Jun Wang

In the rapidly advancing field of software development, the demand for practical code translation tools has surged, driven by the need for interoperability across different programming environments. Existing learning-based approaches often need help with low-resource programming languages that lack sufficient parallel code corpora for training. To address these limitations, we propose a novel training framework that begins with monolingual seed corpora, generating parallel datasets via back-translation and incorporating compiler feedback to optimize the translation model. As a case study, we apply our method to train a code translation model for a new-born low-resource programming language, Cangjie. We also construct a parallel test dataset for Java-to-Cangjie translation and test cases to evaluate the effectiveness of our approach. Experimental results demonstrate that compiler feedback greatly enhances syntactical correctness, semantic accuracy, and test pass rates of the translated Cangjie code. These findings highlight the potential of our method to support code translation in low-resource settings, expanding the capabilities of learning-based models for programming languages with limited data availability.

Machine translation has become an integral part of daily life, with terminology translation playing a crucial role in ensuring the accuracy of translation results. However, existing translation systems, such as Google Translate, have been shown to occasionally produce errors in terminology translation. Current metrics for assessing terminology translation rely on reference translations and bilingual dictionaries, limiting their effectiveness in large-scale automated MT system testing. To address this challenge, we propose a novel method: Metamorphic Testing for Terminology Translation (TermMT), which achieves effective and efficient testing for terminology translation in MT systems without relying on reference translations or bilingual terminology dictionaries. Our approach involves constructing metamorphic relations based on the characteristics of terms: (a) adding an appropriate reference of the term in the given context would not change the translation of the term; (b) if we modify part of a multi-word term, the translation of the revised word combination would change. To evaluate the effectiveness of TermMT, we tested the terminology translation capabilities of three machine translation systems, Google Translate, Bing Microsoft Translator, and mBART, using the English portion of the bilingual UM-corpus dataset. The results show that TermMT detected a total of 3,765 translation errors on Google Translate, 2,351 on Bing Microsoft Translator, and 6,011 on mBART, with precisions of 82.33%, 83.00%, and 86.33%, respectively.

In the wake of developments in the field of Natural Language Processing, Question Answering (QA) software has penetrated our daily lives. Due to the data-driven programming paradigm, QA software inevitably contains bugs, i.e., misbehaving in real-world applications. Current testing techniques for testing QA software include two folds, reference-based testing and metamorphic testing. This paper adopts a different angle to achieve testing for QA software: we notice that answers to questions would have inference relations, i.e., the answers to some questions could be logically inferred from the answers to other questions. If these answers on QA software do not satisfy the inference relations, an inference bug is detected. To generate the questions with the inference relations automatically, we propose a novel testing method Knowledge Graph driven Inference Testing (KGIT), which employs facts in the Knowledge Graph (KG) as the seeds to logically construct test cases containing questions and contexts with inference relations. To evaluate the effectiveness of KGIT, we conduct an extensive empirical study with more than 2.8 million test cases generated from the large-scale KG YAGO4 and three QA models based on the state-of-the-art QA model structure. The experimental results show that our method (a) could detect a considerable number of inference bugs in all three studied QA models and (b) is helpful in retraining QA models to improve their inference ability.

We research how automatically predict the charges and relevant law articles of criminal cases in our work. At first, the distributions of charges, relevant law articles and fact description of criminal cases are analyzed based on CAIL2018. CAIL2018, the first large-scale dataset for legal judgment prediction in China, contains large amounts of criminal cases collected from the Supreme People’s Court of China. By our analysis, we find the distribution of criminal cases is typical 8020 distribution. Then we present our framework to predict criminal cases automatically. In our framework, data enhancement, oversampling, key word extraction are used to optimize data quality, and deep learning is employed to predict charges and relevant articles. In the prediction, single deep learning model is tested firstly, then ensemble of different deep learning models are compared to achieve better performance than that of single model. In our work, we find data enhancement and ensemble strategy can improve the performance of judgment prediction. More differences of joint models and data, better performance of ensemble strategy.

Accurate photovoltaic (PV) power prediction is an important basis for ensuring the safe and stable operation of high-permeability PV power generation systems. There are many environmental factors affecting the output power of PV power generation, and the influence of fog and haze on PV power cannot be ignored. In this paper the main factors of affecting PV power and the necessity of accounting for fog and haze in PV power prediction are described. The Air Quality Index (AQI) is selected as an indicator to measure the severity of fog and haze, and it is included in the historical data source of PV power prediction together with weather conditions. Aiming at the complex nonlinear problem of PV power prediction, a PV power prediction model based on BP neural network algorithm in fog and haze weather is established. In order to ensure the accuracy of PV power prediction, a training sample selection method based on ”similar time” is proposed. The analysis of the example verifies the necessity of considering fog and haze in the PV power prediction and the effectiveness of the proposed method.