Pre-trained autoregressive (AR) language models such as BART and GPTs have dominated OPen-ended Long Text Generation (Open-LTG).However, the AR nature will decrease the inference efficiency along with the increase of generation length, which hinder their application in Open-LTG.To improve inference efficiency, we alternatively explore the potential of the pre-trained masked language models (MLMs) along with a representative iterative non-autoregressive (NAR) decoding strategy for Open-LTG.Our preliminary study shows that pre-trained MLMs can merely generate short text and will collapse for long text modeling.To enhance the long text generation capability of MLMs, we introduce two simple yet effective strategies for the iterative NAR model: dynamic sliding window attention (DSWA) and linear temperature decay (LTD). It can alleviate long-distance collapse problems and achieve longer text generation with a flexible trade-off between performance and inference speedup.Experiments on the storytelling and multi-paragraph opinionated article writing tasks show that pre-trained MLMs can achieve more than 3 ×→ 13 × speedup with better performance than strong AR models.
Despite the excellent performance of Pre-trained Language Models on many text generation tasks, they suffer from inefficient inference on computation and memory due to their large-scale parameters and the universal autoregressive decoding paradigm. In this work, we propose a novel fine-tuning method DEER, which can make a single pre-trained model support Dynamic and Efficient infERence and achieve an adaptive trade-off between model performance and latency. In particular, our critical insight is to jointly utilize the non-autoregressive (NAR) generation and dynamic parameter pruning techniques, which can flexibly control the decoding iteration steps and model sizes according to memory and latency limitations. Besides, we also explore the effectiveness of the pre-trained MLMs (i.e., the BERT family) for text generation tasks since their bidirectional attention nature is more suitable for the NAR training objective. Extensive experiments on both monolingual and multilingual pre-trained MLMs demonstrate the effectiveness of our proposed DEER method by consistently achieving (1) higher BLEU scores than the strong autoregressive Transformer model on three neural machine translation tasks with 3 → 12 times speedup, (2) competitive performance (but with much faster inference speed) compared with the BART model on four GLGE benchmark tasks. Our code will be publicly available at GitHubhttps://github.com/dropreg/DEER.
Establishing retrieval-based dialogue systems that can select appropriate responses from the pre-built index has gained increasing attention. Recent common practice is to construct a two-stage pipeline with a fast retriever (e.g., bi-encoder) for first-stage recall followed by a smart response reranker (e.g., cross-encoder) for precise ranking. However, existing studies either optimize the retriever and reranker in independent ways, or distill the knowledge from a pre-trained reranker into the retriever in an asynchronous way, leading to sub-optimal performance of both modules. Thus, an open question remains about how to train them for a better combination of the best of both worlds. To this end, we present a cooperative training of the response retriever and the reranker whose parameters are dynamically optimized by the ground-truth labels as well as list-wise supervision signals from each other. As a result, the two modules can learn from each other and evolve together throughout the training. Experimental results on two benchmarks demonstrate the superiority of our method.
Hierarchical text classification (HTC) focuses on classifying one text into multiple labels, which are organized as a hierarchical taxonomy.Due to its wide involution in realistic scenarios, HTC attracts long-term attention from both industry and academia.However, the high cost of hierarchical multi-label annotation makes HTC suffer from the data scarcity problem.In view of the difficulty in balancing the controllability of multiple structural labels and text diversity, automatically generating high-quality data for HTC is challenging and under-explored.To fill this blank, we propose a novel data generation framework tailored for HTC, which can achieve both label controllability and text diversity by extracting high-quality semantic-level and phrase-level hierarchical label information.Experimental results on three benchmarks demonstrate that, compared with existing data augmentation methods, the data generated from our method can bring the most significant performance improvements of several strong HTC models.Extensive analysis confirms that the improvements yielded by our proposed method do correlate to the enhancement of label controllability and text diversity.
Diffusion models have been successfully adapted to text generation tasks by mapping the discrete text into the continuous space. However, there exist nonnegligible gaps between training and inference, owing to the absence of the forward process during inference. Thus, the model only predicts based on the previously generated reverse noise rather than the noise computed by the forward process. Besides, the widely-used downsampling strategy in speeding up the inference will cause the mismatch of diffusion trajectories between training and inference. To understand and mitigate the above two types of training-inference discrepancies, we launch a thorough preliminary study. Based on our observations, we propose two simple yet effective methods to bridge the gaps mentioned above, named Distance Penalty and Adaptive Decay Sampling. Extensive experiments on 6 generation tasks confirm the superiority of our methods, which can achieve 100× → 200× speedup with better performance. Our code will be released at https://github.com/CODINNLG/Bridge_Gap_Diffusion.
Dynamic early exit has demonstrated great potential in coping with the sharply increasing number of pre-trained language model parameters, which can achieve a good trade-off between performance and efficiency. The existing early exit paradigm relies on training parametrical internal classifiers at each intermediate layer to complete specific tasks. Based on the predictions of these internal classifiers, different methods are designed to decide when to exit. Under this circumstance, each intermediate layer takes on both generic language representation learning and task-specific feature extraction, which makes each intermediate layer struggle to balance two types of backward loss signals during training. To break this dilemma, we propose an adapter method to decouple the two distinct types of representation and further introduce a non-parametric simplex equiangular tight frame classifier (ETF) for improvement. Extensive experiments on monolingual and multilingual tasks demonstrate that our method gains significant improvements over strong PLM backbones and early exit methods.
The conventional success of textual classification relies on annotated data, and the new paradigm of pre-trained language models (PLMs) still requires a few labeled data for downstream tasks. However, in real-world applications, label noise inevitably exists in training data, damaging the effectiveness, robustness, and generalization of the models constructed on such data. Recently, remarkable achievements have been made to mitigate this dilemma in visual data, while only a few explore textual data. To fill this gap, we present SelfMix, a simple yet effective method, to handle label noise in text classification tasks. SelfMix uses the Gaussian Mixture Model to separate samples and leverages semi-supervised learning. Unlike previous works requiring multiple models, our method utilizes the dropout mechanism on a single model to reduce the confirmation bias in self-training and introduces a textual level mixup training strategy. Experimental results on three text classification benchmarks with different types of text show that the performance of our proposed method outperforms these strong baselines designed for both textual and visual data under different noise ratios and noise types. Our anonymous code is available at https://github.com/noise-learning/SelfMix.
Currently, human-bot symbiosis dialog systems, e.g. pre- and after-sales in E-commerce, are ubiquitous, and the dialog routing component is essential to improve the overall efficiency, reduce human resource cost and increase user experience. To satisfy this requirement, existing methods are mostly heuristic and cannot obtain high-quality performance. In this paper, we investigate the important problem by thoroughly mining both the data-to-task and task-to-task knowledge among various kinds of dialog data. To achieve the above target, we propose a comprehensive and general solution with multi-task learning framework, specifically including a novel dialog encoder and two tailored gated mechanism modules. The proposed Gated Mechanism enhanced Multi-task Model (G3M) can play the role of hierarchical information filtering and is non-invasive to the existing dialog systems. Experiments on two datasets collected from the real world demonstrate our method’s effectiveness and the results achieve the state-of-the-art performance by relatively increasing 8.7%/11.8% on RMSE metric and 2.2%/4.4% on F1 metric.
A deployed question answering (QA) model can easily fail when the test data has a distribution shift compared to the training data. Robustness tuning (RT) methods have been widely studied to enhance model robustness against distribution shifts before model deployment. However, can we improve a model after deployment? To answer this question, we evaluate test-time adaptation (TTA) to improve a model after deployment. We first introduce ColdQA, a unified evaluation benchmark for robust QA against text corruption and changes in language and domain. We then evaluate previous TTA methods on ColdQA and compare them to RT methods. We also propose a novel TTA method called online imitation learning (OIL). Through extensive experiments, we find that TTA is comparable to RT methods, and applying TTA after RT can significantly boost the performance on ColdQA. Our proposed OIL improves TTA to be more robust to variation in hyper-parameters and test distributions over time.
Recent research has revealed that neural language models at scale suffer from poor temporal generalization capability, i.e., language model pre-trained on static data from past years performs worse over time on emerging data. Existing methods mainly perform continual training to mitigate such a misalignment. While effective to some extent but is far from being addressed on both the language modeling and downstream tasks. In this paper, we empirically observe that temporal generalization is closely affiliated with lexical semantic change, which is one of the essential phenomena of natural languages. Based on this observation, we propose a simple yet effective lexical-level masking strategy to post-train a converged language model. Experiments on two pre-trained language models, two different classification tasks, and four benchmark datasets demonstrate the effectiveness of our proposed method over existing temporal adaptation methods, i.e., continual training with new data. Our code is available at https://github.com/zhaochen0110/LMLM.
Transformer-based autoregressive and non-autoregressive models have played an essential role in sequence generation tasks. The autoregressive model can obtain excellent performance, while the non-autoregressive model brings fast decoding speed for inference. In this paper, we propose JANUS, a Joint Autoregressive and Non-autoregressive training method using aUxiliary losS to enhance the model performance in both AR and NAR manner simultaneously and effectively alleviate the problem of distribution discrepancy.Further, we pre-train BART with JANUS on a large corpus with minimal cost (16 GPU days) and make the BART-JANUS capable of non-autoregressive generation, demonstrating that our approach can transfer the AR knowledge to NAR. Empirically, we show our approach and BART-JANUS can achieve significant improvement on multiple generation tasks, including machine translation and GLGE benchmarks. Our code is available at Github.
Sequential information, a.k.a., orders, is assumed to be essential for processing a sequence with recurrent neural network or convolutional neural network based encoders. However, is it possible to encode natural languages without orders? Given a bag of words from a disordered sentence, humans may still be able to understand what those words mean by reordering or reconstructing them. Inspired by such an intuition, in this paper, we perform a study to investigate how “order” information takes effects in natural language learning. By running comprehensive comparisons, we quantitatively compare the ability of several representative neural models to organize sentences from a bag of words under three typical scenarios, and summarize some empirical findings and challenges, which can shed light on future research on this line of work.
Story generation is a challenging task of automatically creating natural languages to describe a sequence of events, which requires outputting text with not only a consistent topic but also novel wordings. Although many approaches have been proposed and obvious progress has been made on this task, there is still a large room for improvement, especially for improving thematic consistency and wording diversity. To mitigate the gap between generated stories and those written by human writers, in this paper, we propose a planning-based conditional variational autoencoder, namely Plan-CVAE, which first plans a keyword sequence and then generates a story based on the keyword sequence. In our method, the keywords planning strategy is used to improve thematic consistency while the CVAE module allows enhancing wording diversity. Experimental results on a benchmark dataset confirm that our proposed method can generate stories with both thematic consistency and wording novelty, and outperforms state-of-the-art methods on both automatic metrics and human evaluations.
Adapting pre-trained language models (PrLMs) (e.g., BERT) to new domains has gained much attention recently. Instead of fine-tuning PrLMs as done in most previous work, we investigate how to adapt the features of PrLMs to new domains without fine-tuning. We explore unsupervised domain adaptation (UDA) in this paper. With the features from PrLMs, we adapt the models trained with labeled data from the source domain to the unlabeled target domain. Self-training is widely used for UDA, and it predicts pseudo labels on the target domain data for training. However, the predicted pseudo labels inevitably include noise, which will negatively affect training a robust model. To improve the robustness of self-training, in this paper we present class-aware feature self-distillation (CFd) to learn discriminative features from PrLMs, in which PrLM features are self-distilled into a feature adaptation module and the features from the same class are more tightly clustered. We further extend CFd to a cross-language setting, in which language discrepancy is studied. Experiments on two monolingual and multilingual Amazon review datasets show that CFd can consistently improve the performance of self-training in cross-domain and cross-language settings.
Variational autoencoders (VAEs) and Wasserstein autoencoders (WAEs) have achieved noticeable progress in open-domain response generation. Through introducing latent variables in continuous space, these models are capable of capturing utterance-level semantics, e.g., topic, syntactic properties, and thus can generate informative and diversified responses. In this work, we improve the WAE for response generation. In addition to the utterance-level information, we also model user-level information in latent continue space. Specifically, we embed user-level and utterance-level information into two multimodal distributions, and combine these two multimodal distributions into a mixed distribution. This mixed distribution will be used as the prior distribution of WAE in our proposed model, named as PersonaWAE. Experimental results on a large-scale real-world dataset confirm the superiority of our model for generating informative and personalized responses, where both automatic and human evaluations outperform state-of-the-art models.
Different from other text generation tasks, in product description generation, it is of vital importance to generate faithful descriptions that stick to the product attribute information. However, little attention has been paid to this problem. To bridge this gap we propose a model named Fidelity-oriented Product Description Generator (FPDG). FPDG takes the entity label of each word into account, since the product attribute information is always conveyed by entity words. Specifically, we first propose a Recurrent Neural Network (RNN) decoder based on the Entity-label-guided Long Short-Term Memory (ELSTM) cell, taking both the embedding and the entity label of each word as input. Second, we establish a keyword memory that stores the entity labels as keys and keywords as values, and FPDG will attend to keywords through attending to their entity labels. Experiments conducted a large-scale real-world product description dataset show that our model achieves the state-of-the-art performance in terms of both traditional generation metrics as well as human evaluations. Specifically, FPDG increases the fidelity of the generated descriptions by 25%.
Due to its potential applications, open-domain dialogue generation has become popular and achieved remarkable progress in recent years, but sometimes suffers from generic responses. Previous models are generally trained based on 1-to-1 mapping from an input query to its response, which actually ignores the nature of 1-to-n mapping in dialogue that there may exist multiple valid responses corresponding to the same query. In this paper, we propose to utilize the multiple references by considering the correlation of different valid responses and modeling the 1-to-n mapping with a novel two-step generation architecture. The first generation phase extracts the common features of different responses which, combined with distinctive features obtained in the second phase, can generate multiple diverse and appropriate responses. Experimental results show that our proposed model can effectively improve the quality of response and outperform existing neural dialogue models on both automatic and human evaluations.
It is a challenging task to automatically compose poems with not only fluent expressions but also aesthetic wording. Although much attention has been paid to this task and promising progress is made, there exist notable gaps between automatically generated ones with those created by humans, especially on the aspects of term novelty and thematic consistency. Towards filling the gap, in this paper, we propose a conditional variational autoencoder with adversarial training for classical Chinese poem generation, where the autoencoder part generates poems with novel terms and a discriminator is applied to adversarially learn their thematic consistency with their titles. Experimental results on a large poetry corpus confirm the validity and effectiveness of our model, where its automatic and human evaluation scores outperform existing models.