Existing vision-language pre-training (VLP) methods primarily rely on paired image-text datasets, which are either annotated by enormous human labors, or crawled from the internet followed by elaborate data cleaning techniques. To reduce the dependency on well-aligned image-text pairs, it is promising to directly leverage the large-scale text-only and image-only corpora. This paper proposes a data augmentation method, namely cross-modal CutMix (CMC), for implicit cross-modal alignment learning in unpaired VLP. Specifically, CMC transforms natural sentences from the textual view into a multi-modal view, where visually-grounded words in a sentence are randomly replaced by diverse image patches with similar semantics. There are several appealing proprieties of the proposed CMC. First, it enhances the data diversity while keeping the semantic meaning intact for tackling problems where the aligned data are scarce; Second, by attaching cross-modal noise on uni-modal data, it guides models to learn token-level interactions across modalities for better denoising. Furthermore, we present a new unpaired VLP method, dubbed as VLMixer, that integrates CMC with contrastive learning to pull together the uni-modal and multi-modal views for better instance-level alignments among different modalities. Extensive experiments on five downstream tasks show that VLMixer could surpass previous state-of-the-art unpaired VLP methods.
Despite advances in generating fluent texts, existing pretraining models tend to attach incoherent event sequences to involved entities when generating narratives such as stories and news. We conjecture that such issues result from representing entities as static embeddings of superficial words, while neglecting to model their ever-changing states, i.e., the information they carry, as the text unfolds. Therefore, we extend the Transformer model to dynamically conduct entity state updates and sentence realization for narrative generation. We propose a contrastive framework to learn the state representations in a discrete space, and insert additional attention layers into the decoder to better exploit these states. Experiments on two narrative datasets show that our model can generate more coherent and diverse narratives than strong baselines with the guidance of meaningful entity states.
A prevailing paradigm in neural text generation is one-shot generation, where text is produced in a single step. The one-shot setting is inadequate, however, when the constraints the user wishes to impose on the generated text are dynamic, especially when authoring longer documents. We address this limitation with an interactive text generation setting in which the user interacts with the system by issuing commands to edit existing text. To this end, we propose a novel text editing task, and introduce WikiDocEdits, a dataset of single-sentence edits crawled from Wikipedia. We show that our Interactive Editor, a transformer-based model trained on this dataset, outperforms baselines and obtains positive results in both automatic and human evaluations. We present empirical and qualitative analyses of this model's performance.
Recently, neural topic models (NTMs) have been incorporated into pre-trained language models (PLMs), to capture the global semantic information for text summarization. However, in these methods, there remain limitations in the way they capture and integrate the global semantic information. In this paper, we propose a novel model, the graph contrastive topic enhanced language model (GRETEL), that incorporates the graph contrastive topic model with the pre-trained language model, to fully leverage both the global and local contextual semantics for long document extractive summarization. To better capture and incorporate the global semantic information into PLMs, the graph contrastive topic model integrates the hierarchical transformer encoder and the graph contrastive learning to fuse the semantic information from the global document context and the gold summary. To this end, GRETEL encourages the model to efficiently extract salient sentences that are topically related to the gold summary, rather than redundant sentences that cover sub-optimal topics. Experimental results on both general domain and biomedical datasets demonstrate that our proposed method outperforms SOTA methods.
Language grounding to vision is an active field of research aiming to enrich text-based representations of word meanings by leveraging perceptual knowledge from vision. Despite many attempts at language grounding, it is still unclear how to effectively inject visual knowledge into the word embeddings of a language in such a way that a proper balance of textual and visual knowledge is maintained. Some common concerns are the following. Is visual grounding beneficial for abstract words or is its contribution only limited to concrete words? What is the optimal way of bridging the gap between text and vision? How much do we gain by visually grounding textual embeddings? The present study addresses these questions by proposing a simple yet very effective grounding approach for pre-trained word embeddings. Our model aligns textual embeddings with vision while largely preserving the distributional statistics that characterize word use in text corpora. By applying a learned alignment, we are able to generate visually grounded embeddings for unseen words, including abstract words. A series of evaluations on word similarity benchmarks shows that visual grounding is beneficial not only for concrete words, but also for abstract words. We also show that our method for visual grounding offers advantages for contextualized embeddings, but only when these are trained on corpora of relatively modest size. Code and grounded embeddings for English are available at https://github.com/Hazel1994/Visually_Grounded_Word_Embeddings_2.
The goal of this work is to build flexible video-language models that can generalize to various video-to-text tasks from few examples, such as domain-specific captioning, question answering, and future event prediction. Existing few-shot video-language learners focus exclusively on the encoder, resulting in the absence of a video-to-text decoder to handle generative tasks. Video captioners have been pretrained on large-scale video-language datasets, but they rely heavily on finetuning and lack the ability to generate text for unseen tasks in a few-shot setting. We propose VidIL, a few-shot Video-language Learner via Image and Language models, which demonstrates strong performance on few-shot video-to-text tasks without the necessity of pretraining or finetuning on any video datasets. We use the image-language models to translate the video content into frame captions, object, attribute, and event phrases, and compose them into a temporal structure template. We then instruct a language model, with a prompt containing a few in-context examples, to generate a target output from the composed content. The flexibility of prompting allows the model to capture any form of text input, such as automatic speech recognition (ASR) transcripts. Our experiments demonstrate the power of language models in understanding videos on a wide variety of video-language tasks, including video captioning, video question answering, video caption retrieval, and video future event prediction. Especially, on video future event prediction, our few-shot model significantly outperforms state-of-the-art supervised models trained on large-scale video datasets. Code and resources are publicly available for research purposes at https://github.com/MikeWangWZHL/VidIL .
An important preliminary step of optical character recognition systems is the detection of text rows. To address this task in the context of historical data with missing labels, we propose a self-paced learning algorithm capable of improving the row detection performance. We conjecture that pages with more ground-truth bounding boxes are less likely to have missing annotations. Based on this hypothesis, we sort the training examples in descending order with respect to the number of ground-truth bounding boxes, and organize them into k batches. Using our self-paced learning method, we train a row detector over k iterations, progressively adding batches with less ground-truth annotations. At each iteration, we combine the ground-truth bounding boxes with pseudo-bounding boxes (bounding boxes predicted by the model itself) using non-maximum suppression, and we include the resulting annotations at the next training iteration. We demonstrate that our self-paced learning strategy brings significant performance gains on two data sets of historical documents, improving the average precision of YOLOv4 with more than 12% on one data set and 39% on the other.
The task of cross-modal retrieval between texts and videos aims to understand the correspondence between vision and language. Existing studies follow a trend of measuring text-video similarity on the basis of textual and video embeddings. In common practice, video representation is constructed by feeding video frames into 2D/3D-CNN for global visual feature extraction or only learning simple semantic relations by using local-level fine-grained frame regions via graph convolutional network. However, these video representations do not fully exploit spatio-temporal relation among visual components in learning video representations, resulting in their inability to distinguish videos with the same visual components but with different relations. To solve this problem, we propose a Visual Spatio-temporal Relation-enhanced Network (VSR-Net), a novel cross-modal retrieval framework that enhances visual representation with spatio-temporal relations among components. Specifically, visual spatio-temporal relations are encoded using a multi-layer spatio-temporal transformer to learn visual relational features. We combine fine-grained local relation and global features in bridging text-video modalities. Extensive experimental are conducted on both MSR-VTT and MSVD datasets. The results demonstrate the effectiveness of our proposed model.
In cross-lingual speech synthesis, the speech in various languages can be synthesized for a monoglot speaker. Normally, only the data of monoglot speakers are available for model training, thus the speaker similarity is relatively low between the synthesized cross-lingual speech and the native language recordings. Based on the multilingual transformer text-to-speech model, this paper studies a multi-task learning framework to improve the cross-lingual speaker similarity. To further improve the speaker similarity, joint training with a speaker classifier is proposed. Here, a scheme similar to parallel scheduled sampling is proposed to train the transformer model efficiently to avoid breaking the parallel training mechanism when introducing joint training. By using multi-task learning and speaker classifier joint training, in subjective and objective evaluations, the cross-lingual speaker similarity can be consistently improved for both the seen and unseen speakers in the training set.
Building scalable models to learn from diverse, multimodal data remains an open challenge. For vision-language data, the dominant approaches are based on contrastive learning objectives that train a separate encoder for each modality. While effective, contrastive learning approaches introduce sampling bias depending on the data augmentations used, which can degrade performance on downstream tasks. Moreover, these methods are limited to paired image-text data, and cannot leverage widely-available unpaired data. In this paper, we investigate whether a large multimodal model trained purely via masked token prediction, without using modality-specific encoders or contrastive learning, can learn transferable representations for downstream tasks. We propose a simple and scalable network architecture, the Multimodal Masked Autoencoder (M3AE), which learns a unified encoder for both vision and language data via masked token prediction. We provide an empirical study of M3AE trained on a large-scale image-text dataset, and find that M3AE is able to learn generalizable representations that transfer well to downstream tasks. Surprisingly, we find that M3AE benefits from a higher text mask ratio (50-90%), in contrast to BERT whose standard masking ratio is 15%, due to the joint training of two data modalities. We also provide qualitative analysis showing that the learned representation incorporates meaningful information from both image and language. Lastly, we demonstrate the scalability of M3AE with larger model size and training time, and its flexibility to train on both paired image-text data as well as unpaired data.