EMNLP 2023 Papers

In this paper, we study multimodal coreference resolution, specifically where a longer descriptive text, i.e., a narration is paired with an image. This poses significant challenges due to fine-grained image-text alignment, inherent ambiguity present in narrative language, and unavailability of large annotated training sets. To tackle these challenges, we present a data efficient semi-supervised approach that utilizes image-narration pairs to resolve coreferences and narrative grounding in a multimodal context. Our approach incorporates losses for both labeled and unlabeled data within a cross-modal framework. Our evaluation shows that the proposed approach outperforms strong baselines both quantitatively and qualitatively, for the tasks of coreference resolution and narrative grounding.

Grounding has been argued to be a crucial component towards the development of more complete and truly semantically competent artificial intelligence systems. Literature has divided into two camps: While some argue that grounding allows for qualitatively different generalizations, others believe it can be compensated by mono-modal data quantity. Limited empirical evidence has emerged for or against either position, which we argue is due to the methodological challenges that come with studying grounding and its effects on NLP systems. In this paper, we establish a methodological framework for studying what the effects are---if any---of providing models with richer input sources than text-only. The crux of it lies in the construction of comparable samples of populations of models trained on different input modalities, so that we can tease apart the qualitative effects of different input sources from quantifiable model performances. Experiments using this framework reveal qualitative differences in model behavior between cross-modally grounded, cross-lingually grounded, and ungrounded models, which we measure both at a global dataset level as well as for specific word representations, depending on how concrete their semantics is.

Various types of social biases have been reported with pretrained Masked Language Models (MLMs) in prior work. However, multiple underlying factors are associated with an MLM such as its model size, size of the training data, training objectives, the domain from which pretraining data is sampled, tokenization, and languages present in the pretrained corpora, to name a few. It remains unclear as to which of those factors influence social biases that are learned by MLMs. To study the relationship between model factors and the social biases learned by an MLM, as well as the downstream task performance of the model, we conduct a comprehensive study over 39 pretrained MLMs covering different model sizes, training objectives, tokenization methods, training data domains and languages. Our results shed light on important factors often neglected in prior literature, such as tokenization or model objectives.

Emotion-Cause analysis has attracted the attention of researchers in recent years. However, most existing datasets are limited in size and number of emotion categories. They often focus on extracting parts of the document that contain the emotion cause and fail to provide more abstractive, generalizable root cause. To bridge this gap, we introduce a large-scale dataset of emotion causes, derived from 9.8 million cleaned tweets over 15 years. We describe our curation process, which includes a comprehensive pipeline for data gathering, cleaning, labeling, and validation, ensuring the dataset's reliability and richness. We extract emotion labels and provide abstractive summarization of the events causing emotions. The final dataset comprises over 700,000 tweets with corresponding emotion-cause pairs spanning 48 emotion classes, validated by human evaluators. The novelty of our dataset stems from its broad spectrum of emotion classes and the abstractive emotion cause that facilitates the development of an emotion-cause knowledge graph for nuanced reasoning. Our dataset will enable the design of emotion-aware systems that account for the diverse emotional responses of different people for the same event.

Parameter-shared pre-trained language models (PLMs) have emerged as a successful approach in resource-constrained environments, enabling substantial reductions in model storage and memory costs without significant performance compromise. However, it is important to note that parameter sharing does not alleviate computational burdens associated with inference, thus impeding its practicality in situations characterized by limited stringent latency requirements or computational resources. Building upon neural ordinary differential equations (ODEs), we introduce a straightforward technique to enhance the inference efficiency of parameter-shared PLMs. Additionally, we propose a simple pre-training technique that leads to fully or partially shared models capable of achieving even greater inference acceleration. The experimental results demonstrate the effectiveness of our methods on both autoregressive and autoencoding PLMs, providing novel insights into more efficient utilization of parameter-shared models in resource-constrained settings.

Fallacies are arguments that employ faulty reasoning. Given their persuasive and seemingly valid nature, fallacious arguments are often used in political debates. Employing these misleading arguments in politics can have detrimental consequences for society, since they can lead to inaccurate conclusions and invalid inferences from the public opinion and the policymakers. Automatically detecting and classifying fallacious arguments represents therefore a crucial challenge to limit the spread of misleading or manipulative claims and promote a more informed and healthier political discourse. Our contribution to address this challenging task is twofold. First, we extend the ElecDeb60To16 dataset of U.S. presidential debates annotated with fallacious arguments, by incorporating the most recent Trump-Biden presidential debate. We include updated token-level annotations, incorporating argumentative components (i.e., claims and premises), the relations between these components (i.e., support and attack), and six categories of fallacious arguments (i.e., Ad Hominem, Appeal to Authority, Appeal to Emotion, False Cause, Slippery Slope, and Slogans). Second, we perform the twofold task of fallacious argument detection and classification by defining neural network architectures based on Transformers models, combining text, argumentative features, and engineered features. Our results show the advantages of complementing transformer-generated text representations with non-text features.

Machine reading comprehension (MRC) is an important area of conversation agents and draws a lot of attention. However, there is a notable limitation to current MRC benchmarks: The labeled answers are mostly either spans extracted from the target corpus or the choices of the given candidates, ignoring the natural aspect of high-quality responses. As a result, MRC models trained on these datasets can not generate human-like responses in real QA scenarios. To this end, we construct a new dataset called \textbf{Penguin} to promote the research of MRC, providing a training and test bed for natural response generation to real scenarios. Concretely, Penguin consists of 200k training data with high-quality fluent, and well-informed responses. Penguin is the first benchmark towards natural response generation in Chinese MRC on a relatively large scale. To address the challenges in Penguin, we develop two strong baselines: end-to-end and two-stage frameworks. Following that, we further design \textit{Prompt-BART}: fine-tuning the pre-trained generative language models with a mixture of prefix prompts in Penguin. Extensive experiments validated the effectiveness of this design.

The field of text-to-image (T2I) generation has garnered significant attention both within the research community and among everyday users. Despite the advancements of T2I models, a common issue encountered by users is the need for repetitive editing of input prompts in order to receive a satisfactory image, which is time-consuming and labor-intensive. Given the demonstrated text generation power of large-scale language models, such as GPT-k, we investigate the potential of utilizing such models to improve the prompt editing process for T2I generation. We conduct a series of experiments to compare the common edits made by humans and GPT-k, evaluate the performance of GPT-k in prompting T2I, and examine factors that may influence this process. We found that GPT-k models focus more on inserting modifiers while humans tend to replace words and phrases, which includes changes to the subject matter. Experimental results show that GPT-k are more effective in adjusting modifiers rather than predicting spontaneous changes in the primary subject matters. Adopting the edit suggested by GPT-k models may reduce the percentage of remaining edits by 20-30%.

Entity linking is a prominent thread of research focused on structured data creation by linking spans of text to an ontology or knowledge source. We revisit the use of structured prediction for entity linking which classifies each individual input token as an entity, and aggregates the token predictions. Our system, called SpEL (Structured prediction for Entity Linking) is a state-of-the-art entity linking system that uses some new ideas to apply structured prediction to the task of entity linking including: two refined fine-tuning steps; a context sensitive prediction aggregation strategy; reduction of the size of the model's output vocabulary, and; we address a common problem in entity-linking systems where there is a training vs. inference tokenization mismatch. Our experiments show that we can outperform the state-of-the-art on the commonly used AIDA benchmark dataset for entity linking to Wikipedia. Our method is also very compute efficient in terms of number of parameters and speed of inference.

Many annotation tasks in natural language processing are highly subjective in that there can be different valid and justified perspectives on what is a proper label for a given example. This also applies to the judgment of argument quality, where the assignment of a single ground truth is often questionable. At the same time, there are generally accepted concepts behind argumentation that form a common ground. To best represent the interplay of individual and shared perspectives, we consider a continuum of approaches ranging from models that fully aggregate perspectives into a majority label to ``share nothing''-architectures in which each annotator is considered in isolation from all other annotators. In between these extremes, inspired by models used in the field of recommender systems, we investigate the extent to which architectures that predict labels for single annotators but include layers that model the relations between different annotators are beneficial. By means of two tasks of argument quality classification (argument concreteness and validity/novelty of conclusions), we show that recommender architectures increase the averaged annotator-individual F1-scores up to 43% over a majority-label model. Our findings indicate that approaches to subjectivity can benefit from relating individual perspectives.

Language models have been shown to perform remarkably well on a wide range of natural language processing tasks. In this paper, we propose LEAP, a novel system that uses language models to perform multi-step logical reasoning and incorporates explicit planning into the inference procedure. Explicit planning enables the system to make more informed reasoning decisions at each step by looking ahead into their future effects. Moreover, we propose a training strategy that safeguards the planning process from being led astray by spurious features. Our full system significantly outperforms other competing methods on multiple standard datasets. When using small T5 models as its core selection and deduction components, our system performs competitively compared to GPT-3 despite having only about 1B parameters (i.e., 175 times smaller than GPT-3). When using GPT-3.5, it significantly outperforms chain-of-thought prompting on the challenging PrOntoQA dataset. We have conducted extensive empirical studies to demonstrate that explicit planning plays a crucial role in the system's performance.

Recent work has proposed a methodology for the systematic evaluation of "Situated Language Understanding Agents" --- agents that operate in rich linguistic and non-linguistic contexts --- through testing them in carefully constructed interactive settings. Other recent work has argued that Large Language Models (LLMs), if suitably set up, can be understood as (simulators of) such agents. A connection suggests itself, which this paper explores: Can LLMs be evaluated meaningfully by exposing them to constrained game-like settings that are built to challenge specific capabilities? As a proof of concept, this paper investigates five interaction settings, showing that current chat-optimised LLMs are, to an extent, capable of following game-play instructions. Both this capability and the quality of the game play, measured by how well the objectives of the different games are met, follows the development cycle, with newer models generally performing better. The metrics even for the comparatively simple example games are far from being saturated, suggesting that the proposed instrument will remain to have diagnostic value.

\emph{Autoregressive} (AR) encoder-decoder neural networks have proved successful in many NLP problems, including \emph{Semantic Parsing} -- a task that translates natural language to machine-readable \emph{parse trees}. However, the sequential prediction process of AR models can be slow. To accelerate AR for semantic parsing, we introduce a new technique called \emph{TreePiece} that tokenizes a parse tree into subtrees and generates one subtree per decoding step. On TOPv2 benchmark, TreePiece shows $4.6$ times faster decoding speed than standard AR, and comparable speed but significantly higher accuracy compared to \emph{Non-Autoregressive} (NAR).

Zero-shot Dialogue State Tracking (DST) addresses the challenge of acquiring and annotating task-oriented dialogues, which can be time-consuming and costly. However, DST extends beyond simple slot-filling and requires effective updating strategies for tracking dialogue state as conversations progress. In this paper, we propose ParsingDST, a new In-Context Learning (ICL) method, to introduce additional intricate updating strategies in zero-shot DST. Our approach reformulates the DST task by leveraging powerful Large Language Models (LLMs) and translating the original dialogue text to JSON through semantic parsing as an intermediate state. We also design a novel framework that includes more modules to ensure the effectiveness of updating strategies in the text-to-JSON process. Experimental results demonstrate that our approach outperforms existing zero-shot DST methods on MultiWOZ, exhibiting significant improvements in Joint Goal Accuracy (JGA) and slot accuracy compared to existing ICL methods.

Framing bias plays a significant role in exacerbating political polarization by distorting the perception of actual events. Media outlets with divergent political stances often use polarized language in their reporting of the same event. We propose a new loss function that encourages the model to minimize the polarity difference between the polarized input articles to reduce framing bias. Specifically, our loss is designed to jointly optimize the model to map polarity ends bidirectionally. Our experimental results demonstrate that incorporating the proposed polarity minimization loss leads to a substantial reduction in framing bias when compared to a BART-based multi-document summarization model. Notably, we find that the effectiveness of this approach is most pronounced when the model is trained to minimize the polarity loss associated with informational framing bias (i.e., skewed selection of information to report).

Explainable NLP techniques primarily explain by answering "Which tokens in the input are responsible for this prediction?". We argue that for NLP models that make predictions by comparing two input texts, it is more useful to explain by answering "What differences between the two inputs explain this prediction?". We introduce a technique to generate contrastive phrasal highlights that explain the predictions of a semantic divergence model via phrase alignment guided erasure. We show that the resulting highlights match human rationales of cross-lingual semantic differences better than popular post-hoc saliency techniques and that they successfully help people detect fine-grained meaning differences in human translations and critical machine translation errors.

Causal reasoning ability is crucial for numerous NLP applications. Despite the impressive emerging ability of ChatGPT in various NLP tasks, it is unclear how well ChatGPT performs in causal reasoning. In this paper, we conduct the first comprehensive evaluation of the ChatGPT's causal reasoning capabilities. Experiments show that ChatGPT is not a good causal reasoner, but a good causal interpreter. Besides, ChatGPT has a serious hallucination on causal reasoning, possibly due to the reporting biases between causal and non-causal relationships in natural language, as well as ChatGPT's upgrading processes, such as RLHF. The In-Context Learning (ICL) and Chain-of-Thought (COT) techniques can further exacerbate such causal hallucination. Additionally, the causal reasoning ability of ChatGPT is sensitive to the words used to express the causal concept in prompts, and close-ended prompts perform better than open-ended prompts. For events in sentences, ChatGPT excels at capturing explicit causality rather than implicit causality, and performs better in sentences with lower event density and smaller lexical distance between events.

Large language models (LLMs) are a promising avenue for machine translation (MT). However, current LLM-based MT systems are brittle: their effectiveness highly depends on the choice of few-shot examples and they often require extra post-processing due to overgeneration. Alternatives such as finetuning on translation instructions are computationally expensive and may weaken in-context learning capabilities, due to overspecialization. In this paper, we provide a closer look at this problem. We start by showing that adapter-based finetuning with LoRA matches the performance of traditional finetuning while reducing the number of training parameters by a factor of 50. This method also outperforms few-shot prompting and eliminates the need for post-processing or in-context examples. However, we show that finetuning generally degrades few-shot performance, hindering adaptation capabilities. Finally, to obtain the best of both worlds, we propose a simple approach that incorporates few-shot examples during finetuning. Experiments on 10 language pairs show that our proposed approach recovers the original few-shot capabilities while keeping the added benefits of finetuning.

Large language models (LLMs) are capable to perform complex reasoning by in-context learning (ICL) when provided with a few input-output demonstrations (demos) and more powerful when intermediate reasoning steps (chain of thoughts (CoT)) of the demos are given. Is it necessary to use multi-demo in ICL? In this paper, we study ICL using fewer demos for each test query on the tasks in~\cite{wei2022chain}. Surprisingly, we do not observe significant degradation when using only one randomly chosen demo. To study this phenomenon, for each test query, we categorize demos into ''positive demos'' leading to the correct answer, and ``negative demos'' resulting in wrong answers. Our analysis reveals an inherent bias in those widely studied datasets and the redundancy of demos: most demos are positive for a majority of test queries, which explains the good performance of ICL with one random demo. Moreover, ICL (with and w/o CoT) using only one positive demo significantly outperforms multi-demo ICL adopted by most previous works, indicating the weakness of LLMs in finding positive demo(s) for input queries, which is difficult to evaluate on the biased datasets. Furthermore, we observe a counterintuitive behavior of ICL using multi-demo, i.e., its accuracy degrades(improves) when given more positive(negative) demos. This implies that ICL can be easily misguided by interference among demos and their spurious correlations. Our analyses highlight several fundamental challenges that need to be addressed in LLMs training, ICL, and benchmark design.