Pre-trained language models (LMs) perform well in In-Topic setups, where training and testing data come from the same topics. However, they face challenges in Cross-Topic scenarios where testing data is derived from distinct topics -- such as Gun Control. This study analyzes various LMs with three probing-based experiments to shed light on the reasons behind the In- vs. Cross-Topic generalization gap. Thereby, we demonstrate, for the first time, that generalization gaps and the robustness of the embedding space vary significantly across LMs. Additionally, we assess larger LMs and underscore the relevance of our analysis for recent models. Overall, diverse pre-training objectives, architectural regularization, or data deduplication contribute to more robust LMs and diminish generalization gaps. Our research contributes to a deeper understanding and comparison of language models across different generalization scenarios.
Long documents often exhibit structure with hierarchically organized elements of different functions, such as section headers and paragraphs. Despite the omnipresence of document structure, its role in natural language processing (NLP) remains opaque. Do long-document Transformer models acquire an internal representation of document structure during pre-training? How can structural information be communicated to a model after pre-training, and how does it influence downstream performance? To answer these questions, we develop a novel suite of probing tasks to assess structure-awareness of long-document Transformers, propose general-purpose structure infusion methods, and evaluate the effects of structure infusion on QASPER and Evidence Inference, two challenging long-document NLP tasks. Results on LED and LongT5 suggest that they acquire implicit understanding of document structure during pre-training, which can be further enhanced by structure infusion, leading to improved end-task performance. To foster research on the role of document structure in NLP modeling, we make our data and code publicly available.
Reasoning is a fundamental component for achieving language understanding. Among the multiple types of reasoning, conditional reasoning, the ability to draw different conclusions depending on some condition, has been understudied in large language models (LLMs). Recent prompting methods, such as chain of thought, have significantly improved LLMs on reasoning tasks. Nevertheless, there is still little understanding of what triggers reasoning abilities in LLMs. We hypothesize that code prompts can trigger conditional reasoning in LLMs trained on text and code. We propose a chain of prompts that transforms a natural language problem into code and prompts the LLM with the generated code. Our experiments find that code prompts exhibit a performance boost between 2.6 and 7.7 points on GPT 3.5 across multiple datasets requiring conditional reasoning. We then conduct experiments to discover how code prompts elicit conditional reasoning abilities and through which features. We observe that prompts need to contain natural language text accompanied by high-quality code that closely represents the semantics of the instance text. Furthermore, we show that code prompts are more efficient, requiring fewer demonstrations, and that they trigger superior state tracking of variables or key entities.
The success of task-oriented and document-grounded dialogue systems depends on users accepting and enjoying using them. To achieve this, recently published work in the field of Human-Computer Interaction suggests that the combination of considering demographic information, user emotions and learning from the implicit feedback in their utterances, is particularly important. However, these findings have not yet been transferred to the field of Natural Language Processing, where these data are primarily studied separately. Accordingly, no sufficiently annotated dataset is available. To address this gap, we introduce FEDI, the first English dialogue dataset for task-oriented document-grounded dialogues annotated with demographic information, user emotions and implicit feedback. Our experiments with FLAN-T5, GPT-2 and LLaMA-2 show that these data have the potential to improve task completion and the factual consistency of the generated responses and user acceptance.
We introduce Adapters, an open-source library that unifies parameter-efficient and modular transfer learning in large language models. By integrating 10 diverse adapter methods into a unified interface, Adapters offers ease of use and flexible configuration. Our library allows researchers and practitioners to leverage adapter modularity through composition blocks, enabling the design of complex adapter setups. We demonstrate the library's efficacy by evaluating its performance against full fine-tuning on various NLP tasks. Adapters provides a powerful tool for addressing the challenges of conventional fine-tuning paradigms and promoting more efficient and modular transfer learning. The library is available via https://adapterhub.ml/adapters.
The increased use of large language models (LLMs) across a variety of real-world applications calls for mechanisms to verify the factual accuracy of their outputs. In this work, we present a holistic end-to-end solution for annotating the factuality of LLM-generated responses, which encompasses a multi-stage annotation scheme designed to yield detailed labels concerning the verifiability and factual inconsistencies found in LLM outputs. We design and build an annotation tool to speed up the labelling procedure and ease the workload of raters. It allows flexible incorporation of automatic results in any stage, e.g. automatically-retrieved evidence. We further construct an open-domain document-level factuality benchmark in three-level granularity: claim, sentence and document. Preliminary experiments show that FacTool, FactScore and Perplexity.ai are struggling to identify false claims with the best F1=0.53. Annotation tool, benchmark and code are available at https://github.com/yuxiaw/Factcheck-GPT.
Language models (LMs) have demonstrated remarkable capabilities across a wide range of tasks in various domains. Despite their impressive performance, the reliability of their output is concerning and questionable regarding the demand for AI safety. Assessing the confidence of LM predictions and calibrating them across different tasks with the aim to align LM confidence with accuracy can help mitigate risks and enable LMs to make better decisions. There have been various works in this respect, but there has been no comprehensive overview of this important research area. The present survey aims to bridge this gap. In particular, we discuss methods and techniques for LM confidence estimation and calibration, encompassing different LMs and various tasks. We further outline the challenges of estimating the confidence for large language models and we suggest some promising directions for future work.
In-context learning (ICL) has become one of the most popular learning paradigms. While there is a growing body of literature focusing on prompt engineering, there is a lack of systematic analysis comparing the effects of prompts across different models and tasks. To address this gap, we present a comprehensive prompt analysis based on the sensitivity of a function. Our analysis reveals that sensitivity is an unsupervised proxy for model performance, as it exhibits a strong negative correlation with accuracy. We use gradient-based saliency scores to empirically demonstrate how different prompts affect the relevance of input tokens to the output, resulting in different levels of sensitivity. Furthermore, we introduce sensitivity-aware decoding which incorporates sensitivity estimation as a penalty term in the standard greedy decoding. We show that this approach is particularly helpful when information in the input is scarce. Our work provides a fresh perspective on the analysis of prompts, and contributes to a better understanding of the mechanism of ICL.
Memes are a modern form of communication and meme templates possess a base semantics that is customizable by whomever posts it on social media. Machine learning systems struggle with memes, which is likely due to such systems having insufficient context to understand memes, as there is more to memes than the obvious image and text. Here, to aid understanding of memes, we release a knowledge base of memes and information found on www.knowyourmeme.com, which we call the Know Your Meme Knowledge Base (KYMKB), composed of more than 54,000 images. The KYMKB includes popular meme templates, examples of each template, and detailed information about the template. We hypothesize that meme templates can be used to inject models with the context missing from previous approaches. To test our hypothesis, we create a non-parametric majority-based classifier, which we call Template-Label Counter (TLC). We find TLC more effective than or competitive with fine-tuned baselines. To demonstrate the power of meme templates and the value of both our knowledge base and method, we conduct thorough classification experiments and exploratory data analysis in the context of five meme analysis tasks.
In many domains of argumentation, people's arguments are driven by so-called attitude roots, i.e., underlying beliefs and world views, and their corresponding attitude themes. Given the strength of these latent drivers of arguments, recent work in psychology suggests that instead of directly countering surface-level reasoning (e.g., falsifying given premises), one should follow an argumentation style inspired by the Jiu-Jitsu 'soft' combat system (Hornsey and Fielding, 2017): first, identify an arguer's attitude roots and themes, and then choose a prototypical rebuttal that is aligned with those drivers instead of invalidating those. In this work, we are the first to explore Jiu-Jitsu argumentation for peer review by proposing the novel task of attitude and theme-guided rebuttal generation. To this end, we enrich an existing dataset for discourse structure in peer reviews with attitude roots, attitude themes, and canonical rebuttals. To facilitate this process, we recast established annotation concepts from the domain of peer reviews (e.g., aspects a review sentence is relating to) and train domain-specific models. We then propose strong rebuttal generation strategies, which we benchmark on our novel dataset for the task of end-to-end attitude and theme-guided rebuttal generation and two subtasks.