We present a method for introducing a text encoder into pre-trained end-to-end speech translation systems. It enhances the ability of adapting one modality (i.e., source-language speech) to another (i.e., source-language text). Thus, the speech translation model can learn from both unlabeled and labeled data, especially when the source-language text data is abundant. Beyond this, we present a denoising method to build a robust text encoder that can deal with both normal and noisy text data. Our system sets new state-of-the-arts on the MuST-C En-De, En-Fr, and LibriSpeech En-Fr tasks.
Image captioning, an important vision-language task, often requires a tremendous number of finely labeled image-caption pairs for learning the underlying alignment between images and texts. In this paper, we proposed a multimodal data augmentation method, leveraging a recent text-to-image model called Stable Diffusion, to expand the training set via high-quality generation of image-caption pairs. Extensive experiments on the MS COCO dataset demonstrate the advantages of our approach over several benchmark methods, and particularly a significant boost when having fewer training instances. In addition, models trained on our augmented datasets also outperform prior unpaired image captioning methods by a large margin. Finally, further improvement regarding the training efficiency and effectiveness can be obtained after intentionally filtering the generated data based on quality assessment.
Multimodal speech emotion recognition aims to detect speakers' emotions from audio and text. Prior works mainly focus on exploiting advanced networks to model and fuse different modality information to facilitate performance, while neglecting the effect of different fusion strategies on emotion recognition. In this work, we consider a simple yet important problem: how to fuse audio and text modality information is more helpful for this multimodal task. Further, we propose a multimodal emotion recognition model improved by perspective loss. Empirical results show our method obtained new state-of-the-art results on the IEMOCAP dataset. The in-depth analysis explains why the improved model can achieve improvements and outperforms baselines.
Fine-tuning a pre-trained model can leverage the semantic information from large-scale pre-training data and mitigate the over-fitting problem on downstream tasks with limited training examples. While the problem of catastrophic forgetting in backbone has been extensively studied, the potential bias existing in a pre-trained model due to the corresponding pre-training task and data, attracts less attention. In this work, we investigate this problem by demonstrating that the obtained classifier after fine-tuning will be close to that induced by the pre-trained model. To reduce the bias in the classifier effectively, we introduce a reference distribution obtained from a fixed text classifier, which can help regularize the learned vision classifier. The proposed method, Text Supervised fine-tuning (TeS), is evaluated with diverse pre-trained vision models including ResNet and ViT, and text encoders including BERT and CLIP, on 11 downstream tasks. The consistent improvement with a clear margin over distinct scenarios confirms the effectiveness of our proposal.
Visual question answering (VQA) is a challenging task to provide an accurate natural language answer given an image and a natural language question about the image. It involves multi-modal learning, i.e., computer vision (CV) and natural language processing (NLP), as well as flexible answer prediction for free-form and open-ended answers. Existing approaches often fail in cases that require reading and understanding text in images to answer questions. In practice, they cannot effectively handle the answer sequence derived from text tokens because the visual features are not text-oriented. To address the above issues, we propose a Text-Aware Dual Routing Network (TDR) which simultaneously handles the VQA cases with and without understanding text information in the input images. Specifically, we build a two-branch answer prediction network that contains a specific branch for each case and further develop a dual routing scheme to dynamically determine which branch should be chosen. In the branch that involves text understanding, we incorporate the Optical Character Recognition (OCR) features into the model to help understand the text in the images. Extensive experiments on the VQA v2.0 dataset demonstrate that our proposed TDR outperforms existing methods, especially on the ''number'' related VQA questions.
Text summarization has a wide range of applications in many scenarios. The evaluation of the quality of the generated text is a complex problem. A big challenge to language evaluation is that there is a clear divergence between existing metrics and human evaluation. For example, the quality of a document summary can be measured by human annotators from both objective aspects, such as grammatical and semantic correctness, as well as subjective dimensions, such as comprehensiveness, succinctness, and interestingness. Most of the automatic evaluation methods like BLUE/ROUGE may be not able to capture the above dimensions well. In this paper, we propose a new evaluation framework based on LLMs, which provides a comprehensive evaluation framework by comparing generated text and reference text from both objective and subjective aspects. First, we propose to model objective and subjective dimensions of generated text based on roleplayers prompting mechanism. Furthermore, we introduce a context-based prompting mechanism that is able to generate dynamic roleplayer profiles based on input context. Finally, we design a multi-roleplayer prompting technology based on batch prompting to integrate multiple evaluation results into evaluation results. Experimental results on two real datasets for summarization show that our model is highly competitive and has a very high consistency with human annotators.
Recent studies show that, despite being effective on numerous tasks, text processing algorithms may be vulnerable to deliberate attacks. However, the question of whether such weaknesses can directly lead to security threats is still under-explored. To bridge this gap, we conducted vulnerability tests on Text-to-SQL, a technique that builds natural language interfaces for databases. Empirically, we showed that the Text-to-SQL modules of two commercial black boxes (Baidu-UNIT and Codex-powered Ai2sql) can be manipulated to produce malicious code, potentially leading to data breaches and Denial of Service. This is the first demonstration of the danger of NLP models being exploited as attack vectors in the wild. Moreover, experiments involving four open-source frameworks verified that simple backdoor attacks can achieve a 100% success rate on Text-to-SQL systems with almost no prediction performance impact. By reporting these findings and suggesting practical defences, we call for immediate attention from the NLP community to the identification and remediation of software security issues.
Neural language models (LMs) represent facts about the world described by text. Sometimes these facts derive from training data (in most LMs, a representation of the word banana encodes the fact that bananas are fruits). Sometimes facts derive from input text itself (a representation of the sentence "I poured out the bottle" encodes the fact that the bottle became empty). Tools for inspecting and modifying LM fact representations would be useful almost everywhere LMs are used: making it possible to update them when the world changes, to localize and remove sources of bias, and to identify errors in generated text. We describe REMEDI, an approach for querying and modifying factual knowledge in LMs. REMEDI learns a map from textual queries to fact encodings in an LM's internal representation system. These encodings can be used as knowledge editors: by adding them to LM hidden representations, we can modify downstream generation to be consistent with new facts. REMEDI encodings can also be used as model probes: by comparing them to LM representations, we can ascertain what properties LMs attribute to mentioned entities, and predict when they will generate outputs that conflict with background knowledge or input text. REMEDI thus links work on probing, prompting, and model editing, and offers steps toward general tools for fine-grained inspection and control of knowledge in LMs.
Current video generation models usually convert signals indicating appearance and motion received from inputs (e.g., image, text) or latent spaces (e.g., noise vectors) into consecutive frames, fulfilling a stochastic generation process for the uncertainty introduced by latent code sampling. However, this generation pattern lacks deterministic constraints for both appearance and motion, leading to uncontrollable and undesirable outcomes. To this end, we propose a new task called Text-driven Video Prediction (TVP). Taking the first frame and text caption as inputs, this task aims to synthesize the following frames. Specifically, appearance and motion components are provided by the image and caption separately. The key to addressing the TVP task depends on fully exploring the underlying motion information in text descriptions, thus facilitating plausible video generation. In fact, this task is intrinsically a cause-and-effect problem, as the text content directly influences the motion changes of frames. To investigate the capability of text in causal inference for progressive motion information, our TVP framework contains a Text Inference Module (TIM), producing step-wise embeddings to regulate motion inference for subsequent frames. In particular, a refinement mechanism incorporating global motion semantics guarantees coherent generation. Extensive experiments are conducted on Something-Something V2 and Single Moving MNIST datasets. Experimental results demonstrate that our model achieves better results over other baselines, verifying the effectiveness of the proposed framework.
The goal of a speech-to-image transform is to produce a photo-realistic picture directly from a speech signal. Recently, various studies have focused on this task and have achieved promising performance. However, current speech-to-image approaches are based on a stacked modular framework that suffers from three vital issues: 1) Training separate networks is time-consuming as well as inefficient and the convergence of the final generative model strongly depends on the previous generators; 2) The quality of precursor images is ignored by this architecture; 3) Multiple discriminator networks are required to be trained. To this end, we propose an efficient and effective single-stage framework called Fusion-S2iGan to yield perceptually plausible and semantically consistent image samples on the basis of given spoken descriptions. Fusion-S2iGan introduces a visual+speech fusion module (VSFM), constructed with a pixel-attention module (PAM), a speech-modulation module (SMM) and a weighted-fusion module (WFM), to inject the speech embedding from a speech encoder into the generator while improving the quality of synthesized pictures. Fusion-S2iGan spreads the bimodal information over all layers of the generator network to reinforce the visual feature maps at various hierarchical levels in the architecture. We conduct a series of experiments on four benchmark data sets, i.e., CUB birds, Oxford-102, Flickr8k and Places-subset. The experimental results demonstrate the superiority of the presented Fusion-S2iGan compared to the state-of-the-art models with a multi-stage architecture and a performance level that is close to traditional text-to-image approaches.