NAACL 2022 Papers

The massive amount of trainable parameters in the pre-trained language models (PLMs) makes them hard to be deployed to multiple downstream tasks. To address this issue, parameter-efficient transfer learning methods have been proposed to tune only a few parameters during fine-tuning while freezing the rest. This paper looks at existing methods along this line through the kernel lens. Motivated by the connection between self-attention in transformer-based PLMs and kernel learning, we propose kernel-wise adapters, namely Kernel-mix, that utilize the kernel structure in self-attention to guide the assignment of the tunable parameters. These adapters use guidelines found in classical kernel learning and enable separate parameter tuning for each attention head. Our empirical results, over a diverse set of natural language generation and understanding tasks, show that our proposed adapters can attain or improve the strong performance of existing baselines.

We present a method to control the emotional prosody of Text to Speech (TTS) systems by using phoneme-level intermediate features (pitch, energy, and duration) as levers. As a key idea, we propose Differential Scaling (DS) to disentangle features relating to affective prosody from those arising due to acoustics conditions and speaker identity. With thorough experimental studies, we show that the proposed method improves over the prior art in accurately emulating the desired emotions while retaining the naturalness of speech. We extend the traditional evaluation of using individual sentences for a more complete evaluation of HCI systems. We present a novel experimental setup by replacing an actor with a TTS system in offline and live conversations. The emotion to be rendered is either predicted or manually assigned. The results show that the proposed method is strongly preferred over the state-of-the-art TTS system and adds the much-coveted “human touch” in machine dialogue. Audio samples from our experiments and the code are available at: https://emtts.github.io/tts-demo/

Persuasion is an intricate process involving empathetic connection between two individuals. Plain persuasive responses may make a conversation non-engaging. Even the most well-intended and reasoned persuasive conversations can fall through in the absence of empathetic connection between the speaker and listener. In this paper, we propose a novel task of incorporating empathy when generating persuasive responses. We develop an empathetic persuasive dialogue system by fine-tuning a maximum likelihood Estimation (MLE)-based language model in a reinforcement learning (RL) framework. To design feedbacks for our RL-agent, we define an effective and efficient reward function considering consistency, repetitiveness, emotion and persuasion rewards to ensure consistency, non-repetitiveness, empathy and persuasiveness in the generated responses. Due to lack of emotion annotated persuasive data, we first annotate the existing Persuaion For Good dataset with emotions, then build transformer based classifiers to provide emotion based feedbacks to our RL agent. Experimental results confirm that our proposed model increases the rate of generating persuasive responses as compared to the available state-of-the-art dialogue models while making the dialogues empathetically more engaging and retaining the language quality in responses.

In empathetic conversations, humans express their empathy to others with empathetic intents. However, most existing empathetic conversational methods suffer from a lack of empathetic intents, which leads to monotonous empathy. To address the bias of the empathetic intents distribution between empathetic dialogue models and humans, we propose a novel model to generate empathetic responses with human-consistent empathetic intents, EmpHi for short. Precisely, EmpHi learns the distribution of potential empathetic intents with a discrete latent variable, then combines both implicit and explicit intent representation to generate responses with various empathetic intents. Experiments show that EmpHi outperforms state-of-the-art models in terms of empathy, relevance, and diversity on both automatic and human evaluation. Moreover, the case studies demonstrate the high interpretability and outstanding performance of our model.

Each utterance in multi-turn empathetic dialogues has features such as emotion, keywords, and utterance-level meaning. Feature transitions between utterances occur naturally. However, existing approaches fail to perceive the transitions because they extract features for the context at the coarse-grained level. To solve the above issue, we propose a novel approach of recognizing feature transitions between utterances, which helps understand the dialogue flow and better grasp the features of utterance that needs attention. Also, we introduce a response generation strategy to help focus on emotion and keywords related to appropriate features when generating responses. Experimental results show that our approach outperforms baselines and especially, achieves significant improvements on multi-turn dialogues.

Few-shot language learners adapt knowledge from a pre-trained model to recognize novel classes from a few-labeled sentences. In such settings, fine-tuning a pre-trained language model can cause severe over-fitting. In this paper, we propose an Embedding Hallucination (EmbedHalluc) method, which generates auxiliary embedding-label pairs to expand the fine-tuning dataset. The hallucinator is trained by playing an adversarial game with the discriminator, such that the hallucinated embedding is indiscriminative to the real ones in the fine-tuning dataset. By training with the extended dataset, the language learner effectively learns from the diverse hallucinated embeddings to overcome the over-fitting issue. Experiments demonstrate that our proposed method is effective in a wide range of language tasks, outperforming current fine-tuning methods. Further, we show that EmbedHalluc outperforms other methods that address this over-fitting problem, such as common data augmentation, semi-supervised pseudo-labeling, and regularization.

The task of natural language inference (NLI), to decide if a hypothesis entails or contradicts a premise, received considerable attention in recent years. All competitive systems build on top of contextualized representations and make use of transformer architectures for learning an NLI model. When somebody is faced with a particular NLI task, they need to select the best model that is available. This is a time-consuming and resource-intense endeavour. To solve this practical problem, we propose a simple method for predicting the performance without actually fine-tuning the model. We do this by testing how well the pre-trained models perform on the aNLI task when just comparing sentence embeddings with cosine similarity to what kind of performance is achieved when training a classifier on top of these embeddings. We show that the accuracy of the cosine similarity approach correlates strongly with the accuracy of the classification approach with a Pearson correlation coefficient of 0.65. Since the similarity is orders of magnitude faster to compute on a given dataset (less than a minute vs. hours), our method can lead to significant time savings in the process of model selection.

Multi-triple extraction is a challenging task due to the existence of informative inter-triple correlations, and consequently rich interactions across the constituent entities and relations. While existing works only explore entity representations, we propose to explicitly introduce relation representation, jointly represent it with entities, and novelly align them to identify valid triples.We perform comprehensive experiments on document-level relation extraction and joint entity and relation extraction along with ablations to demonstrate the advantage of the proposed method.

Nowadays, pretrained language models (PLMs) have dominated the majority of NLP tasks. While, little research has been conducted on systematically evaluating the language abilities of PLMs. In this paper, we present a large-scale empirical study on general language ability evaluation of PLMs (ElitePLM). In our study, we design four evaluation dimensions, memory, comprehension, reasoning, and composition, to measure ten widely-used PLMs within five categories. Our empirical results demonstrate that: (1) PLMs with varying training objectives and strategies are good at different ability tests; (2) fine-tuning PLMs in downstream tasks is usually sensitive to the data size and distribution; (3) PLMs have excellent transferability between similar tasks. Moreover, the prediction results of PLMs in our experiments are released as an open resource for more deep and detailed analysis on the language abilities of PLMs. This paper can guide the future work to select, apply, and design PLMs for specific tasks. We have made all the details of experiments publicly available at https://github.com/RUCAIBox/ElitePLM.

Manually labeled training data is expensive, noisy, and often scarce, such as when developing new features or localizing existing features for a new region. In cases where labeled data is limited but unlabeled data is abundant, semi-supervised learning methods such as consistency training can be used to improve model performance, by training models to output consistent predictions between original and augmented versions of unlabeled data. In this work, we explore different data augmentation methods for consistency training (CT) on Natural Language Understanding (NLU) domain classification (DC) in the limited labeled data regime. We explore three types of augmentation techniques (human paraphrasing, back-translation, and dropout) for unlabeled data and train DC models to jointly minimize both the supervised loss and the consistency loss on unlabeled data. Our results demonstrate that DC models trained with CT methods and dropout based augmentation on only 0.1% (2,998 instances) of labeled data with the remainder as unlabeled can achieve a top-1 relative accuracy reduction of 12.25% compared to fully supervised model trained with 100% of labeled data, outperforming fully supervised models trained on 10x that amount of labeled data. The dropout-based augmentation achieves similar performance compare to back-translation based augmentation with much less computational resources. This paves the way for applications of using large scale unlabeled data for semi-supervised learning in production NLU systems.

The Transformer architecture continues to show remarkable performance gains in many Natural Language Processing tasks. However, obtaining such state-of-the-art performance in different tasks requires fine-tuning the same model separately for each task. Clearly, such an approach is demanding in terms of both memory requirements and computing power. In this paper, aiming to improve training efficiency across multiple tasks, we propose to collectively factorize the weighs of the multi-head attention module of a pre-trained Transformer. We test our proposed method on finetuning multiple natural language understanding tasks by employing BERT-Large as an instantiation of the Transformer and the GLUE as the evaluation benchmark. Experimental results show that our method requires training and storing only 1% of the initial model parameters for each task and matches or improves the original fine-tuned model’s performance for each task while effectively decreasing the parameter requirements by two orders of magnitude. Furthermore, compared to well-known adapter-based alternatives on the GLUE benchmark, our method consistently reaches the same levels of performance while requiring approximately four times fewer total and trainable parameters per task.

The remarkable success of large language models has been driven by dense models trained on massive unlabeled, unstructured corpora. These corpora typically contain text from diverse, heterogeneous sources, but information about the source of the text is rarely used during training. Transferring their knowledge to a target domain is typically done by continuing training in-domain. In this paper, we introduce a method to permit domain adaptation to many diverse domains using a computationally efficient adapter approach. Our method is based on the observation that textual domains are partially overlapping, and we represent domains as a hierarchical tree structure where each node in the tree is associated with a set of adapter weights. When combined with a frozen pretrained language model, this approach enables parameter sharing among related domains, while avoiding negative interference between unrelated ones. Experimental results with GPT-2 and a large fraction of the 100 most represented websites in C4 show across-the-board improvements in-domain. We additionally provide an inference time algorithm for a held-out domain and show that averaging over multiple paths through the tree enables further gains in generalization, while adding only a marginal cost to inference.

Abstractive summarization models are typically pre-trained on large amounts of generic texts, then fine-tuned on tens or hundreds of thousands of annotated samples. However, in opinion summarization, large annotated datasets of reviews paired with reference summaries are not available and would be expensive to create. This calls for fine-tuning methods robust to overfitting on small datasets. In addition, generically pre-trained models are often not accustomed to the specifics of customer reviews and, after fine-tuning, yield summaries with disfluencies and semantic mistakes. To address these problems, we utilize an efficient few-shot method based on adapters which, as we show, can easily store in-domain knowledge. Instead of fine-tuning the entire model, we add adapters and pre-train them in a task-specific way on a large corpus of unannotated customer reviews, using held-out reviews as pseudo summaries. Then, fine-tune the adapters on the small available human-annotated dataset. We show that this self-supervised adapter pre-training improves summary quality over standard fine-tuning by 2.0 and 1.3 ROUGE-L points on the Amazon and Yelp datasets, respectively. Finally, for summary personalization, we condition on aspect keyword queries, automatically created from generic datasets. In the same vein, we pre-train the adapters in a query-based manner on customer reviews and then fine-tune them on annotated datasets. This results in better-organized summary content reflected in improved coherence and fewer redundancies.

Bash is a Unix command language used for interacting with the Operating System. Recent works on natural language to Bash translation have made significant advances, but none of the previous methods utilize the problem’s inherent structure. We identify this structure andpropose a Segmented Invocation Transformer (SIT) that utilizes the information from the constituency parse tree of the natural language text. Our method is motivated by the alignment between segments in the natural language text and Bash command components. Incorporating the structure in the modelling improves the performance of the model. Since such systems must be universally accessible, we benchmark the inference times on a CPU rather than a GPU. We observe a 1.8x improvement in the inference time and a 5x reduction in model parameters. Attribution analysis using Integrated Gradients reveals that the proposed method can capture the problem structure.

Recently, large-scale transformer-based models have been proven to be effective over various tasks across many domains. Nevertheless, applying them in industrial production requires tedious and heavy works to reduce inference costs. To fill such a gap, we introduce a scalable inference solution: Easy and Efficient Transformer (EET), including a series of transformer inference optimization at the algorithm and implementation levels. First, we design highly optimized kernels for long inputs and large hidden sizes. Second, we propose a flexible CUDA memory manager to reduce the memory footprint when deploying a large model. Compared with the state-of-the-art transformer inference library (Faster Transformer v4.0), EET can achieve an average of 1.40-4.20x speedup on the transformer decoder layer with an A100 GPU.

Existing approaches to mitigate demographic biases evaluate on monolingual data, however, multilingual data has not been examined. In this work, we treat the gender as domains (e.g., male vs. female) and present a standard domain adaptation model to reduce the gender bias and improve performance of text classifiers under multilingual settings. We evaluate our approach on two text classification tasks, hate speech detection and rating prediction, and demonstrate the effectiveness of our approach with three fair-aware baselines.

Little attention has been paid on EArly Rumor Detection (EARD), and EARD performance was evaluated inappropriately on a few datasets where the actual early-stage information is largely missing. To reverse such situation, we construct BEARD, a new Benchmark dataset for EARD, based on claims from fact-checking websites by trying to gather as many early relevant posts as possible. We also propose HEARD, a novel model based on neural Hawkes process for EARD, which can guide a generic rumor detection model to make timely, accurate and stable predictions. Experiments show that HEARD achieves effective EARD performance on two commonly used general rumor detection datasets and our BEARD dataset.

Knowledge-based authentication is crucial for task-oriented spoken dialogue systems that offer personalised and privacy-focused services. Such systems should be able to enrol (E), verify (V), and identify (I) new and recurring users based on their personal information, e.g. postcode, name, and date of birth. In this work, we formalise the three authentication tasks and their evaluation protocols, and we present EVI, a challenging spoken multilingual dataset with 5,506 dialogues in English, Polish, and French. Our proposed models set the first competitive benchmarks, explore the challenges of multilingual natural language processing of spoken dialogue, and set directions for future research.

Recent works have empirically shown the effectiveness of data augmentation (DA) in NLP tasks, especially for those suffering from data scarcity. Intuitively, given the size of generated data, their diversity and quality are crucial to the performance of targeted tasks. However, to the best of our knowledge, most existing methods consider only either the diversity or the quality of augmented data, thus cannot fully mine the potential of DA for NLP. In this paper, we present an easy and plug-in data augmentation framework EPiDA to support effective text classification. EPiDA employs two mechanisms: relative entropy maximization (REM) and conditional entropy minimization (CEM) to control data generation, where REM is designed to enhance the diversity of augmented data while CEM is exploited to ensure their semantic consistency. EPiDA can support efficient and continuous data generation for effective classifier training. Extensive experiments show that EPiDA outperforms existing SOTA methods in most cases, though not using any agent networks or pre-trained generation networks, and it works well with various DA algorithms and classification models.