ACLSTM: A Novel Method for CQA Answer Quality Prediction Based on Question-Answer Joint Learning

Given the limitations of the community question answering (CQA) answer quality prediction method in measuring the semantic information of the answer text, this paper proposes an answer quality prediction model based on the question-answer joint learning (ACLSTM). The attention mechanism is used to obtain the dependency relationship between the Question-and-Answer (Q&A) pairs. Convolutional Neural Network (CNN) and Long Short-term Memory Network (LSTM) are used to extract semantic features of Q&A pairs and calculate their matching degree. Besides, answer semantic representation is combined with other effective extended features as the input representation of the fully connected layer. Compared with other quality prediction models, the ACLSTM model can effectively improve the prediction effect of answer quality. In particular, the mediumquality answer prediction, and its prediction effect is improved after adding effective extended features. Experiments prove that after the ACLSTM model learning, the Q&A pairs can better measure the semantic match between each other, fully reflecting the model’s superior performance in the semantic information processing of the answer text.     

Arabic Named Entity Recognition: A BERT-BGRU Approach

Named Entity Recognition (NER) is one of the fundamental tasks in Natural Language Processing (NLP), which aims to locate, extract, and classify named entities into a predefined category such as person, organization and location. Most of the earlier research for identifying named entities relied on using handcrafted features and very large knowledge resources, which is time consuming and not adequate for resource-scarce languages such as Arabic. Recently, deep learning achieved state-of-the-art performance on many NLP tasks including NER without requiring hand-crafted features. In addition, transfer learning has also proven its efficiency in several NLP tasks by exploiting pretrained language models that are used to transfer knowledge learned from large-scale datasets to domain-specific tasks. Bidirectional Encoder Representation from Transformer (BERT) is a contextual language model that generates the semantic vectors dynamically according to the context of the words. BERT architecture relay on multi-head attention that allows it to capture global dependencies between words. In this paper, we propose a deep learning-based model by fine-tuning BERT model to recognize and classify Arabic named entities. The pre-trained BERT context embeddings were used as input features to a Bidirectional Gated Recurrent Unit (BGRU) and were fine-tuned using two annotated Arabic Named Entity Recognition (ANER) datasets. Experimental results demonstrate that the proposed model outperformed state-of-the-art ANER models achieving 92.28% and 90.68% F-measure values on the ANERCorp dataset and the merged ANERCorp and AQMAR dataset, respectively.     

BERT-CNN: A Deep Learning Model for Detecting Emotions from Text

Due to the widespread usage of social media in our recent daily lifestyles, sentiment analysis becomes an important field in pattern recognition and Natural Language Processing (NLP). In this field, users’ feedback data on a specific issue are evaluated and analyzed. Detecting emotions within the text is therefore considered one of the important challenges of the current NLP research. Emotions have been widely studied in psychology and behavioral science as they are an integral part of the human nature. Emotions describe a state of mind of distinct behaviors, feelings, thoughts and experiences. The main objective of this paper is to propose a new model named BERT-CNN to detect emotions from text. This model is formed by a combination of the Bidirectional Encoder Representations from Transformer (BERT) and the Convolutional Neural networks (CNN) for textual classification. This model embraces the BERT to train the word semantic representation language model. According to the word context, the semantic vector is dynamically generated and then placed into the CNN to predict the output. Results of a comparative study proved that the BERT-CNN model overcomes the state-of-art baseline performance produced by different models in the literature using the semeval 2019 task3 dataset and ISEAR datasets. The BERT-CNN model achieves an accuracy of 94.7% and an F1-score of 94% for semeval2019 task3 dataset and an accuracy of 75.8% and an F1-score of 76% for ISEAR dataset.     

Fine-Grained Features for Image Captioning

Image captioning involves two different major modalities (image and sentence) that convert a given image into a language that adheres to visual semantics. Almost all methods first extract image features to reduce the difficulty of visual semantic embedding and then use the caption model to generate fluent sentences. The Convolutional Neural Network (CNN) is often used to extract image features in image captioning, and the use of object detection networks to extract region features has achieved great success. However, the region features retrieved by this method are object-level and do not pay attention to fine-grained details because of the detection model’s limitation. We offer an approach to address this issue that more properly generates captions by fusing fine-grained features and region features. First, we extract fine-grained features using a panoramic segmentation algorithm. Second, we suggest two fusion methods and contrast their fusion outcomes. An X-linear Attention Network (X-LAN) serves as the foundation for both fusion methods. According to experimental findings on the COCO dataset, the two-branch fusion approach is superior. It is important to note that on the COCO Karpathy test split, CIDEr is increased up to 134.3% in comparison to the baseline, highlighting the potency and viability of our method.     

Sentence Similarity Measurement with Convolutional Neural Networks Using Semantic and Syntactic Features

Calculating the semantic similarity of two sentences is an extremely challenging problem. We propose a solution based on convolutional neural networks (CNN) using semantic and syntactic features of sentences. The similarity score between two sentences is computed as follows. First, given a sentence, two matrices are constructed accordingly, which are called the syntax model input matrix and the semantic model input matrix; one records some syntax features, and the other records some semantic features. By experimenting with different arrangements of representing the syntactic and semantic features of the sentences in the matrices, we adopt the most effective way of constructing the matrices. Second, these two matrices are given to two neural networks, which are called the sentence model and the semantic model, respectively. The convolution process of the neural networks of the two models is carried out in multiple perspectives. The outputs of the two models are combined as a vector, which is the representation of the sentence. Third, given the representation vectors of two sentences, the similarity score of these representations is computed by a layer in the CNN. Experiment results show that our algorithm (SSCNN) surpasses the performance MPCPP, which noticeably the best recent work of using CNN for sentence similarity computation. Comparing with MPCNN, the convolution computation in SSCNN is considerably simpler. Based on the results of this work, we suggest that by further utilization of semantic and syntactic features, the performance of sentence similarity measurements has considerable potentials to be improved in the future.     

Improving Chinese Word Representation with Conceptual Semantics

The meaning of a word includes a conceptual meaning and a distributive meaning. Word embedding based on distribution suffers from insufficient conceptual semantic representation caused by data sparsity, especially for low-frequency words. In knowledge bases, manually annotated semantic knowledge is stable and the essential attributes of words are accurately denoted. In this paper, we propose a Conceptual Semantics Enhanced Word Representation (CEWR) model, computing the synset embedding and hypernym embedding of Chinese words based on the Tongyici Cilin thesaurus, and aggregating it with distributed word representation to have both distributed information and the conceptual meaning encoded in the representation of words. We evaluate the CEWR model on two tasks: word similarity computation and short text classification. The Spearman correlation between model results and human judgement are improved to 64.71%, 81.84%, and 85.16% on Wordsim297, MC30, and RG65, respectively. Moreover, CEWR improves the F1 score by 3% in the short text classification task. The experimental results show that CEWR can represent words in a more informative approach than distributed word embedding. This proves that conceptual semantics, especially hypernymous information, is a good complement to distributed word representation.     

A multi-criteria context-sensitive approach for social image collection summarization

Recent increase in the number of digital photos in the content sharing and social networking websites has created an endless demand for techniques to analyze, navigate, and summarize these images. In this paper, we focus on image collection summarization. Earlier methods in image collection summarization consider representativeness and diversity criteria while recent ones also consider other criteria such as image quality, aesthetic or appeal. In this paper, we propose a multi-criteria context-sensitive approach for social image collection summarization. In the proposed method, two different sets of features are combined while each one looks at different criteria for image collection summarization: social attractiveness features and semantic features. The first feature set considers different aspects that make an image appealing such as image quality, aesthetic, and emotion to create attractiveness score for input images while the second one covers semantic content of images and assigns semantic score to them. We use social network infrastructure to identify attractiveness features and domain ontology for extracting ontology features. The final summarization is provided by integrating the attractiveness and semantic features of input images. The experimental results on a collection of human generated summaries on a set of Flickr images demonstrate the effectiveness of the proposed image collection summarization approach.     

MII: A Novel Text Classification Model Combining Deep Active Learning with BERT

Active learning has been widely utilized to reduce the labeling cost of supervised learning. By selecting specific instances to train the model, the performance of the model was improved within limited steps. However, rare work paid attention to the effectiveness of active learning on it. In this paper, we proposed a deep active learning model with bidirectional encoder representations from transformers (BERT) for text classification. BERT takes advantage of the self-attention mechanism to integrate contextual information, which is beneficial to accelerate the convergence of training. As for the process of active learning, we design an instance selection strategy based on posterior probabilities Margin, Intra-correlation and Inter-correlation (MII). Selected instances are characterized by small margin, low intra-cohesion and high inter-cohesion. We conduct extensive experiments and analytics with our methods. The effect of learner is compared while the effect of sampling strategy and text classification is assessed from three real datasets. The results show that our method outperforms the baselines in terms of accuracy.     

Why was this cited? Explainable machine learning applied to COVID-19 research literature

Multiple studies have investigated bibliometric factors predictive of the citation count a research article will receive. In this article, we go beyond bibliometric data by using a range of machine learning techniques to find patterns predictive of citation count using both article content and available metadata. As the input collection, we use the CORD-19 corpus containing research articles—mostly from biology and medicine—applicable to the COVID-19 crisis. Our study employs a combination of state-of-the-art machine learning techniques for text understanding, including embeddings-based language model BERT, several systems for detection and semantic expansion of entities: ConceptNet, Pubtator and ScispaCy. To interpret the resulting models, we use several explanation algorithms: random forest feature importance, LIME, and Shapley values. We compare the performance and comprehensibility of models obtained by “black-box” machine learning algorithms (neural networks and random forests) with models built with rule learning (CORELS, CBA), which are intrinsically explainable. Multiple rules were discovered, which referred to biomedical entities of potential interest. Of the rules with the highest lift measure, several rules pointed to dipeptidyl peptidase4 (DPP4), a known MERS-CoV receptor and a critical determinant of camel to human transmission of the camel coronavirus (MERS-CoV). Some other interesting patterns related to the type of animal investigated were found. Articles referring to bats and camels tend to draw citations, while articles referring to most other animal species related to coronavirus are lowly cited. Bat coronavirus is the only other virus from a non-human species in the betaB clade along with the SARS-CoV and SARS-CoV-2 viruses. MERS-CoV is in a sister betaC clade, also close to human SARS coronaviruses. Thus both species linked to high citation counts harbor coronaviruses which are more phylogenetically similar to human SARS viruses. On the other hand, feline (FIPV, FCOV) and canine coronaviruses (CCOV) are in the alpha coronavirus clade and more distant from the betaB clade with human SARS viruses. Other results include detection of apparent citation bias favouring authors with western sounding names. Equal performance of TF-IDF weights and binary word incidence matrix was observed, with the latter resulting in better interpretability. The best predictive performance was obtained with a “black-box” method—neural network. The rule-based models led to most insights, especially when coupled with text representation using semantic entity detection methods. Follow-up work should focus on the analysis of citation patterns in the context of phylogenetic trees, as well on patterns referring to DPP4, which is currently considered as a SARS-Cov-2 therapeutic target.

     

A deep-learning based citation count prediction model with paper metadata semantic features

Predicting the impact of academic papers can help scholars quickly identify the high-quality papers in the field. How to develop efficient predictive model for evaluating potential papers has attracted increasing attention in academia. Many studies have shown that early citations contribute to improving the performance of predicting the long-term impact of a paper. Besides early citations, some bibliometric features and altmetric features have also been explored for predicting the impact of academic papers. Furthermore, paper metadata text such as title, abstract and keyword contains valuable information which has effect on its citation count. However, present studies ignore the semantic information contained in the metadata text. In this paper, we propose a novel citation prediction model based on paper metadata text to predict the long-term citation count, and the core of our model is to obtain the semantic information from the metadata text. We use deep learning techniques to encode the metadata text, and then further extract high-level semantic features for learning the citation prediction task. We also integrate early citations for improving the prediction performance of the model. We show that our proposed model outperforms the state-of-the-art models in predicting the long-term citation count of the papers, and metadata semantic features are effective for improving the accuracy of the citation prediction models.

     

Jointly Part-of-Speech Tagging and Semantic Role Labeling Using Auxiliary Deep Neural Network Model

Previous studies have shown that there is potential semantic dependency between part-of-speech and semantic roles. At the same time, the predicate-argument structure in a sentence is important information for semantic role labeling task. In this work, we introduce the auxiliary deep neural network model, which models semantic dependency between part-of-speech and semantic roles and incorporates the information of predicate-argument into semantic role labeling. Based on the framework of joint learning, part-of-speech tagging is used as an auxiliary task to improve the result of the semantic role labeling. In addition, we introduce the argument recognition layer in the training process of the main task-semantic role labeling, so the argument-related structural information selected by the predicate through the attention mechanism is used to assist the main task. Because the model makes full use of the semantic dependency between part-of-speech and semantic roles and the structural information of predicateargument, our model achieved the F1 value of 89.0% on the WSJ test set of CoNLL2005, which is superior to existing state-of-the-art model about 0.8%.     

面向传统服饰的细粒度跨模态检索算法

目的 由于跨模态数据集有限和模态异构表征问题,利用跨模态检索算法解决实际应用问题一直是当前多模态研究中的一大研究方向.方法 提出了一种面向传统服饰的细粒度跨模态检索算法,解决传统服饰跨模态检索的单模态表征和跨模态表征一致的问题.在单模态特征表征方面,沿用DCMH使用深度学习的方法对初始数据进行特征提取;在跨模态表征一致方面,新增自监督语义网络,以自监督的方式对应标签信息提取细粒度信息,并将其用于图文哈希学习的监督,从而得到更好的图文哈希表征.通过在传统服饰数据集上与其他方法进行对比实验,验证算法的有效性.结论 有关此方面的应用探索,有利于解决互联网时代中国传统服饰文本、图像处理等的保护性难题,为未来纹样检索中的工作做铺垫,实现中国传统服饰的创新性传承和发展.     

Semantic Segmentation by Using Down-Sampling and Subpixel Convolution: DSSC-UNet

Recently, semantic segmentation has been widely applied to image processing, scene understanding, and many others. Especially, in deep learning-based semantic segmentation, the U-Net with convolutional encoder-decoder architecture is a representative model which is proposed for image segmentation in the biomedical field. It used max pooling operation for reducing the size of image and making noise robust. However, instead of reducing the complexity of the model, max pooling has the disadvantage of omitting some information about the image in reducing it. So, this paper used two diagonal elements of down-sampling operation instead of it. We think that the down-sampling feature maps have more information intrinsically than max pooling feature maps because of keeping the Nyquist theorem and extracting the latent information from them. In addition, this paper used two other diagonal elements for the skip connection. In decoding, we used Subpixel Convolution rather than transposed convolution to efficiently decode the encoded feature maps. Including all the ideas, this paper proposed the new encoder-decoder model called Down-Sampling and Subpixel Convolution U-Net (DSSC-UNet). To prove the better performance of the proposed model, this paper measured the performance of the U-Net and DSSC-UNet on the Cityscapes. As a result, DSSC-UNet achieved 89.6% Mean Intersection Over Union (Mean-IoU) and U-Net achieved 85.6% Mean-IoU, confirming that DSSC-UNet achieved better performance.     

A Phrase Topic Model Based on Distributed Representation

Traditional topic models have been widely used for analyzing semantic topics from electronic documents. However, the obvious defects of topic words acquired by them are poor in readability and consistency. Only the domain experts are possible to guess their meaning. In fact, phrases are the main unit for people to express semantics. This paper presents a Distributed Representation-Phrase Latent Dirichlet Allocation (DRPhrase LDA) which is a phrase topic model. Specifically, we reasonably enhance the semantic information of phrases via distributed representation in this model. The experimental results show the topics quality acquired by our model is more readable and consistent than other similar topic models.     

English to Tamil machine translation system using universal networking language

This paper proposes English to Tamil machine translation system, using the universal networking language (UNL) as the intermediate representation. The UNL approach is a hybrid approach of the rule and knowledge-based approaches to machine translation. UNL is a declarative formal language, specifically designed to represent semantic data extracted from a natural language text. The input English sentence is converted to UNL (enconversion), which is then converted to a Tamil sentence (deconversion) by ensuring that the meaning of the input sentence is preserved. The representation of UNL was modified to suit the translation process. A new sentence formation algorithm was also proposed to rearrange the translated Tamil words to sentences. The translation system was evaluated using bilingual evaluation understudy (BLEU) score. A BLEU score of 0.581 was achieved, which is an indication that most of the information in the input sentence is retained in the translated sentence. The scores obtained using the UNL based approach were compared with existing approaches to translation, and it can be concluded that the UNL is a more suited approach to machine translation.     

Patent classification by fine-tuning BERT language model

In this work we focus on fine-tuning a pre-trained BERT model and applying it to patent classification. When applied to large datasets of over two million patents, our approach outperforms the state of the art by an approach using CNN with word embeddings. Besides, we focus on patent claims without other parts in patent documents. Our contributions include: (1) a new state-of-the-art result based on pre-trained BERT model and fine-tuning for patent classification, (2) a large dataset USPTO-3M at the CPC subclass level with SQL statements that can be used by future researchers, (3) showing that patent claims alone are sufficient to achieve state-of-the-art results for classification task, in contrast to conventional wisdom.     

A Novel Optimized Language-Independent Text Summarization Technique

A substantial amount of textual data is present electronically in several languages. These texts directed the gear to information redundancy. It is essential to remove this redundancy and decrease the reading time of these data. Therefore, we need a computerized text summarization technique to extract relevant information from group of text documents with correlated subjects. This paper proposes a language-independent extractive summarization technique. The proposed technique presents a clustering-based optimization technique. The clustering technique determines the main subjects of the text, while the proposed optimization technique minimizes redundancy, and maximizes significance. Experiments are devised and evaluated using BillSum dataset for the English language, MLSUM for German and Russian and Mawdoo3 for the Arabic language. The experiments are evaluated using ROUGE metrics. The results showed the effectiveness of the proposed technique compared to other language-dependent and language-independent summarization techniques. Our technique achieved better ROUGE metrics for all the utilized datasets. The technique accomplished an F-measure of 41.9% for Rouge-1, 18.7% for Rouge-2, 39.4% for Rouge-3, and 16.8% for Rouge-4 on average for all the dataset using all three objectives. Our system also exhibited an improvement of 26.6%, 35.5%, 34.65%, and 31.54% w.r.t. The recent model contributed in the summarization of BillSum in terms of ROUGE metric evaluation. Our model’s performance is higher than the compared models, especially in the metric results of ROUGE_2 which is bi-gram matching.     

Keyphrase Generation Based on Self-Attention Mechanism

Keyphrase greatly provides summarized and valuable information. This information can help us not only understand text semantics, but also organize and retrieve text content effectively. The task of automatically generating it has received considerable attention in recent decades. From the previous studies, we can see many workable solutions for obtaining keyphrases. One method is to divide the content to be summarized into multiple blocks of text, then we rank and select the most important content. The disadvantage of this method is that it cannot identify keyphrase that does not include in the text, let alone get the real semantic meaning hidden in the text. Another approach uses recurrent neural networks to generate keyphrases from the semantic aspects of the text, but the inherently sequential nature precludes parallelization within training examples, and distances have limitations on context dependencies. Previous works have demonstrated the benefits of the self-attention mechanism, which can learn global text dependency features and can be parallelized. Inspired by the above observation, we propose a keyphrase generation model, which is based entirely on the self-attention mechanism. It is an encoder-decoder model that can make up the above disadvantage effectively. In addition, we also consider the semantic similarity between keyphrases, and add semantic similarity processing module into the model. This proposed model, which is demonstrated by empirical analysis on five datasets, can achieve competitive performance compared to baseline methods.     

Automated Arabic Text Classification Using Hyperparameter Tuned Hybrid Deep Learning Model

The text classification process has been extensively investigated in various languages, especially English. Text classification models are vital in several Natural Language Processing (NLP) applications. The Arabic language has a lot of significance. For instance, it is the fourth mostly-used language on the internet and the sixth official language of the United Nations. However, there are few studies on the text classification process in Arabic. A few text classification studies have been published earlier in the Arabic language. In general, researchers face two challenges in the Arabic text classification process: low accuracy and high dimensionality of the features. In this study, an Automated Arabic Text Classification using Hyperparameter Tuned Hybrid Deep Learning (AATC-HTHDL) model is proposed. The major goal of the proposed AATC-HTHDL method is to identify different class labels for the Arabic text. The first step in the proposed model is to pre-process the input data to transform it into a useful format. The Term Frequency-Inverse Document Frequency (TF-IDF) model is applied to extract the feature vectors. Next, the Convolutional Neural Network with Recurrent Neural Network (CRNN) model is utilized to classify the Arabic text. In the final stage, the Crow Search Algorithm (CSA) is applied to fine-tune the CRNN model’s hyperparameters, showing the work’s novelty. The proposed AATC-HTHDL model was experimentally validated under different parameters and the outcomes established the supremacy of the proposed AATC-HTHDL model over other approaches.