Tokenization and Morphology in Multilingual Language Models: A~Comparative Analysis of mT5 and ByT5

Morphology is a crucial factor for multilingual language modeling as it poses direct challenges for tokenization. Here, we seek to understand how tokenization influences the morphological knowledge encoded in multilingual language models. Specifically, we capture the impact of tokenization by contrasting two multilingual language models: mT5 and ByT5. The two models share the same architecture, training objective, and training data and only differ in their tokenization strategies: subword tokenization vs. character-level tokenization. Probing the morphological knowledge encoded in these models on four tasks and 17 languages, our analyses show that multilingual language models learn the morphological systems of some languages better than others despite similar average performance and that morphological information is encoded in the middle and late layers, where characted-based models need a few more layers to yield commensurate probing accuracy. Finally, we show that languages with more irregularities benefit more from having a higher share of the pre-training data.

POWN: Prototypical Open-World Node Classification

We consider the problem of \textit{true} open-world semi-supervised node classification, in which nodes in a graph either belong to known or new classes, with the latter not present during training. Existing methods detect and reject new classes but fail to distinguish between different new classes. We adapt existing methods and show they do not solve the problem sufficiently. We introduce a novel end-to-end approach for classification into known classes and new classes based on class prototypes, which we call Prototypical Open-World Learning for Node Classification (POWN). Our method combines graph semi-supervised learning, self-supervised learning, and pseudo-labeling to learn prototype representations of new classes in a zero-shot way. In contrast to existing solutions from the vision domain, POWN does not require data augmentation techniques for node classification. Experiments on benchmark datasets demonstrate the effectiveness of POWN, where it outperforms baselines by up to $20\%$ accuracy on the small and up to $30\%$ on the large datasets. Source code is available at https://github.com/Bobowner/POWN.

Emergent communication and learning pressures in language models: a language evolution perspective

Language models and humans are two types of learning systems. Finding or facilitating commonalities could enable major breakthroughs in our understanding of the acquisition and evolution of language. Many theories of language evolution rely heavily on learning biases and learning pressures. Yet due to substantial differences in learning pressures, it is questionable whether the similarity between humans and machines is sufficient for insights to carry over and to be worth testing with human participants. Here, we review the emergent communication literature, a subfield of multi-agent reinforcement learning, from a language evolution perspective. We find that the emergent communication literature excels at designing and adapting models to recover initially absent linguistic phenomena of natural languages. Based on a short literature review, we identify key pressures that have recovered initially absent human patterns in emergent communication models: communicative success, efficiency, learnability, and other psycho-/sociolinguistic factors. We argue that this may serve as inspiration for how to design language models for language acquisition and language evolution research.

GenCodeSearchNet: A Benchmark Test Suite for Evaluating Generalization in Programming Language Understanding

Language models can serve as a valuable tool for software developers to increase productivity. Large generative models can be used for code generation and code completion, while smaller encoder-only models are capable of performing code search tasks using natural language queries.These capabilities are heavily influenced by the quality and diversity of the available training data. Source code datasets used for training usually focus on the most popular languages and testing is mostly conducted on the same distributions, often overlooking low-resource programming languages. Motivated by the NLP generalization taxonomy proposed by Hupkes et.\,al., we propose a new benchmark dataset called GenCodeSearchNet (GeCS) which builds upon existing natural language code search datasets to systemically evaluate the programming language understanding generalization capabilities of language models. As part of the full dataset, we introduce a new, manually curated subset StatCodeSearch that focuses on R, a popular but so far underrepresented programming language that is often used by researchers outside the field of computer science. For evaluation and comparison, we collect several baseline results using fine-tuned BERT-style models and GPT-style large language models in a zero-shot setting.

Open-World Lifelong Graph Learning

We study the problem of lifelong graph learning in an open-world scenario, where a model needs to deal with new tasks and potentially unknown classes. We utilize Out-of-Distribution (OOD) detection methods to recognize new classes and adapt existing non-graph OOD detection methods to graph data. Crucially, we suggest performing new class detection by combining OOD detection methods with information aggregated from the graph neighborhood. Most OOD detection methods avoid determining a crisp threshold for deciding whether a vertex is OOD. To tackle this problem, we propose a Weakly-supervised Relevance Feedback (Open-WRF) method, which decreases the sensitivity to thresholds in OOD detection. We evaluate our approach on six benchmark datasets. Our results show that the proposed neighborhood aggregation method for OOD scores outperforms existing methods independent of the underlying graph neural network. Furthermore, we demonstrate that our Open-WRF method is more robust to threshold selection and analyze the influence of graph neighborhood on OOD detection. The aggregation and threshold methods are compatible with arbitrary graph neural networks and OOD detection methods, making our approach versatile and applicable to many real-world applications.

What makes a language easy to deep-learn?

Neural networks drive the success of natural language processing. A fundamental property of natural languages is their compositional structure, allowing us to describe new meanings systematically. However, neural networks notoriously struggle with systematic generalization and do not necessarily benefit from a compositional structure in emergent communication simulations. Here, we test how neural networks compare to humans in learning and generalizing a new language. We do this by closely replicating an artificial language learning study (conducted originally with human participants) and evaluating the memorization and generalization capabilities of deep neural networks with respect to the degree of structure in the input language. Our results show striking similarities between humans and deep neural networks: More structured linguistic input leads to more systematic generalization and better convergence between humans and neural network agents and between different neural agents. We then replicate this structure bias found in humans and our recurrent neural networks with a Transformer-based large language model (GPT-3), showing a similar benefit for structured linguistic input regarding generalization systematicity and memorization errors. These findings show that the underlying structure of languages is crucial for systematic generalization. Due to the correlation between community size and linguistic structure in natural languages, our findings underscore the challenge of automated processing of low-resource languages. Nevertheless, the similarity between humans and machines opens new avenues for language evolution research.

Emergent Communication for Understanding Human Language Evolution: What's Missing?

Emergent communication protocols among humans and artificial neural network agents do not yet share the same properties and show some critical mismatches in results. We describe three important phenomena with respect to the emergence and benefits of compositionality: ease-of-learning, generalization, and group size effects (i.e., larger groups create more systematic languages). The latter two are not fully replicated with neural agents, which hinders the use of neural emergent communication for language evolution research. We argue that one possible reason for these mismatches is that key cognitive and communicative constraints of humans are not yet integrated. Specifically, in humans, memory constraints and the alternation between the roles of speaker and listener underlie the emergence of linguistic structure, yet these constraints are typically absent in neural simulations. We suggest that introducing such communicative and cognitive constraints would promote more linguistically plausible behaviors with neural agents.

Bag-of-Words vs. Sequence vs. Graph vs. Hierarchy for Single- and Multi-Label Text Classification

Graph neural networks have triggered a resurgence of graph-based text classification methods, defining today's state of the art. We show that a simple multi-layer perceptron (MLP) using a Bag of Words (BoW) outperforms the recent graph-based models TextGCN and HeteGCN in an inductive text classification setting and is comparable with HyperGAT in single-label classification. We also run our own experiments on multi-label classification, where the simple MLP outperforms the recent sequential-based gMLP and aMLP models. Moreover, we fine-tune a sequence-based BERT and a lightweight DistilBERT model, which both outperform all models on both single-label and multi-label settings in most datasets. These results question the importance of synthetic graphs used in modern text classifiers. In terms of parameters, DistilBERT is still twice as large as our BoW-based wide MLP, while graph-based models like TextGCN require setting up an $\mathcal{O}(N^2)$ graph, where $N$ is the vocabulary plus corpus size.

Lifelong Learning in Evolving Graphs with Limited Labeled Data and Unseen Class Detection

Large-scale graph data in the real-world are often dynamic rather than static. The data are changing with new nodes, edges, and even classes appearing over time, such as in citation networks and research-and-development collaboration networks. Graph neural networks (GNNs) have emerged as the standard method for numerous tasks on graph-structured data. In this work, we employ a two-step procedure to explore how GNNs can be incrementally adapted to new unseen graph data. First, we analyze the verge between transductive and inductive learning on standard benchmark datasets. After inductive pretraining, we add unlabeled data to the graph and show that the models are stable. Then, we explore the case of continually adding more and more labeled data, while considering cases, where not all past instances are annotated with class labels. Furthermore, we introduce new classes while the graph evolves and explore methods that automatically detect instances from previously unseen classes. In order to deal with evolving graphs in a principled way, we propose a lifelong learning framework for graph data along with an evaluation protocol. In this framework, we evaluate representative GNN architectures. We observe that implicit knowledge within model parameters becomes more important when explicit knowledge, i.e., data from past tasks, is limited. We find that in open-world node classification, the data from surprisingly few past tasks are sufficient to reach the performance reached by remembering data from all past tasks. In the challenging task of unseen class detection, we find that using a weighted cross-entropy loss is important for stability.

Forget me not: A Gentle Reminder to Mind the Simple Multi-Layer Perceptron Baseline for Text Classification

Graph neural networks have triggered a resurgence of graph-based text classification. We show that already a simple MLP baseline achieves comparable performance on benchmark datasets, questioning the importance of synthetic graph structures. When considering an inductive scenario, i. e., when adding new documents to a corpus, a simple MLP even outperforms the recent graph-based models TextGCN and HeteGCN and is comparable with HyperGAT. We further fine-tune DistilBERT and find that it outperforms all state-of-the-art models. We suggest that future studies use at least an MLP baseline to contextualize the results. We provide recommendations for the design and training of such a baseline.