Domain-specific Training Data Research Articles

Ancient Chinese Poetry Collation Based on BERT

The rapid advancements in intelligent knowledge management technologies, exemplified by generative large language models, have yet to be fully explored and applied in the field of collation of ancient Chinese poetry. This study investigates the application of BERT-based pre-trained models, namely bert-base-chinese and SikuBERT, in the specialized task of ancient Chinese poetry collation. Focusing on the poetry of Li Bai, we employed a meticulously curated dataset to fine-tune these models, with the objective of enhancing their ability to identify and rectify errors in classical verse. Through a systematic approach to model adaptation, our research aimed to bridge the gap between generic language understanding and the nuanced complexities of ancient poetry.Results indicate that both models, after fine-tuning, exhibit substantial improvement in accurately addressing textual issues in the poetry. Specifically, SikuBERT, with its background in classical Chinese literature, achieved an impressive accuracy rate exceeding 40% post-fine-tuning, reflecting a marked increase from its base performance, thereby validating the significance of domain-specific training data. Meanwhile, bert-base-chinese also displayed notable enhancements, underscoring the models’ adaptability to specialized tasks. The investigation further emphasizes the potential for artificial intelligence to contribute to the precision and efficiency of ancient literature studies. We highlight future directions including refining fine-tuning methodologies, expanding the models’ capability to generalize across diverse poetic styles and periods, and integrating multi-modal data to deepen the understanding of historical context and authorial intent. This work underscores the transformative role of AI in the digital preservation and scholarly analysis of ancient poetry, paving the way for innovative approaches in the field of classical literature collation.

A detailed library perspective on nearly unsupervised information extraction workflows in digital libraries

Information extraction can support novel and effective access paths for digital libraries. Nevertheless, designing reliable extraction workflows can be cost-intensive in practice. On the one hand, suitable extraction methods rely on domain-specific training data. On the other hand, unsupervised and open extraction methods usually produce not-canonicalized extraction results. This paper is an extension of our original work and tackles the question of how digital libraries can handle such extractions and whether their quality is sufficient in practice. We focus on unsupervised extraction workflows by analyzing them in case studies in the domains of encyclopedias (Wikipedia), Pharmacy, and Political Sciences. As an extension, we analyze the extractions in more detail, verify our findings on a second extraction method, discuss another canonicalizing method, and give an outlook on how non-English texts can be handled. Therefore, we report on opportunities and limitations. Finally, we discuss best practices for unsupervised extraction workflows.

Multi-level feature disentanglement network for cross-dataset face forgery detection

Synthesizing videos with forged faces is a fundamental yet important safety-critical task that has caused severe security issues in recent years. Although many existing face forgery detection methods have achieved superior performance on such synthetic videos, they are severely limited by the domain-specific training data and generally perform unsatisfied when transferred to the cross-dataset scenario due to the domain gaps. Based on this observation, in this paper, we propose a multi-level feature disentanglement network to be robust to this domain bias induced by the different types of fake artifacts in different datasets. Specifically, we first detect the face image and transform it into both color-aware and frequency-aware inputs for multi-modal contextual representation learning. Then, we introduce a novel feature disentangling module that mainly utilizes a pair of complementary attention maps, to disentangle the synthetic features into separate realistic features and the features of fake artifacts. Since the features of fake artifacts are indirectly obtained from the latent features instead of the dataset-specific distribution, our forgery detection model is robust to the dataset-specific domain gaps. By applying the disentangling module to multi-levels of the feature extraction network with multi-modal inputs, we can obtain more robust feature representations. In addition, a realistic-aware adversary loss and a domain-aware adversary loss are adopted to facilitate the network for better feature disentanglement and extraction. Extensive experiments on four datasets verify the generalization of our method and present the state-of-the-art performance.

Transfer Learning for Spam Text Classification

Today’s world is mostly data-driven. To deal with the humongous amount of data, Machine Learning and Data Mining strategies are put into usage. Traditional ML approaches presume that the model is tested on a dataset extracted from the same domain from where the training data has been taken from. Nevertheless, some real-world situations require machines to provide good results with very little domain-specific training data. This creates room for the development of machines that are capable of predicting accurately by being trained on easily found data. Transfer Learning is the key to it. It is the scientific art of applying the knowledge gained while learning a task to another task that is similar to the previous one in some or another way. This article focuses on building a model that is capable of differentiating text data into binary classes; one roofing the text data that is spam and the other not containing spam using BERT’s pre-trained model (bert-base-uncased). This pre-trained model has been trained on Wikipedia and Book Corpus data and the goal of this paper is to highlight the pre-trained model’s capabilities to transfer the knowledge that it has learned from its training (Wiki and Book Corpus) to classifying spam texts from the rest.

Reinforced Transformer with Cross-Lingual Distillation for Cross-Lingual Aspect Sentiment Classification

Though great progress has been made in the Aspect-Based Sentiment Analysis(ABSA) task through research, most of the previous work focuses on English-based ABSA problems, and there are few efforts on other languages mainly due to the lack of training data. In this paper, we propose an approach for performing a Cross-Lingual Aspect Sentiment Classification (CLASC) task which leverages the rich resources in one language (source language) for aspect sentiment classification in a under-resourced language (target language). Specifically, we first build a bilingual lexicon for domain-specific training data to translate the aspect category annotated in the source-language corpus and then translate sentences from the source language to the target language via Machine Translation (MT) tools. However, most MT systems are general-purpose, it non-avoidably introduces translation ambiguities which would degrade the performance of CLASC. In this context, we propose a novel approach called Reinforced Transformer with Cross-Lingual Distillation (RTCLD) combined with target-sensitive adversarial learning to minimize the undesirable effects of translation ambiguities in sentence translation. We conduct experiments on different language combinations, treating English as the source language and Chinese, Russian, and Spanish as target languages. The experimental results show that our proposed approach outperforms the state-of-the-art methods on different target languages.

Evaluation of Automated Public De-Identification Tools on a Corpus of Radiology Reports.

To evaluate publicly available de-identification tools on a large corpus of narrative-text radiology reports. In this retrospective study, 21 categories of protected health information (PHI) in 2503 radiology reports were annotated from a large multihospital academic health system, collected between January 1, 2012 and January 8, 2019. A subset consisting of 1023 reports served as a test set; the remainder were used as domain-specific training data. The types and frequencies of PHI present within the reports were tallied. Five public de-identification tools were evaluated: MITRE Identification Scrubber Toolkit, U.S. National Library of Medicine‒Scrubber, Massachusetts Institute of Technology de-identification software, Emory Health Information DE-identification (HIDE) software, and Neuro named-entity recognition (NeuroNER). The tools were compared using metrics including recall, precision, and F1 score (the harmonic mean of recall and precision) for each category of PHI. The annotators identified 3528 spans of PHI text within the 2503 reports. Cohen κ for interrater agreement was 0.938. Dates accounted for the majority of PHI found in the dataset of radiology reports (n = 2755 [78%]). The two best-performing tools both used machine learning methods-NeuroNER (precision, 94.5%; recall, 92.6%; microaveraged F1 score [F1], 93.6%) and Emory HIDE (precision, 96.6%; recall, 88.2%; F1, 92.2%)-but none exceeded 50% F1 on the important patient names category. PHI appeared infrequently within the corpus of reports studied, which created difficulties for training machine learning systems. Out-of-the-box de-identification tools achieved limited performance on the corpus of radiology reports, suggesting the need for further advancements in public datasets and trained models.Supplemental material is available for this article.See also the commentary by Tenenholtz and Wood in this issue.© RSNA, 2020.

Low Resource Sequence Tagging with Weak Labels

Current methods for sequence tagging depend on large quantities of domain-specific training data, limiting their use in new, user-defined tasks with few or no annotations. While crowdsourcing can be a cheap source of labels, it often introduces errors that degrade the performance of models trained on such crowdsourced data. Another solution is to use transfer learning to tackle low resource sequence labelling, but current approaches rely heavily on similar high resource datasets in different languages. In this paper, we propose a domain adaptation method using Bayesian sequence combination to exploit pre-trained models and unreliable crowdsourced data that does not require high resource data in a different language. Our method boosts performance by learning the relationship between each labeller and the target task and trains a sequence labeller on the target domain with little or no gold-standard data. We apply our approach to labelling diagnostic classes in medical and educational case studies, showing that the model achieves strong performance though zero-shot transfer learning and is more effective than alternative ensemble methods. Using NER and information extraction tasks, we show how our approach can train a model directly from crowdsourced labels, outperforming pipeline approaches that first aggregate the crowdsourced data, then train on the aggregated labels.

Deep neural network acoustic models for spoken assessment applications

In this paper, we investigate the effectiveness of applying deep neural network hidden Markov models, or DNN-HMMs, for acoustic modeling in the context of educational applications. Specifically, we focus on spoken responses from non-native and child speech that tend to show great acoustic variability. We perform comprehensive experiments to compare the performance between traditional Gaussian mixture model (GMM)-HMMs and DNN-HMMs in three large language assessment datasets that contain various spoken tasks, classified broadly as constrained and open-ended tasks. Our experimental results suggest useful conclusions that can help guide the design of real-life educational applications. DNN-HMMs outperform conventional GMM-HMMs by a large margin for all spoken tasks commonly used in spoken assessment applications. In our experiments, DNN-HMMs trained using 25h of data can outperform GMM-HMMs trained with 6.7–9 times data. Specifically regarding overall performance, when all available training data were used (175, 227, 169h respectively), we achieved a relative word error rate decrease of 20.4% for adult English and 29.3% for child English, and a relative character error rate decrease of 14.3% for adult Chinese, when switching from GMMs to DNNs. In comparing between types of tasks, we notice that the more challenging open-ended tasks benefit significantly more than constrained item types by the use of DNN-HMMs. For open-ended tasks, having large amounts of training data is the key, as DNN-HMMs can take full advantage of the added training data and further push performance. In contrast, the performance of constrained spoken tasks saturates at around 25h of training data. At the same time, constrained spoken tasks require only a few hours of data (1 or 5h) to build well-performing acoustic models. This is an encouraging observation, that indicates the potential to build reliable spoken assessment applications based on constrained tasks, when few domain specific training data are available.

Method and apparatus for phonetic context adaptation for improved speech recognition

The present invention provides a computerized method and apparatus for automatically generating from a first speech recognizer a second speech recognizer which can be adapted to a specific domain. The first speech recognizer can include a first acoustic model with a first decision network and corresponding first phonetic contexts. The first acoustic model can be used as a starting point for the adaptation process. A second acoustic model with a second decision network and corresponding second phonetic contexts for the second speech recognizer can be generated by re-estimating the first decision network and the corresponding first phonetic contexts based on domain-specific training data.