EMNLP 2023 论文列表

Originally proposed as a method for knowledge transfer from one model to another, some recent studies have suggested that knowledge distillation (KD) is in fact a form of regularization. Perhaps the strongest argument of all for this new perspective comes from its apparent similarities with label smoothing (LS). Here we re-examine this stated equivalence between the two methods by comparing the predictive confidences of the models they train. Experiments on four text classification tasks involving models of different sizes show that: (a) In most settings, KD and LS drive model confidence in completely opposite directions, and (b) In KD, the student inherits not only its knowledge but also its confidence from the teacher, reinforcing the classical knowledge transfer view.

Subword tokenization has become the de-facto standard for tokenization although comparative evaluations of their quality across languages are scarce. Existing evaluation studies focus on the effect of a tokenization algorithm on the performance in downstream tasks, or on engineering criteria such as the compression rate. We present a new evaluation paradigm that focuses on the cognitive plausibility of subword tokenization. We analyze the correlation of the tokenizer output with the reading time and accuracy of human responses on a lexical decision task. We compare three tokenization algorithms across several languages and vocabulary sizes. Our results indicate that the Unigram algorithm yields less cognitively plausible tokenization behavior and a worse coverage of derivational morphemes, in contrast with prior work.

Language models (LMs) are capable of conducting in-context learning for multiple choice reasoning tasks, but the options in these tasks are treated equally. As humans often first eliminate wrong options before picking the final correct answer, we argue a similar two-step strategy can make LMs better at these tasks. To this end, we present the Process of Elimination (POE), a two-step scoring method. In the first step, POE scores each option, and eliminates seemingly wrong options. In the second step, POE masks these wrong options, and makes the final prediction from the remaining options. Zero-shot experiments on 8 reasoning tasks illustrate the effectiveness of POE, and a following analysis finds our method to be especially performant on logical reasoning tasks. We further analyze the effect of masks, and show that POE applies to few-shot settings and large language models (LLMs) like ChatGPT.

Neuro-symbolic (NS) models for knowledge graph completion (KGC) combine the benefits of symbolic models (interpretable inference) with those of distributed representations (parameter sharing, high accuracy). While several NS models exist for KGs with static facts, there is limited work on temporal KGC (TKGC) for KGs where a fact is associated with a time interval. In response, we propose a novel NS model for TKGC called NeuSTIP, which performs link prediction and time interval prediction in a TKG. NeuSTIP learns temporal rules with Allen predicates, which ensure temporal consistency between neighboring predicates in the rule body. We further design a unique scoring function that evaluates the confidence of the candidate answers while performing link and time interval predictions by utilizing the learned rules. Our empirical evaluation on two time interval based TKGC datasets shows that our model shows competitive performance on link prediction and establishes a new state of the art on time prediction.

The recent success of SimCSE has greatly advanced state-of-the-art sentence representations. However, the original formulation of SimCSE does not fully exploit the potential of hard negative samples in contrastive learning. This study introduces an unsupervised contrastive learning framework that combines SimCSE with hard negative mining, aiming to enhance the quality of sentence embeddings. The proposed focal-InfoNCE function introduces self-paced modulation terms in the contrastive objective, downweighting the loss associated with easy negatives and encouraging the model focusing on hard negatives. Experimentation on various STS benchmarks shows that our method improves sentence embeddings in terms of Spearman's correlation and representation alignment and uniformity.

We present The Vault, an open-source dataset of high quality code-text pairs in multiple programming languages for training large language models to understand and generate code. We propose methods for thoroughly extracting samples that use both rules and deep learning to ensure that they contain high-quality pairs of code and text, resulting in a dataset of 43 million high-quality code-text pairs. We thoroughly evaluated this dataset and discovered that when used to train common code language models (such as CodeT5, CodeBERT, and CodeGen), it outperforms the same models train on other datasets such as CodeSearchNet. These evaluations included common coding tasks such as code generation, code summarization, and code search. The Vault can be used by researchers and practitioners to train a wide range of big language models that understand code. Alternatively, researchers can use our data cleaning methods and scripts to improve their own datasets. We anticipate that using The Vault to train large language models will improve their ability to understand and generate code, propelling AI research and software development forward. We are releasing our source code and a framework to make it easier for others to replicate our results.

Social and behavioral determinants of health (SDOH) play a significant role in shaping health outcomes, and extracting these determinants from clinical notes is a first step to help healthcare providers systematically identify opportunities to provide appropriate care and address disparities. Progress on using NLP methods for this task has been hindered by the lack of high-quality publicly available labeled data, largely due to the privacy and regulatory constraints on the use of real patients' information. This paper introduces a new dataset, SDOH-NLI, that is based on publicly available notes and which we release publicly. We formulate SDOH extraction as a natural language inference task, and provide binary textual entailment labels obtained from human raters for a cross product of a set of social history snippets as premises and SDOH factors as hypotheses. Our dataset differs from standard NLI benchmarks in that our premises and hypotheses are obtained independently. We evaluate both "off-the-shelf" entailment models as well as models fine-tuned on our data, and highlight the ways in which our dataset appears more challenging than commonly used NLI datasets.

Due to its growing impact on public opinion, hate speech on social media has garnered increased attention. While automated methods for identifying hate speech have been presented in the past, they have mostly been limited to analyzing textual content. The interpretability of such models has received very little attention, despite the social and legal consequences of erroneous predictions. In this work, we present a novel problem of \textit{Distress Identification and Cause Extraction (DICE)} from multimodal online posts. We develop a multi-task deep framework for the simultaneous detection of distress content and identify connected causal phrases from the text using emotional information. The emotional information is incorporated into the training process using a zero-shot strategy, and a novel mechanism is devised to fuse the features from the multimodal inputs. Furthermore, we introduce the first-of-its-kind \textit{Distress and Cause annotated Multimodal (DCaM)} dataset of 20,764 social media posts. We thoroughly evaluate our proposed method by comparing it to several existing benchmarks. Empirical assessment and comprehensive qualitative analysis demonstrate that our proposed method works well on distress detection and cause extraction tasks, improving F1 and ROS scores by 1.95\% and 3\%, respectively, relative to the best-performing baseline. The code and the dataset can be accessed from the following link: \url{https://www.iitp.ac.in/~ai-nlp-ml/resources.html\#DICE}.

The rampant proliferation of large language models, fluent enough to generate text indistinguishable from human-written language, gives unprecedented importance to the detection of machine-generated text. This work is motivated by an important research question: How will the detectors of machine-generated text perform on outputs of a new generator, that the detectors were not trained on? We begin by collecting generation data from a wide range of LLMs, and train neural detectors on data from each generator and test its performance on held-out generators. While none of the detectors can generalize to all generators, we observe a consistent and interesting pattern that the detectors trained on data from a medium-size LLM can zero-shot generalize to the larger version. As a concrete application, we demonstrate that robust detectors can be built on an ensemble of training data from medium-sized models.

Machine learning (ML) systems in natural language processing (NLP) face significant challenges in generalizing to out-of-distribution (OOD) data, where the test distribution differs from the training data distribution. This poses important questions about the robustness of NLP models and their high accuracy, which may be artificially inflated due to their underlying sensitivity to systematic biases. Despite these challenges, there is a lack of comprehensive surveys on the generalization challenge from an OOD perspective in natural language understanding. Therefore, this paper aims to fill this gap by presenting the first comprehensive review of recent progress, methods, and evaluations on this topic. We further discuss the challenges involved and potential future research directions. By providing convenient access to existing work, we hope this survey will encourage future research in this area.

Recent advances in prompt engineering enable large language models (LLMs) to solve multi-hop logical reasoning problems with impressive accuracy. However, there is little existing work investigating the robustness of LLMs with few-shot prompting techniques. Therefore, we introduce a systematic approach to test the robustness of LLMs in multi-hop reasoning tasks via domain-agnostic perturbations. We include perturbations at multiple levels of abstractions (e.g. lexical perturbations such as typos, and semantic perturbations such as the inclusion of intermediate reasoning steps in the questions) to conduct behavioral analysis on the LLMs. Throughout our experiments, we find that models are more sensitive to certain perturbations such as replacing words with their synonyms. We also demonstrate that increasing the proportion of perturbed exemplars in the prompts improves the robustness of few-shot prompting methods.

The concept of a complex event schema pertains to the graph structure that represents real-world knowledge of events and their multi-dimensional relationships. However, previous studies on event schema induction have been hindered by challenges such as error propagation and data quality issues. To tackle these challenges, we propose a knowledge-enriched discrete diffusion model. Specifically, we distill the abundant event scenario knowledge of Large Language Models (LLMs) through an object-oriented Python style prompt. We incorporate this knowledge into the training data, enhancing its quality. Subsequently, we employ a discrete diffusion process to generate all nodes and links simultaneously in a non-auto-regressive manner to tackle the problem of error propagation. Additionally, we devise an entity relationship prediction module to complete entity relationships between event arguments. Experimental results demonstrate that our approach achieves outstanding performance across a range of evaluation metrics.

Empathetic response generation aims to generate empathetic responses by understanding the speaker's emotional feelings from the language of dialogue. Recent methods capture emotional words in the language of communicators and construct them as static vectors to perceive nuanced emotions. However, linguistic research has shown that emotional words in language are dynamic and have correlations with other grammar semantic roles, i.e., words with semantic meanings, in grammar. Previous methods overlook these two characteristics, which easily lead to misunderstandings of emotions and neglect of key semantics. To address this issue, we propose a dynamical Emotion-Semantic Correlation Model (ESCM) for empathetic dialogue generation tasks. ESCM constructs dynamic emotion-semantic vectors through the interaction of context and emotions. We introduce dependency trees to reflect the correlations between emotions and semantics. Based on dynamic emotion-semantic vectors and dependency trees, we propose a dynamic correlation graph convolutional network to guide the model in learning context meanings in dialogue and generating empathetic responses. Experimental results on the EMPATHETIC-DIALOGUES dataset show that ESCM understands semantics and emotions more accurately and expresses fluent and informative empathetic responses. Our analysis results also indicate that the correlations between emotions and semantics are frequently used in dialogues, which is of great significance for empathetic perception and expression.

Large language models (LLMs) have shown impressive achievements in solving a broad range of tasks. Augmented by instruction fine-tuning, LLMs have also been shown to generalize in zero-shot settings as well. However, whether LLMs closely align with the human disagreement distribution has not been well-studied, especially within the scope of natural language inference (NLI). In this paper, we evaluate the performance and alignment of LLM distribution with humans using two different techniques to estimate the multinomial distribution: Monte Carlo Estimation (MCE) and Log Probability Estimation (LPE). As a result, we show LLMs exhibit limited ability in solving NLI tasks and simultaneously fail to capture human disagreement distribution. The inference and human alignment performances plunge even further on data samples with high human disagreement levels, raising concerns about their natural language understanding (NLU) ability and their representativeness to a larger human population.

The computational cost of transformer-based language models grows quadratically with the sequence length. In this paper, we introduce the selective cache, which stores the selected key-value pairs from the previous context. By selecting important key-value pairs the model makes better use of the cache so that in limited cache size, a longer context history can be stored. We design three kinds of selection methods. The first is based on human language processing. The key-value pairs are selected if they correspond to tokens that are fixated longer, as recorded in eye-tracking-while-reading experiments. We also incorporate the cognitively-inspired selection process into the language model as a trainable process, resulting in two additional methods with improved performance. The selection task is converted into a pruning task so they can be trained with differentiable masks. We demonstrate that the proposed selective cache improves the language modeling performance across different datasets. With the same number of stored key-value pairs (cache size), our selective cache outperforms XL cache and compressive cache by considerable margins.

This study investigates machine translation between related languages i.e., languages within the same family that share linguistic characteristics such as word order and lexical similarity. Machine translation through few-shot prompting leverages a small set of translation pair examples to generate translations for test sentences. This procedure requires the model to learn how to generate translations while simultaneously ensuring that token ordering is maintained to produce a fluent and accurate translation. We propose that for related languages, the task of machine translation can be simplified by leveraging the monotonic alignment characteristic of such languages. We introduce DecoMT, a novel approach of few-shot prompting that decomposes the translation process into a sequence of word chunk translations. Through automatic and human evaluation conducted on multiple related language pairs across various language families, we demonstrate that our proposed approach of decomposed prompting surpasses multiple established few-shot baseline approaches. For example, DecoMT outperforms the strong few-shot prompting BLOOM model with an average improvement of 8 chrF++ scores across the examined languages.

Text simplification has emerged as an increasingly useful application of AI for bridging the communication gap in specialized fields such as medicine, where the lexicon is often dominated by technical jargon and complex constructs. Despite notable progress, methods in medical simplification sometimes result in the generated text having lower quality and diversity. In this work, we explore ways to further improve the readability of text simplification in the medical domain. We propose (1) a new unlikelihood loss that encourages generation of simpler terms and (2) a reranked beam search decoding method that optimizes for simplicity, which achieve better performance on readability metrics on three datasets. This study's findings offer promising avenues for improving text simplification in the medical field.

In this work, we study the language model backbone replacement problem for personalized downstream tasks in a non-stationary on-device scenario. In real world, company may periodically update the knowledge and architectures of backbones to keep the competitive in the market, meanwhile, to accommodate the users' own preference, models are personalized to fit users' own distribution locally. Traditional full model tuning or transfer learning for such replacements often incur considerable local device training costs and necessitate extensive backpropagation within deep transformer layers. Addressing this issue, we propose a novel, lightweight tuning method for personalized NLP classification tasks post-backbone replacement. Our approach leverages a personalized matrix calculated from documents corresponding to users' old and new backbones. This matrix facilitates top-layer parameter tuning, drastically reducing backpropagation computation. To further mitigate training costs associated with matrix linear optimization, we employ correlation clustering to curate a few examples from personalized cluster sets for individuals. Our method achieves over 1000 times computation reduction in Flops for backpropagation and brings the user-specific initialization for personal matrix yielding significant performance boost compared with popular transfer learning methods.

Parameter-efficient fine-tuning (PEFT) has shown its effectiveness in adapting the pre-trained language models to downstream tasks while only updating a small number of parameters. Despite the success, most existing methods independently adapt to each task without considering knowledge transfer between tasks and are limited to low-data regimes. To overcome this issue, we propose Prototype-based HyperAdapter (PHA), a novel framework built on the adapter-tuning and hypernetwork. It introduces an instance-dense retriever and a prototypical hypernetwork to generate the conditional modules in a sample-efficient manner. This leads to comparable performance improvements against existing PEFT methods on multi-task learning and few-shot transfer learning. More importantly, when the available data size gets smaller, our method outperforms other strong baselines by a large margin. Based on our extensive empirical experiments across various datasets, we demonstrate that PHA strikes a better trade-off between trainable parameters, accuracy on stream tasks, and sample efficiency. Our code is publicly available at https://github.com/Bumble666/PHA