Despite recent advances in the field of supervised deep learning for text line segmentation, unsupervised deep learning solutions are beginning to gain popularity. In this paper, we present an unsupervised deep learning method that embeds document image patches to a compact Euclidean space where distances correspond to a coarse text line pattern similarity. Once this space has been produced, text line segmentation can be easily implemented using standard techniques with the embedded feature vectors. To train the model, we extract random pairs of document image patches with the assumption that neighbour patches contain a similar coarse trend of text lines, whereas if one of them is rotated, they contain different coarse trends of text lines. Doing well on this task requires the model to learn to recognize the text lines and their salient parts. The benefit of our approach is zero manual labelling effort. We evaluate the method qualitatively and quantitatively on several variants of text line segmentation datasets to demonstrate its effectivity.
Finding online research papers relevant to one's interests is very challenging due to the increasing number of publications. Therefore, personalized research paper recommendation has become a significant and timely research topic. Collaborative filtering is a successful recommendation approach, which exploits the ratings given to items by users as a source of information for learning to make accurate recommendations. However, the ratings are often very sparse as in the research paper domain, due to the huge number of publications growing every year. Therefore, more attention has been drawn to hybrid methods that consider both ratings and content information. Nevertheless, most of the hybrid recommendation approaches that are based on text embedding have utilized bag-of-words techniques, which ignore word order and semantic meaning. In this paper, we propose a hybrid approach that leverages deep semantic representation of research papers based on social tags assigned by users. The experimental evaluation is performed on CiteULike, a real and publicly available dataset. The obtained findings show that the proposed model is effective in recommending research papers even when the rating data is very sparse.
Language models have proven to be very useful when adapted to specific domains. Nonetheless, little research has been done on the adaptation of domain-specific BERT models in the French language. In this paper, we focus on creating a language model adapted to French legal text with the goal of helping law professionals. We conclude that some specific tasks do not benefit from generic language models pre-trained on large amounts of data. We explore the use of smaller architectures in domain-specific sub-languages and their benefits for French legal text. We prove that domain-specific pre-trained models can perform better than their equivalent generalised ones in the legal domain. Finally, we release JuriBERT, a new set of BERT models adapted to the French legal domain.
Numerous scene text detection methods have been proposed in recent years. Most of them declare they have achieved state-of-the-art performances. However, the performance comparison is unfair, due to lots of inconsistent settings (e.g., training data, backbone network, multi-scale feature fusion, evaluation protocols, etc.). These various settings would dissemble the pros and cons of the proposed core techniques. In this paper, we carefully examine and analyze the inconsistent settings, and propose a unified framework for the bottom-up based scene text detection methods. Under the unified framework, we ensure the consistent settings for non-core modules, and mainly investigate the representations of describing arbitrary-shape scene texts, e.g., regressing points on text contours, clustering pixels with predicted auxiliary information, grouping connected components with learned linkages, etc. With the comprehensive investigations and elaborate analyses, it not only cleans up the obstacle of understanding the performance differences between existing methods but also reveals the advantages and disadvantages of previous models under fair comparisons.
Pre-training a large transformer model on a massive amount of unlabeled data and fine-tuning it on labeled datasets for diverse downstream tasks has proven to be a successful strategy, for a variety of vision and natural language processing tasks. However, direct fine-tuning of the pre-trained model may be suboptimal if there exist large discrepancies across data domains for pre-training and fine-tuning. To tackle this issue, several previous studies have proposed further pre-training strategies, where we continue to pre-train the model on the target unlabeled dataset before fine-tuning. However, all of them solely focus on language models and we empirically find that a Vision Transformer is vulnerable to overfitting as we continue to pretrain the model on target unlabeled data. In order to tackle this limitation, we propose self-distillation as a regularization for a further pre-training stage. Specifically, we first further pre-train the initial pre-trained model on the target unlabeled data and then consider it as a teacher for self-distillation. Then we take the same initial pre-trained model as a student and enforce its hidden representations to be close to those of the teacher while optimizing the student with a masked auto-encoding objective. We empirically validate the efficacy of self-distillation on a variety of benchmark datasets for image and text classification tasks. Experimentally, we show that our proposed method outperforms all the relevant baselines. Theoretically, we analyze the proposed method with a simplified model to understand how self-distillation for further pre-training can potentially help improve the performance of the downstream tasks.
In this work, we explore sources of inter-annotator disagreement in narrative analysis, in light of the question of whether or not a narrative plot exists in the text. For this purpose, we present a method for a conceptual decomposition of an existing annotation into two separate levels: (1) \textbf{whether} or not a narrative plot exists in the text, and (2) \textbf{which} plot elements exist in the text. We apply this method to an existing dataset of sentences annotated with three different narrative plot elements: \textit{Complication}, \textit{Resolution} and \textit{Success}. We then employ statistical analysis in order to quantify how much of the inter-annotator disagreement can be explained by each of the two levels. We further perform a qualitative analysis of disagreement cases in each level, observing several sources of disagreement, such as text ambiguity, scheme definition and personal differences between the annotators. The insights gathered on the dataset may serve to reduce inter-annotator disagreement in future annotation endeavors. We conclude with a broader discussion on the potential implications of our approach in studying and evaluating inter-annotator disagreement in other settings.
Prompt learning has been shown to achieve near-Fine-tune performance in most text classification tasks with very few training examples. It is advantageous for NLP tasks where samples are scarce. In this paper, we attempt to apply it to a practical scenario, i.e resume information extraction, and to enhance the existing method to make it more applicable to the resume information extraction task. In particular, we created multiple sets of manual templates and verbalizers based on the textual characteristics of resumes. In addition, we compared the performance of Masked Language Model (MLM) pre-training language models (PLMs) and Seq2Seq PLMs on this task. Furthermore, we improve the design method of verbalizer for Knowledgeable Prompt-tuning in order to provide a example for the design of Prompt templates and verbalizer for other application-based NLP tasks. In this case, we propose the concept of Manual Knowledgeable Verbalizer(MKV). A rule for constructing the Knowledgeable Verbalizer corresponding to the application scenario. Experiments demonstrate that templates and verbalizers designed based on our rules are more effective and robust than existing manual templates and automatically generated prompt methods. It is established that the currently available automatic prompt methods cannot compete with manually designed prompt templates for some realistic task scenarios. The results of the final confusion matrix indicate that our proposed MKV significantly resolved the sample imbalance issue.
Forms are a widespread type of template-based document used in a great variety of fields including, among others, administration, medicine, finance, or insurance. The automatic extraction of the information included in these documents is greatly demanded due to the increasing volume of forms that are generated in a daily basis. However, this is not a straightforward task when working with scanned forms because of the great diversity of templates with different location of form entities, and the quality of the scanned documents. In this context, there is a feature that is shared by all forms: they contain a collection of interlinked entities built as key-value (or label-value) pairs, together with other entities such as headers or images. In this work, we have tacked the problem of entity linking in forms by combining image processing techniques and a text classification model based on the BERT architecture. This approach achieves state-of-the-art results with a F1-score of 0.80 on the FUNSD dataset, a 5% improvement regarding the best previous method. The code of this project is available at https://github.com/mavillot/FUNSD-Entity-Linking.
We present a new approach to perform zero-shot cross-modal transfer between speech and text for translation tasks. Multilingual speech and text are encoded in a joint fixed-size representation space. Then, we compare different approaches to decode these multimodal and multilingual fixed-size representations, enabling zero-shot translation between languages and modalities. All our models are trained without the need of cross-modal labeled translation data. Despite a fixed-size representation, we achieve very competitive results on several text and speech translation tasks. In particular, we significantly improve the state-of-the-art for zero-shot speech translation on Must-C. Incorporating a speech decoder in our framework, we introduce the first results for zero-shot direct speech-to-speech and text-to-speech translation.
Speech synthesis has come a long way as current text-to-speech (TTS) models can now generate natural human-sounding speech. However, most of the TTS research focuses on using adult speech data and there has been very limited work done on child speech synthesis. This study developed and validated a training pipeline for fine-tuning state-of-the-art (SOTA) neural TTS models using child speech datasets. This approach adopts a multi-speaker TTS retuning workflow to provide a transfer-learning pipeline. A publicly available child speech dataset was cleaned to provide a smaller subset of approximately 19 hours, which formed the basis of our fine-tuning experiments. Both subjective and objective evaluations were performed using a pretrained MOSNet for objective evaluation and a novel subjective framework for mean opinion score (MOS) evaluations. Subjective evaluations achieved the MOS of 3.95 for speech intelligibility, 3.89 for voice naturalness, and 3.96 for voice consistency. Objective evaluation using a pretrained MOSNet showed a strong correlation between real and synthetic child voices. Speaker similarity was also verified by calculating the cosine similarity between the embeddings of utterances. An automatic speech recognition (ASR) model is also used to provide a word error rate (WER) comparison between the real and synthetic child voices. The final trained TTS model was able to synthesize child-like speech from reference audio samples as short as 5 seconds.