Synthesizing high-quality, realistic images from text-descriptions is a challenging task, and current methods synthesize images from text in a multi-stage manner, typically by first generating a rough initial image and then refining image details at subsequent stages. However, existing methods that follow this paradigm suffer from three important limitations. Firstly, they synthesize initial images without attempting to separate image attributes at a word-level. As a result, object attributes of initial images (that provide a basis for subsequent refinement) are inherently entangled and ambiguous in nature. Secondly, by using common text-representations for all regions, current methods prevent us from interpreting text in fundamentally different ways at different parts of an image. Different image regions are therefore only allowed to assimilate the same type of information from text at each refinement stage. Finally, current methods generate refinement features only once at each refinement stage and attempt to address all image aspects in a single shot. This single-shot refinement limits the precision with which each refinement stage can learn to improve the prior image. Our proposed method introduces three novel components to address these shortcomings: (1) An initial generation stage that explicitly generates separate sets of image features for each word n-gram. (2) A spatial dynamic memory module for refinement of images. (3) An iterative multi-headed mechanism to make it easier to improve upon multiple image aspects. Experimental results demonstrate that our Multi-Headed Spatial Dynamic Memory image refinement with our Multi-Tailed Word-level Initial Generation (MSMT-GAN) performs favourably against the previous state of the art on the CUB and COCO datasets.
Deep learning has been widely used in medical image segmentation and other aspects. However, the performance of existing medical image segmentation models has been limited by the challenge of obtaining sufficient number of high-quality data with the high cost of data annotation. To overcome the limitation, we propose a new vision-language medical image segmentation model LViT (Language meets Vision Transformer). In our model, medical text annotation is introduced to compensate for the quality deficiency in image data. In addition, the text information can guide the generation of pseudo labels to a certain extent and further guarantee the quality of pseudo labels in semi-supervised learning. We also propose the Exponential Pseudo label Iteration mechanism (EPI) to help extend the semi-supervised version of LViT and the Pixel-Level Attention Module (PLAM) to preserve local features of images. In our model, LV (Language-Vision) loss is designed to supervise the training of unlabeled images using text information directly. To validate the performance of LViT, we construct multimodal medical segmentation datasets (image + text) containing pathological images, X-rays,etc. Experimental results show that our proposed LViT has better segmentation performance in both fully and semi-supervised conditions. Code and datasets are available at https://github.com/HUANGLIZI/LViT.
Scene text recognition is a challenging task due to diverse variations of text instances in natural scene images. Conventional methods based on CNN-RNN-CTC or encoder-decoder with attention mechanism may not fully investigate stable and efficient feature representations for multi-oriented scene texts. In this paper, we propose a primitive representation learning method that aims to exploit intrinsic representations of scene text images. We model elements in feature maps as the nodes of an undirected graph. A pooling aggregator and a weighted aggregator are proposed to learn primitive representations, which are transformed into high-level visual text representations by graph convolutional networks. A Primitive REpresentation learning Network (PREN) is constructed to use the visual text representations for parallel decoding. Furthermore, by integrating visual text representations into an encoder-decoder model with the 2D attention mechanism, we propose a framework called PREN2D to alleviate the misalignment problem in attention-based methods. Experimental results on both English and Chinese scene text recognition tasks demonstrate that PREN keeps a balance between accuracy and efficiency, while PREN2D achieves state-of-the-art performance.
In this paper, we propose a data augmentation framework for Optical Character Recognition (OCR). The proposed framework is able to synthesize new viewing angles and illumination scenarios, effectively enriching any available OCR dataset. Its modular structure allows to be modified to match individual user requirements. The framework enables to comfortably scale the enlargement factor of the available dataset. Furthermore, the proposed method is not restricted to single frame OCR but can also be applied to video OCR. We demonstrate the performance of our framework by augmenting a 15% subset of the common Brno Mobile OCR dataset. Our proposed framework is capable of leveraging the performance of OCR applications especially for small datasets. Applying the proposed method, improvements of up to 2.79 percentage points in terms of Character Error Rate (CER), and up to 7.88 percentage points in terms of Word Error Rate (WER) are achieved on the subset. Especially the recognition of challenging text lines can be improved. The CER may be decreased by up to 14.92 percentage points and the WER by up to 18.19 percentage points for this class. Moreover, we are able to achieve smaller error rates when training on the 15% subset augmented with the proposed method than on the original non-augmented full dataset.
Modern neural text generation systems can produce remarkably fluent and grammatical texts. While earlier language models suffered from repetition and syntactic errors, the errors made by contemporary models are often semantic, narrative, or discourse failures. To facilitate research of these complex error types, we introduce a new structured, crowdsourced error annotation schema called Scarecrow. The error categories used in Scarecrow -- such as redundancy, commonsense errors, and incoherence -- were identified by combining expert analysis with several pilot rounds of ontology-free crowd annotation to arrive at a schema which covers the error phenomena found in real machine generated text. We use Scarecrow to collect 13k annotations of 1.3k human and machine generate paragraphs of English language news text, amounting to over 41k spans each labeled with its error category, severity, a natural language explanation, and antecedent span (where relevant). We collect annotations for text generated by state-of-the-art systems with varying known performance levels, from GPT-2 Small through the largest GPT-3. We isolate several factors for detailed analysis, including parameter count, training data, and decoding technique. Our results show both expected and surprising differences across these settings. These findings demonstrate the value of Scarecrow annotations in the assessment of current and future text generation systems. We release our complete annotation toolkit and dataset at https://yao-dou.github.io/scarecrow/.
We propose a new learning framework that captures the tiered structure of many real-world user-interaction applications, where the users can be divided into two groups based on their different tolerance on exploration risks and should be treated separately. In this setting, we simultaneously maintain two policies $\pi^{\text{O}}$ and $\pi^{\text{E}}$: $\pi^{\text{O}}$ ("O" for "online") interacts with more risk-tolerant users from the first tier and minimizes regret by balancing exploration and exploitation as usual, while $\pi^{\text{E}}$ ("E" for "exploit") exclusively focuses on exploitation for risk-averse users from the second tier utilizing the data collected so far. An important question is whether such a separation yields advantages over the standard online setting (i.e., $\pi^{\text{E}}=\pi^{\text{O}}$) for the risk-averse users. We individually consider the gap-independent vs.~gap-dependent settings. For the former, we prove that the separation is indeed not beneficial from a minimax perspective. For the latter, we show that if choosing Pessimistic Value Iteration as the exploitation algorithm to produce $\pi^{\text{E}}$, we can achieve a constant regret for risk-averse users independent of the number of episodes $K$, which is in sharp contrast to the $\Omega(\log K)$ regret for any online RL algorithms in the same setting, while the regret of $\pi^{\text{O}}$ (almost) maintains its online regret optimality and does not need to compromise for the success of $\pi^{\text{E}}$.
We investigate models that can generate arbitrary natural language text (e.g. all English sentences) from a bounded, convex and well-behaved control space. We call them universal vec2text models. Such models would allow making semantic decisions in the vector space (e.g. via reinforcement learning) while the natural language generation is handled by the vec2text model. We propose four desired properties: universality, diversity, fluency, and semantic structure, that such vec2text models should possess and we provide quantitative and qualitative methods to assess them. We implement a vec2text model by adding a bottleneck to a 250M parameters Transformer model and training it with an auto-encoding objective on 400M sentences (10B tokens) extracted from a massive web corpus. We propose a simple data augmentation technique based on round-trip translations and show in extensive experiments that the resulting vec2text model surprisingly leads to vector spaces that fulfill our four desired properties and that this model strongly outperforms both standard and denoising auto-encoders.
A well-known challenge associated with the multi-label classification problem is modelling dependencies between labels. Most attempts at modelling label dependencies focus on co-occurrences, ignoring the valuable information that can be extracted by detecting label subsets that rarely occur together. For example, consider customer product reviews; a product probably would not simultaneously be tagged by both "recommended" (i.e., reviewer is happy and recommends the product) and "urgent" (i.e., the review suggests immediate action to remedy an unsatisfactory experience). Aside from the consideration of positive and negative dependencies, the direction of a relationship should also be considered. For a multi-label image classification problem, the "ship" and "sea" labels have an obvious dependency, but the presence of the former implies the latter much more strongly than the other way around. These examples motivate the modelling of multiple types of bi-directional relationships between labels. In this paper, we propose a novel method, entitled Multi-relation Message Passing (MrMP), for the multi-label classification problem. Experiments on benchmark multi-label text classification datasets show that the MrMP module yields similar or superior performance compared to state-of-the-art methods. The approach imposes only minor additional computational and memory overheads.
Data-to-text (D2T) generation in the biomedical domain is a promising - yet mostly unexplored - field of research. Here, we apply neural models for D2T generation to a real-world dataset consisting of package leaflets of European medicines. We show that fine-tuned transformers are able to generate realistic, multisentence text from data in the biomedical domain, yet have important limitations. We also release a new dataset (BioLeaflets) for benchmarking D2T generation models in the biomedical domain.
In the last few years, the ML community has created a number of new NLP models based on transformer architecture. These models have shown great performance for various NLP tasks on benchmark datasets, often surpassing SOTA results. Buoyed with this success, one often finds industry practitioners actively experimenting with fine-tuning these models to build NLP applications for industry use cases. However, for most datasets that are used by practitioners to build industrial NLP applications, it is hard to guarantee the presence of any noise in the data. While most transformer based NLP models have performed exceedingly well in transferring the learnings from one dataset to another, it remains unclear how these models perform when fine-tuned on noisy text. We address the open question by Kumar et al. (2020) to explore the sensitivity of popular transformer based NLP models to noise in the text data. We continue working with the noise as defined by them -- spelling mistakes & typos (which are the most commonly occurring noise). We show (via experimental results) that these models perform badly on most common NLP tasks namely text classification, textual similarity, NER, question answering, text summarization on benchmark datasets. We further show that as the noise in data increases, the performance degrades. Our findings suggest that one must be vary of the presence of noise in their datasets while fine-tuning popular transformer based NLP models.