EMNLP 2022 论文列表

Prompt learning with immensely large Casual Language Models (CLMs) has been shown promising for attribute-controllable text generation (CTG). However, vanilla prompt tuning tends to imitate training corpus characteristics beyond the control attributes, resulting in a poor generalization ability. Moreover, it is less able to capture the relationship between different attributes, further limiting the control performance. In this paper, we propose a new CTG approach, namely DisCup, which incorporates the attribute knowledge of discriminator to optimize the control-prompts, steering a frozen CLM to produce attribute-specific texts. Specifically, the frozen CLM model, capable of producing multitudinous texts, is first used to generate the next-token candidates based on the context, so as to ensure the diversity of tokens to be predicted. Then, we leverage an attribute-discriminator to select desired/undesired tokens from those candidates, providing the inter-attribute knowledge. Finally, we bridge the above two traits by an unlikelihood objective for prompt-tuning. Extensive experimental results show that DisCup can achieve a new state-of-the-art control performance while maintaining an efficient and high-quality text generation, only relying on around 10 virtual tokens.

Addressing hate speech in online spaces has been conceptualized as a classification task that uses Natural Language Processing (NLP) techniques. Through this conceptualization, the hate speech detection task has relied on common conventions and practices from NLP. For instance, inter-annotator agreement is conceptualized as a way to measure dataset quality and certain metrics and benchmarks are used to assure model generalization. However, hate speech is a deeply complex and situated concept that eludes such static and disembodied practices. In this position paper, we critically reflect on these methodologies for hate speech detection, we argue that many conventions in NLP are poorly suited for the problem and encourage researchers to develop methods that are more appropriate for the task.

Dense retrievers encode queries and documents and map them in an embedding space using pre-trained language models. These embeddings need to be high-dimensional to fit training signals and guarantee the retrieval effectiveness of dense retrievers. However, these high-dimensional embeddings lead to larger index storage and higher retrieval latency. To reduce the embedding dimensions of dense retrieval, this paper proposes a Conditional Autoencoder (ConAE) to compress the high-dimensional embeddings to maintain the same embedding distribution and better recover the ranking features. Our experiments show that ConAE is effective in compressing embeddings by achieving comparable ranking performance with its teacher model and making the retrieval system more efficient. Our further analyses show that ConAE can alleviate the redundancy of the embeddings of dense retrieval with only one linear layer. All codes of this work are available at https://github.com/NEUIR/ConAE.

Despite the potential of federated learning, it is known to be vulnerable to backdoor attacks. Many robust federated aggregation methods are proposed to reduce the potential backdoor risk. However, they are mainly validated in the CV field. In this paper, we find that NLP backdoors are hard to defend against than CV, and we provide a theoretical analysis that the malicious update detection error probabilities are determined by the relative backdoor strengths. NLP attacks tend to have small relative backdoor strengths, which may result in the failure of robust federated aggregation methods for NLP attacks. Inspired by the theoretical results, we can choose some dimensions with higher backdoor strengths to settle this issue. We propose a novel federated aggregation algorithm, Dim-Krum, for NLP tasks, and experimental results validate its effectiveness.

Solid evaluation of neural machine translation (NMT) is key to its understanding and improvement. Current evaluation of an NMT system is usually built upon a heuristic decoding algorithm (e.g., beam search) and an evaluation metric assessing similarity between the translation and golden reference. However, this system-level evaluation framework is limited by evaluating only one best hypothesis and search errors brought by heuristic decoding algorithms. To better understand NMT models, we propose a novel evaluation protocol, which defines model errors with model’s ranking capability over hypothesis space. To tackle the problem of exponentially large space, we propose two approximation methods, top region evaluation along with an exact top-k decoding algorithm, which finds top-ranked hypotheses in the whole hypothesis space, and Monte Carlo sampling evaluation, which simulates hypothesis space from a broader perspective. To quantify errors, we define our NMT model errors by measuring distance between the hypothesis array ranked by the model and the ideally ranked hypothesis array. After confirming the strong correlation with human judgment, we apply our evaluation to various NMT benchmarks and model architectures. We show that the state-of-the-art Transformer models face serious ranking issues and only perform at the random chance level in the top region. We further analyze model errors on architectures with different depths and widths, as well as different data-augmentation techniques, showing how these factors affect model errors. Finally, we connect model errors with the search algorithms and provide interesting findings of beam search inductive bias and correlation with Minimum Bayes Risk (MBR) decoding.

To protect the privacy of individuals whose data is being shared, it is of high importance to develop methods allowing researchers and companies to release textual data while providing formal privacy guarantees to its originators. In the field of NLP, substantial efforts have been directed at building mechanisms following the framework of local differential privacy, thereby anonymizing individual text samples before releasing them. In practice, these approaches are often dissatisfying in terms of the quality of their output language due to the strong noise required for local differential privacy. In this paper, we approach the problem at hand using global differential privacy, particularly by training a generative language model in a differentially private manner and consequently sampling data from it. Using natural language prompts and a new prompt-mismatch loss, we are able to create highly accurate and fluent textual datasets taking on specific desired attributes such as sentiment or topic and resembling statistical properties of the training data. We perform thorough experiments indicating that our synthetic datasets do not leak information from our original data and are of high language quality and highly suitable for training models for further analysis on real-world data. Notably, we also demonstrate that training classifiers on private synthetic data outperforms directly training classifiers with DP-SGD.

Fine-tuning a pre-trained language model via the contrastive learning framework with a large amount of unlabeled sentences or labeled sentence pairs is a common way to obtain high-quality sentence representations. Although the contrastive learning framework has shown its superiority on sentence representation learning over previous methods, the potential of such a framework is under-explored so far due to the simple method it used to construct positive pairs. Motivated by this, we propose a method that makes hard positives from the original training examples. A pivotal ingredient of our approach is the use of prefix that attached to a pre-trained language model, which allows for differentiable data augmentation during contrastive learning. Our method can be summarized in two steps: supervised prefix-tuning followed by joint contrastive fine-tuning with unlabeled or labeled examples. Our experiments confirm the effectiveness of our data augmentation approach. The proposed method yields significant improvements over existing methods under both semi-supervised and supervised settings. Our experiments under a low labeled data setting also show that our method is more label-efficient than the state-of-the-art contrastive learning methods.

Delta tuning (DET, also known as parameter-efficient tuning) is deemed as the new paradigm for using pre-trained language models (PLMs). Up to now, various DETs with distinct design elements have been proposed, achieving performance on par with fine-tuning. However, the mechanisms behind the above success are still under-explored, especially the connections among various DETs. To fathom the mystery, we hypothesize that the adaptations of different DETs could all be reparameterized as low-dimensional optimizations in a unified optimization subspace, which could be found by jointly decomposing independent solutions of different DETs. Then we explore the connections among different DETs by conducting optimization within the subspace. In experiments, we find that, for a certain DET, conducting optimization simply in the subspace could achieve comparable performance to its original space, and the found solution in the subspace could be transferred to another DET and achieve non-trivial performance. We also visualize the performance landscape of the subspace, and find that, there exists a substantial region where different DETs all perform well. Finally, we extend our analysis and show the strong connections between fine-tuning and DETs. The codes are publicly available at https://github.com/thunlp/Unified-DeltaTuning.

Taking into account background knowledge as the context has always been an important part of solving tasks that involve natural language. One representative example of such tasks is text-based games, where players need to make decisions based on both description text previously shown in the game, and their own background knowledge about the language and common sense. In this work, we investigate not simply giving common sense, as can be seen in prior research, but also its effective usage. We assume that a part of the environment states different from common sense should constitute one of the grounds for action selection. We propose a novel agent, DiffG-RL, which constructs a Difference Graph that organizes the environment states and common sense by means of interactive objects with a dedicated graph encoder. DiffG-RL also contains a framework for extracting the appropriate amount and representation of common sense from the source to support the construction of the graph. We validate DiffG-RL in experiments with text-based games that require common sense and show that it outperforms baselines by 17% of scores. We will make our code publicly available.

Text analysis in the social sciences often involves using specialized dictionaries to reason with abstract concepts, such as perceptions about the economy or abuse on social media. These dictionaries allow researchers to impart domain knowledge and note subtle usages of words relating to a concept(s) of interest. We introduce the dictionary-assisted supervised contrastive learning (DASCL) objective, allowing researchers to leverage specialized dictionaries when fine-tuning pretrained language models. The text is first keyword simplified: a common, fixed token replaces any word in the corpus that appears in the dictionary(ies) relevant to the concept of interest. During fine-tuning, a supervised contrastive objective draws closer the embeddings of the original and keyword-simplified texts of the same class while pushing further apart the embeddings of different classes. The keyword-simplified texts of the same class are more textually similar than their original text counterparts, which additionally draws the embeddings of the same class closer together. Combining DASCL and cross-entropy improves classification performance metrics in few-shot learning settings and social science applications compared to using cross-entropy alone and alternative contrastive and data augmentation methods.

Financial risk prediction is an essential task for risk management in capital markets. While traditional prediction models are built based on the hard information of numerical data, recent studies have shown that the soft information of verbal cues in earnings conference calls is significant for predicting market risk due to its less constrained fashion and direct interaction between managers and analysts. However, most existing models mainly focus on extracting useful semantic information from the textual conference call transcripts but ignore their subtle yet important information of dialogue structures. To bridge this gap, we develop a graph attention network called DialogueGAT for financial risk prediction by simultaneously modeling the speakers and their utterances in dialogues in conference calls. Different from previous studies, we propose a new method for constructing the graph of speakers and utterances in a dialogue, and design contextual attention at both speaker and utterance levels for disentangling their effects on the downstream prediction task. For model evaluation, we extend an existing dataset of conference call transcripts by adding the dialogue structure and speaker information. Empirical results on our dataset of S&P1500 companies demonstrate the superiority of our proposed model over competitive baselines from the extant literature.

Dialogue meaning representation formulates natural language utterance semantics in their conversational context in an explicit and machine-readable form. Previous work typically follows the intent-slot framework, which is easy for annotation yet limited in scalability for complex linguistic expressions. A line of works alleviates the representation issue by introducing hierarchical structures but challenging to express complex compositional semantics, such as negation and coreference. We propose Dialogue Meaning Representation (DMR), a pliable and easily extendable representation for task-oriented dialogue. Our representation contains a set of nodes and edges to represent rich compositional semantics. Moreover, we propose an inheritance hierarchy mechanism focusing on domain extensibility. Additionally, we annotated DMR-FastFood, a multi-turn dialogue dataset with more than 70k utterances, with DMR. We propose two evaluation tasks to evaluate different dialogue models and a novel coreference resolution model GNNCoref for the graph-based coreference resolution task. Experiments show that DMR can be parsed well with pre-trained Seq2Seq models, and GNNCoref outperforms the baseline models by a large margin.The dataset and code are available at https://github.com/amazon-research/dialogue-meaning-representation

While interacting with chatbots, users may elicit multiple intents in a single dialogue utterance. Instead of training a dedicated multi-intent detection model, we propose DialogUSR, a dialogue utterance splitting and reformulation task that first splits multi-intent user query into several single-intent sub-queries and then recovers all the coreferred and omitted information in the sub-queries. DialogUSR can serve as a plug-in and domain-agnostic module that empowers the multi-intent detection for the deployed chatbots with minimal efforts. We collect a high-quality naturally occurring dataset that covers 23 domains with a multi-step crowd-souring procedure. To benchmark the proposed dataset, we propose multiple action-based generative models that involve end-to-end and two-stage training, and conduct in-depth analyses on the pros and cons of the proposed baselines.

Current end-to-end retrieval-based dialogue systems are mainly based on Recurrent Neural Networks or Transformers with attention mechanisms. Although promising results have been achieved, these models often suffer from slow inference or huge number of parameters. In this paper, we propose a novel lightweight fully convolutional architecture, called DialogConv, for response selection. DialogConv is exclusively built on top of convolution to extract matching features of context and response. Dialogues are modeled in 3D views, where DialogConv performs convolution operations on embedding view, word view and utterance view to capture richer semantic information from multiple contextual views. On the four benchmark datasets, compared with state-of-the-art baselines, DialogConv is on average about 8.5x smaller in size, and 79.39x and 10.64x faster on CPU and GPU devices, respectively. At the same time, DialogConv achieves the competitive effectiveness of response selection.

In this paper, we introduce the task of learning unsupervised dialogue embeddings.Trivial approaches such as combining pre-trained word or sentence embeddings and encoding through pre-trained language models (PLMs) have been shown to be feasible for this task.However, these approaches typically ignore the conversational interactions between interlocutors, resulting in poor performance.To address this issue, we proposed a self-guided contrastive learning approach named dial2vec.Dial2vec considers a dialogue as an information exchange process.It captures the interaction patterns between interlocutors and leverages them to guide the learning of the embeddings corresponding to each interlocutor.Then the dialogue embedding is obtained by an aggregation of the embeddings from all interlocutors.To verify our approach, we establish a comprehensive benchmark consisting of six widely-used dialogue datasets.We consider three evaluation tasks: domain categorization, semantic relatedness, and dialogue retrieval.Dial2vec achieves on average 8.7, 9.0, and 13.8 points absolute improvements in terms of purity, Spearman’s correlation, and mean average precision (MAP) over the strongest baseline on the three tasks respectively.Further analysis shows that dial2vec obtains informative and discriminative embeddings for both interlocutors under the guidance of the conversational interactions and achieves the best performance when aggregating them through the interlocutor-level pooling strategy.All codes and data are publicly available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/dial2vec.

Hierarchical text classification (HTC) is a key problem and task in many industrial applications, which aims to predict labels organized in a hierarchy for given input text. For example, HTC can group the descriptions of online products into a taxonomy or organizing customer reviews into a hierarchy of categories. In real-life applications, while Pre-trained Language Models (PLMs) have dominated many NLP tasks, they face significant challenges too—the conventional fine-tuning process needs to modify and save models with a huge number of parameters. This is becoming more critical for HTC in both global and local modelling—the latter needs to learn multiple classifiers at different levels/nodes in a hierarchy. The concern will be even more serious since PLM sizes are continuing to increase in order to attain more competitive performances. Most recently, prefix tuning has become a very attractive technology by only tuning and saving a tiny set of parameters. Exploring prefix turning for HTC is hence highly desirable and has timely impact. In this paper, we investigate prefix tuning on HTC in two typical setups: local and global HTC. Our experiment shows that the prefix-tuning model only needs less than 1% of parameters and can achieve performance comparable to regular full fine-tuning. We demonstrate that using contrastive learning in learning prefix vectors can further improve HTC performance.

Providing better language tools for low-resource and endangered languages is imperative for equitable growth.Recent progress with massively multilingual pretrained models has proven surprisingly effective at performing zero-shot transfer to a wide variety of languages.However, this transfer is not universal, with many languages not currently understood by multilingual approaches.It is estimated that only 72 languages possess a “small set of labeled datasets” on which we could test a model’s performance, the vast majority of languages not having the resources available to simply evaluate performances on.In this work, we attempt to clarify which languages do and do not currently benefit from such transfer.To that end, we develop a general approach that requires only unlabelled text to detect which languages are not well understood by a cross-lingual model.Our approach is derived from the hypothesis that if a model’s understanding is insensitive to perturbations to text in a language, it is likely to have a limited understanding of that language.We construct a cross-lingual sentence similarity task to evaluate our approach empirically on 350, primarily low-resource, languages.

We show that large pre-trained language models are inherently highly capable of identifying label errors in natural language datasets: simply examining out-of-sample data points in descending order of fine-tuned task loss significantly outperforms more complex error-detection mechanisms proposed in previous work. To this end, we contribute a novel method for introducing realistic, human-originated label noise into existing crowdsourced datasets such as SNLI and TweetNLP. We show that this noise has similar properties to real, hand-verified label errors, and is harder to detect than existing synthetic noise, creating challenges for model robustness.We argue that human-originated noise is a better standard for evaluation than synthetic noise. Finally, we use crowdsourced verification to evaluate the detection of real errors on IMDB, Amazon Reviews, and Recon, and confirm that pre-trained models perform at a 9–36% higher absolute Area Under the Precision-Recall Curve than existing models.

Neuropsychological exams are commonly used to diagnose various kinds of cognitive impairment. They typically involve a trained examiner who conducts a series of cognitive tests with a subject. In recent years, there has been growing interest in developing machine learning methods to extract speech and language biomarkers from exam recordings to provide automated input for cognitive assessment. Inspired by recent findings suggesting that the examiner’s language can influence cognitive impairment classifications, in this paper, we study the influence of the examiner on automatic dementia identification decisions in real-world neuropsychological exams. To mitigate the influence of the examiner, we propose a systematic three-stage pipeline for detecting dementia from exam recordings. In the first stage, we perform audio-based speaker diarization (i.e., estimating who spoke when?) by incorporating speaker discriminative features. In the second stage, we employ text-based language models to identify the role of the speaker (i.e., examiner or subject). Finally, in the third stage, we employ text- and audio-based models to detect cognitive impairment from hypothesized subject segments. Our studies suggest that incorporating audio-based diarization followed by text-based role identification helps mitigate the influences from the examiner’s segments. Further, we found that the text and audio modalities complement each other, and the performance improves when we use both modalities. We also perform several carefully designed experimental studies to assess the performance of each stage.