Combining Self-supervised Learning and Active Learning for Disfluency Detection

Spoken language is fundamentally different from the written language in that it contains frequent disfluencies or parts of an utterance that are corrected by the speaker. Disfluency detection (removing these disfluencies) is desirable to clean the input for use in downstream NLP tasks. Most existing approaches to disfluency detection heavily rely on human-annotated data, which is scarce and expensive to obtain in practice. To tackle the training data bottleneck, in this work, we investigate methods for combining self-supervised learning and active learning for disfluency detection. First, we construct large-scale pseudo training data by randomly adding or deleting words from unlabeled data and propose two self-supervised pre-training tasks: (i) a tagging task to detect the added noisy words and (ii) sentence classification to distinguish original sentences from grammatically incorrect sentences. We then combine these two tasks to jointly pre-train a neural network. The pre-trained neural network is then fine-tuned using human-annotated disfluency detection training data. The self-supervised learning method can capture task-special knowledge for disfluency detection and achieve better performance when fine-tuning on a small annotated dataset compared to other supervised methods. However, limited in that the pseudo training data are generated based on simple heuristics and cannot fully cover all the disfluency patterns, there is still a performance gap compared to the supervised models trained on the full training dataset. We further explore how to bridge the performance gap by integrating active learning during the fine-tuning process. Active learning strives to reduce annotation costs by choosing the most critical examples to label and can address the weakness of self-supervised learning with a small annotated dataset. We show that by combining self-supervised learning with active learning, our model is able to match state-of-the-art performance with just about 10% of the original training data on both the commonly used English Switchboard test set and a set of in-house annotated Chinese data.

Ship Detection in Sentinel 2 Multi-Spectral Images with Self-Supervised Learning

Automatic ship detection provides an essential function towards maritime domain awareness for security or economic monitoring purposes. This work presents an approach for training a deep learning ship detector in Sentinel-2 multi-spectral images with few labeled examples. We design a network architecture for detecting ships with a backbone that can be pre-trained separately. By using self supervised learning, an emerging unsupervised training procedure, we learn good features on Sentinel-2 images, without requiring labeling, to initialize our network’s backbone. The full network is then fine-tuned to learn to detect ships in challenging settings. We evaluated this approach versus pre-training on ImageNet and versus a classical image processing pipeline. We examined the impact of variations in the self-supervised learning step and we show that in the few-shot learning setting self-supervised pre-training achieves better results than ImageNet pre-training. When enough training data are available, our self-supervised approach is as good as ImageNet pre-training. We conclude that a better design of the self-supervised task and bigger non-annotated dataset sizes can lead to surpassing ImageNet pre-training performance without any annotation costs.

A Visual and VAE Based Hierarchical Indoor Localization Method

Precise localization and pose estimation in indoor environments are commonly employed in a wide range of applications, including robotics, augmented reality, and navigation and positioning services. Such applications can be solved via visual-based localization using a pre-built 3D model. The increase in searching space associated with large scenes can be overcome by retrieving images in advance and subsequently estimating the pose. The majority of current deep learning-based image retrieval methods require labeled data, which increase data annotation costs and complicate the acquisition of data. In this paper, we propose an unsupervised hierarchical indoor localization framework that integrates an unsupervised network variational autoencoder (VAE) with a visual-based Structure-from-Motion (SfM) approach in order to extract global and local features. During the localization process, global features are applied for the image retrieval at the level of the scene map in order to obtain candidate images, and are subsequently used to estimate the pose from 2D-3D matches between query and candidate images. RGB images only are used as the input of the proposed localization system, which is both convenient and challenging. Experimental results reveal that the proposed method can localize images within 0.16 m and 4° in the 7-Scenes data sets and 32.8% within 5 m and 20° in the Baidu data set. Furthermore, our proposed method achieves a higher precision compared to advanced methods.

Automatic generation of lexica for sentiment polarity shifters

Alleviating pain is good and abandoning hope is bad. We instinctively understand how words like alleviate and abandon affect the polarity of a phrase, inverting or weakening it. When these words are content words, such as verbs, nouns, and adjectives, we refer to them as polarity shifters. Shifters are a frequent occurrence in human language and an important part of successfully modeling negation in sentiment analysis; yet research on negation modeling has focused almost exclusively on a small handful of closed-class negation words, such as not, no, and without. A major reason for this is that shifters are far more lexically diverse than negation words, but no resources exist to help identify them. We seek to remedy this lack of shifter resources by introducing a large lexicon of polarity shifters that covers English verbs, nouns, and adjectives. Creating the lexicon entirely by hand would be prohibitively expensive. Instead, we develop a bootstrapping approach that combines automatic classification with human verification to ensure the high quality of our lexicon while reducing annotation costs by over 70%. Our approach leverages a number of linguistic insights; while some features are based on textual patterns, others use semantic resources or syntactic relatedness. The created lexicon is evaluated both on a polarity shifter gold standard and on a polarity classification task.

Automated Conversation Review to Surface Virtual Assistant Misunderstandings: Reducing Cost and Increasing Privacy

With the rise of Intelligent Virtual Assistants (IVAs), there is a necessary rise in human effort to identify conversations containing misunderstood user inputs. These conversations uncover error in natural language understanding and help prioritize and expedite improvements to the IVA. As human reviewer time is valuable and manual analysis is time consuming, prioritizing the conversations where misunderstanding has likely occurred reduces costs and speeds improvement. In addition, less conversations reviewed by humans mean less user data is exposed, increasing privacy. We present a scalable system for automated conversation review that can identify potential miscommunications. Our system provides IVA designers with suggested actions to fix errors in IVA understanding, prioritizes areas of language model repair, and automates the review of conversations where desired.Verint - Next IT builds IVAs on behalf of other companies and organizations, and therefore analyzes large volumes of conversational data. Our review system has been in production for over three years and saves our company roughly $1.5 million in annotation costs yearly, as well as shortened the refinement cycle of production IVAs. In this paper, the system design is discussed and performance in identifying errors in IVA understanding is compared to that of human reviewers.

Cost-Effective Active Learning for Hierarchical Multi-Label Classification

Active learning reduces the labeling cost by actively querying labels for the most valuable data. It is particularly important for multi-label learning, where the annotation cost is rather high because each instance may have multiple labels simultaneously. In many multi-label tasks, the labels are organized into hierarchies from coarse to fine. The labels at different levels of the hierarchy contribute differently to the model training, and also have diverse annotation costs. In this paper, we propose a multi-label active learning approach to exploit the label hierarchies for cost-effective queries. By incorporating the potential contribution of ancestor and descendant labels, a novel criterion is proposed to estimate the informativeness of each candidate query. Further, a subset selection method is introduced to perform active batch selection by balancing the informativeness and cost of each instance-label pair. Experimental results validate the effectiveness of both the proposed criterion and the selection method.

Evaluating active learning methods for annotating semantic predications

Objectives This study evaluated and compared a variety of active learning strategies, including a novel strategy we proposed, as applied to the task of filtering incorrect semantic predications in SemMedDB. Materials and methods We evaluated 8 active learning strategies covering 3 types—uncertainty, representative, and combined—on 2 datasets of 6,000 total semantic predications from SemMedDB covering the domains of substance interactions and clinical medicine, respectively. We also designed a novel combined strategy called dynamic β that does not use hand-tuned hyperparameters. Each strategy was assessed by the Area under the Learning Curve (ALC) and the number of training examples required to achieve a target Area Under the ROC curve. We also visualized and compared the query patterns of the query strategies. Results All types of active learning (AL) methods beat the baseline on both datasets. Combined strategies outperformed all other methods in terms of ALC, outperforming the baseline by over 0.05 ALC for both datasets and reducing 58% annotation efforts in the best case. While representative strategies performed well, their performance was matched or outperformed by the combined methods. Our proposed AL method dynamic β shows promising ability to achieve near-optimal performance across 2 datasets. Discussion Our visual analysis of query patterns indicates that strategies which efficiently obtain a representative subsample perform better on this task. Conclusion Active learning is shown to be effective at reducing annotation costs for filtering incorrect semantic predications from SemMedDB. Our proposed AL method demonstrated promising performance.