EMNLP 2023 Papers

Generative models have been widely applied to solve extractive tasks, where parts of the input is extracted to form the desired output, and achieved significant success. For example, in extractive question answering (QA), generative models have constantly yielded state-of-the-art results. In this work, we study the issue of tokenization inconsistency that is commonly neglected in training these models. This issue damages the extractive nature of these tasks after the input and output are tokenized inconsistently by the tokenizer, and thus leads to performance drop as well as hallucination. We propose a simple yet effective fix to this issue and conduct a case study on extractive QA. We show that, with consistent tokenization, the model performs better in both in-domain and out-of-domain datasets, with a notable average of +1.7 F1 gain when a BART model is trained on SQuAD and evaluated on 8 QA datasets. Further, the model converges faster, and becomes less likely to generate out-of-context answers. Our results demonstrate the need for increased scrutiny regarding how tokenization is done in extractive tasks and the benefits of consistent tokenization during training.

In this work we investigate the capacity of language models to generate explicit, interpretable, and interactive world models of scientific and common-sense reasoning tasks. We operationalize this as a task of generating text games, expressed as hundreds of lines of Python code. To facilitate this task, we introduce ByteSized32, a corpus of 32 reasoning-focused text games totalling 20k lines of Python code. We empirically demonstrate that GPT-4 can use these games as templates for single-shot in-context learning, successfully producing runnable games on unseen topics in 28% of cases. When allowed to self-reflect on program errors, game runnability substantially increases to 58%. While evaluating simulation fidelity is labor intensive, we introduce a suite of automated metrics to assess game fidelity, technical validity, adherence to task specifications, and winnability, showing a high-degree of agreement with expert human ratings. We pose this as a challenge task to spur further development at the juncture of world modeling and code generation.

*In-context learning* is the paradigm that adapts large language models to downstream tasks by providing a few examples. *Few-shot selection*---selecting appropriate examples for each test instance separately---is important for in-context learning. In this paper, we propose **Skill-KNN**, a skill-based few-shot selection method for in-context learning. The key advantages of Skill-KNN include: (1) it addresses the problem that existing methods based on pre-trained embeddings can be easily biased by surface natural language features that are not important for the target task; (2) it does not require training or fine-tuning of any models, making it suitable for frequently expanding or changing example banks. The key insight is to optimize the inputs fed into the embedding model, rather than tuning the model itself. Technically, Skill-KNN generates the skill-based descriptions for each test case and candidate example by utilizing a pre-processing few-shot prompting, thus eliminating unimportant surface features. Experimental results across five cross-domain semantic parsing datasets and six backbone models show that Skill-KNN significantly outperforms existing methods.

The multi-turn doctor-patient dialogue includes rich medical knowledge, like the symptoms of the patient, the diagnosis and medication suggested by the doctor. If mined and represented properly, such medical knowledge can benefit a large range of clinical applications, including diagnosis assistance and medication recommendation. To derive structured knowledge from free text dialogues, we target a critical task: the Dialogue Medical Information Extraction (DMIE). DMIE aims to detect pre-defined clinical meaningful medical items (symptoms, surgery, etc.) as well as their statuses (positive, negative, etc.) from the dialogue. Existing approaches mainly formulate DMIE as a multi-label classification problem and ignore the relationships among medical items and statuses. Different from previous approaches, we propose a heterogeneous graph to model the relationship between items. We further propose two consecutive attention based modules to enrich the item representation with the dialogue and status. In this manner, we are able to model the relationships among medical items and statuses in the DMIE task. Experimental results on the public benchmark data set show that the proposed model outperforms previous works and achieves the state-of-the-art performance.

Understanding the internal reasoning behind the predictions of machine learning systems is increasingly vital, given their rising adoption and acceptance. While previous approaches, such as LIME generate algorithmic explanations by attributing importance to input features for individual examples, recent research indicates that practitioners prefer examining language explanations that explain sub-groups of examples (Lakkaraju et al., 2022). In this paper, we introduce MaNtLE, a model-agnostic natural language explainer that analyzes a set of classifier predictions and generates faithful natural language explanations of classifier rationale for structured classification tasks. MaNtLE uses multi-task training on thousands of synthetic classification tasks to generate faithful explanations. Our experiments indicate that, on average, MaNtLE-generated explanations are at least 11% more faithful compared to LIME and Anchors explanations across three tasks. Human evaluations demonstrate that users can better predict model behavior using explanations from MaNtLE compared to other techniques.

Recently Large Language Models (LLMs) such as GPT-3, ChatGPT, and FLAN have led to impressive progress in Natural Language Inference (NLI) tasks. However, these models may rely on simple heuristics or artifacts in the evaluation data to achieve their high performance, which suggests that they still suffer from logical inconsistency. To assess the logical consistency of these models, we propose a LogicAttack, a method to attack NLI models using diverse logical forms of premise and hypothesis, providing a more robust evaluation of their performance. Our approach leverages a range of inference rules from propositional logic, such as Modus Tollens and Bidirectional Dilemma, to generate effective adversarial attacks and identify common vulnerabilities across multiple NLI models. We achieve an average ~53% Attack Success Rate (ASR) across multiple logic-based attacks. Moreover, we demonstrate that incorporating generated attack samples into training enhances the logical reasoning ability of the target model and decreases its vulnerability to logic-based attacks. Data and source code are available at https://github.com/msantoshmadhav/LogicAttack.

While prompt tuning approaches have achieved competitive performance with high efficiency, we observe that they invariably employ the same initialization process, wherein the soft prompt is either randomly initialized or derived from an existing embedding vocabulary. In contrast to these conventional methods, this study aims to investigate an alternative way to derive soft prompt. Our empirical studies show that the soft prompt typically exhibits a low ``intrinsic rank'' characteristic. With such observations, we propose decomposed prompt tuning, a novel approach that utilizes low-rank matrices to initialize the soft prompt. Through the low-rank reparameterization, our method significantly reduces the number of trainable parameters while maintaining effectiveness. Experimental results on the SuperGLUE benchmark in both high-resource and low-resource scenarios demonstrate the effectiveness of the proposed method.

Building and maintaining end-to-end task bots using minimal human effort is a long-standing challenge in dialog research. In this work, we introduce SGP-TOD, Schema-Guided Prompting for building Task-Oriented Dialog systems effortlessly based on large language models (LLMs). Utilizing the predefined task schema, i.e., belief instruction and dialog policy, we instruct fixed LLMs to generate appropriate responses on novel tasks, without the need for training data. Specifically, SGP-TOD comprises three components: an LLM for interacting with users, a Dialog State Tracking (DST) Prompter to aid the LLM in tracking dialog states with the given belief instruction, and a Policy Prompter to direct the LLM to generate proper responses adhering to the provided dialog policy. Experimental results on Multiwoz, RADDLE, and STAR datasets show that our training-free strategy, SGP-TOD, yields state-of-the-art (SOTA) zero-shot performance, significantly surpassing the few-shot approaches. In a domain-extension setting, SGP-TOD aptly adapts to new functionalities by merely adding supplementary schema rules. We make our code and data publicly available.

Recent studies show that prompt tuning can better leverage the power of large language models than fine-tuning on downstream natural language understanding tasks. However, the existing prompt tuning methods have training instability issues, as the variance of scores under different random seeds is quite large. To address this critical problem, we first investigate and find that the loss landscape of vanilla prompt tuning is precipitous when it is visualized, where a slight change of input data can cause a big fluctuation in the loss landscape. This is an essential factor that leads to the instability of prompt tuning. Based on this observation, we introduce perturbation-based regularizers, which can smooth the loss landscape, into prompt tuning. We propose a new algorithm, called Prompt Tuning with Perturbation-based regularizer~(PTP), which can not only alleviate training instability dramatically but also boost the performance of prompt tuning. We design two kinds of perturbation-based regularizers, including random-noise-based and adversarial-based. In particular, our proposed perturbations are flexible on both text space and embedding space. Extensive experiments show the effectiveness of our proposed methods in stabilizing the training. Our new algorithms improve the state-of-the-art prompt tuning methods by 1.94\% and 2.34\% on SuperGLUE and FewGLUE benchmarks, respectively.

We employ a characterization of linguistic complexity from psycholinguistic and language acquisition research to develop data-driven curricula to understand the underlying linguistic knowledge that models learn to address NLP tasks. The novelty of our approach is in the development of linguistic curricula derived from data, existing knowledge about linguistic complexity, and model behavior during training. Through the evaluation of several benchmark NLP datasets, our curriculum learning approaches identify sets of linguistic metrics (indices) that inform the challenges and reasoning required to address each task. Our work will inform future research in all NLP areas, allowing linguistic complexity to be considered early in the research and development process. In addition, our work prompts an examination of gold standards and fair evaluation in NLP.

In this position paper, we argue that instead of morally aligning LLMs to specific set of ethical principles, we should infuse generic ethical reasoning capabilities into them so that they can handle value pluralism at a global scale. When provided with an ethical policy, an LLM should be capable of making decisions that are ethically consistent to the policy. We develop a framework that integrates moral dilemmas with moral principles pertaining to different foramlisms of normative ethics, and at different levels of abstractions. Initial experiments with GPT-x models shows that while GPT-4 is a nearly perfect ethical reasoner, the models still have bias towards the moral values of Western and English speaking societies.

Automatically highlighting words that cause semantic differences between two documents could be useful for a wide range of applications. We formulate recognizing semantic differences (RSD) as a token-level regression task and study three unsupervised approaches that rely on a masked language model. To assess the approaches, we begin with basic English sentences and gradually move to more complex, cross-lingual document pairs. Our results show that an approach based on word alignment and sentence-level contrastive learning has a robust correlation to gold labels. However, all unsupervised approaches still leave a large margin of improvement.

Deep generative modeling of natural languages has achieved many successes, such as producing fluent sentences and translating from one language into another. However, the development of generative modeling techniques for paraphrase generation still lags behind largely due to the challenges in addressing the complex conflicts between expression diversity and semantic preservation. This paper proposes to generate diverse and high-quality paraphrases by exploiting the pre-trained models with instance-dependent prompts. To learn generalizable prompts, we assume that the number of abstract transforming patterns of paraphrase generation (governed by prompts) is finite and usually not large. Therefore, we present vector-quantized prompts as the cues to control the generation of pre-trained models. Extensive experiments demonstrate that the proposed method achieves new state-of-art results on three benchmark datasets, including Quora, Wikianswers, and MSCOCO. We will release all the code upon acceptance.

Multilingual Neural Machine Translation (MNMT) facilitates knowledge sharing but often suffers from poor zero-shot (ZS) translation qualities. While prior work has explored the causes of overall low zero-shot translation qualities, our work introduces a fresh perspective: the presence of significant variations in zero-shot performance. This suggests that MNMT does not uniformly exhibit poor zero-shot capability; instead, certain translation directions yield reasonable results. Through systematic experimentation, spanning 1,560 language directions across 40 languages, we identify three key factors contributing to high variations in ZS NMT performance: 1) target-side translation quality, 2) vocabulary overlap, and 3) linguistic properties. Our findings highlight that the target side translation quality is the most influential factor, with vocabulary overlap consistently impacting zero-shot capabilities. Additionally, linguistic properties, such as language family and writing system, play a role, particularly with smaller models. Furthermore, we suggest that the off-target issue is a symptom of inadequate performance, emphasizing that zero-shot translation challenges extend beyond addressing the off-target problem. To support future research, we release the data and models as a benchmark for the study of ZS NMT.

Pretrained language models are expected to effectively map input text to a set of vectors while preserving the inherent relationships within the text. Consequently, designing a white-box model to compute metrics that reflect the presence of specific internal relations in these vectors has become a common approach for post-hoc interpretability analysis of pretrained language models. However, achieving interpretability in white-box models and ensuring the rigor of metric computation becomes challenging when the source model lacks inherent interpretability. Therefore, in this paper, we discuss striking a balance in this trade-off and propose a novel line to constructing metrics for understanding the mechanisms of pretrained language models. We have specifically designed a family of metrics along this line of investigation, and the model used to compute these metrics is referred to as the tree topological probe. We conducted measurements on BERT-large by using these metrics. Based on the experimental results, we propose a speculation regarding the working mechanism of BERT-like pretrained language models, as well as a strategy for enhancing fine-tuning performance by leveraging the topological probe to improve specific submodules.

Narrative comprehension is a challenging task that requires a deep understanding of the foundational elements of narratives. Acquiring this skill requires extensive annotated data. To mitigate the burden of data annotation, we present Parrot, a zero-shot approach for narrative reading comprehension through parallel reading, which involves two parallel narratives that tell the same story. By leveraging one narrative as a source of supervision signal to guide the understanding of the other, Parrot abstracts the textual content and develops genuine narrative understanding. Evaluation conducted on two narrative comprehension benchmarks demonstrates that Parrot surpasses previous zero-shot approaches and achieves comparable performance to fully supervised models. The code will be available at https://github.com/zhaochaocs/Parrot.

Question answering over hybrid contexts is a complex task, which requires the combination of information extracted from unstructured texts and structured tables in various ways. Recently, In-Context Learning demonstrated significant performance advances for reasoning tasks. In this paradigm, a large language model performs predictions based on a small set of supporting exemplars. The performance of In-Context Learning depends heavily on the selection procedure of the supporting exemplars, particularly in the case of HybridQA, where considering the diversity of reasoning chains and the large size of the hybrid contexts becomes crucial. In this work, we present Selection of ExEmplars for hybrid Reasoning (SEER), a novel method for selecting a set of exemplars that is both representative and diverse. The key novelty of SEER is that it formulates exemplar selection as a Knapsack Integer Linear Program. The Knapsack framework provides the flexibility to incorporate diversity constraints that prioritize exemplars with desirable attributes, and capacity constraints that ensure that the prompt size respects the provided capacity budgets. The effectiveness of SEER is demonstrated on FinQA and TAT-QA, two real-world benchmarks for HybridQA, where it outperforms previous exemplar selection methods.

Timely and accurate extraction of Adverse Drug Events (ADE) from biomedical literature is paramount for public safety, but involves slow and costly manual labor. We set out to improve drug safety monitoring (pharmacovigilance, PV) through the use of Natural Language Processing (NLP). We introduce BioDEX, a large-scale resource for Biomedical adverse Drug Event eXtraction, rooted in the historical output of drug safety reporting in the U.S. BioDEX consists of 65k abstracts and 19k full-text biomedical papers with 256k associated document-level safety reports created by medical experts. The core features of these reports include the reported weight, age, and biological sex of a patient, a set of drugs taken by the patient, the drug dosages, the reactions experienced, and whether the reaction was life threatening. In this work, we consider the task of predicting the core information of the report given its originating paper. We estimate human performance to be 72.0% F1, whereas our best model achieves 59.1% F1 (62.3 validation), indicating significant headroom. We also begin to explore ways in which these models could help professional PV reviewers. Our code and data are available at https://github.com/KarelDO/BioDEX.

Frame identification aims to find semantic frames associated with target words in a sentence. Recent researches measure the similarity or matching score between targets and candidate frames by modeling frame definitions. However, they either lack sufficient representation learning of the definitions or face challenges in efficiently selecting the most suitable frame from over 1000 candidate frames. Moreover, commonly used lexicon filtering ($lf$) to obtain candidate frames for the target may ignore out-of-vocabulary targets and cause inadequate frame modeling. In this paper, we propose CoFFTEA, a $\underline{Co}$arse-to-$\underline{F}$ine $\underline{F}$rame and $\underline{T}$arget $\underline{E}$ncoders $\underline{A}$rchitecture. With contrastive learning and dual encoders, CoFFTEA efficiently and effectively models the alignment between frames and targets. By employing a coarse-to-fine curriculum learning procedure, CoFFTEA gradually learns to differentiate frames with varying degrees of similarity. Experimental results demonstrate that CoFFTEA outperforms previous models by 0.93 overall scores and 1.53 R@1 without $lf$. Further analysis suggests that CoFFTEA can better model the relationships between frame and frame, as well as target and target. The code for our approach is available at https://github.com/pkunlp-icler/COFFTEA.