EMNLP 2022 论文列表

Are Large Pre-Trained Language Models Leaking Your Personal Information? In this paper, we analyze whether Pre-Trained Language Models (PLMs) are prone to leaking personal information. Specifically, we query PLMs for email addresses with contexts of the email address or prompts containing the owner’s name. We find that PLMs do leak personal information due to memorization. However, since the models are weak at association, the risk of specific personal information being extracted by attackers is low. We hope this work could help the community to better understand the privacy risk of PLMs and bring new insights to make PLMs safe.

Recent work on explainable NLP has shown that few-shot prompting can enable large pre-trained language models (LLMs) to generate grammatical and factual natural language explanations for data labels. In this work, we study the connection between explainability and sample hardness by investigating the following research question – “Are LLMs and humans equally good at explaining data labels for both easy and hard samples?” We answer this question by first collecting human-written explanations in the form of generalizable commonsense rules on the task of Winograd Schema Challenge (Winogrande dataset). We compare these explanations with those generated by GPT-3 while varying the hardness of the test samples as well as the in-context samples. We observe that (1) GPT-3 explanations are as grammatical as human explanations regardless of the hardness of the test samples, (2) for easy examples, GPT-3 generates highly supportive explanations but human explanations are more generalizable, and (3) for hard examples, human explanations are significantly better than GPT-3 explanations both in terms of label-supportiveness and generalizability judgements. We also find that hardness of the in-context examples impacts the quality of GPT-3 explanations. Finally, we show that the supportiveness and generalizability aspects of human explanations are also impacted by sample hardness, although by a much smaller margin than models.

The term ‘spurious correlations’ has been used in NLP to informally denote any undesirable feature-label correlations. However, a correlation can be undesirable because (i) the feature is irrelevant to the label (e.g. punctuation in a review), or (ii) the feature’s effect on the label depends on the context (e.g. negation words in a review), which is ubiquitous in language tasks. In case (i), we want the model to be invariant to the feature, which is neither necessary nor sufficient for prediction. But in case (ii), even an ideal model (e.g. humans) must rely on the feature, since it is necessary (but not sufficient) for prediction. Therefore, a more fine-grained treatment of spurious features is needed to specify the desired model behavior. We formalize this distinction using a causal model and probabilities of necessity and sufficiency, which delineates the causal relations between a feature and a label. We then show that this distinction helps explain results of existing debiasing methods on different spurious features, and demystifies surprising results such as the encoding of spurious features in model representations after debiasing.

Quality control is essential for creating extractive question answering (EQA) datasets via crowdsourcing. Aggregation across answers, i.e. word spans within passages annotated, by different crowd workers is one major focus for ensuring its quality. However, crowd workers cannot reach a consensus on a considerable portion of questions. We introduce a simple yet effective answer aggregation method that takes into account the relations among the answer, question, and context passage. We evaluate answer quality from both the view of question answering model to determine how confident the QA model is about each answer and the view of the answer verification model to determine whether the answer is correct. Then we compute aggregation scores with each answer’s quality and its contextual embedding produced by pre-trained language models. The experiments on a large real crowdsourced EQA dataset show that our framework outperforms baselines by around 16% on precision and effectively conduct answer aggregation for extractive QA task.

The emergent cross-lingual transfer seen in multilingual pretrained models has sparked significant interest in studying their behavior. However, because these analyses have focused on fully trained multilingual models, little is known about the dynamics of the multilingual pretraining process. We investigate when these models acquire their in-language and cross-lingual abilities by probing checkpoints taken from throughout XLM-R pretraining, using a suite of linguistic tasks. Our analysis shows that the model achieves high in-language performance early on, with lower-level linguistic skills acquired before more complex ones. In contrast, the point in pretraining when the model learns to transfer cross-lingually differs across language pairs. Interestingly, we also observe that, across many languages and tasks, the final model layer exhibits significant performance degradation over time, while linguistic knowledge propagates to lower layers of the network. Taken together, these insights highlight the complexity of multilingual pretraining and the resulting varied behavior for different languages over time.

Prompting inputs with natural language task descriptions has emerged as a popular mechanism to elicit reasonably accurate outputs from large-scale generative language models with little to no in-context supervision. This also helps gain insight into how well language models capture the semantics of a wide range of downstream tasks purely from self-supervised pre-training on massive corpora of unlabeled text. Such models have naturally also been exposed to a lot of undesirable content like racist and sexist language and there is only some work on awareness of models along these dimensions. In this paper, we define and comprehensively evaluate how well such language models capture the semantics of four tasks for bias: diagnosis, identification, extraction and rephrasing. We define three broad classes of task descriptions for these tasks: statement, question, and completion, with numerous lexical variants within each class. We study the efficacy of prompting for each task using these classes and the null task description across several decoding methods and few-shot examples. Our analyses indicate that language models are capable of performing these tasks to widely varying degrees across different bias dimensions, such as gender and political affiliation. We believe our work is an important step towards unbiased language models by quantifying the limits of current self-supervision objectives at accomplishing such sociologically challenging tasks.

Dialogue summarization is abstractive in nature, making it suffer from factual errors. The factual correctness of summaries has the highest priority before practical applications. Many efforts have been made to improve faithfulness in text summarization. However, there is a lack of systematic study on dialogue summarization systems. In this work, we first perform the fine-grained human analysis on the faithfulness of dialogue summaries and observe that over 35% of generated summaries are faithfully inconsistent respective the source dialogues. Furthermore, we present a new model-level faithfulness evaluation method. It examines generation models with multi-choice questions created by rule-based transformations. Experimental results show that our evaluation schema is a strong proxy for the factual correctness of summarization models. The human-annotated faithfulness samples and the evaluation toolkit are released to facilitate future research toward faithful dialogue summarization.

Math word problem (MWP) solving is an important task in question answering which requires human-like reasoning ability. Analogical reasoning has long been used in mathematical education, as it enables students to apply common relational structures of mathematical situations to solve new problems. In this paper, we propose to build a novel MWP solver by leveraging analogical MWPs, which advance the solver’s generalization ability across different kinds of MWPs. The key idea, named analogy identification, is to associate the analogical MWP pairs in a latent space, i.e., encoding an MWP close to another analogical MWP, while leaving away from the non-analogical ones. Moreover, a solution discriminator is integrated into the MWP solver to enhance the association between an MWP and its true solution. The evaluation results verify that our proposed analogical learning strategy promotes the performance of MWP-BERT on Math23k over the state-of-the-art model Generate2Rank, with 5 times fewer parameters in the encoder. We also find that our model has a stronger generalization ability in solving difficult MWPs due to the analogical learning from easy MWPs.

Polysemy is the phenomenon where a single word form possesses two or more related senses. It is an extremely ubiquitous part of natural language and analyzing it has sparked rich discussions in the linguistics, psychology and philosophy communities alike. With scarce attention paid to polysemy in computational linguistics, and even scarcer attention toward quantifying polysemy, in this paper, we propose a novel, unsupervised framework to compute and estimate polysemy scores for words in multiple languages. We infuse our proposed quantification with syntactic knowledge in the form of dependency structures. This informs the final polysemy scores of the lexicon motivated by recent linguistic findings that suggest there is an implicit relation between syntax and ambiguity/polysemy. We adopt a graph based approach by computing the discrete Ollivier Ricci curvature on a graph of the contextual nearest neighbors. We test our framework on curated datasets controlling for different sense distributions of words in 3 typologically diverse languages - English, French and Spanish. The effectiveness of our framework is demonstrated by significant correlations of our quantification with expert human annotated language resources like WordNet. We observe a 0.3 point increase in the correlation coefficient as compared to previous quantification studies in English. Our research leverages contextual language models and syntactic structures to empirically support the widely held theoretical linguistic notion that syntax is intricately linked to ambiguity/polysemy.

Although existing neural network approaches have achieved great progress on Chinese spelling correction, there is still room to improve. The model is required to avoid over-correction and to distinguish a correct token from its phonological and visual similar ones. In this paper, we propose an error-guided correction model to address these issues. By borrowing the powerful ability of the pre-trained BERT model, we propose a novel zero-shot error detection method to do a preliminary detection, which guides our model to attend more on the probably wrong tokens in encoding and to avoid modifying the correct tokens in generating. Furthermore, we introduce a new loss function to integrate the error confusion set, which enables our model to distinguish similar tokens. Moreover, our model supports highly parallel decoding to meet real applications. Experiments are conducted on widely used benchmarks. Our model achieves superior performance against state-of-the-art approaches by a remarkable margin, on both the quality and computation speed.

Parameter-efficient fine-tuning has garnered lots of attention in recent studies.On this subject, we investigate the capability of different transformer modules in transferring knowledge from a pre-trained model to a downstream task. Our empirical results suggest that every transformer module is a winning ticket such that fine-tuning the specific module while the rest of the network is frozen achieves a comparable performance to the full fine-tuning case. Among different modules in LMs, LayerNorms exhibit a significant capacity for transfer learning to the extent that with only 0.003% updateable parameters in the layer-wise analysis, they can show acceptable performance on various target tasks.We argue that the performance of LayerNorms could be attributed to their high-magnitude weights compared to other components in a pre-trained model.

We present an empirical study on methods for span finding, the selection of consecutive tokens in text for some downstream tasks. We focus on approaches that can be employed in training end-to-end information extraction systems, and find there is no definitive solution without considering task properties, and provide our observations to help with future design choices: 1) a tagging approach often yields higher precision while span enumeration and boundary prediction provide higher recall; 2) span type information can benefit a boundary prediction approach; 3) additional contextualization does not help span finding in most cases.

Though linguistic knowledge emerges during large-scale language model pretraining, recent work attempt to explicitly incorporate human-defined linguistic priors into task-specific fine-tuning. Infusing language models with syntactic or semantic knowledge from parsers has shown improvements on many language understanding tasks. To further investigate the effectiveness of structural linguistic priors, we conduct empirical study of replacing parsed graphs or trees with trivial ones (rarely carrying linguistic knowledge e.g., balanced tree) for tasks in the GLUE benchmark. Encoding with trivial graphs achieves competitive or even better performance in fully-supervised and few-shot settings. It reveals that the gains might not be significantly attributed to explicit linguistic priors but rather to more feature interactions brought by fusion layers. Hence we call for attention to using trivial graphs as necessary baselines to design advanced knowledge fusion methods in the future.

Large language models are shown to present privacy risks through memorization of training data, andseveral recent works have studied such risks for the pre-training phase. Little attention, however, has been given to the fine-tuning phase and it is not well understood how different fine-tuning methods (such as fine-tuning the full model, the model head, and adapter) compare in terms of memorization risk. This presents increasing concern as the “pre-train and fine-tune” paradigm proliferates. In this paper, we empirically study memorization of fine-tuning methods using membership inference and extraction attacks, and show that their susceptibility to attacks is very different. We observe that fine-tuning the head of the model has the highest susceptibility to attacks, whereas fine-tuning smaller adapters appears to be less vulnerable to known extraction attacks.

Access to external knowledge is essential for many natural language processing tasks, such as question answering and dialogue. Existing methods often rely on a parametric model that stores knowledge in its parameters, or use a retrieval-augmented model that has access to an external knowledge source. Parametric and retrieval-augmented models have complementary strengths in terms of computational efficiency and predictive accuracy. To combine the strength of both approaches, we propose the Efficient Memory-Augmented Transformer (EMAT) – it encodes external knowledge into a key-value memory and exploits the fast maximum inner product search for memory querying. We also introduce pre-training tasks that allow EMAT to encode informative key-value representations, and to learn an implicit strategy to integrate multiple memory slots into the transformer. Experiments on various knowledge-intensive tasks such as question answering and dialogue datasets show that, simply augmenting parametric models (T5-base) using our method produces more accurate results (e.g., 25.8 → 44.3 EM on NQ) while retaining a high throughput (e.g., 1000 queries/s on NQ). Compared to retrieval-augmented models, EMAT runs substantially faster across the board and produces more accurate results on WoW and ELI5.

Position Embedding (PE) is essential for transformer to capture the sequence ordering of input tokens. Despite its general effectiveness verified in Natural Language Processing (NLP) and Computer Vision (CV), its application in cross-modal tasks remains unexplored and suffers from two challenges: 1) the input text tokens and image patches are not aligned, 2) the encoding space of each modality is different, making it unavailable for feature comparison. In this paper, we propose a unified position embedding method for these problems, called AnChor-basEd Relative Position Embedding (ACE-RPE), in which we first introduce an anchor locating mechanism to bridge the semantic gap and locate anchors from different modalities. Then we conduct the distance calculation of each text token and image patch by computing their shortest paths from the located anchors. Last, we embed the anchor-based distance to guide the computation of cross-attention. In this way, it calculates cross-modal relative position embedding for cross-modal transformer. Benefiting from ACE-RPE, our method obtains new SOTA results on a wide range of benchmarks, such as Image-Text Retrieval on MS-COCO and Flickr30K, Visual Entailment on SNLI-VE, Visual Reasoning on NLVR2 and Weakly-supervised Visual Grounding on RefCOCO+.

Temporal knowledge graphs (TKGs) extrapolation reasoning predicts future events based on historical information, which has great research significance and broad application value. Existing methods can be divided into embedding-based methods and logical rule-based methods. Embedding-based methods rely on learned entity and relation embeddings to make predictions and thus lack interpretability. Logical rule-based methods bring scalability problems due to being limited by the learned logical rules. We combine the two methods to capture deep causal logic by learning rule embeddings, and propose an interpretable model for temporal knowledge graph reasoning called adaptive logical rule embedding model for inductive reasoning (ALRE-IR). ALRE-IR can adaptively extract and assess reasons contained in historical events, and make predictions based on causal logic. Furthermore, we propose a one-class augmented matching loss for optimization. When evaluated on the ICEWS14, ICEWS0515 and ICEWS18 datasets, the performance of ALRE-IR outperforms other state-of-the-art baselines. The results also demonstrate that ALRE-IR still shows outstanding performance when transferred to related dataset with common relation vocabulary, indicating our proposed model has good zero-shot reasoning ability.

There are growing interests in adapting large-scale language models using parameter-efficient fine-tuning methods. However, accelerating the model itself and achieving better inference efficiency through model compression has not been thoroughly explored yet.Model compression could provide the benefits of reducing memory footprints, enabling low-precision computations, and ultimately achieving cost-effective inference.To combine parameter-efficient adaptation and model compression, we propose AlphaTuning consisting of post-training quantization of the pre-trained language model and fine-tuning only some parts of quantized parameters for a target task.Specifically, AlphaTuning works by employing binary-coding quantization, which factorizes the full-precision parameters into binary parameters and a separate set of scaling factors.During the adaptation phase, the binary values are frozen for all tasks, while the scaling factors are fine-tuned for the downstream task.We demonstrate that AlphaTuning, when applied to GPT-2 and OPT, performs competitively with full fine-tuning on a variety of downstream tasks while achieving >10x compression ratio under 4-bit quantization and >1,000x reduction in the number of trainable parameters.

Sparsity of formal knowledge and roughness of non-ontological construction make sparsity problem particularly prominent in Open Knowledge Graphs (OpenKGs). Due to sparse links, learning effective representation for few-shot entities becomes difficult. We hypothesize that by introducing negative samples, a contrastive learning (CL) formulation could be beneficial in such scenarios. However, existing CL methods model KG triplets as binary objects of entities ignoring the relation-guided ternary propagation patterns and they are too generic, i.e., they ignore zero-shot, few-shot and synonymity problems that appear in OpenKGs. To address this, we propose TernaryCL, a CL framework based on ternary propagation patterns among head, relation and tail. TernaryCL designs Contrastive Entity and Contrastive Relation to mine ternary discriminative features with both negative entities and relations, introduces Contrastive Self to help zero- and few-shot entities learn discriminative features, Contrastive Synonym to model synonymous entities, and Contrastive Fusion to aggregate graph features from multiple paths. Extensive experiments on benchmarks demonstrate the superiority of TernaryCL over state-of-the-art models.