Recent talking avatar generation models have made strides in achieving realistic and accurate lip synchronization with the audio, but often fall short in controlling and conveying detailed expressions and emotions of the avatar, making the generated video less vivid and controllable. In this paper, we propose a text-guided approach for generating emotionally expressive 2D avatars, offering fine-grained control, improved interactivity, and generalizability to the resulting video. Our framework, named InstructAvatar, leverages a natural language interface to control the emotion as well as the facial motion of avatars. Technically, we utilize GPT-4V to design an automatic annotation pipeline, constructing an instruction-video paired training dataset. This is combined with a novel two-branch diffusion-based generator to predict avatars using both audio and text instructions simultaneously. Experimental results demonstrate that InstructAvatar produces results that align well with both conditions, and outperforms existing methods in fine-grained emotion control, lip-sync quality, and naturalness.

Recent advancements in video autoencoders (Video AEs) have significantly improved the quality and efficiency of video generation. In this paper, we propose a novel and compact video autoencoder, VidTwin, that decouples video into two distinct latent spaces: Structure latent vectors, which capture overall content and global movement, and Dynamics latent vectors, which represent fine-grained details and rapid movements. Specifically, our approach leverages an Encoder-Decoder backbone, augmented with two submodules for extracting these latent spaces, respectively. The first submodule employs a Q-Former to extract low-frequency motion trends, followed by downsampling blocks to remove redundant content details. The second averages the latent vectors along the spatial dimension to capture rapid motion. Extensive experiments show that VidTwin achieves a high compression rate of 0.20% with high reconstruction quality (PSNR of 28.14 on the MCL-JCV dataset), and performs efficiently and effectively in downstream generative tasks. Moreover, our model demonstrates explainability and scalability, paving the way for future research in video latent representation and generation.

People interact with the real-world largely dependent on visual signal, which are ubiquitous and illustrate detailed demonstrations. In this paper, we explore utilizing visual signals as a new interface for models to interact with the environment. Specifically, we choose videos as a representative visual signal. And by training autoregressive Transformers on video datasets in a self-supervised objective, we find that the model emerges a zero-shot capability to infer the semantics from a demonstration video, and imitate the semantics to an unseen scenario. This allows the models to perform unseen tasks by watching the demonstration video in an in-context manner, without further fine-tuning. To validate the imitation capacity, we design various evaluation metrics including both objective and subjective measures. The results show that our models can generate high-quality video clips that accurately align with the semantic guidance provided by the demonstration videos, and we also show that the imitation capacity follows the scaling law.

Recent advances in text-guided video editing have showcased promising results in appearance editing (e.g., stylization). However, video motion editing in the temporal dimension (e.g., from eating to waving), which distinguishes video editing from image editing, is underexplored. In this work, we present UniEdit, a tuning-free framework that supports both video motion and appearance editing by harnessing the power of a pre-trained text-to-video generator within an inversion-then-generation framework. To realize motion editing while preserving source video content, based on the insights that temporal and spatial self-attention layers encode inter-frame and intra-frame dependency, we introduce auxiliary motion-reference and reconstruction branches to produce text-guided motion and source features respectively. The obtained features are then injected into the main editing path via temporal and spatial self-attention layers. We also validate the effectiveness and flexibility of UniEdit by deploying it on three T2V generative models with different architectures. Experiments demonstrate that UniEdit covers video motion editing and various appearance editing scenarios, and surpasses the state-of-the-art methods. Our code is publicly available.

Due to the rapid development of computer vision, single-modal (RGB) object tracking has made significant progress in recent years. Considering the limitation of single imaging sensor, multi-modal images (RGB, infrared, etc.) are introduced to compensate for this deficiency for all-weather object tracking in complex environments. However, as acquiring sufficient multi-modal tracking data is hard while the dominant modality changes with the open environment, most existing techniques fail to extract multi-modal complementary information dynamically, yielding unsatisfactory tracking performance. To handle this problem, we propose a novel multi-modal visual prompt tracking model based on a universal bi-directional adapter, cross-prompting multiple modalities mutually. Our model consists of a universal bi-directional adapter and multiple modality-specific transformer encoder branches with sharing parameters. The encoders extract features of each modality separately by using a frozen, pre-trained foundation model. We develop a simple but effective light feature adapter to transfer modality-specific information from one modality to another, performing visual feature prompt fusion in an adaptive manner. With adding fewer (0.32M) trainable parameters, our model achieves superior tracking performance in comparison with both the full fine-tuning methods and the prompt learning-based methods. Our code is available: https://github.com/SparkTempest/BAT.

Large Language Models (LLMs) excel in various tasks, but they rely on carefully crafted prompts that often demand substantial human effort. To automate this process, in this paper, we propose a novel framework for discrete prompt optimization, called EvoPrompt, which borrows the idea of evolutionary algorithms (EAs) as they exhibit good performance and fast convergence. To enable EAs to work on discrete prompts, which are natural language expressions that need to be coherent and human-readable, we connect LLMs with EAs. This approach allows us to simultaneously leverage the powerful language processing capabilities of LLMs and the efficient optimization performance of EAs. Specifically, abstaining from any gradients or parameters, EvoPrompt starts from a population of prompts and iteratively generates new prompts with LLMs based on the evolutionary operators, improving the population based on the development set. We optimize prompts for both closed- and open-source LLMs including GPT-3.5 and Alpaca, on 31 datasets covering language understanding, generation tasks, as well as BIG-Bench Hard (BBH) tasks. EvoPrompt significantly outperforms human-engineered prompts and existing methods for automatic prompt generation (e.g., up to 25% on BBH). Furthermore, EvoPrompt demonstrates that connecting LLMs with EAs creates synergies, which could inspire further research on the combination of LLMs and conventional algorithms.

Zero-shot talking avatar generation aims at synthesizing natural talking videos from speech and a single portrait image. Previous methods have relied on domain-specific heuristics such as warping-based motion representation and 3D Morphable Models, which limit the naturalness and diversity of the generated avatars. In this work, we introduce GAIA (Generative AI for Avatar), which eliminates the domain priors in talking avatar generation. In light of the observation that the speech only drives the motion of the avatar while the appearance of the avatar and the background typically remain the same throughout the entire video, we divide our approach into two stages: 1) disentangling each frame into motion and appearance representations; 2) generating motion sequences conditioned on the speech and reference portrait image. We collect a large-scale high-quality talking avatar dataset and train the model on it with different scales (up to 2B parameters). Experimental results verify the superiority, scalability, and flexibility of GAIA as 1) the resulting model beats previous baseline models in terms of naturalness, diversity, lip-sync quality, and visual quality; 2) the framework is scalable since larger models yield better results; 3) it is general and enables different applications like controllable talking avatar generation and text-instructed avatar generation.

Zhujin Gao, Junliang Guo, Xu Tan, Yongxin Zhu, Fang Zhang, Jiang Bian, Linli Xu. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Residual networks, as discrete approximations of Ordinary Differential Equations (ODEs), have inspired significant advancements in neural network design, including multistep methods, high-order methods, and multi-particle dynamical systems. The precision of the solution to ODEs significantly affects parameter optimization, thereby impacting model performance. In this work, we present a series of advanced explorations of Transformer architecture design to minimize the error compared to the true ``solution.'' First, we introduce a predictor-corrector learning framework to minimize truncation errors, which consists of a high-order predictor and a multistep corrector. Second, we propose an exponential moving average-based coefficient learning method to strengthen our higher-order predictor. Extensive experiments on large-scale machine translation, abstractive summarization, language modeling, and natural language understanding benchmarks demonstrate the superiority of our approach. On the WMT'14 English-German and English-French tasks, our model achieved BLEU scores of 30.95 and 44.27, respectively. Furthermore, on the OPUS multilingual machine translation task, our model surpasses a robust 3.8B DeepNet by an average of 2.9 SacreBLEU, using only 1/3 parameters. Notably, it also beats LLama models by 5.7 accuracy points on the LM Harness Evaluation.

Significant progress has been made in text-to-video generation through the use of powerful generative models and large-scale internet data. However, substantial challenges remain in precisely controlling individual elements within the generated video, such as the movement and appearance of specific characters and the manipulation of viewpoints. In this work, we propose a novel paradigm that generates each element in 3D representation separately and then composites them with priors from Large Language Models (LLMs) and 2D diffusion models. Specifically, given an input textual query, our scheme consists of four stages: 1) we leverage the LLMs as the director to first decompose the complex query into several sub-queries, where each sub-query describes each element of the generated video; 2) to generate each element, pre-trained models are invoked by the LLMs to obtain the corresponding 3D representation; 3) to composite the generated 3D representations, we prompt multi-modal LLMs to produce coarse guidance on the scale, location, and trajectory of different objects; 4) to make the results adhere to natural distribution, we further leverage 2D diffusion priors and use score distillation sampling to refine the composition. Extensive experiments demonstrate that our method can generate high-fidelity videos from text with flexible control over each element.

Retrosynthesis, which predicts the reactants of a given target molecule, is an essential task for drug discovery. In recent years, the machine learing based retrosynthesis methods have achieved promising results. In this work, we introduce RetroKNN, a local reaction template retrieval method to further boost the performance of template-based systems with non-parametric retrieval. We first build an atom-template store and a bond-template store that contains the local templates in the training data, then retrieve from these templates with a k-nearest-neighbor (KNN) search during inference. The retrieved templates are combined with neural network predictions as the final output. Furthermore, we propose a lightweight adapter to adjust the weights when combing neural network and KNN predictions conditioned on the hidden representation and the retrieved templates. We conduct comprehensive experiments on two widely used benchmarks, the USPTO-50K and USPTO-MIT. Especially for the top-1 accuracy, we improved 7.1% on the USPTO-50K dataset and 12.0% on the USPTO-MIT dataset.These results demonstrate the effectiveness of our method.

Video dubbing aims to translate the original speech in a film or television program into the speech in a target language, which can be achieved with a cascaded system consisting of speech recognition, machine translation and speech synthesis. To ensure the translated speech to be well aligned with the corresponding video, the length/duration of the translated speech should be as close as possible to that of the original speech, which requires strict length control. Previous works usually control the number of words or characters generated by the machine translation model to be similar to the source sentence, without considering the isochronicity of speech as the speech duration of words/characters in different languages varies. In this paper, we propose VideoDubber, a machine translation system tailored for the task of video dubbing, which directly considers the speech duration of each token in translation, to match the length of source and target speech. Specifically, we control the speech length of generated sentence by guiding the prediction of each word with the duration information, including the speech duration of itself as well as how much duration is left for the remaining words. We design experiments on four language directions (German -> English, Spanish -> English, Chinese English), and the results show that VideoDubber achieves better length control ability on the generated speech than baseline methods. To make up the lack of real-world datasets, we also construct a real-world test set collected from films to provide comprehensive evaluations on the video dubbing task.

Non-autoregressive (NAR) generation, which is first proposed in neural machine translation (NMT) to speed up inference, has attracted much attention in both machine learning and natural language processing communities. While NAR generation can significantly accelerate inference speed for machine translation, the speedup comes at the cost of sacrificed translation accuracy compared to its counterpart, autoregressive (AR) generation. In recent years, many new models and algorithms have been designed/proposed to bridge the accuracy gap between NAR generation and AR generation. In this paper, we conduct a systematic survey with comparisons and discussions of various non-autoregressive translation (NAT) models from different aspects. Specifically, we categorize the efforts of NAT into several groups, including data manipulation, modeling methods, training criterion, decoding algorithms, and the benefit from pre-trained models. Furthermore, we briefly review other applications of NAR models beyond machine translation, such as grammatical error correction, text summarization, text style transfer, dialogue, semantic parsing, automatic speech recognition, and so on. In addition, we also discuss potential directions for future exploration, including releasing the dependency of KD, reasonable training objectives, pre-training for NAR, and wider applications, etc. We hope this survey can help researchers capture the latest progress in NAR generation, inspire the design of advanced NAR models and algorithms, and enable industry practitioners to choose appropriate solutions for their applications.

The task of Grammatical Error Correction (GEC) has received remarkable attention with wide applications in Natural Language Processing (NLP) in recent years. While one of the key principles of GEC is to keep the correct parts unchanged and avoid over-correction, previous sequence-to-sequence (seq2seq) models generate results from scratch, which are not guaranteed to follow the original sentence structure and may suffer from the over-correction problem. In the meantime, the recently proposed sequence tagging models can overcome the over-correction problem by only generating edit operations, but are conditioned on human designed language-specific tagging labels. In this paper, we combine the pros and alleviate the cons of both models by proposing a novel Sequence-to-Action (S2A) module. The S2A module jointly takes the source and target sentences as input, and is able to automatically generate a token-level action sequence before predicting each token, where each action is generated from three choices named SKIP, COPY and GENerate. Then the actions are fused with the basic seq2seq framework to provide final predictions. We conduct experiments on the benchmark datasets of both English and Chinese GEC tasks. Our model consistently outperforms the seq2seq baselines, while being able to significantly alleviate the over-correction problem as well as holding better generality and diversity in the generation results compared to the sequence tagging models.

Kexun Zhang, Rui Wang, Xu Tan, Junliang Guo, Yi Ren, Tao Qin, Tie-Yan Liu. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Binaural audio plays a significant role in constructing immersive augmented and virtual realities. As it is expensive to record binaural audio from the real world, synthesizing them from mono audio has attracted increasing attention. This synthesis process involves not only the basic physical warping of the mono audio, but also room reverberations and head/ear related filtration, which, however, are difficult to accurately simulate in traditional digital signal processing. In this paper, we formulate the synthesis process from a different perspective by decomposing the binaural audio into a common part that shared by the left and right channels as well as a specific part that differs in each channel. Accordingly, we propose BinauralGrad, a novel two-stage framework equipped with diffusion models to synthesize them respectively. Specifically, in the first stage, the common information of the binaural audio is generated with a single-channel diffusion model conditioned on the mono audio, based on which the binaural audio is generated by a two-channel diffusion model in the second stage. Combining this novel perspective of two-stage synthesis with advanced generative models (i.e., the diffusion models), the proposed BinauralGrad is able to generate accurate and high-fidelity binaural audio samples. Experiment results show that on a benchmark dataset, BinauralGrad outperforms the existing baselines by a large margin in terms of both object and subject evaluation metrics (Wave L2: $0.128$ vs. $0.157$, MOS: $3.80$ vs. $3.61$). The generated audio samples\footnote{\url{https://speechresearch.github.io/binauralgrad}} and code\footnote{\url{https://github.com/microsoft/NeuralSpeech/tree/master/BinauralGrad}} are available online.

In this paper, we propose to formulate the task-oriented dialogue system as the purely natural language generation task, so as to fully leverage the large-scale pre-trained models like GPT-2 and simplify complicated delexicalization prepossessing. However, directly applying this method heavily suffers from the dialogue entity inconsistency caused by the removal of delexicalized tokens, as well as the catastrophic forgetting problem of the pre-trained model during fine-tuning, leading to unsatisfactory performance. To alleviate these problems, we design a novel GPT-Adapter-CopyNet network, which incorporates the lightweight adapter and CopyNet modules into GPT-2 to achieve better performance on transfer learning and dialogue entity generation. Experimental results conducted on the DSTC8 Track 1 benchmark and MultiWOZ dataset demonstrate that our proposed approach significantly outperforms baseline models with a remarkable performance on automatic and human evaluations.

Hierarchical multi-label text classification (HMTC) deals with the challenging task where an instance can be assigned to multiple hierarchically structured categories at the same time.The majority of prior studies either focus on reducing the HMTC task into a flat multi-label problem ignoring the vertical category correlations or exploiting the dependencies across different hierarchical levels without considering the horizontal correlations among categories at the same level, which inevitably leads to fundamental information loss.In this paper, we propose a novel HMTC framework that considers both vertical and horizontal category correlations.Specifically, we first design a loosely coupled graph convolutional neural network as the representation extractor to obtain representations for words, documents, and, more importantly, level-wise representations for categories, which are not considered in previous works.Then, the learned category representations are adopted to capture the vertical dependencies among levels of category hierarchy and model the horizontal correlations.Finally, based on the document embeddings and category embeddings, we design a hybrid algorithm to predict the categories of the entire hierarchical structure.Extensive experiments conducted on real-world HMTC datasets validate the effectiveness of the proposed framework with significant improvements over the baselines.

Large-scale pre-trained language models (e.g., BERT) have attracted great attention in recent years. It is straightforward to fine-tune them on natural language understanding tasks such as text classification, however, effectively and efficiently incorporating them into natural language generation tasks such as neural machine translation remains a challenging problem. In this paper, we integrate two pre-trained BERT models from the source and target language domains into a sequence-to-sequence model by introducing light-weight adapter modules. The adapters are inserted between BERT layers and tuned on downstream tasks, while the parameters of BERT models are fixed during fine-tuning. As pre-trained language models are usually very deep, inserting adapters into all layers will result in a considerable scale of new parameters. To deal with this problem, we introduce latent variables to decide whether using adapters or not in each layer, which are learned during fine-tuning. In this way, the model is able to automatically determine which adapters to use, therefore hugely promoting the parameter efficiency and decoding speed. We evaluate the proposed framework on various neural machine translation tasks. Equipped with parallel sequence decoding, our model consistently outperforms autoregressive baselines while reducing the inference latency by half. With automatic adapter selection, the proposed model further achieves 20% speedup while still outperforming autoregressive baselines. When applied to autoregressive decoding, the proposed model can also achieve comparable performance with the state-of-the-art baseline models.

Non-autoregressive translation (NAT) models remove the dependence on previous target tokens and generate all target tokens in parallel, resulting in significant inference speedup but at the cost of inferior translation accuracy compared to autoregressive translation (AT) models. Considering that AT models have higher accuracy and are easier to train than NAT models, and both of them share the same model configurations, a natural idea to improve the accuracy of NAT models is to transfer a well-trained AT model to an NAT model through fine-tuning. However, since AT and NAT models differ greatly in training strategy, straightforward fine-tuning does not work well. In this work, we introduce curriculum learning into fine-tuning for NAT. Specifically, we design a curriculum in the fine-tuning process to progressively switch the training from autoregressive generation to non-autoregressive generation. Experiments on four benchmark translation datasets show that the proposed method achieves good improvement (more than 1 BLEU score) over previous NAT baselines in terms of translation accuracy, and greatly speed up (more than 10 times) the inference process over AT baselines.

We propose a novel introspective model for variational neural machine translation (IntroVNMT) in this paper, inspired by the recent successful application of introspective variational autoencoder (IntroVAE) in high quality image synthesis. Different from the vanilla variational NMT model, IntroVNMT is capable of improving itself introspectively by evaluating the quality of the generated target sentences according to the high-level latent variables of the real and generated target sentences. As a consequence of introspective training, the proposed model is able to discriminate between the generated and real sentences of the target language via the latent variables generated by the encoder of the model. In this way, IntroVNMT is able to generate more realistic target sentences in practice. In the meantime, IntroVNMT inherits the advantages of the variational autoencoders (VAEs), and the model training process is more stable than the generative adversarial network (GAN) based models. Experimental results on different translation tasks demonstrate that the proposed model can achieve significant improvements over the vanilla variational NMT model.

The masked language model has received remarkable attention due to its effectiveness on various natural language processing tasks. However, few works have adopted this technique in the sequence-to-sequence models. In this work, we introduce a jointly masked sequence-to-sequence model and explore its application on non-autoregressive neural machine translation~(NAT). Specifically, we first empirically study the functionalities of the encoder and the decoder in NAT models, and find that the encoder takes a more important role than the decoder regarding the translation quality. Therefore, we propose to train the encoder more rigorously by masking the encoder input while training. As for the decoder, we propose to train it based on the consecutive masking of the decoder input with an n-gram loss function to alleviate the problem of translating duplicate words. The two types of masks are applied to the model jointly at the training stage. We conduct experiments on five benchmark machine translation tasks, and our model can achieve 27.69/32.24 BLEU scores on WMT14 English-German/German-English tasks with 5+ times speed up compared with an autoregressive model.

Graph Neural Networks (GNNs) have achieved great success on graph-structured data, and their applications on traditional data structures such as natural language processing and semi-supervised text classification have been extensively explored in recent years. While previous works only consider the text information while building the graph, heterogeneous information such as labels is ignored. In this paper, we consider incorporating the label information while building the graph by adding text-label-text paths, through which the supervision information will propagate among the graph more directly. Specifically, we treat labels as nodes in the graph which also contains text and word nodes, and then connect labels with texts belonging to that label. Through graph convolutions, label embeddings are jointly learned with text embeddings in the same latent semantic space. The newly incorporated label nodes will facilitate learning more accurate text embeddings by introducing the label information, and thus benefit the downstream text classification tasks. Extensive results on several benchmark datasets show that the proposed framework outperforms baseline methods by a significant margin.

While large scale pre-trained language models such as BERT have achieved great success on various natural language understanding tasks, how to efficiently and effectively incorporate them into sequence-to-sequence models and the corresponding text generation tasks remains a non-trivial problem. In this paper, we propose to address this problem by taking two different BERT models as the encoder and decoder respectively, and fine-tuning them by introducing simple and lightweight adapter modules, which are inserted between BERT layers and tuned on the task-specific dataset. In this way, we obtain a flexible and efficient model which is able to jointly leverage the information contained in the source-side and target-side BERT models, while bypassing the catastrophic forgetting problem. Each component in the framework can be considered as a plug-in unit, making the framework flexible and task agnostic. Our framework is based on a parallel sequence decoding algorithm named Mask-Predict considering the bi-directional and conditional independent nature of BERT, and can be adapted to traditional autoregressive decoding easily. We conduct extensive experiments on neural machine translation tasks where the proposed method consistently outperforms autoregressive baselines while reducing the inference latency by half, and achieves $36.49$/$33.57$ BLEU scores on IWSLT14 German-English/WMT14 German-English translation. When adapted to autoregressive decoding, the proposed method achieves $30.60$/$43.56$ BLEU scores on WMT14 English-German/English-French translation, on par with the state-of-the-art baseline models.

Non-autoregressive translation (NAT) models, which remove the dependence on previous target tokens from the inputs of the decoder, achieve significantly inference speedup but at the cost of inferior accuracy compared to autoregressive translation (AT) models. Previous work shows that the quality of the inputs of the decoder is important and largely impacts the model accuracy. In this paper, we propose two methods to enhance the decoder inputs so as to improve NAT models. The first one directly leverages a phrase table generated by conventional SMT approaches to translate source tokens to target tokens, which are then fed into the decoder as inputs. The second one transforms source-side word embeddings to target-side word embeddings through sentence-level alignment and word-level adversary learning, and then feeds the transformed word embeddings into the decoder as inputs. Experimental results show our method largely outperforms the NAT baseline (Gu et al. 2017) by 5.11 BLEU scores on WMT14 English-German task and 4.72 BLEU scores on WMT16 English-Romanian task.

Recent advances in the field of network embedding have shown that low-dimensional network representation is playing a critical role in network analysis. Most existing network embedding methods encode the local proximity of a node, such as the first- and second-order proximities. While being efficient, these methods are short of leveraging the global structural information between nodes distant from each other. In addition, most existing methods learn embeddings on one single fixed network, and thus cannot be generalized to unseen nodes or networks without retraining. In this paper we present SPINE, a method that can jointly capture the local proximity and proximities at any distance, while being inductive to efficiently deal with unseen nodes or networks. Extensive experimental results on benchmark datasets demonstrate the superiority of the proposed framework over the state of the art.

Crowdfunding is an emerging online fundraising mechanism for creators to launch campaigns (projects) to solicit funds or expand their influence. Tracking the dynamics, i.e., daily funding amounts can be of great help to campaign creators as well as contributors. Previous works on this subject either fit the fluctuations of time-series with predefined stochastic process or apply a regularization term to constrain learned tendencies, resulting in limited generalization abilities. Patterns of funding-amount sequences in crowdfunding are often exclusive and non-linear, making previous predictors suboptimal. To tackle this problem, we propose a novel method based on synthesized bases which can be composed into arbitrary patterns. Concretely, we build a large set of candidate basis from which we select based on reliability, diversity and latent structures. We use representations of sequences in this basis space as a predictor, and adopt a dual-graph to exploit neighbouring information to enhance its prediction quality. Experimental results demonstrate the effectiveness of our method.