EMNLP 2023 Papers

The translation of ambiguous text presents a challenge for translation systems, as it requires using the surrounding context to disambiguate the intended meaning as much as possible. While prior work has studied ambiguities that result from different grammatical features of the source and target language, we study semantic ambiguities that exist in the source (English in this work) itself. In particular, we focus on idioms that are open to both literal and figurative interpretations (e.g., goose egg), and collect TIDE, a dataset of 512 pairs of English sentences containing idioms with disambiguating context such that one is literal (it laid a goose egg) and another is figurative (they scored a goose egg, as in a score of zero). In experiments, we compare MT-specific models and language models for (i) their preference when given an ambiguous subsentence, (ii) their sensitivity to disambiguating context, and (iii) the performance disparity between figurative and literal source sentences. We find that current MT models consistently translate English idioms literally, even when the context suggests a figurative interpretation. On the other hand, LMs are far more context-aware, although there remain disparities across target languages. Our findings underline the potential of LMs as a strong backbone for context-aware translation.

Theory of Mind (ToM) is a critical component of intelligence but its assessment remains the subject of heated debates. Prior research applied human ToM assessments to natural language processing models using either human-created standardized tests or rule-based templates. However, these methods primarily focus on simplistic reasoning and require further validation. Here, we leverage dynamic epistemic logic to isolate a particular component of ToM and to generate controlled problems. We also introduce new verbalization techniques to express these problems in English natural language. Our findings indicate that some language model scaling (from 70M to 6B and 350M to 174B) does not consistently yield results better than random chance. While GPT-4 demonstrates superior epistemic reasoning capabilities, there is still room for improvement. Our code and datasets are publicly available.

Learning logic rules for knowledge graph reasoning is essential as such rules provide interpretable explanations for reasoning and can be generalized to different domains. However, existing methods often face challenges such as searching in a vast search space (e.g., enumeration of relational paths or multiplication of high-dimensional matrices) and inefficient optimization (e.g., techniques based on reinforcement learning or EM algorithm). To address these limitations, this paper proposes a novel framework called LatentLogic to efficiently mine logic rules by controllable generation in the latent space. Specifically, to map the discrete relational paths into the latent space, we leverage a pre-trained VAE and employ a discriminator to establish an energy-based distribution. Additionally, we incorporate a sampler based on ordinary differential equations, enabling the efficient generation of logic rules in our approach. Extensive experiments on benchmark datasets demonstrate the effectiveness and efficiency of our proposed method.

The technology of text-to-SQL has significantly enhanced the efficiency of accessing and manipulating databases. However, limited research has been conducted to study its vulnerabilities emerging from malicious user interaction. By proposing TrojanSQL, a backdoor-based SQL injection framework for text-to-SQL systems, we show how state-of-the-art text-to-SQL parsers can be easily misled to produce harmful SQL statements that can invalidate user queries or compromise sensitive information about the database. The study explores two specific injection attacks, namely $\textit{boolean-based injection}$ and $\textit{union-based injection}$, which use different types of triggers to achieve distinct goals in compromising the parser. Experimental results demonstrate that both medium-sized models based on fine-tuning and LLM-based parsers using prompting techniques are vulnerable to this type of attack, with attack success rates as high as 99\% and 89\%, respectively. We hope that this study will raise more concerns about the potential security risks of building natural language interfaces to databases.

Pre-trained language models (PLMs) have demonstrated their exceptional performance across a wide range of natural language processing tasks. The utilization of PLM-based sentence embeddings enables the generation of contextual representations that capture rich semantic information. However, despite their success with unseen samples, current PLM-based representations suffer from poor robustness in adversarial scenarios. In this paper, we propose RobustEmbed, a self-supervised sentence embedding framework that enhances both generalization and robustness in various text representation tasks and against diverse adversarial attacks. By generating high-risk adversarial perturbations to promote higher invariance in the embedding space and leveraging the perturbation within a novel contrastive objective approach, RobustEmbed effectively learns high-quality sentence embeddings. Our extensive experiments validate the superiority of RobustEmbed over previous state-of-the-art self-supervised representations in adversarial settings, while also showcasing relative improvements in seven semantic textual similarity (STS) tasks and six transfer tasks. Specifically, our framework achieves a significant reduction in attack success rate from 75.51% to 39.62% for the BERTAttack attack technique, along with enhancements of 1.20% and 0.40% in STS tasks and transfer tasks, respectively.

Understanding the prevalence and dynamics of justice partisanship and ideology in the US Supreme Court is critical in studying jurisdiction. Most research quantifies partisanship based on voting behavior, and oral arguments in the courtroom --- the last essential procedure before the final case outcome --- have not been well studied for this purpose. To address this gap, we present a framework for analyzing the language of justices in the courtroom for partisan signals, and study how partisanship in speech aligns with voting patterns. Our results show that the affiliated party of justices can be predicted reliably from their oral contributions. We further show a strong correlation between language partisanship and voting ideology.

A recent focus of large language model (LLM) development, as exemplified by generative search engines, is to incorporate external references to generate and support its claims. However, evaluating the attribution, i.e., verifying whether the generated statement is fully supported by the cited reference, remains an open problem. Although human evaluation is common practice, it is costly and time-consuming. In this paper, we investigate automatic evaluation of attribution given by LLMs. We begin by defining different types of attribution errors, and then explore two approaches for automatic evaluation: prompting LLMs and fine-tuning smaller LMs. The fine-tuning data is repurposed from related tasks such as question answering, fact-checking, natural language inference, and summarization. We manually curate a set of test examples covering 12 domains from a generative search engine, New Bing. Our results on this curated test set and simulated examples from existing benchmarks highlight both promising signals and challenges. We hope our problem formulation, testbeds, and findings will help lay the foundation for future studies on this important problem.

Metaphorical language, such as ``spending time together'', projects meaning from a source domain (here, $\textit{money}$) to a target domain ($\textit{time}$). Thereby, it highlights certain aspects of the target domain, such as the $\textit{effort}$ behind the time investment. Highlighting aspects with metaphors (while hiding others) bridges the two domains and is the core of metaphorical meaning construction. For metaphor interpretation, linguistic theories stress that identifying the highlighted aspects is important for a better understanding of metaphors. However, metaphor research in NLP has not yet dealt with the phenomenon of highlighting. In this paper, we introduce the task of identifying the main aspect highlighted in a metaphorical sentence. Given the inherent interaction of source domains and highlighted aspects, we propose two multitask approaches - a joint learning approach and a continual learning approach - based on a finetuned contrastive learning model to jointly predict highlighted aspects and source domains. We further investigate whether (predicted) information about a source domain leads to better performance in predicting the highlighted aspects, and vice versa. Our experiments on an existing corpus suggest that, with the corresponding information, the performance to predict the other improves in terms of model accuracy in predicting highlighted aspects and source domains notably compared to the single-task baselines.

With the rapid development of large language models (LLMs), it is highly demanded that LLMs can be adopted to make decisions to enable the artificial general intelligence. Most approaches leverage manually crafted examples to prompt the LLMs to imitate the decision process of human. However, designing optimal prompts is difficult and the patterned prompts can hardly be generalized to more complex environments. In this paper, we propose a novel model named Large Decision Model with Memory (LDM$^2$), which leverages a dynamic memory mechanism to construct dynamic prompts, guiding the LLMs in making proper decisions according to the faced state. LDM$^2$ consists of two stages: memory formation and memory refinement. In the former stage, human behaviors are decomposed into state-action tuples utilizing the powerful summarizing ability of LLMs. Then, these tuples are stored in the memory, whose indices are generated by the LLMs, to facilitate the retrieval of the most relevant subset of memorized tuples based on the current state. In the latter stage, our LDM$^2$ employs tree exploration to discover more suitable decision processes and enrich the memory by adding valuable state-action tuples. The dynamic circle of exploration and memory enhancement provides LDM$^2$ a better understanding of the global environment. Extensive experiments conducted in two interactive environments have shown that our LDM$^2$ outperforms the baselines in terms of both score and success rate, which demonstrates its effectiveness.

Language models (LMs) that jointly generate end-task answers as well as free-text rationales are known as self-rationalization models. Recent works demonstrate great performance gain for self-rationalization by few-shot prompting LMs with rationale-augmented exemplars. However, the ability to benefit from explanations only emerges with large-scale LMs, which have poor accessibility. In this work, we explore the less-studied setting of leveraging explanations for small LMs to improve few-shot self-rationalization. We first revisit the relationship between rationales and answers. Inspired by the implicit mental process of how human beings assess explanations, we present a novel approach, Zero-shot Augmentation of Rationale-Answer pairs (ZARA), to automatically construct pseudo-parallel data for self-training by reducing the problem of plausibility judgement to natural language inference. Experimental results show ZARA achieves SOTA performance on the FEB benchmark, for both the task accuracy and the explanation metric. In addition, we conduct human and quantitative evaluation validating ZARA's ability to automatically identify plausible and accurate rationale-answer pairs.

Despite remarkable advances that large language models have achieved in chatbots nowadays, maintaining a non-toxic user-AI interactive environment has become increasingly critical nowadays. However, previous efforts in toxicity detection have been mostly based on benchmarks derived from social media contents, leaving the unique challenges inherent to real-world user-AI interactions insufficiently explored. In this work, we introduce ToxicChat, a novel benchmark constructed based on real user queries from an open-source chatbot. This benchmark contains the rich, nuanced phenomena that can be tricky for current toxicity detection models to identify, revealing a significant domain difference when compared to social media contents. Our systematic evaluation of models trained on existing toxicity datasets has shown their shortcomings when applied to this unique domain of ToxicChat. Our work illuminates the potentially overlooked challenges of toxicity detection in real-world user-AI conversations. In the future, ToxicChat can be a valuable resource to drive further advancements toward building a safe and healthy environment for user-AI interactions.

Large Language models (LLMs) are trained on vast amounts of data, including sensitive information that poses a risk to personal privacy if exposed. LLMs have shown the ability to memorize and reproduce portions of their training data when prompted by adversaries. Prior research has focused on addressing this memorization issue and preventing verbatim replication through techniques like knowledge unlearning and data pre-processing. However, these methods have limitations regarding the number of protected samples, limited privacy types, and potentially lower-quality generative models. To tackle this challenge more effectively, we propose “DeMem,” a novel unlearning approach that utilizes an efficient reinforcement learning feedback loop via proximal policy optimization. By fine-tuning the language model with a negative similarity score as a reward signal, we incentivize the LLMs to learn a paraphrasing policy to unlearn the pre-training data. Our experiments demonstrate that DeMem surpasses strong baselines and state-of-the-art methods in terms of its ability to generalize and strike a balance between maintaining privacy and LLM performance.

In Chinese studies, understanding the nuanced traits of historical figures, often not explicitly evident in biographical data, has been a key interest. However, identifying these traits can be challenging due to the need for domain expertise, specialist knowledge, and context-specific insights, making the process time-consuming and difficult to scale. Our focus on studying officials from China's Ming Dynasty is no exception. To tackle this challenge, we propose MingOfficial, a large-scale multi-modal dataset consisting of both structured (career records, annotated personnel types) and text (historical texts) data for $9,376$ officials. We further couple the dataset with a a graph neural network (GNN) to combine both modalities in order to allow investigation of social structures and provide features to boost down-stream tasks. Experiments show that our proposed MingOfficial could enable exploratory analysis of official identities, and also significantly boost performance in tasks such as identifying nuance identities (e.g.\ civil officials holding military power) from $24.6\%$ to $98.2\%$ F$_1$ score in hold-out test set. By making MingOfficial publicly available (see main text for the URL) as both a dataset and an interactive tool, we aim to stimulate further research into the role of social context and representation learning in identifying individual characteristics, and hope to provide inspiration for computational approaches in other fields beyond Chinese studies.

With the rapid advancement in deep generative models, recent neural Text-To-Speech(TTS) models have succeeded in synthesizing human-like speech. There have been some efforts to generate speech with various prosody beyond monotonous prosody patterns. However, previous works have several limitations. First, typical TTS models depend on the scaled sampling temperature for boosting the diversity of prosody. Speech samples generated at high sampling temperatures often lack perceptual prosodic diversity, which can adversely affect the naturalness of the speech. Second, the diversity among samples is neglected since the sampling procedure often focuses on a single speech sample rather than multiple ones. In this paper, we propose DPP-TTS: a text-to-speech model based on Determinantal Point Processes (DPPs) with a prosody diversifying module. Our TTS model is capable of generating speech samples that simultaneously consider perceptual diversity in each sample and among multiple samples. We demonstrate that DPP-TTS generates speech samples with more diversified prosody than baselines in the side-by-side comparison test considering the naturalness of speech at the same time.

Large language models encode impressively broad world knowledge in their parameters. However, the knowledge in static language models falls out of date, limiting the model’s effective “shelf life.” While online fine-tuning can reduce this degradation, we find that naively fine-tuning on a stream of documents leads to a low level of information uptake. We hypothesize that online fine-tuning does not sufficiently attend to important information. That is, the gradient signal from important tokens representing factual information is drowned out by the gradient from inherently noisy tokens, suggesting that a dynamic, context-aware learning rate may be beneficial. We therefore propose learning which tokens to upweight. We meta-train a small, autoregressive model to reweight the language modeling loss for each token during online fine-tuning, with the objective of maximizing the out-of-date base question-answering model’s ability to answer questions about a document after a single weighted gradient step. We call this approach Context-aware Meta-learned Loss Scaling (CaMeLS). Across three different distributions of documents, our experiments find that CaMeLS provides substantially improved information uptake on streams of thousands of documents compared with standard fine-tuning and baseline heuristics for reweighting token losses.

Writing strong arguments can be challenging for learners. It requires to select and arrange multiple argumentative discourse units (ADUs) in a logical and coherent way as well as to decide which ADUs to leave implicit, so called enthymemes. However, when important ADUs are missing, readers might not be able to follow the reasoning or understand the argument's main point. This paper introduces two new tasks for learner arguments: to identify gaps in arguments (enthymeme detection) and to fill such gaps (enthymeme reconstruction). Approaches to both tasks may help learners improve their argument quality. We study how corpora for these tasks can be created automatically by deleting ADUs from an argumentative text that are central to the argument and its quality, while maintaining the text's naturalness. Based on the ICLEv3 corpus of argumentative learner essays, we create 40,089 argument instances for enthymeme detection and reconstruction. Through manual studies, we provide evidence that the proposed corpus creation process leads to the desired quality reduction, and results in arguments that are similarly natural to those written by learners. Finally, first baseline approaches to enthymeme detection and reconstruction demonstrate the corpus' usefulness.

Current variational dialog models have employed pre-trained language models (PLMs) to parameterize the likelihood and posterior distributions. However, the Gaussian assumption made on the prior distribution is incompatible with these distributions, thus restricting the diversity of generated responses. These models also suffer from posterior collapse, i.e., the decoder tends to ignore latent variables and directly access information captured in the encoder through the cross-attention mechanism. In this work, we propose Dior-CVAE, a hierarchical conditional variational autoencoder (CVAE) with diffusion priors to address these challenges. We employ a diffusion model to increase the complexity of the prior distribution and its compatibility with the distributions produced by a PLM. Also, we propose memory dropout to the cross-attention mechanism, which actively encourages the use of latent variables for response generation. Overall, experiments across two commonly used open-domain dialog datasets show that our method can generate more diverse responses without large-scale dialog pre-training. Code is available at https://github.com/UKPLab/dior-cvae.

As Large Language Models (LLMs) become popular, there emerged an important trend of using multimodality to augment the LLMs' generation ability, which enables LLMs to better interact with the world. However, there lacks a unified perception of at which stage and how to incorporate different modalities. In this survey, we review methods that assist and augment generative models by retrieving multimodal knowledge, whose formats range from images, codes, tables, graphs, to audio. Such methods offer a promising solution to important concerns such as factuality, reasoning, interpretability, and robustness. By providing an in-depth review, this survey is expected to provide scholars with a deeper understanding of the methods' applications and encourage them to adapt existing techniques to the fast-growing field of LLMs.

Language model detoxification aims to minimize the risk of generating offensive or harmful content in pretrained language models (PLMs) for safer deployment. Existing methods can be roughly categorized as finetuning-based and decoding-based. However, the former is often resource-intensive, while the latter relies on additional components and potentially compromises the generation fluency. In this paper, we propose a more lightweight approach that enables the PLM itself to achieve ``self-detoxification''. Our method is built upon the observation that prepending a negative steering prompt can effectively induce PLMs to generate toxic content. At the same time, we are inspired by the recent research in the interpretability field, which formulates the evolving contextualized representations within the PLM as an information stream facilitated by the attention layers. Drawing on this idea, we devise a method to identify the toxification direction from the normal generation process to the one prompted with the negative prefix, and then steer the generation to the reversed direction by manipulating the information movement within the attention layers. Experimental results show that our approach, without any fine-tuning or extra components, can achieve comparable performance with state-of-the-art methods.