Leveraging vast training data, multimodal large language models (MLLMs) have demonstrated formidable general visual comprehension capabilities and achieved remarkable performance across various tasks. However, their performance in visual document understanding still leaves much room for improvement. This discrepancy is primarily attributed to the fact that visual document understanding is a fine-grained prediction task. In natural scenes, MLLMs typically use low-resolution images, leading to a substantial loss of visual information. Furthermore, general-purpose MLLMs do not excel in handling document-oriented instructions. In this paper, we propose a High-Resolution Visual Document Assistant (HRVDA), which bridges the gap between MLLMs and visual document understanding. This model employs a content filtering mechanism and an instruction filtering module to separately filter out the content-agnostic visual tokens and instruction-agnostic visual tokens, thereby achieving efficient model training and inference for high-resolution images. In addition, we construct a document-oriented visual instruction tuning dataset and apply a multi-stage training strategy to enhance the model's document modeling capabilities. Extensive experiments demonstrate that our model achieves state-of-the-art performance across multiple document understanding datasets, while maintaining training efficiency and inference speed comparable to low-resolution models.
Gradient compression with error compensation has attracted significant attention with the target of reducing the heavy communication overhead in distributed learning. However, existing compression methods either perform only unidirectional compression in one iteration with higher communication cost, or bidirectional compression with slower convergence rate. In this work, we propose the Local Immediate Error Compensated SGD (LIEC-SGD) optimization algorithm to break the above bottlenecks based on bidirectional compression and carefully designed compensation approaches. Specifically, the bidirectional compression technique is to reduce the communication cost, and the compensation technique compensates the local compression error to the model update immediately while only maintaining the global error variable on the server throughout the iterations to boost its efficacy. Theoretically, we prove that LIEC-SGD is superior to previous works in either the convergence rate or the communication cost, which indicates that LIEC-SGD could inherit the dual advantages from unidirectional compression and bidirectional compression. Finally, experiments of training deep neural networks validate the effectiveness of the proposed LIEC-SGD algorithm.
To protect privacy and meet legal regulations, federated learning (FL) has gained significant attention for training speech-to-text (S2T) systems, including automatic speech recognition (ASR) and speech translation (ST). However, the commonly used FL approach (i.e., \textsc{FedAvg}) in S2T tasks typically suffers from extensive communication overhead due to multi-round interactions based on the whole model and performance degradation caused by data heterogeneity among clients.To address these issues, we propose a personalized federated S2T framework that introduces \textsc{FedLoRA}, a lightweight LoRA module for client-side tuning and interaction with the server to minimize communication overhead, and \textsc{FedMem}, a global model equipped with a $k$-nearest-neighbor ($k$NN) classifier that captures client-specific distributional shifts to achieve personalization and overcome data heterogeneity. Extensive experiments based on Conformer and Whisper backbone models on CoVoST and GigaSpeech benchmarks show that our approach significantly reduces the communication overhead on all S2T tasks and effectively personalizes the global model to overcome data heterogeneity.
While Diffusion Generative Models have achieved great success on image generation tasks, how to efficiently and effectively incorporate them into speech generation especially translation tasks remains a non-trivial problem. Specifically, due to the low information density of speech data, the transformed discrete speech unit sequence is much longer than the corresponding text transcription, posing significant challenges to existing auto-regressive models. Furthermore, it is not optimal to brutally apply discrete diffusion on the speech unit sequence while disregarding the continuous space structure, which will degrade the generation performance significantly. In this paper, we propose a novel diffusion model by applying the diffusion forward process in the \textit{continuous} speech representation space, while employing the diffusion backward process in the \textit{discrete} speech unit space. In this way, we preserve the semantic structure of the continuous speech representation space in the diffusion process and integrate the continuous and discrete diffusion models. We conduct extensive experiments on the textless direct speech-to-speech translation task, where the proposed method achieves comparable results to the computationally intensive auto-regressive baselines (500 steps on average) with significantly fewer decoding steps (50 steps).
Multimodal entity linking (MEL) task, which aims at resolving ambiguous mentions to a multimodal knowledge graph, has attracted wide attention in recent years. Though large efforts have been made to explore the complementary effect among multiple modalities, however, they may fail to fully absorb the comprehensive expression of abbreviated textual context and implicit visual indication. Even worse, the inevitable noisy data may cause inconsistency of different modalities during the learning process, which severely degenerates the performance. To address the above issues, in this paper, we propose a novel Multi-GraIned Multimodal InteraCtion Network $\textbf{(MIMIC)}$ framework for solving the MEL task. Specifically, the unified inputs of mentions and entities are first encoded by textual/visual encoders separately, to extract global descriptive features and local detailed features. Then, to derive the similarity matching score for each mention-entity pair, we device three interaction units to comprehensively explore the intra-modal interaction and inter-modal fusion among features of entities and mentions. In particular, three modules, namely the Text-based Global-Local interaction Unit (TGLU), Vision-based DuaL interaction Unit (VDLU) and Cross-Modal Fusion-based interaction Unit (CMFU) are designed to capture and integrate the fine-grained representation lying in abbreviated text and implicit visual cues. Afterwards, we introduce a unit-consistency objective function via contrastive learning to avoid inconsistency and model degradation. Experimental results on three public benchmark datasets demonstrate that our solution outperforms various state-of-the-art baselines, and ablation studies verify the effectiveness of designed modules.
Pronunciation assessment is a major challenge in the computer-aided pronunciation training system, especially at the word (phoneme)-level. To obtain word (phoneme)-level scores, current methods usually rely on aligning components to obtain acoustic features of each word (phoneme), which limits the performance of assessment to the accuracy of alignments. Therefore, to address this problem, we propose a simple yet effective method, namely \underline{M}asked pre-training for \underline{P}ronunciation \underline{A}ssessment (MPA). Specifically, by incorporating a mask-predict strategy, our MPA supports end-to-end training without leveraging any aligning components and can solve misalignment issues to a large extent during prediction. Furthermore, we design two evaluation strategies to enable our model to conduct assessments in both unsupervised and supervised settings. Experimental results on SpeechOcean762 dataset demonstrate that MPA could achieve better performance than previous methods, without any explicit alignment. In spite of this, MPA still has some limitations, such as requiring more inference time and reference text. They expect to be addressed in future work.
In this paper, we propose a novel multi-modal framework for Scene Text Visual Question Answering (STVQA), which requires models to read scene text in images for question answering. Apart from text or visual objects, which could exist independently, scene text naturally links text and visual modalities together by conveying linguistic semantics while being a visual object in an image simultaneously. Different to conventional STVQA models which take the linguistic semantics and visual semantics in scene text as two separate features, in this paper, we propose a paradigm of "Locate Then Generate" (LTG), which explicitly unifies this two semantics with the spatial bounding box as a bridge connecting them. Specifically, at first, LTG locates the region in an image that may contain the answer words with an answer location module (ALM) consisting of a region proposal network and a language refinement network, both of which can transform to each other with one-to-one mapping via the scene text bounding box. Next, given the answer words selected by ALM, LTG generates a readable answer sequence with an answer generation module (AGM) based on a pre-trained language model. As a benefit of the explicit alignment of the visual and linguistic semantics, even without any scene text based pre-training tasks, LTG can boost the absolute accuracy by +6.06% and +6.92% on the TextVQA dataset and the ST-VQA dataset respectively, compared with a non-pre-training baseline. We further demonstrate that LTG effectively unifies visual and text modalities through the spatial bounding box connection, which is underappreciated in previous methods.
Diffusion models have achieved state-of-the-art synthesis quality on visual and audio tasks, and recent works adapt them to textual data by diffusing on the embedding space. But the difference between the continuous data space and the embedding space raises challenges to the diffusion model, which have not been carefully explored. In this paper, we conduct systematic studies and analyze the challenges threefold. Firstly, the data distribution is learnable for embeddings, which may lead to the collapse of the loss function. Secondly, as the norm of embedding varies between popular and rare words, adding the same noise scale will lead to sub-optimal results. In addition, we find that noises sampled from a standard Gaussian distribution may distract the diffusion process. To solve the above challenges, we propose Difformer, a denoising diffusion probabilistic model based on Transformer, which consists of three techniques including utilizing an anchor loss function, a layer normalization module for embeddings, and a norm factor to the Gaussian noise. All techniques are complementary to each other and critical to boosting the model performance together. Experiments are conducted on benchmark datasets over two seminal text generation tasks including machine translation and text summarization. The results show that Difformer significantly outperforms the embedding diffusion baselines, while achieving competitive results with strong autoregressive baselines.
We consider a novel task of automatically generating text descriptions of music. Compared with other well-established text generation tasks such as image caption, the scarcity of well-paired music and text datasets makes it a much more challenging task. In this paper, we exploit the crowd-sourced music comments to construct a new dataset and propose a sequence-to-sequence model to generate text descriptions of music. More concretely, we use the dilated convolutional layer as the basic component of the encoder and a memory based recurrent neural network as the decoder. To enhance the authenticity and thematicity of generated texts, we further propose to fine-tune the model with a discriminator as well as a novel topic evaluator. To measure the quality of generated texts, we also propose two new evaluation metrics, which are more aligned with human evaluation than traditional metrics such as BLEU. Experimental results verify that our model is capable of generating fluent and meaningful comments while containing thematic and content information of the original music.
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.