Search

Generating emotional responses plays an important role in human–computer conversational system. Adopting emotional information to the generation process improves user's satisfaction and makes the generated responses more human-like. Furthermore using fixed and unrelated emotion dictionary limits the overall performance of recent models. In order to generate and utilize the emotional words related to our datasets and sensitive to semantic importance, we propose a dynamic emotion dictionary module to generate a proper emotional expression. Considering the response sometimes focuses on the certain words of the input text, we also introduce dual-copying mechanism for taking advantages of not only emotion dictionary but also words appearing in the input text. Experimental results demonstrate that our model outperforms strong alternatives in several metrics, e.g. BLEU, embedding similarity, Rouge-L and the quality of human evaluation. [ABSTRACT FROM AUTHOR]

Nowadays many neural networks have been widely employed for semantic parsing problems especially for Structured Query Language (SQL) parsing, which aims at transforming natural language sentences into SQL representations. Selecting proper table headers in SQL tasks is extremely important, and the main cause of performance drop is that attention mechanism in neural models sometimes obtains wrong word-level distribution over headers. In order to obtain better header selection, we propose a match matrix aggregation enhanced SQL parser to consider the outer character-level ROUGE-L match information between the question and headers, then dynamically combine it with the inner generated attention matrix. We also introduce customized BERT and extra semantic information of the question and headers to further improve the performance. The results on two SQL datasets demonstrate that our method achieves an inspiring performance and highly outperforms other state-of-the-art alternatives. [ABSTRACT FROM AUTHOR]

Slot filling and intent detection are the basic and crucial fields of natural language processing (NLP) for understanding and analyzing human language, owing to their wide applications in real-world scenarios. Most existing methods of slot filling and intent detection tasks utilize linear chain conditional random field (CRF) for only optimizing slot filling, no matter the method is a pipeline or a joint model. In order to describe and exploit the implicit connections which indicate the appearance compatibility of different tag pairs, we introduce a graph-based CRF for a joint optimization of tag distribution of the slots and the intents. Instead of applying the complex inference algorithm of traditional graph-based CRF, we use an end-to-end method to implement the inference, which is formulated as a specialized multi-layer graph convolutional network (GCN). Furthermore, mask mechanism is introduced to our model for addressing multi-task problems with different tag-sets. Experimental results show the superiority of our model compared with other alternative methods. Our code is available at https://github.com/tomsonsgs/e2e-mask-graph-crf. [ABSTRACT FROM AUTHOR]

Emotion detection (ED) and emotion-cause pair extraction (ECPE) have drawn extensive research interests due to their wide applications in real-world scenarios. However, existing work fails to capture the implicit connection between two tasks. This limits the performances of these tasks. To address this issue, we propose a novel joint framework to take full advantage of the clause-level and word-level information about ED and ECPE tasks. Specifically, we explore a multi-level attentional module to model the relationship between two clauses in an emotion-cause pair. Results on two benchmark datasets show that our proposed model achieves the current best performance, outperforming the previous methods and strong neural baselines by a large margin. Our code is available at https://github.com/tomsonsgs/LVE-joint-MANN-master. [ABSTRACT FROM AUTHOR]

Sarcasm detection on twitter has received increasing research in recent years. However, existing work has two limitations. First, existing work mainly uses discrete models, requiring a large number of manual features, which can be expensive to obtain. Second, most existing work focuses on feature engineering according to the tweet itself, and does not utilize contextual information regarding the target tweet. However, contextual information (e.g. a conversation or the history tweets of the target tweet author) may be available for the target tweet. To address the above two issues, we explore the neural network models for twitter sarcasm detection. Based on convolutional neural network, we propose two different context-augmented neural models for this task. Results on the dataset show that neural models can achieve better performance compared to state-of-the-art discrete models. Meanwhile, the proposed context-augmented neural models can effectively decode sarcastic clues from contextual information, and give a relative improvement in the detection performance. [ABSTRACT FROM AUTHOR]

The task of event extraction consists of three subtasks namely entity recognition, trigger identification and argument role classification. Recent work tackles these subtasks jointly with the method of multi-task learning for better extraction performance. Despite being effective, existing attempts typically treat labels of event subtasks as uninformative and independent one-hot vectors, ignoring the potential loss of useful label information, thereby making it difficult for these models to incorporate interactive features on the label level. In this paper, we propose a joint label space framework to improve Chinese event extraction. Specifically, the model converts labels of all subtasks into a dense matrix, giving each Chinese character a shared label distribution via an incrementally refined attention mechanism. Then the learned label embeddings are also used as the weight of the output layer for each subtask, hence adjusted along with model training. In addition, we incorporate the word lexicon into the character representation in a soft probabilistic manner, hence alleviating the impact of word segmentation errors. Extensive experiments on Chinese and English benchmarks demonstrate that our model outperforms state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Implicit sentiment analysis is a challenging task because the sentiment of a text is expressed in a connotative manner. To tackle this problem, we propose to use textual events as a knowledge source to enrich network representations. To consider task interactions, we present a novel lightweight joint learning paradigm that can pass task-related messages between tasks during training iterations. This is distinct from previous methods that involve multi-task learning by simple parameter sharing. Besides, a human-annotated corpus with implicit sentiment labels and event labels is scarce, which hinders practical applications of deep neural models. Therefore, we further investigate a back-translation approach to expand training instances. Experiment results on a public benchmark demonstrate the effectiveness of both the proposed multi-task architecture and data augmentation strategy. [ABSTRACT FROM AUTHOR]

Chinese information extraction is traditionally performed in the process of word segmentation, entity recognition, relation extraction and event detection. This pipelined approach suffers from two limitations: 1) It is prone to introduce propagated errors from upstream tasks to subsequent applications; 2) Mutual benefits of cross-task dependencies are hard to be introduced in non-overlapping models. To address these two challenges, we propose a novel transition-based model that jointly performs entity recognition, relation extraction and event detection as a single task. In addition, we incorporate subword-level information into character sequence with the use of a hybrid lattice structure, removing the reliance of external word tokenizers. Results on standard ACE benchmarks show the benefits of the proposed joint model and lattice network, which gives the best result in the literature. [ABSTRACT FROM AUTHOR]

Scientific information extraction is a crucial step for understanding scientific publications. In this paper, we focus on scientific keyphrase extraction, which aims to identify keyphrases from scientific articles and classify them into predefined categories. We present a neural network based approach for this task, which employs the bidirectional long short-memory (LSTM) to represent the sentences in the article. On top of the bidirectional LSTM layer in our neural model, conditional random field (CRF) is used to predict the label sequence for the whole sentence. Considering the expensive annotated data for supervised learning methods, we introduce self-training method into our neural model to leverage the unlabeled articles. Experimental results on the ScienceIE corpus and ACL keyphrase corpus show that our neural model achieves promising performance without any hand-designed features and external knowledge resources. Furthermore, it efficiently incorporates the unlabeled data and achieve competitive performance compared with previous state-of-the-art systems. [ABSTRACT FROM AUTHOR]

Negation and speculation scope detection is an important task in natural language processing. Previous studies all show that syntactic information is crucial to the task. However, these work mainly focuses on human-designed discrete features and local features extracted from dependency tree, limiting the performance of the task. In this paper, we propose a recursive neural network sequence labeling model, representing whole dependency tree globally and learning automatically syntactic features, for the task. Specifically, recursive neural network first learns a high-level representation for the words in the context of each sentence, and captures the cue word with its target scope through the global dependency structure. Then, CRFs layer takes the representation from recursive neural network as input to jointly decode labels for the whole sentence. Experimental results on English dataset BioScope and Chinese dataset CNeSp show that our model outperforms the state-of-the-art systems. [ABSTRACT FROM AUTHOR]

Transfer learning has shown promising results for transferring knowledge ofsource tasks to target tasks in natural language processing (NLP). In this paper, we investigate a multi-task and multi-view learning (MTMVL) framework for end-to-end neural relation extraction, using large noisy data as one auxiliary task for improving manually labeled data, and a dependency parsing objective as a second auxiliary task for leveraging syntax. Two views are taken for relation extraction, consisting of a joint tagging view and a novel context relation view. To capture sentence multi-level information, we explore a weighted average approach to represent shared deep bi-directional recurrent neural networks and encourage auxiliary tasks to accommodate relation extraction task via a time-related scheduled sampling strategy. We evaluate MTMVL framework on manual-labeled ACE2005 dataset, and experimental results show that our model outperforms the state-of-the-art methods, which indicates the effectiveness of multi-auxiliary information for knowledge transferring. [ABSTRACT FROM AUTHOR]

Extracting knowledge from time series provides important tools for many real applications. However, many challenging problems still open due to the stochastic nature of large amount of time series. Considering this scenario, new data mining and machine learning techniques have continuously developed. In this paper, we study time series based on its topological features, observed on a complex network generated from the time series data. Specifically, we present a trend detection algorithm for stochastic time series based on community detection and network metrics. The proposed model presents some advantages over traditional time series analysis, such as adaptive number of classes with measurable strength and better noise absorption. The appealing feature of this work is to pave a new way to represent time series trends by communities of complex networks in topological space instead of physical space (spatial–temporal space or frequency spectral) as traditional techniques do. Experimental results on artificial and real data-sets shows that the proposed method is able to classify the time series into local and global patterns. As a consequence, it improves the predictability on time series. [ABSTRACT FROM AUTHOR]

Topic and sentiment joint modelling has been successfully used in sentiment analysis for product reviews. However, the problem of text sparse is universal with the widespread smart devices and the shorter product reviews. In this paper, we propose a joint sentiment-topic model WSTM (Word-pair Sentiment-Topic Model) for the short text reviews, detecting sentiments and topics simultaneously from the text, especially considering the text sparse problem. Unlike other topic models modelling the generative process of each document, our directly models the generation of the word-pair set from the whole global corpus. In the generative process of WSTM, all of the words in a sentence have the same sentiment polarity, and two words in a word-pair have the same topic. We apply WSTM to two real-life Chinese product review datasets to verify its performance. In three experiments, compared with the existing approaches, the results demonstrate WSTM is quantitatively effective on both topic discovery and document level sentiment. [ABSTRACT FROM AUTHOR]

We investigate distantly supervised relation extraction with knowledge-guided latent graphs and an iterative graph learner. For the relation extraction tasks, we assume that the input sentences contain latent graphs with useful structural information between mentions and relations of the distantly supervised data. We first embed input sentences with default initialized graphs to utilize this information. We also used the pre-trained Knowledge Bases (KB) to guide the latent space of a Variational Graph Auto-Encoders (VGAE) module. Then the VGAE applies the re-parameterization mechanism to reconstruct the initial graph, injecting extra knowledge into the initial latent graph and reducing its unsteadiness. Subsequently, we optimize the latent graph structures and their corresponding node embeddings simultaneously by the iterative graph learner, and the better latent graphs can improve the downstream relation extraction task. Experiment results on three datasets (NYT10, WIKIDISTANT and GIDS) show that latent graph learning helps our model perform better than previous works. [ABSTRACT FROM AUTHOR]

Click-through rate (CTR) prediction can provide considerable economic and social benefits. Few studies have considered the importance of low-order features, usually employing a simple feature interaction method. To address these issues, we propose a novel model called Senet and extreme deep field-embedded factorization machine (Se-xDFEFM) for more effective CTR prediction. We first embed the squeeze-excitation network (Senet) module into Se-xDFEFM to complete low-order feature refinement, which can better filter noisy information. Then, we implement our field-embedded factorization machine (FEFM) to learn the symmetric matrix embeddings for each field pair, along with the single-vector embeddings for each feature, which builds a firm foundation for the subsequent feature interaction. Finally, we design a compressed interaction network (CIN) to realize feature construction with definite order through a vector-wise interaction. We use a deep neural network (DNN) with the CIN to simultaneously implement effective but complementary explicit and implicit feature interactions. Experimental results demonstrate that the Se-xDFEFM model outperforms other state-of-the-art baselines. Our model is effective and robust for CTR prediction. Importantly, our model variants also achieve competitive recommendation performance, demonstrating their scalability. [ABSTRACT FROM AUTHOR]

The task of Cross-Lingual Event Extraction (CLEE) aims to transfer event knowledge from rich-resourced languages to low-resourced languages under a zero-shot cross-lingual setting. Despite presenting successful transfer abilities, existing CLEE methods face two limitations. Firstly, a majority of works concentrate on single source CLEE ignoring the benefits of multi-source potentials. Secondly, universal dependency trees have been adopted for learning shared syntactic features across languages while leaving exploration of language-specific features (e.g. word order) insufficient. In this work, we investigate the effectiveness of multi-source CLEE and propose a Syntax-enhanced Parameter Generation Network (SPGN) for the task. SPGN mainly consists of two components, i.e., a parameter generation tree network for determining language-private syntactic knowledge, and an adversarial network for learning language-shared representations. Experiments on widely used ACE2005 and recently released MINION dataset show that our proposed method significantly outperforms existing baseline systems on total 10 languages. Specifically, we observed an improvement of 1.9% in average F1-score over the best-performing baseline on ACE2005 and 3.5% on MINION, highlighting the benefits of leveraging both language-specific and shared features in CLEE. Further analysis confirms the individual effectiveness of each component of the SPGN model, suggesting its potential applicability in other low-resourced language processing tasks. [ABSTRACT FROM AUTHOR]

The pair-wise aspect and opinion term extraction (PAOTE) task aims to extract aspect terms and opinion terms from reviews in the form of opinion pairs, which provides a global profile for reviews of goods or users. Up-to-date studies ignore the interaction between term detection and term pairing, which may be crucial for the PAOTE task. Other studies use syntactic dependency structures to enhance their models, which cannot better provide task-specific structural information. In this work, we design an aspect-to-opinion graph and transform PAOTE into a graph parsing task. To exploit the interaction between term detection and pairing, we propose a novel mutually-aware interaction network (MAIN), which interactively updates the representations for term detection and pairing via graph sampling and convolution. Further, the word-word graph learned during training can be iteratively refined and gradually approaches the aspect-to-opinion graph. Experimental results on four benchmark datasets show that our proposed method significantly outperforms strong baselines with state-of-the-art performance and achieves a maximum increase of 2.01 points on the F1 metric. Further analysis demonstrates the advance of the aspect-to-opinion graph and the effectiveness of the mutually-aware interaction mechanism. [ABSTRACT FROM AUTHOR]

Currently, there are many researches on new word discovering, tagging the orientation of sentiment word and building sentiment word dictionary, but few about the way of finding sentiment word. Therefore this paper proposed a method for discovering new sentiment words based on OC-SVM. The method firstly acquired the words set by using the seed words extending algorithm, then filtered it through word frequency, stop word and an old dictionary, so as to get new words. Under the appropriate F value, the experimental evaluation result can reach correct rate at 45.5%.Due to the imbalance of sentiment words and other words in the training set, so it used the word frequency, adjacent words and the part of speech, etc. as features of OC-SVM (one-class support vector machine) classification, built an effective new sentiment word discovery system. When experimental results of recall rate is 26.6%, the correct rate can reach 45.7%, which demonstrates the effectiveness of the algorithm. [ABSTRACT FROM AUTHOR]

This paper developed easy arc first algorithm combined with SVM, and used it in dependency parsing of Chinese complex nominal phrase. In each iteration step, the algorithm found out the optimized non-directional arc among arcs linking neighboring subtrees according to their features, and determined the arc' s direction using SVM. Results of experiments show that accurate rate of the algorithm is significantly higher than easy first arc algorithm, and higher than Chinese parser based on MST algorithm. Time complexity of the algorithm is O(n² log n). [ABSTRACT FROM AUTHOR]

Training an effective image sentiment analysis model using high-quality samples and the implicit cross-modal semantics among heterogeneous features is still challenging. To address this problem, we propose an active sample refinement (ASR) strategy to obtain sufficient high-quality images with definite sentiment semantics. We mine the cluster correlation among the heterogeneous SENet features. Discriminative cross-modal semantics is generated to train an effective but robust image classifier. Ensemble learning is employed to further boost performance. Our method outperforms other competitive baselines, demonstrating its effectiveness and robustness. Meanwhile, the ASR strategy is a useful supplement to the current data augmentation method. [ABSTRACT FROM AUTHOR]

The aspect-based sentiment triplet extraction (ASTE), as a complete sentiment analysis task, aims to recognize the aspect term, the opinion expression, and the sentiment polarity in a sentence. Current state-of-the-art ASTE models employ a joint extracting scheme for better task improvements. However, how to better solve the triplet overlap issues in the task, and effectively model the mutual interactions between the triplet structures remain challenging. In this work, we explore a neural transition model for end-to-end ASTE. We model the triplet prediction as a graph structure, based on which we implement a transition system with neural design. We further propose a triplet memory mechanism to fully leverage the underlying interactions from the previously recognized triplets relevant to the current parse. We experiment on the benchmark datasets, and the results show that our model achieved state-of-the-art performances against current baselines, meanwhile being more efficient on decoding. Further analysis is conducted to verify the effectiveness of our transition framework. [ABSTRACT FROM AUTHOR]

Material images are susceptible to changes, depending on the light intensity, visual angle, shooting distance, and other conditions. Feature learning has shown great potential for addressing this issue. However, the knowledge achieved using a simple feature fusion method is insufficient to fully represent the material images. In this study, we aimed to exploit the diverse knowledge learned by a novel progressive feature fusion method to improve the recognition performance. To obtain implicit cross-modal knowledge, we perform early feature fusion and capture the cluster canonical correlations among the state-of-the-art (SOTA) heterogeneous squeeze-and-excitation network (SENet) features. A set of more discriminative deep-level visual semantics (DVSs) is obtained. We then perform gene selection-based middle feature fusion to thoroughly exploit the feature-shared knowledge among the generated DVSs. Finally, any type of general classifier can use the feature-shared knowledge to perform the final material recognition. Experimental results on two public datasets (Fabric and MattrSet) showed that our method outperformed other SOTA baseline methods in terms of accuracy and real-time efficiency. Even most traditional classifiers were able to obtain a satisfactory performance using our method, thus demonstrating its high practicality. [ABSTRACT FROM AUTHOR]

Motivation Entity relation extraction is one of the fundamental tasks in biomedical text mining, which is usually solved by the models from natural language processing. Compared with traditional pipeline methods, joint methods can avoid the error propagation from entity to relation, giving better performances. However, the existing joint models are built upon sequential scheme, and fail to detect overlapping entity and relation, which are ubiquitous in biomedical texts. The main reason is that sequential models have relatively weaker power in capturing long-range dependencies, which results in lower performance in encoding longer sentences. In this article, we propose a novel span-graph neural model for jointly extracting overlapping entity relation in biomedical texts. Our model treats the task as relation triplets prediction, and builds the entity-graph by enumerating possible candidate entity spans. The proposed model captures the relationship between the correlated entities via a span scorer and a relation scorer, respectively, and finally outputs all valid relational triplets. Results Experimental results on two biomedical entity relation extraction tasks, including drug–drug interaction detection and protein–protein interaction detection, show that the proposed method outperforms previous models by a substantial margin, demonstrating the effectiveness of span-graph-based method for overlapping relation extraction in biomedical texts. Further in-depth analysis proves that our model is more effective in capturing the long-range dependencies for relation extraction compared with the sequential models. Availability and implementation Related codes are made publicly available at http://github.com/Baxelyne/SpanBioER. [ABSTRACT FROM AUTHOR]

Biomedical information extraction (BioIE) is an important task. The aim is to analyze biomedical texts and extract structured information such as named entities and semantic relations between them. In recent years, pre-trained language models have largely improved the performance of BioIE. However, they neglect to incorporate external structural knowledge, which can provide rich factual information to support the underlying understanding and reasoning for biomedical information extraction. In this paper, we first evaluate current extraction methods, including vanilla neural networks, general language models and pre-trained contextualized language models on biomedical information extraction tasks, including named entity recognition, relation extraction and event extraction. We then propose to enrich a contextualized language model by integrating a large scale of biomedical knowledge graphs (namely, BioKGLM). In order to effectively encode knowledge, we explore a three-stage training procedure and introduce different fusion strategies to facilitate knowledge injection. Experimental results on multiple tasks show that BioKGLM consistently outperforms state-of-the-art extraction models. A further analysis proves that BioKGLM can capture the underlying relations between biomedical knowledge concepts, which are crucial for BioIE. [ABSTRACT FROM AUTHOR]

Reference 1 Huang W, Zhang J, Ji D (2020) A transition-based neural framework for Chinese information extraction. Junchi Zhang should no longer be denoted as a corresponding author, and Donghong Ji should be denoted as the sole corresponding author. [Extracted from the article]

Event extraction in the biomedical domain has many complex situations, such as nested events, overlapping events, and multiple processing streams. For the problem of massive nested and multi-level nested events in the corpus, existing approaches generally transform these nested events into judgments of the relationships between triggers, resulting ineffective utilization of the complete event information. Similarly, when confronted with the issue of overlapping events in the corpus, the prevailing approach often recognizes only one of the overlapping events while neglecting the other. Due to the complexity of the biomedical events, their effectiveness is unsatisfactory. Biomedical events, which contain nested and overlapping events, can be conceptualized as a tree structure. In light of this, we present a novel tree-like structured perceptron designed for transition-based biomedical event extraction, which converts event predictions into sequences of transition operations. This approach enables the incremental construction of nested events, with each operation relying on the already constructed event and surrounding contextual features. We evaluate the proposed model on two tasks: Genia 2011 and Genia 2013. The experimental results show a 0.22% and 0.04% improvement in F1 score, respectively, compared to the currently better performing baseline models without using external resources. • A structured perceptron for transition-based biomedical event extraction is proposed. • The model can solve various complex structures of biomedical events. • Beam search and global contrastive learning to enhance the model's capabilities. [ABSTRACT FROM AUTHOR]

The extraction of opinions and their corresponding targets has gained significant interest recently, as it offers valuable insights into Opinion Mining (OM) at a granular level. Opinion and target terms to be extracted by existing OM tasks need to be explicitly present in reviews. Targets that are not present but implied in contextual semantics, are neglected by existing OM tasks, even though an investigation reported that about 60% of reviews contain implicit targets. To enable implicit target extraction, a novel task named Mining Opinions towards Implicit Targets (MOIT) under the fine-grained OM, is proposed to extract both opinions and their corresponding implicit targets, enabling a more comprehensive analysis of reviews. To set up the basis for follow-up research on MOIT, two large-scale datasets were constructed as resources in two languages, where the Chinese dataset was built from scratch via a standard human annotation process, and the English dataset was built semi-automatically through machine translation and manual checking. Furthermore, three baseline models adapting three representative paradigms of information extraction, namely sequence labeling, question answering, and text generation, were proposed to solve MOIT. Extensive experiments demonstrated the effectiveness of the models. The proposed MOIT task extends the field of OM research, and the datasets and models establish a foundation for future studies in this area. [Display omitted] • A novel task focused on extracting implicit targets is introduced. • Two large-scale datasets were constructed to facilitate the study. • Three baseline models were proposed; the generation-based model performed best. [ABSTRACT FROM AUTHOR]

While existing studies have established the relationship between electronic word-of-mouth (eWOM) and studio performance, limited research has been conducted to demonstrate how the attention-based model applies to the motion picture industry. In this study, examining a review corpus of seven Hollywood studios, we proved that deep learning with the attention mechanism has the best accuracy in both eWOM and stock price movement. We present both a hierarchical two-layer attention network and hierarchical convoluted attention network (HCAN), which quantify the importance of crucial eWOM features in capturing valuable information from audience members' reviews. Further, comparing the two case studies, we determined that the HCAN model is superior to both machine learning and attention-based models. Our work helps to highlight the business value of the attention-based model and has implications for studio business decisions. [ABSTRACT FROM AUTHOR]

Automatic caption generation from images is an interesting and mainstream direction in the field of machine learning. This method enables us to build a powerful computer model that can interpret the implicit semantic information of images. However, the current state of research faces significant challenges such as those related to extracting robust image features, suppressing noisy words, and improving a caption's coherence. For the first problem, a novel computer vision system is presented to create a new image feature called MK–KDES-1 (MK–KDES represents Multiple Kernel–Kernel Descriptors) after extracting three KDES features and fusing them by MKL (Multiple Kernel Learning) model. The MK–KDES-1 feature captures both textural characteristics and shape characteristics of images, which contribute to improving the BLEU_1 (BLEU represents Bilingual Evaluation Understudy) scores of captions. For the second problem, an effective newly designed two-layer TR (Tag Refinement) strategy is integrated into our NLG (Natural Language Generation) algorithm. Words that are most relevant semantically to images are summarized to generate N-gram phrases. Noisy words are suppressed using the innovative TR strategy. For the last problem, on the one hand, a pop WE (Word Embeddings) model and a novel metric called PDI (Positive Distance Information) are introduced together to generate N-gram phrases. The phrases are evaluated by the AWSC (Accumulated Word Semantic Correlation) metric. On the other hand, the phrases are fused to generate captions by the ST (Syntactic Trees). Experimental results demonstrate that informative captions with high BLEU_3 scores can be obtained to describe images. [ABSTRACT FROM AUTHOR]

Named entity recognition (NER) is an important task for biomedical text mining. Most prior work focused on recognizing regular entities that consist of continuous word sequences and are not overlapped with each other. In this paper, we propose a neural network model called Bi-LSTM-CRF that consists of bidirectional (Bi) long short-term memories (LSTMs) and conditional random fields (CRFs) to identify regular entities and the components of irregular entities. Then the components are combined to build final irregular entities according to manually designed rules. Furthermore, we propose a novel model called NerOne that consists of the Bi-LSTM-CRF network and another Bi-LSTM network. The Bi-LSTM-CRF network performs the same task as the aforementioned model, and the Bi-LSTM network determines whether two components should be combined. Therefore, NerOne automatically combines the components instead of using manually designed rules. We evaluate our models on two datasets for recognizing regular and irregular biomedical entities. Experimental results show that, with less feature engineering, the performances of our models are comparable with those of state-of-the-art systems. We show that the method of automatically combining the components is as effective as the method of manually designing rules. Our work can facilitate the research on biomedical text mining. [ABSTRACT FROM AUTHOR]

Existing studies learn sentiment-specific word representations to boost the performance of Twitter sentiment classification, via encoding both n-gram and distant supervised tweet sentiment information in learning process. Pioneer efforts explicitly or implicitly assume that all words within a tweet have the same sentiment polarity as that of the whole tweet, which basically ignores the word its own sentiment polarity. To alleviate this problem, we propose to learn sentiment-specific word embedding by exploiting both the lexicon resource and distant supervised information. In particular, we develop a multi-level sentiment-enriched word embedding learning method, which employs a parallel asymmetric neural network to model n-gram, word-level sentiment, and tweet-level sentiment in the learning process. Extensive experiments on standard benchmarks demonstrate our approach outperforms state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Conventional models relying on similarity utilizing low-level surface statistical, syntactic and lexical semantic features are suboptimal in contradiction recognition, especially for the large-scale data, such as the sensor and traffic data. To tackle this problem, this work treats the text and hypothesis pair as event graph and proposes a novel model based on event graph to recognize contradiction in big data collection. The proposed model is capable of seamlessly sewing the high-level event semantic features corresponding to the conflicting linguistic phenomena to identify contradictory construction. Experimental results show that our event graph based contradiction recognition model achieves significant improvement as compared to state-of-the-art competitors. [ABSTRACT FROM AUTHOR]

Code-switched emotion detection, a task analyzing the emotion in code-switched texts, has gain increasing research attention within recent years. Prior works utilize various neural models with sophisticated features to pursuit high performance of the task, while they still overlook some crucial characteristics of the code-switched texts. In this work, we present an innovative approach for improving code-switched emotion detection. We first consider a bilingual-view parallel translation for code-switched text enhancement, i.e., translating the code-switched texts into two languages. Then we propose an adversarial dual-channel encoder architecture, where two private encoders take as inputs the parallel texts in two languages, respectively. The private encoders and the shared encoder work collaboratively, and effectively retrieve the features from monolingual and bilingual perspectives under adversarial training. We conduct extensive experiments on five code-switched benchmark datasets. Results show that our model outperforms the strongly-performing baselines that leverage external code-switched or bilingual word embedding with over 1.5% F1 score on the Chinese–English, Spanish–English and Hindi–English code-mixed data, becoming the new state-of-the-art system. Further analyses including ablation, qualitative and error studies, demonstrate the effectiveness of our proposed encoder for code-switched texts, as well as the bilingual-view parallel translation strategy. [ABSTRACT FROM AUTHOR]

The recognition of textual entailment (RTE) as the main text understanding task is crucial to the application in biomedical and clinical field, however, the developing of which has been hindered, due to the scarcity of the data annotation. In this work, we propose a domain adaptation framework for the cross-domain clinical RTE. We first construct a hierarchical feature encoder architecture for fully exploring the interactions between the input sentence pair. We then establish shared and private feature extractors based on the feature encoder, for capturing both the domain-specific and domain-invariant features. We further introduce a domain discriminator with the adversarial training algorithm for enhancing the cross-domain transferring. Based on the real-world Chinese dataset, our framework achieves significantly enhanced performances against baseline domain adaptation methods, on the few-shot and zero-shot transferring settings. Further analysis reveals that our model is effective for the cross-domain clinical RTE. [ABSTRACT FROM AUTHOR]

Background: Disease prediction based on Electronic Health Records (EHR) has become one hot research topic in biomedical community. Existing work mainly focuses on the prediction of one target disease, and little work is proposed for multiple associated diseases prediction. Meanwhile, a piece of EHR usually contains two main information: the textual description and physical indicators. However, existing work largely adopts statistical models with discrete features from numerical physical indicators in EHR, and fails to make full use of textual description information.Methods: In this paper, we study the problem of kidney disease prediction in hypertension patients by using neural network model. Specifically, we first model the prediction problem as a binary classification task. Then we propose a hybrid neural network which incorporates Bidirectional Long Short-Term Memory (BiLSTM) and Autoencoder networks to fully capture the information in EHR.Results: We construct a dataset based on a large number of raw EHR data. The dataset consists of totally 35,332 records from hypertension patients. Experimental results show that the proposed neural model achieves 89.7% accuracy for the task.Conclusions: A hybrid neural network model was presented. Based on the constructed dataset, the comparison results of different models demonstrated the effectiveness of the proposed neural model. The proposed model outperformed traditional statistical models with discrete features and neural baseline systems. [ABSTRACT FROM AUTHOR]