Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

The ability to track entities is fundamental for language understanding, yet the internal mechanisms governing this capability in Small Language Models (SLMs) are poorly understood. Previous studies often rely on indirect probing or complex interpretability methods, leaving a gap for lightweight diagnostics that connect model behavior to performance. To bridge this gap, we introduce a framework to analyze entity tracking by measuring the attention flow between entity and non-entity tokens within SLMs. We apply this to analyze models both before and after Parameter-Efficient Fine-Tuning (PEFT). Our analysis reveals two key findings. First, SLMs' attentional strategies vary significantly with text type, but entities consistently receive a high degree of focus. Second, we show that PEFT - specifically QLoRA - dramatically improves classification performance on entity-centric tasks by increasing the model's attentional focus on entity-related tokens. Our work provides direct evidence for how PEFT can refine a model's internal mechanisms and establishes attention analysis as a valuable, lightweight diagnostic tool for interpreting and improving SLMs1.

In linguistics, coherence can be achieved by different means, such as by maintaining reference to the same set of entities across sentences and by establishing discourse relations between them. However, most existing work on coherence modeling focuses exclusively on either entity features or discourse relation features, with little attention given to combining the two. In this study, we explore two methods for jointly modeling entities and discourse relations for coherence assessment. Experiments on three benchmark datasets show that integrating both types of features significantly enhances the performance of coherence models, highlighting the benefits of modeling both simultaneously for coherence evaluation.

Discourse coherence theories posit relations between text spans as a key feature of coherent texts. However, existing work on coherence modeling has paid little attention to discourse relations. In this paper, we provide empirical evidence to demonstrate that relation features are correlated with text coherence. Then, we investigate a novel fusion model that uses position-aware attention and a visible matrix to combine text- and relation-based features for coherence assessment. Experimental results on two benchmarks show that our approaches can significantly improve baselines, demonstrating the importance of relation features for coherence modeling.

Understanding temporal relations between events in a text is essential for determining its temporal structure. Recent advancements in large language models (LLMs) have spurred research on temporal relation extraction. However, LLMs perform poorly in zero-shot and few-shot settings, often underperforming smaller fine-tuned models. Despite these limitations, little attention has been given to improving LLMs in temporal structure extraction tasks. This study systematically examines LLMs’ ability to extract and infer discourse-level temporal relations, identifying factors influencing their reasoning and extraction capabilities, including input context, reasoning process and ensuring consistency. We propose a three-step framework to improve LLMs’ temporal relation extraction capabilities: context selection, prompts inspired by Allen’s interval algebra (Allen, 1983), and reflection-based consistency learning (Shinn et al., 2024). Our results show the effectiveness of our method in guiding LLMs towards structured processing of temporal structure in discourse.

With the increasing use of generative Artificial Intelligence (AI) methods to support science workflows, we are interested in the use of discourse-level information to find supporting evidence for AI generated scientific claims. A first step towards this objective is to examine the task of inferring discourse structure in scientific writing. In this work, we present a preliminary investigation of pretrained language model (PLM) and Large Language Model (LLM) approaches for Discourse Relation Classification (DRC), focusing on scientific publications, an under-studied genre for this task. We examine how context can help with the DRC task, with our experiments showing that context, as defined by discourse structure, is generally helpful. We also present an analysis of which scientific discourse relation types might benefit most from context.

Theoretical results from discrete geometry suggest that normed spaces can abstractly embed finite metric spaces with surprisingly low theoretical bounds on distortion in low dimensions. In this paper, inspired by this theoretical insight, we highlight normed spaces as a more flexible and computationally efficient alternative to several popular Riemannian manifolds for learning graph embeddings. Normed space embeddings significantly outperform several popular manifolds on a large range of synthetic and real-world graph reconstruction benchmark datasets while requiring significantly fewer computational resources. We also empirically verify the superiority of normed space embeddings on growing families of graphs associated with negative, zero, and positive curvature, further reinforcing the flexibility of normed spaces in capturing diverse graph structures as graph sizes increase. Lastly, we demonstrate the utility of normed space embeddings on two applied graph embedding tasks, namely, link prediction and recommender systems. Our work highlights the potential of normed spaces for geometric graph representation learning, raises new research questions, and offers a valuable tool for experimental mathematics in the field of finite metric space embeddings. We make our code and data publically available.

Despite the predominance of contextualized embeddings in NLP, approaches to detect semantic change relying on these embeddings and clustering methods underperform simpler counterparts based on static word embeddings. This stems from the poor quality of the clustering methods to produce sense clusters-which struggle to capture word senses, especially those with low frequency. This issue hinders the next step in examining how changes in word senses in one language influence another. To address this issue, we propose a graph-based clustering approach to capture nuanced changes in both high- and low-frequency word senses across time and languages, including the acquisition and loss of these senses over time. Our experimental results show that our approach substantially surpasses previous approaches in the SemEval2020 binary classification task across four languages. Moreover, we showcase the ability of our approach as a versatile visualization tool to detect semantic changes in both intra-language and inter-language setups. We make our code and data available(1).

We consider an unanswered question in the discourse processing community: why do relation classifiers trained on explicit examples (with connectives removed) perform poorly in real implicit scenarios? Prior work claimed this is due to linguistic dissimilarity between explicit and implicit examples but provided no empirical evidence. In this study, we show that one cause for such failure is a label shift after connectives are eliminated. Specifically, we find that the discourse relations expressed by some explicit instances will change when connectives disappear. Unlike previous work manually analyzing a few examples, we present empirical evidence at the corpus level to prove the existence of such shift. Then, we analyze why label shift occurs by considering factors such as the syntactic role played by connectives, ambiguity of connectives, and more. Finally, we investigate two strategies to mitigate the label shift: filtering out noisy data and joint learning with connectives. Experiments on PDTB 2.0, PDTB 3.0, and the GUM dataset demonstrate that classifiers trained with our strategies outperform strong baselines.

Theoretical results from discrete geometry suggest that normed spaces can abstractly embed finite metric spaces with surprisingly low theoretical bounds on distortion in low dimensions. In this paper, inspired by this theoretical insight, we propose normed spaces as a more flexible and computationally efficient alternative to several popular Riemannian manifolds for learning graph embeddings. Our normed space embeddings significantly outperform several popular manifolds on a large range of synthetic and real-world graph reconstruction benchmark datasets while requiring significantly fewer computational resources. We also empirically verify the superiority of normed space embeddings on growing families of graphs associated with negative, zero, and positive curvature, further reinforcing the flexibility of normed spaces in capturing diverse graph structures as graph sizes increase. Lastly, we demonstrate the utility of normed space embeddings on two applied graph embedding tasks, namely, link prediction and recommender systems. Our work highlights the potential of normed spaces for geometric graph representation learning, raises new research questions, and offers a valuable tool for experimental mathematics in the field of finite metric space embeddings.

Automating Cross-lingual Science Journalism (CSJ) aims to generate popular science summaries from English scientific texts for non-expert readers in their local language. We introduce CSJ as a downstream task of text simplification and cross-lingual scientific summarization to facilitate science journalists’ work. We analyze the performance of possible existing solutions as baselines for the CSJ task. Based on these findings, we propose to combine the three components - SELECT, SIMPLIFY and REWRITE (SSR) to produce cross-lingual simplified science summaries for non-expert readers. Our empirical evaluation on the Wikipedia dataset shows that SSR significantly outperforms the baselines for the CSJ task and can serve as a strong baseline for future work. We also perform an ablation study investigating the impact of individual components of SSR. Further, we analyze the performance of SSR on a high-quality, real-world CSJ dataset with human evaluation and in-depth analysis, demonstrating the superior performance of SSR for CSJ.

Recently, there has been a growing interest in designing text generation systems from a discourse coherence perspective, e.g., modeling the interdependence between sentences. Still, recent BERT-based evaluation metrics are weak in recognizing coherence, and thus are not reliable in a way to spot the discourselevel improvements of those text generation systems. In this work, we introduce DiscoScore, a parametrized discourse metric, which uses BERT to model discourse coherence from different perspectives, driven by Centering theory. Our experiments encompass 16 non-discourse and discourse metrics, including DiscoScore and popular coherence models, evaluated on summarization and document-level machine translation (MT). We find that (i) the majority of BERT-based metrics correlate much worse with human rated coherence than early discourse metrics, invented a decade ago; (ii) the recent state-of-the-art BARTScore is weak when operated at system level-which is particularly problematic as systems are typically compared in this manner. DiscoScore, in contrast, achieves strong system-level correlation with human ratings, not only in coherence but also in factual consistency and other aspects, and surpasses BARTScore by over 10 correlation points on average. Further, aiming to understand DiscoScore, we provide justifications to the importance of discourse coherence for evaluation metrics, and explain the superiority of one variant over another. Our code is available at https://github.com/AIPHES/ DiscoScore.

Coherence is an important aspect of text quality, and various approaches have been applied to coherence modeling. However, existing methods solely focus on a single document's coherence patterns, ignoring the underlying correlation between documents. We investigate a GCN-based coherence model that is capable of capturing structural similarities between documents. Our model first creates a graph structure for each document, from where we mine different subgraph patterns. We then construct a heterogeneous graph for the training corpus, connecting documents based on their shared subgraphs. Finally, a GCN is applied to the heterogeneous graph to model the connectivity relationships. We evaluate our method on two tasks, assessing discourse coherence and automated essay scoring. Results show that our GCN-based model outperforms all baselines, achieving a new state-of-the-art on both tasks.

Cross-lingual science journalism generates popular science stories of scientific articles different from the source language for a non-expert audience. Hence, a cross-lingual popular summary must contain the salient content of the input document, and the content should be coherent, comprehensible, and in a local language for the targeted audience. We improve these aspects of cross-lingual summary generation by joint training of two high-level NLP tasks, simplification and cross-lingual summarization. The former task reduces linguistic complexity, and the latter focuses on cross-lingual abstractive summarization. We propose a novel multi-task architecture - SimCSum consisting of one shared encoder and two parallel decoders jointly learning simplification and cross-lingual summarization. We empirically investigate the performance of SimCSum by comparing it with several strong baselines over several evaluation metrics and by human evaluation. Overall, SimCSum demonstrates statistically significant improvements over the state-of-the-art on two non-synthetic cross-lingual scientific datasets. Furthermore, we conduct an in-depth investigation into the linguistic properties of generated summaries and an error analysis.

HITS participated in the Discourse Segmentation (DS, Task 1) and Connective Detection (CD, Task 2) tasks at the DISRPT 2023.Task 1 focuses on segmenting the text into discourse units, while Task 2 aims to detect the discourse connectives.We deployed a framework based on different pre-trained models according to the target language for these two tasks.HITS also participated in the Relation Classification track (Task 3).The main task was recognizing the discourse relation between text spans from different languages.We designed a joint model for languages with a small corpus while separate models for large corpora.The adversarial training strategy is applied to enhance the robustness of relation classifiers.The implementation of our models for three tasks is available at https://github.com/liuwei1206/disrpt2023. * Equal contribution.Wei is responsible for relation classification, while Yi works on discourse segmentation and connective detection.

Recent research has shown that alignment between the structure of graph data and the geometry of an embedding space is crucial for learning high-quality representations of the data. The uniform geometry of Euclidean and hyperbolic spaces allows for representing graphs with uniform geometric and topological features, such as grids and hierarchies, with minimal distortion. However, real-world graph data is characterized by multiple types of geometric and topological features, necessitating more sophisticated geometric embedding spaces. In this work, we utilize the Riemannian symmetric space of symmetric positive definite matrices ( SPD ) to construct graph neural networks that can robustly handle complex graphs. To do this, we develop an innovative library that leverages the SPD gyrocalculus tools [ 26 ] to implement the building blocks of five popular graph neural networks in SPD . Experimental results demonstrate that our graph neural networks in SPD substantially outperform their counterparts in Euclidean and hyperbolic spaces, as well as the Cartesian product thereof, on complex graphs for node and graph classification tasks. We release the library and datasets at https://github.com/andyweizhao/SPD4GNNs .

Implicit discourse relation classification is a challenging task due to the absence of discourse connectives. To overcome this issue, we design an end-to-end neural model to explicitly generate discourse connectives for the task, inspired by the annotation process of PDTB. Specifically, our model jointly learns to generate discourse connectives between arguments and predict discourse relations based on the arguments and the generated connectives. To prevent our relation classifier from being misled by poor connectives generated at the early stage of training while alleviating the discrepancy between training and inference, we adopt Scheduled Sampling to the joint learning. We evaluate our method on three benchmarks, PDTB 2.0, PDTB 3.0, and PCC. Results show that our joint model significantly outperforms various baselines on three datasets, demonstrating its superiority for the task.