean graph are two tasks naturally interdependent: selecting high-quality data requires clean graph structure while cleaning noisy graph structure requires sufficient labeled data. Considering the complexity mentioned above, we propose an active learning framework, GALClean, which has been specifically designed to adopt an iterative approach for conducting both data selection and graph purification simultaneously with best information learned from the prior iteration. Importantly, we summarize GALClean as an instance of the Expectation-Maximization algorithm, which provides a theoretical understanding of its design and mechanisms. This theory naturally leads to an enhanced version, GALClean+. Extensive experiments have demonstrated the effectiveness and robustness of our proposed method across various types and levels of noisy graphs.

Type-based Neural Link Prediction Adapter for Complex Query Answering

Answering complex logical queries on incomplete knowledge graphs (KGs) is a fundamental and challenging task in multi-hop reasoning. Recent work defines this task as an end-to-end optimization problem, which significantly reduces the training cost and enhances the generalization of the model by a pretrained link predictors for query answering. However, most existing proposals ignore the critical semantic knowledge inherently available in KGs, such as type information, which could help answer complex logical queries. To this end, we propose TypE-based Neural Link Prediction Adapter (TENLPA), a novel model that constructs type-based entity-relation graphs to discover the latent relationships between entities and relations by leveraging type information in KGs. Meanwhile, in order to effectively combine type information with complex logical queries, an adaptive learning mechanism is introduced, which is trained by back-propagating during the complex query answering process to achieve adaptive adjustment of neural link predictors. Experiments on 3 standard datasets show that TENLPA model achieves state-of-the-art performance on complex query answering with good generalization and robustness.

Lifelogging As An Extreme Form of Personal Information Management -- What Lessons To Learn

Personal data includes the digital footprints that we leave behind as part of our everyday activities, both online and offline in the real world. It includes data we collect ourselves, such as from wearables, as well as the data collected by others about our online behaviour and activities. Sometimes we are able to use the personal data we ourselves collect, in order to examine some parts of our lives but for the most part, our personal data is leveraged by third parties including internet companies, for services like targeted advertising and recommendations. Lifelogging is a form of extreme personal data gathering and in this article we present an overview of the tools used to manage access to lifelogs as demonstrated at the most recent of the annual Lifelog Search Challenge benchmarking workshops. Here, experimental systems are showcased in live, real time information seeking tasks by real users. This overview of these systems' capabilities show the range of possibilities for accessing our own personal data which may, in time, become more easily available as consumer-level services.

A Comparative Analysis of Gene Expression Profiling by Statistical and Machine Learning Approaches

Many machine learning models have been proposed to classify phenotypes from gene expression data. In addition to their good performance, these models can potentially provide some understanding of phenotypes by extracting explanations for their decisions. These explanations often take the form of a list of genes ranked in order of importance for the predictions, the highest-ranked genes being interpreted as linked to the phenotype. We discuss the biological and the methodological limitations of such explanations. Experiments are performed on several datasets gathering cancer and healthy tissue samples from the TCGA, GTEx and TARGET databases. A collection of machine learning models including logistic regression, multilayer perceptron, and graph neural network are trained to classify samples according to their cancer type. Gene rankings are obtained from explainability methods adapted to these models, and compared to the ones from classical statistical feature selection methods such as mutual information, DESeq2, and EdgeR. Interestingly, on simple tasks, we observe that the information learned by black-box neural networks is related to the notion of differential expression. In all cases, a small set containing the best-ranked genes is sufficient to achieve a good classification. However, these genes differ significantly between the methods and similar classification performance can be achieved with numerous lower ranked genes. In conclusion, although these methods enable the identification of biomarkers characteristic of certain pathologies, our results question the completeness of the selected gene sets and thus of explainability by the identification of the underlying biological processes.

Multilingual Gradient Word-Order Typology from Universal Dependencies

While information from the field of linguistic typology has the potential to improve performance on NLP tasks, reliable typological data is a prerequisite. Existing typological databases, including WALS and Grambank, suffer from inconsistencies primarily caused by their categorical format. Furthermore, typological categorisations by definition differ significantly from the continuous nature of phenomena, as found in natural language corpora. In this paper, we introduce a new seed dataset made up of continuous-valued data, rather than categorical data, that can better reflect the variability of language. While this initial dataset focuses on word-order typology, we also present the methodology used to create the dataset, which can be easily adapted to generate data for a broader set of features and languages.

Dynamic Occupancy Grids for Object Detection: A Radar-Centric Approach

Dynamic Occupancy Grid Mapping is a technique used to generate a local map of the environment containing both static and dynamic information. Typically, these maps are primarily generated using lidar measurements. However, with improvements in radar sensing, resulting in better accuracy and higher resolution, radar is emerging as a viable alternative to lidar as the primary sensor for mapping. In this paper, we propose a radar-centric dynamic occupancy grid mapping algorithm with adaptations to the state computation, inverse sensor model, and field-of-view computation tailored to the specifics of radar measurements. We extensively evaluate our approach using real data to demonstrate its effectiveness and establish the first benchmark for radar-based dynamic occupancy grid mapping using the publicly available Radarscenes dataset.

Different Tastes of Entities: Investigating Human Label Variation in Named Entity Annotations

Named Entity Recognition (NER) is a key information extraction task with a long-standing tradition. While recent studies address and aim to correct annotation errors via re-labeling efforts, little is known about the sources of human label variation, such as text ambiguity, annotation error, or guideline divergence. This is especially the case for high-quality datasets and beyond English CoNLL03. This paper studies disagreements in expert-annotated named entity datasets for three languages: English, Danish, and Bavarian. We show that text ambiguity and artificial guideline changes are dominant factors for diverse annotations among high-quality revisions. We survey student annotations on a subset of difficult entities and substantiate the feasibility and necessity of manifold annotations for understanding named entity ambiguities from a distributional perspective.