Knowledge Distillation (KD) is one of the most effective approaches to deploying large-scale pre-trained language models in low-latency environments by transferring the knowledge contained in the large-scale models to smaller student models. Prior KD approaches use the soft labels and intermediate activations generated by the teacher to transfer knowledge to the student model parameters alone. In this paper, we show that having access to non-parametric memory in the form of a knowledge base with the teacher’s soft labels and predictions can further improve student generalization. To enable the student to retrieve from the knowledge base effectively, we propose a new framework and loss function that preserves the semantic similarities of teacher and student training examples. We show through extensive experiments that our retrieval mechanism can achieve state-of-the-art performance for task-specific knowledge distillation on the GLUE benchmark.

Multimedia content, such as advertisements and story videos, exhibit a rich blend of creativity and multiple modalities. They incorporate elements like text, visuals, audio, and storytelling techniques, employing devices like emotions, symbolism, and slogans to convey meaning. There is a dearth of large annotated training datasets in the multimedia domain hindering the development of supervised learning models with satisfactory performance for real-world applications. On the other hand, the rise of large language models (LLMs) has witnessed remarkable zero-shot performance in various natural language processing (NLP) tasks, such as emotion classification, question answering, and topic classification. To leverage such advanced techniques to bridge this performance gap in multimedia understanding, we propose verbalizing long videos to generate their descriptions in natural language, followed by performing video-understanding tasks on the generated story as opposed to the original video. Through extensive experiments on fifteen video-understanding tasks, we demonstrate that our method, despite being zero-shot, achieves significantly better results than supervised baselines for video understanding. Furthermore, to alleviate a lack of story understanding benchmarks, we publicly release the first dataset on a crucial task in computational social science on persuasion strategy identification.

Music captioning has gained significant attention in the wake of the rising prominence of streaming media platforms. Traditional approaches often prioritize either the audio or lyrics aspect of the music, inadvertently ignoring the intricate interplay between the two. However, a comprehensive understanding of music necessitates the integration of both these elements. In this study, we delve into this overlooked realm by introducing a method to systematically learn multimodal alignment between audio and lyrics through contrastive learning. This not only recognizes and emphasizes the synergy between audio and lyrics but also paves the way for models to achieve deeper cross-modal coherence, thereby producing high-quality captions. We provide both theoretical and empirical results demonstrating the advantage of the proposed method, which achieves new state-of-the-art on two music captioning datasets.

Domain adaptation is an effective solution to data scarcity in low-resource scenarios. However, when applied to token-level tasks such as bioNER, domain adaptation methods often suffer from the challenging linguistic characteristics that clinical narratives possess, which leads to unsatsifactory performance. In this paper, we present a simple yet effective hardness-guided domain adaptation framework for bioNER tasks that can effectively leverage the domain hardness information to improve the adaptability of the learnt model in the low-resource scenarios. Experimental results on biomedical datasets show that our model can achieve significant performance improvement over the recently published state-of-the-art (SOTA) MetaNER model.

Though remarkable efforts have been made in non-parallel text style transfer, the evaluation system is unsatisfactory. It always evaluates over samples from only one checkpoint of the model and compares three metrics, i.e., transfer accuracy, BLEU score, and PPL score. In this paper, we argue the inappropriateness of both existing evaluation metrics and the evaluation method. Specifically, for evaluation metrics, we make a detailed analysis and comparison from three aspects: style transfer, content preservation, and naturalness; for the evaluation method, we reiterate the fallacy of picking one checkpoint for model comparison. As a result, we establish a robust evaluation method by examining the trade-off between style transfer and naturalness, and between content preservation and naturalness. Notably, we elaborate the human evaluation and automatically identify the inaccurate measurement of content preservation computed by the BLEU score. To overcome this issue, we propose a graph-based method to extract attribute content and attribute-independent content from input sentences in the YELP dataset and IMDB dataset. With the modified datasets, we design a new evaluation metric called “attribute hit” and propose an efficient regularization to leverage the attribute-dependent content and attribute-independent content as guiding signals. Experimental results have demonstrated the effectiveness of the proposed strategy.

With the need of fast retrieval speed and small memory footprint, document hashing has been playing a crucial role in large-scale information retrieval. To generate high-quality hashing code, both semantics and neighborhood information are crucial. However, most existing methods leverage only one of them or simply combine them via some intuitive criteria, lacking a theoretical principle to guide the integration process. In this paper, we encode the neighborhood information with a graph-induced Gaussian distribution, and propose to integrate the two types of information with a graph-driven generative model. To deal with the complicated correlations among documents, we further propose a tree-structured approximation method for learning. Under the approximation, we prove that the training objective can be decomposed into terms involving only singleton or pairwise documents, enabling the model to be trained as efficiently as uncorrelated ones. Extensive experimental results on three benchmark datasets show that our method achieves superior performance over state-of-the-art methods, demonstrating the effectiveness of the proposed model for simultaneously preserving semantic and neighborhood information.

The neural attention mechanism plays an important role in many natural language processing applications. In particular, multi-head attention extends single-head attention by allowing a model to jointly attend information from different perspectives. However, without explicit constraining, multi-head attention may suffer from attention collapse, an issue that makes different heads extract similar attentive features, thus limiting the model’s representation power. In this paper, for the first time, we provide a novel understanding of multi-head attention from a Bayesian perspective. Based on the recently developed particle-optimization sampling techniques, we propose a non-parametric approach that explicitly improves the repulsiveness in multi-head attention and consequently strengthens model’s expressiveness. Remarkably, our Bayesian interpretation provides theoretical inspirations on the not-well-understood questions: why and how one uses multi-head attention. Extensive experiments on various attention models and applications demonstrate that the proposed repulsive attention can improve the learned feature diversity, leading to more informative representations with consistent performance improvement on multiple tasks.

In sequence-to-sequence models, classical optimal transport (OT) can be applied to semantically match generated sentences with target sentences. However, in non-parallel settings, target sentences are usually unavailable. To tackle this issue without losing the benefits of classical OT, we present a semantic matching scheme based on the Optimal Partial Transport (OPT). Specifically, our approach partially matches semantically meaningful words between source and partial target sequences. To overcome the difficulty of detecting active regions in OPT (corresponding to the words needed to be matched), we further exploit prior knowledge to perform partial matching. Extensive experiments are conducted to evaluate the proposed approach, showing consistent improvements over sequence-to-sequence tasks.

Generative semantic hashing is a promising technique for large-scale information retrieval thanks to its fast retrieval speed and small memory footprint. For the tractability of training, existing generative-hashing methods mostly assume a factorized form for the posterior distribution, enforcing independence among the bits of hash codes. From the perspectives of both model representation and code space size, independence is always not the best assumption. In this paper, to introduce correlations among the bits of hash codes, we propose to employ the distribution of Boltzmann machine as the variational posterior. To address the intractability issue of training, we first develop an approximate method to reparameterize the distribution of a Boltzmann machine by augmenting it as a hierarchical concatenation of a Gaussian-like distribution and a Bernoulli distribution. Based on that, an asymptotically-exact lower bound is further derived for the evidence lower bound (ELBO). With these novel techniques, the entire model can be optimized efficiently. Extensive experimental results demonstrate that by effectively modeling correlations among different bits within a hash code, our model can achieve significant performance gains.

Text generation from a knowledge base aims to translate knowledge triples to natural language descriptions. Most existing methods ignore the faithfulness between a generated text description and the original table, leading to generated information that goes beyond the content of the table. In this paper, for the first time, we propose a novel Transformer-based generation framework to achieve the goal. The core techniques in our method to enforce faithfulness include a new table-text optimal-transport matching loss and a table-text embedding similarity loss based on the Transformer model. Furthermore, to evaluate faithfulness, we propose a new automatic metric specialized to the table-to-text generation problem. We also provide detailed analysis on each component of our model in our experiments. Automatic and human evaluations show that our framework can significantly outperform state-of-the-art by a large margin.

Auto-regressive text generation models usually focus on local fluency, and may cause inconsistent semantic meaning in long text generation. Further, automatically generating words with similar semantics is challenging, and hand-crafted linguistic rules are difficult to apply. We consider a text planning scheme and present a model-based imitation-learning approach to alleviate the aforementioned issues. Specifically, we propose a novel guider network to focus on the generative process over a longer horizon, which can assist next-word prediction and provide intermediate rewards for generator optimization. Extensive experiments demonstrate that the proposed method leads to improved performance.

Deep latent variable models (LVM) such as variational auto-encoder (VAE) have recently played an important role in text generation. One key factor is the exploitation of smooth latent structures to guide the generation. However, the representation power of VAEs is limited due to two reasons: (1) the Gaussian assumption is often made on the variational posteriors; and meanwhile (2) a notorious “posterior collapse” issue occurs. In this paper, we advocate sample-based representations of variational distributions for natural language, leading to implicit latent features, which can provide flexible representation power compared with Gaussian-based posteriors. We further develop an LVM to directly match the aggregated posterior to the prior. It can be viewed as a natural extension of VAEs with a regularization of maximizing mutual information, mitigating the “posterior collapse” issue. We demonstrate the effectiveness and versatility of our models in various text generation scenarios, including language modeling, unaligned style transfer, and dialog response generation. The source code to reproduce our experimental results is available on GitHub.

Hashing is promising for large-scale information retrieval tasks thanks to the efficiency of distance evaluation between binary codes. Generative hashing is often used to generate hashing codes in an unsupervised way. However, existing generative hashing methods only considered the use of simple priors, like Gaussian and Bernoulli priors, which limits these methods to further improve their performance. In this paper, two mixture-prior generative models are proposed, under the objective to produce high-quality hashing codes for documents. Specifically, a Gaussian mixture prior is first imposed onto the variational auto-encoder (VAE), followed by a separate step to cast the continuous latent representation of VAE into binary code. To avoid the performance loss caused by the separate casting, a model using a Bernoulli mixture prior is further developed, in which an end-to-end training is admitted by resorting to the straight-through (ST) discrete gradient estimator. Experimental results on several benchmark datasets demonstrate that the proposed methods, especially the one using Bernoulli mixture priors, consistently outperform existing ones by a substantial margin.

We propose a topic-guided variational auto-encoder (TGVAE) model for text generation. Distinct from existing variational auto-encoder (VAE) based approaches, which assume a simple Gaussian prior for latent code, our model specifies the prior as a Gaussian mixture model (GMM) parametrized by a neural topic module. Each mixture component corresponds to a latent topic, which provides a guidance to generate sentences under the topic. The neural topic module and the VAE-based neural sequence module in our model are learned jointly. In particular, a sequence of invertible Householder transformations is applied to endow the approximate posterior of the latent code with high flexibility during the model inference. Experimental results show that our TGVAE outperforms its competitors on both unconditional and conditional text generation, which can also generate semantically-meaningful sentences with various topics.

Recurrent neural networks (RNNs) have shown promising performance for language modeling. However, traditional training of RNNs using back-propagation through time often suffers from overfitting. One reason for this is that stochastic optimization (used for large training sets) does not provide good estimates of model uncertainty. This paper leverages recent advances in stochastic gradient Markov Chain Monte Carlo (also appropriate for large training sets) to learn weight uncertainty in RNNs. It yields a principled Bayesian learning algorithm, adding gradient noise during training (enhancing exploration of the model-parameter space) and model averaging when testing. Extensive experiments on various RNN models and across a broad range of applications demonstrate the superiority of the proposed approach relative to stochastic optimization.