ACM Transactions on Asian and Low-Resource Language Information Processing最新文献

Aspect-based sentiment classification (ABSC) is a crucial subtask of fine-grained sentiment analysis (SA), which aims to predict the sentiment polarity of the given aspects in a sentence as positive, negative, or neutral. Most existing ABSC methods based on supervised learning. However, these methods rely heavily on fine-grained labeled training data, which can be scarce in low-resource domains, limiting their effectiveness. To overcome this challenge, we propose a low-resource cross-domain aspect-based sentiment classification (CDABSC) approach based on a pre-training and fine-tuning strategy. This approach applies the pre-training and fine-tuning strategy to an advanced deep learning method designed for ABSC, namely the attention-based encoding graph convolutional network (AEGCN) model. Specifically, a high-resource domain is selected as the source domain, and the AEGCN model is pre-trained using a large amount of fine-grained annotated data from the source domain. The optimal parameters of the model are preserved. Subsequently, a low-resource domain is used as the target domain, and the pre-trained model parameters are used as the initial parameters of the target domain model. The target domain is fine-tuned using a small amount of annotated data to adapt the parameters to the target domain model, improving the accuracy of sentiment classification in the low-resource domain. Finally, experimental validation on two domain benchmark datasets, restaurant and laptop, demonstrates that significant outperformance of our approach over the baselines in CDABSC Micro-F1.

Token-level data augmentation generates text samples by modifying the words of the sentences. However, data that are not easily classified can negatively affect the model. In particular, not considering the role of keywords when performing random augmentation operations on samples may lead to the generation of low-quality supplementary samples. Therefore, we propose a supervised contrast learning text classification model based on data quality augment (DQA). First, dynamic training is used to screen high-quality datasets containing beneficial information for model training. The selected data is then augmented with data based on important words with tag information. To obtain a better text representation to serve the downstream classification task, we employ a standard supervised contrast loss to train the model. Finally, we conduct experiments on five text classification datasets to validate the effectiveness of our model. In addition, ablation experiments are conducted to verify the impact of each module on classification.

This paper studies entity linking (EL) in Web tables, which aims to link the string mentions in table cells to their referent entities in a knowledge base. Two main problems exist in previous studies: 1) contextual information is not well utilized in mention-entity similarity computation; 2) the assumption on entity coherence that all entities in the same row or column are highly related to each other is not always correct. In this paper, we propose NPEL, a new Neural Paired Entity Linking framework, to overcome the above problems. In NPEL, we design a deep learning model with different neural networks and an attention mechanism, to model different kinds of contextual information of mentions and entities, for mention-entity similarity computation in Web tables. NPEL also relaxes the above assumption on entity coherence by a new paired entity linking algorithm, which iteratively selects two mentions with the highest confidence for EL. Experiments on real-world datasets exhibit that NPEL has the best performance compared with state-of-the-art baselines in different evaluation metrics.

During the last decade, social media has gained significant popularity as a medium for individuals to express their views on various topics. However, some individuals also exploit the social media platforms to spread hatred through their comments and posts, some of which target individuals, communities or religions. Given the deep emotional connections people have to their religious beliefs, this form of hate speech can be divisive and harmful, and may result in issues of mental health as social disorder. Therefore, there is a need of algorithmic approaches for the automatic detection of instances of hate speech. Most of the existing studies in this area focus on social media content in English, and as a result several low-resource languages lack computational resources for the task. This study attempts to address this research gap by providing a high-quality annotated dataset designed specifically for identifying hate speech against religions in the Hindi-English code-mixed language. This dataset “Targeted Hate Speech Against Religion” (THAR)) consists of 11,549 comments and has been annotated by five independent annotators. It comprises two subtasks: (i) Subtask-1 (Binary classification), (ii) Subtask-2 (multi-class classification). To ensure the quality of annotation, the Fleiss Kappa measure has been employed. The suitability of the dataset is then further explored by applying different standard deep learning, and transformer-based models. The transformer-based model, namely Multilingual Representations for Indian Languages (MuRIL), is found to outperform the other implemented models in both subtasks, achieving macro average and weighted average F1 scores of 0.78 and 0.78 for Subtask-1, and 0.65 and 0.72 for Subtask-2, respectively. The experimental results obtained not only confirm the suitability of the dataset but also advance the research towards automatic detection of hate speech, particularly in the low-resource Hindi-English code-mixed language.

The purpose of the study is to solve an extreme mathematical problem – semantic analysis of natural language, which can be used in various fields, including marketing research, online translators, and search engines. When training the neural network, data training methods based on the LDA model and vector representation of words were used. This study presents the development of a neurocomputer system used for the purpose of semantic analysis of the text in the Kazakh language, based on machine learning and the use of the LDA model. In the course of the study, the stages of system development were considered, regarding the text recognition algorithm. The Python programming language was used as a tool using libraries that greatly simplify the process of creating neural networks, including the Keras library. An experiment was conducted with the involvement of experts to test the effectiveness of the system, the results of which confirmed the reliability of the data provided by the system. The papers of modern computer linguists dealing with the problems of natural language processing using various technologies and methods are considered.

The method of translation from one language to another without human intervention is known as Machine Translation (MT). Multilingual neural machine translation (MNMT) is a technique for MT that builds a single model for multiple languages. It is preferred over other approaches since it decreases training time and improves translation in low-resource contexts, i.e. for languages that have insufficient corpus. However, good-quality MT models are yet to be built for many scenarios such as for Indic-to-Indic Languages (IL-IL). Hence, this paper is an attempt to address and develop the baseline models for low-resource languages i.e. IL-IL (for 11 Indic Languages (ILs)) in a multilingual environment. The models are built on the Samanantar corpus and analyzed on the Flores-200 corpus. All the models are evaluated using standard evaluation metrics i.e. Bilingual Evaluation Understudy (BLEU) score (with the range of 0 to 100). This paper examines the effect of the grouping of related languages, namely East Indo-Aryan (EI), Dravidian (DR), and West Indo-Aryan (WI) on the MNMT model. From the experiments, the results reveal that related language grouping is beneficial for the WI group only while it is detrimental for the EI group and it shows an inconclusive effect on the DR group. The role of pivot-based MNMT models in enhancing translation quality is also investigated in this paper. Owing to the presence of large good-quality corpora from English (EN) to ILs, MNMT IL-IL models using EN as a pivot are built and examined. To achieve this, English-Indic Language (EN-IL) models are developed with and without the usage of related languages. Results show that the use of related language grouping is advantageous specifically for EN to ILs. Thus, related language groups are used for the development of pivot MNMT models. It is also observed that the usage of pivot models greatly improves MNMT baselines. Furthermore, the effect of transliteration on ILs is also analyzed in this paper. To explore transliteration, the best MNMT models from the previous approaches (in most of cases pivot model using related groups) are determined and built on corpus transliterated from the corresponding scripts to a modified Indian language Transliteration script (ITRANS). The outcome of the experiments indicates that transliteration helps the models built for lexically rich languages, with the best increment of BLEU scores observed in Malayalam (ML) and Tamil (TA), i.e. 6.74 and 4.72, respectively. The BLEU score using transliteration models ranges from 7.03 to 24.29. The best model obtained is the Punjabi (PA)-Hindi (HI) language pair trained on PA-WI transliterated corpus.

By summarizing longer consumer health questions into shorter and essential ones, medical question-answering systems can more accurately understand consumer intentions and retrieve suitable answers. However, medical question summarization is very challenging due to obvious distinctions in health trouble descriptions from patients and doctors. Although deep learning has been applied to successfully address the medical question summarization (MQS) task, two challenges remain: how to correctly capture question focus to model its semantic intention, and how to obtain reliable datasets to fairly evaluate performance. To address these challenges, this paper proposes a novel medical question summarization framework based on entity-driven contrastive learning (ECL). ECL employs medical entities present in frequently asked questions (FAQs) as focuses and devises an effective mechanism to generate hard negative samples. This approach compels models to focus on essential information and consequently generate more accurate question summaries. Furthermore, we have discovered that some MQS datasets, such as the iCliniq dataset with a 33% duplicate rate, have significant data leakage issues. To ensure an impartial evaluation of the related methods, this paper carefully examines leaked samples to reorganize more reasonable datasets. Extensive experiments demonstrate that our ECL method outperforms the existing methods and achieves new state-of-the-art performance, i.e., 52.85, 43.16, 41.31, 43.52 in terms of ROUGE-1 metric on MeQSum, CHQ-Summ, iCliniq, HealthCareMagic dataset, respectively. The code and datasets are available at https://github.com/yrbobo/MQS-ECL.

Unsupervised machine translation (UMT) has recently attracted more attention from researchers, enabling models to translate when languages lack parallel corpora. However, the current works mainly consider close language pairs (e.g., English-German and English-French), and the effectiveness of visual content for distant language pairs has yet to be investigated. This paper proposes a unsupervised multimodal machine translation (UMMT) model for low-resource distant language pairs. Specifically, we first employ adequate measures such as transliteration and re-ordering to bring distant language pairs closer together. We then use visual content to extend masked language modeling (MLM) and generate visual masked language modeling (VMLM) for UMT. Finally, empirical experiments are conducted on our distant language pair dataset and the public Multi30k dataset. Experimental results demonstrate the superior performance of our model, with BLEU score improvements of 2.5 and 2.6 on translation for distant language pairs English-Uyghur and Chinese-Uyghur. Moreover, our model also brings remarkable results for close language pairs, improving 2.3 BLEU compared with the existing models in English-German.

Drug-drug interactions (DDIs) are an important biological phenomenon which can result in medical errors from medical practitioners. Drug interactions can change the molecular structure of interacting agents which may prove to be fatal in the worst case. Finding drug interactions early in diagnosis can be pivotal in side-effect prevention. The growth of big data provides a rich source of information for clinical studies to investigate DDIs. We propose a hierarchical classification model which is double-pass in nature. The first pass predicts the occurrence of an interaction and then the second pass further predicts the type of interaction such as effect, advice, mechanism, and int. We applied different deep learning algorithms with Convolutional Bi-LSTM (ConvBLSTM) proving to be the best. The results show that pre-trained vector embeddings prove to be the most appropriate features. The F1-score of the ConvBLSTM algorithm turned out to be 96.39% and 98.37% in Russian and English language respectively which is greater than the state-of-the-art systems. According to the results, it can be concluded that adding a convolution layer before the bi-directional pass improves model performance in the automatic classification and extraction of drug interactions, using pre-trained vector embeddings such as Fasttext and Bio-Bert.

Multilingualism in India is widespread due to its long history of foreign acquaintances. This leads to the presence of an audience familiar with conversing using more than one language. Additionally, due to the social media boom, the usage of multiple languages to communicate has become extensive. Hence, the need for a translation system that can serve the novice and monolingual user is the need of the hour. Such translation systems can be developed by methods such as statistical machine translation and neural machine translation, where each approach has its advantages as well as disadvantages. In addition, the parallel corpus needed to build a translation system, with code-mixed data, is not readily available. In the present work, we present two translation frameworks that can leverage the individual advantages of these pre-existing approaches by building an ensemble model that takes a consensus of the final outputs of the preceding approaches and generates the target output. The developed models were used for translating English-Bengali code-mixed data (written in Roman script) into their equivalent monolingual Bengali instances. A code-mixed to monolingual parallel corpus was also developed to train the preceding systems. Empirical results show improved BLEU and TER scores of 17.23 and 53.18 and 19.12 and 51.29, respectively, for the developed frameworks.