EMNLP 2023 论文列表

Augmenting pretrained language models with retrievers has shown promise in effectively solving common NLP problems, such as language modeling and question answering. In this paper, we evaluate the strengths and weaknesses of popular retriever-augmented language models, namely kNN-LM, REALM, DPR + FiD, Contriever + ATLAS, and Contriever + Flan-T5, in reasoning over retrieved statements across different tasks. Our findings indicate that the simple similarity metric employed by retrievers is insufficient for retrieving all the necessary statements for reasoning. Additionally, the language models do not exhibit strong reasoning even when provided with only the required statements. Furthermore, when combined with imperfect retrievers, the performance of the language models becomes even worse, e.g., Flan-T5's performance drops by 28.6% when retrieving 5 statements using Contriever. While larger language models improve performance, there is still a substantial room for enhancement. Our further analysis indicates that multihop retrieve-and-read is promising for large language models like GPT-3.5, but does not generalize to other language models like Flan-T5-xxl. The code is available at https://github.com/McGill-NLP/retriever-lm-reasoning.

Real-world NLP applications often deal with nonstandard text (e.g., dialectal, informal, or misspelled text). However, language models like BERT deteriorate in the face of dialect variation or noise. How do we push BERT’s modeling capabilities to encompass nonstandard text? Fine-tuning helps, but it is designed for specializing a model to a task and does not seem to bring about the deeper, more pervasive changes needed to adapt a model to nonstandard language. In this paper, we introduce the novel idea of sandwiching BERT's encoder stack between additional encoder layers trained to perform masked language modeling on noisy text. We find that our approach, paired with recent work on including character-level noise in fine-tuning data, can promote zero-shot transfer to dialectal text, as well as reduce the distance in the embedding space between words and their noisy counterparts.

The Universum class, often known as the *other* class or the*miscellaneous* class, is defined as a collection of samples that do not belong to any class of interest. It is a typical class that exists in many classification-based tasks in NLP, such as relation extraction, named entity recognition, sentiment analysis, etc. The Universum class exhibits very different properties, namely heterogeneity and lack of representativeness in training data; however, existing methods often treat the Universum class equally with the classes of interest, leading to problems such as overfitting, misclassification, and diminished model robustness. In this work, we propose a closed boundary learning method that applies closed decision boundaries to classes of interest and designates the area outside all closed boundaries in the feature space as the space of the Universum class. Specifically, we formulate closed boundaries as arbitrary shapes, propose the inter-class rule-based probability estimation for the Universum class to cater to its unique properties, and propose a boundary learning loss to adjust decision boundaries based on the balance of misclassified samples inside and outside the boundary. In adherence to the natural properties of the Universum class, our method enhances both accuracy and robustness of classification models, demonstrated by improvements on six state-of-the-art works across three different tasks. Our code is available at https://github.com/hzzhou01/Closed-Boundary-Learning.

The dramatic increase in the use of social media platforms for information sharing has also fueled a steep growth in online abuse. A simple yet effective way of abusing individuals or communities is by creating memes, which often integrate an image with a short piece of text layered on top of it. Such harmful elements are in rampant use and are a threat to online safety. Hence it is necessary to develop efficient models to detect and flag abusive memes. The problem becomes more challenging in a low-resource setting (e.g., Bengali memes, i.e., images with Bengali text embedded on it) because of the absence of benchmark datasets on which AI models could be trained. In this paper we bridge this gap by building a Bengali meme dataset. To setup an effective benchmark we implement several baseline models for classifying abusive memes using this dataset. We observe that multimodal models that use both textual and visual information outperform unimodal models. Our best-performing model achieves a macro F1 score of 70.51. Finally, we perform a qualitative error analysis of the misclassified memes of the best-performing text-based, image-based and multimodal models.

Eye movements in reading play a crucial role in psycholinguistic research studying the cognitive mechanisms underlying human language processing. More recently, the tight coupling between eye movements and cognition has also been leveraged for language-related machine learning tasks such as the interpretability, enhancement, and pre-training of language models, as well as the inference of reader- and text-specific properties. However, scarcity of eye movement data and its unavailability at application time poses a major challenge for this line of research. Initially, this problem was tackled by resorting to cognitive models for synthesizing eye movement data. However, for the sole purpose of generating human-like scanpaths, purely data-driven machine-learning-based methods have proven to be more suitable. Following recent advances in adapting diffusion processes to discrete data, we propose ScanDL, a novel discrete sequence-to-sequence diffusion model that generates synthetic scanpaths on texts. By leveraging pre-trained word representations and jointly embedding both the stimulus text and the fixation sequence, our model captures multi-modal interactions between the two inputs. We evaluate ScanDL within- and across-dataset and demonstrate that it significantly outperforms state-of-the-art scanpath generation methods. Finally, we provide an extensive psycholinguistic analysis that underlines the model's ability to exhibit human-like reading behavior. Our implementation is made available at https://github.com/DiLi-Lab/ScanDL.

Being able to predict people's opinions on issues and behaviors in realistic scenarios can be helpful in various domains, such as politics and marketing. However, conducting large-scale surveys like the European Social Survey to solicit people's opinions on individual issues can incur prohibitive costs. Leveraging prior research showing influence of core human values on individual decisions and actions, we propose to use value-injected large language models (LLM) to predict opinions and behaviors. To this end, we present Value Injection Method (VIM), a collection of two methods---argument generation and question answering---designed to inject targeted value distributions into LLMs via fine-tuning. We then conduct a series of experiments on four tasks to test the effectiveness of VIM and the possibility of using value-injected LLMs to predict opinions and behaviors of people. We find that LLMs value-injected with variations of VIM substantially outperform the baselines. Also, the results suggest that opinions and behaviors can be better predicted using value-injected LLMs than the baseline approaches.

Novels are often adapted into feature films, but the differences between the two media usually require dropping sections of the source text from the movie script. Here we study this screen adaptation process by constructing narrative alignments using the Smith-Waterman local alignment algorithm coupled with SBERT embedding distance to quantify text similarity between scenes and book units. We use these alignments to perform an automated analysis of 40 adaptations, revealing insights into the screenwriting process concerning (i) faithfulness of adaptation, (ii) importance of dialog, (iii) preservation of narrative order, and (iv) gender representation issues reflective of the Bechdel test.

Scaling up language models has been empirically shown to improve performance on a wide range of downstream tasks. However, if we were to observe worse performance as a function of scale ("inverse scaling") on certain tasks, this would indicate that scaling can also encourage behaviors that are misaligned with human preferences. The Inverse Scaling Prize (McKenzie et al. 2023) identified eleven such inverse scaling tasks, evaluated on models of up to 280B parameters and up to 500 zettaFLOPs of training compute. This paper takes a closer look at these inverse scaling tasks. In this paper, we evaluate models of up to 540B parameters, trained on five times more compute than those evaluated in the Inverse Scaling Prize. With this increased range of model sizes and compute, only four out of the eleven tasks remain inverse scaling. Six tasks exhibit "U-shaped scaling", where performance decreases up to a certain size, and then increases again up to the largest model evaluated (the one remaining task displays positive scaling). In addition, 1-shot examples and chain-of-thought can help mitigate undesirable scaling patterns even further. U-shaped scaling suggests that the inverse scaling trend observed in McKenzie et al. (2023) may not continue to hold for larger models, which we attribute to the presence of distractor tasks that only sufficiently large models can avoid.

Nearest Neighbor Machine Translation ($k$NN-MT) has achieved great success in domain adaptation tasks by integrating pre-trained Neural Machine Translation (NMT) models with domain-specific token-level retrieval. However, the reasons underlying its success have not been thoroughly investigated. In this paper, we comprehensively analyze $k$NN-MT through theoretical and empirical studies. Initially, we provide new insights into the working mechanism of $k$NN-MT as an efficient technique to implicitly execute gradient descent on the output projection layer of NMT, indicating that it is a specific case of model fine-tuning. Subsequently, we conduct multi-domain experiments and word-level analysis to examine the differences in performance between $k$NN-MT and entire-model fine-tuning. Our findings suggest that: ($i$) Incorporating $k$NN-MT with adapters yields comparable translation performance to fine-tuning on in-domain test sets, while achieving better performance on out-of-domain test sets; ($ii$) Fine-tuning significantly outperforms $k$NN-MT on the recall of in-domain low-frequency words, but this gap could be bridged by optimizing the context representations with additional adapter layers.

In this paper, we propose methods for discovering semantic differences in words appearing in two corpora. The key idea is to measure the coverage of meanings of a word in a corpus through the norm of its mean word vector, which is equivalent to examining a kind of variance of the word vector distribution. The proposed methods do not require alignments between words and/or corpora for comparison that previous methods do. All they require are to compute variance (or norms of mean word vectors) for each word type. Nevertheless, they rival the best-performing system in the SemEval-2020 Task 1. In addition, they are (i) robust for the skew in corpus sizes; (ii) capable of detecting semantic differences in infrequent words; and (iii) effective in pinpointing word instances that have a meaning missing in one of the two corpora under comparison. We show these advantages for historical corpora and also for native/non-native English corpora.

Language Models (LMs) have demonstrated impressive molecule understanding ability on various 1D text-related tasks. However, they inherently lack 2D graph perception — a critical ability of human professionals in comprehending molecules' topological structures. To bridge this gap, we propose MolCA: Molecular Graph-Language Modeling with Cross-Modal Projector and Uni-Modal Adapter. MolCA enables an LM (i.e., Galactica) to understand both text- and graph-based molecular contents via the cross-modal projector. Specifically, the cross-modal projector is implemented as a Q-Former to connect a graph encoder's representation space and an LM's text space. Further, MolCA employs a uni-modal adapter (i.e., LoRA) for the LM's efficient adaptation to downstream tasks. Unlike previous studies that couple an LM with a graph encoder via cross-modal contrastive learning, MolCA retains the LM's ability of open-ended text generation and augments it with 2D graph information. To showcase its effectiveness, we extensively benchmark MolCA on tasks of molecule captioning, IUPAC name prediction, and molecule-text retrieval, on which MolCA significantly outperforms the baselines.

Unintended dataset biases typically exist in existing Emotion Recognition in Conversations (ERC) datasets, including label bias, where models favor the majority class due to imbalanced training data, as well as the speaker and neutral word bias, where models make unfair predictions because of excessive correlations between specific neutral words or speakers and classes. However, previous studies in ERC generally focus on capturing context-sensitive and speaker-sensitive dependencies, ignoring the unintended dataset biases of data, which hampers the generalization and fairness in ERC. To address this issue, we propose a Training-Free Debiasing framework (TFD) that operates during prediction without additional training. To ensure compatibility with various ERC models, it does not balance data or modify the model structure. Instead, TFD extracts biases from the model by generating counterfactual utterances and contexts and mitigates them using simple yet empirically robust element-wise subtraction operations. Extensive experiments on three public datasets demonstrate that TFD effectively improves generalization ability and fairness across different ERC models.

The uniform information density (UID) hypothesis states that humans tend to distribute information roughly evenly across an utterance or discourse. Early evidence in support of the UID hypothesis came from Genzel and Charniak (2002), which proposed an entropy rate constancy principle based on the probability of English text under $n$-gram language models. We re-evaluate the claims of Genzel and Charniak (2002) with neural language models, failing to find clear evidence in support of entropy rate constancy. We conduct a range of experiments across datasets, model sizes, and languages and discuss implications for the uniform information density hypothesis and linguistic theories of efficient communication more broadly.

Ultra-fine entity typing plays a crucial role in information extraction by predicting fine-grained semantic types for entity mentions in text. However, this task poses significant challenges due to the massive number of entity types in the output space. The current state-of-the-art approaches, based on standard multi-label classifiers or cross-encoder models, suffer from poor generalization performance or inefficient inference speed. In this paper, we present CASENT, a seq2seq model designed for ultra-fine entity typing that predicts ultra-fine types with calibrated confidence scores. Our model takes an entity mention as input and employs constrained beam search to generate multiple types autoregressively. The raw sequence probabilities associated with the predicted types are then transformed into confidence scores using a novel calibration method. We conduct extensive experiments on the UFET dataset which contains over $10k$ types. Our method outperforms the previous state-of-the-art in terms of F1 score and calibration error, while achieving an inference speedup of over $50$ times. Additionally, we demonstrate the generalization capabilities of our model by evaluating it in zero-shot and few-shot settings on five specialized domain entity typing datasets that are unseen during training. Remarkably, our model outperforms large language models with 10 times more parameters in the zero-shot setting, and when fine-tuned on 50 examples, it significantly outperforms ChatGPT on all datasets.

Semantic role labeling (SRL) has multiple disjoint label sets, e.g., VerbNet and PropBank. Creating these datasets is challenging, therefore a natural question is how to use each one to help the other. Prior work has shown that cross-task interaction helps, but only explored multitask learning so far. A common issue with multi-task setup is that argument sequences are still separately decoded, running the risk of generating structurally inconsistent label sequences (as per lexicons like Semlink). In this paper, we eliminate such issue with a framework that jointly models VerbNet and PropBank labels as one sequence. In this setup, we show that enforcing Semlink constraints during decoding constantly improves the overall F1. With special input constructions, our joint model infers VerbNet arguments from given PropBank arguments with over 99 F1. For learning, we propose a constrained marginal model that learns with knowledge defined in Semlink to further benefit from the large amounts of PropBank-only data. On the joint benchmark based on CoNLL05, our models achieve state-of-the-art F1's, outperforming the prior best in-domain model by 3.5 (VerbNet) and 0.8 (PropBank). For out-of-domain generalization, our models surpass the prior best by 3.4 (VerbNet) and 0.2 (PropBank).

An effective approach to design automated Question Answering (QA) systems is to efficiently retrieve answers from pre-computed databases containing question/answer pairs. One of the main challenges to this design is the lack of training/testing data. Existing resources are limited in size and topics and either do not consider answers (question-question similarity only) or their quality in the annotation process. To fill this gap, we introduce a novel open-domain annotated resource to train and evaluate models for this task. The resource consists of 15,211 input questions. Each question is paired with 30 similar question/answer pairs, resulting in a total of 443,000 annotated examples. The binary label associated with each pair indicates the relevance with respect to the input question. Furthermore, we report extensive experimentation to test the quality and properties of our resource with respect to various key aspects of QA systems, including answer relevance, training strategies, and models input configuration.

Large language models (LLMs) have exhibited striking in-context learning (ICL) ability to adapt to target tasks with a few input-output demonstrations. For better ICL, different methods are proposed to select representative demonstrations from existing training corpora. However, such settings are not aligned with real-world practices, as end-users usually query LMs without access to demonstration pools. In this work, we introduce Self-ICL---a simple framework which bootstraps LMs' intrinsic capabilities to perform zero-shot ICL. Given a test input, Self-ICL first prompts the model to generate pseudo-inputs. Next, the model predicts pseudo-labels for the pseudo-inputs via zero-shot prompting. Finally, we perform ICL for the test input with the pseudo-input-label pairs as demonstrations. Evaluation on 23 BIG-Bench Hard tasks shows Self-ICL outperforms zero-shot baselines on both average accuracy and head-to-head comparison. Moreover, with zero-shot chain-of-thought, Self-ICL achieves results comparable to using real demonstrations. Additionally, we conduct a range of analyses to validate Self-ICL's effectiveness and provide insights for its behaviors under different settings.

Knowledge-Enhanced Pre-trained Language Models (KEPLMs) improve the performance of various downstream NLP tasks by injecting knowledge facts from large-scale Knowledge Graphs (KGs). However, existing methods for pre-training KEPLMs with relational triples are difficult to be adapted to close domains due to the lack of sufficient domain graph semantics. In this paper, we propose a Knowledge-enhanced language representation learning framework for various closed domains (KANGAROO) via capturing the implicit graph structure among the entities. Specifically, since the entity coverage rates of closed-domain KGs can be relatively low and may exhibit the global sparsity phenomenon for knowledge injection, we consider not only the shallow relational representations of triples but also the hyperbolic embeddings of deep hierarchical entity-class structures for effective knowledge fusion. Moreover, as two closed-domain entities under the same entity-class often havel locally dense neighbor subgraphs counted by max point biconnected component, we further propose a data augmentation strategy based on contrastive learning over subgraphs to construct hard negative samples of higher quality. It makes the underlying KELPMs better distinguish the semantics of these neighboring entities to further complement the global semantic sparsity. In the experiments, we evaluate KANGAROO over various knowledge-aware and general NLP tasks in both full and few-shot learning settings, outperforming various KEPLM training paradigms performance in closed-domains significantly.

Pre-trained Language Models (PLMs) are trained on vast unlabeled data, rich in world knowledge. This fact has sparked the interest of the community in quantifying the amount of factual knowledge present in PLMs, as this explains their performance on downstream tasks, and potentially justifies their use as knowledge bases. In this work, we survey methods and datasets that are used to probe PLMs for factual knowledge. Our contributions are: (1) We propose a categorization scheme for factual probing methods that is based on how their inputs, outputs and the probed PLMs are adapted; (2) We provide an overview of the datasets used for factual probing; (3) We synthesize insights about knowledge retention and prompt optimization in PLMs, analyze obstacles to adopting PLMs as knowledge bases and outline directions for future work.