2019 Volume 33 Issue 9
Published: 20 September 2019
-
Select all|
-
Language Analysis and Calculation2019, 33(9): 1-8.In recent years, neural Chinese word segmentation (WS) models have achieved very high performance on closed domain texts. However, the performance drops dramatically in the domain adaptation scenario, where the test data is different from the training data. This paper attempts to improve cross-domain WS performance by employing the automatically collected WS data with partial annotation. Firstly, we extend the currently state-of-the-art BiLSTM-CRF WS model by introducing a new loss function to accommodate partially annotated data. Then, we propose a simple yet effective data selection method to extract target-domain related data from large amounts of partially annotated data. Finally, we employ a data preprocessing method and integrate traditional sparse features into the neural model, both leading to improved performance. The experimental results on the benchmark SIGHAN Bakeoff 2010 and ZhuXian datasets show that our proposed approaches effectively improve the performance of our baseline model by 3.6% in F1.
-
Language Analysis and Calculation2019, 33(9): 9-16,23.To improve the performance of Chinese word segmentation on specific domain, a domain adaption method of word segmentation is proposed based on deep learning and transfer learning. Firstly, a deep learning neural network of bidirectional long short-term memory CRF (BI-LSTM-CRF) model including a dictionary feature is constructed for Chinese word segmentation and trained on the general field corpus to obtain the model parameters. Secondly, the parameters of BI-LSTM-CRF model trained in a common domain corpus are fine-tuned using a small size of training corpus in construction law domain. The domain dictionary information is added to the dictionary feature. The experimental results show that transfer learning decreases the epochs for optimization. Compared with the BI-LSTM-CRF model trained in common domain, the proposed model increases the F1 by 7.02% in construction law domain. Compared with the BI-LSTM-CRF model using a domain dictionary in prediction process, the proposed model increases the F1 by 4.22%.
-
Language Analysis and Calculation2019, 33(9): 17-23.To deal with legal documents involving multi-domain texts, this paper proposes a cross-domain approach on Chinese word segmentation with joint learning. In the method, a large number of source domain samples are used to assist word segmentation in target domain through joint learning, which improves the performance of word segmentation. Experimental results demonstrate that, even with a few annotation samples from target domain, the performance of proposed method is obviously better than that of the traditional method.
-
Language Analysis and Calculation2019, 33(9): 24-30.At present, many deep neural network models deal with Chinese word segmentation tasks with bidirectional long short term memory neural network structure. Issues remain in the aspects that the input features are not rich enough, the semantic understanding is not complete, and the calculation speed is slow. In this paper, a dilated convolution neural networks for Chinese Word Segmentation is proposed. Chinese radical information is integrated to enrich the input and the convolution neural network is applied to extract the feature. The dilated convolution neural networks with residual structure can better understand semantic information and improve training efficiency. Experimented on the four datasets in Bakeoff 2005, the proposed method achieves better performance in terms of accuracy and efficiency compared with the bidirectional long short term memory neural network method.
-
Language Analysis and Calculation2019, 33(9): 31-38.Event coreference resolution is a complicated task in natural language understanding, which is to detect coreferential events in the text. This paper introduces a framework to resolve document-level event coreference, including the ENS_NN for event extraction and event realis detection, and the AGCNN for event coreference resolution. The AGCNN utilizes the attention pooling to capture global features in event sentences, and uses gated CNN to extract complicated semantic features, to improve performance of event coreference resolution. The experiments on the KBP2015 corpora and KBP2016 corpora show that our method achieves the state-of-the-art results.
-
Language Resources Construction2019, 33(9): 39-49.HSK is an international standardized test of Chinese language proficiency. The 650 “default words” listed in the appendix of the new HSK syllabus are often expanded via expert knowledge. Based on the Modern Chinese Dictionary, the Grammatical Knowledge-base of Contemporary Chinese and other data resources, this paper presents a systematic analogy for the vocabulary level of the new HSK syllabus using the method of knowledge engineering and iterative analogy rule of vocabulary association. By realizing the explicitization of implicit knowledge and the systematicization of decentralized knowledge, each procedure of the vocabulary level analogy can be assigned corresponding rules. Then, the results of the analogy are filtered by the constructed Chinese lexical knowledge base, resulting total 23762 words with the analogy level of. Finally, the statistical analysis shows that the vocabulary completed in this paper better demonstrate the guiding role of the new HSK syllabus in Chinese vocabulary grading and text difficulty grading.
-
Language Resources Construction2019, 33(9): 50-59,78.To construct the corpus of logical structure in re-flowable documents for machine learning, this paper proposed a three-stage semi-automatic annotation method based on the logical structure features and editing semantic features. In the first stage, document metadata is identified and annotated aided by the machine; in the second stage, the logical structure of the document is reconstructed automatically; finally, the feature vectors are automatically produced in the third stage. The experimental result shows that the proposed method can reduce manual costs, and the document corpus achieved can improve the accuracy of document structure recognition using machine learning algorithm up to 97.5% F-score.
-
Other Language in/around China2019, 33(9): 60-68.This paper proposes an Uyghur nouns anaphora resolution model ATT-IndRNN-CNN based on Attention Mechanism (ATT), Independently Recurrent Neural Network (IndRNN) and Convolutional Neural Network (CNN). According to the grammar and semantic structure of Uyghur, 17 rules and semantic information features are extracted. The attention mechanism is applied to select the features via the correlation between the features and the resolution results. The results are input into IndRNN and CNN to obtain the global features and local features in the context, respectively. Finally, the two types of features are merged and softmax is used to classify the resolution task. The experimental results show that the proposed method is better than the classical models, achieving the precision of 87.23%, the recall of 88.80%, and the F-measure of 88.04%.
-
Information Extraction and Text Mining2019, 33(9): 69-78.Keywords extraction is an important technique for web page retrieval,knowledge comprehension,and document classification,etc. In this paper,a novel keywords extraction method of combining graph structure with nodes association(GSNA) is proposed,which is able to locate keywords without a corpus. Firstly,the frequent closed itemset are exploited and the strong association rules are generated. Secondly,an association graph is constructed based on association rules,where the head and the body of the rules represent nodes,and an edge exists if and only if there is a strong association rule between two nodes and value of lift are adopted to represent weight. Thirdly,three node factors (i.e. graph structure,node semantics and associations) are unified under the same keyword extraction framework for random walking. Finally,a trustworthy sematic clustering algorithm is employed to avoid the semantic overlapping among terms. Three experiments conducted on the Chinese and English data sets show that GSNA is effective for keywords extraction.
-
Information Extraction and Text Mining2019, 33(9): 79-87.Event detection is one of the important tasks in the field of information extraction. In this paper, Chinese event detection is regarded as a sequence labeling rather than a classification problem. A Chinese event detection model ATT-BiLSTM is proposed, which integrates attention mechanism and long short-term memory neural network. The attention mechanism is used to better capture global features and BiLSTM layers are employed to capture sequence features more effectively. Experiments on the ACE 2005 Chinese dataset show that the performance of the proposed method significantly outperforms other existing Chinese event detection methods.
-
Information Extraction and Text Mining2019, 33(9): 88-95,106.Current research on automatic event extraction focuses on sentence-level corpus. However, due to the complexity and the diversity of event description in texts, a complete event is mentioned by multiple sentences in many cases. This paper first proposes an Attention-based Sequence Labeling model for joint extraction of entities and events. Compared with the pipeline of entity extraction plus event recognition, this joint labeling model improves the F-score by 1%. Then, we use Multi-Layer Perception to label the entities in the events and identify their roles. Finally, based on the labeling and identification results, this paper leverages integer linear programming for global reasoning, improving the F-score of document-level event extraction by 3% compared to the baseline.
-
Machine Reading Comprehension and Text Generation2019, 33(9): 96-106.Current image captioning is challenged the veracity of captions, i.e. an exact caption with tangible and specific entities is generated with a crude and monotonous captions ( e.g. “Messi takes the penalty kick” vs “a person is playing a ball.”). Focused on the identification and filling of person entities, this paper transform this task into a cloze issue with syntactic vacancy by removing the common person representation(e.g.“man”“player”) in the generated image caption. To introduce reading comprehension famework to address Who problem, this paper uses the R-Net to realize the acquisition and filling of the person name entity. In addition, we attempt to use the local and the global information to extract the person name entity, with local information indicating the source document that the image is located and the global information indicating the related documents from external links. Experiments show that the proposed method can effectively improve the quality of image caption generation and increase the BLEU by 2.93%.
-
Sentiment Analysis and Social Computing2019, 33(9): 107-114,140.With the rapid development of the Internet, more and more user comments appear on social networking sites. Review quality prediction aims to judge the quality of online reviews. To better build the representation of the text and study the user-based association between the text, this paper proposes a review quality prediction method of constructing the user and text representations based on neural network model. To properly emphasize the role of user information, we further integrate user information based on attention mechanism into the text to improve the effect of review quality prediction. Experiments on Yelp 2013 dataset show that our model can effectively improve the performance of online review quality detection.
-
Sentiment Analysis and Social Computing2019, 33(9): 115-122.Most of the existing information diffusion models in online social network (OSN) are focused on the network structure, with little attention on the influence of user attributes and information characteristics. This paper proposes an information diffusion model based on user attributes. Firstly, this paper extracts the characteristics of user influence, user attitude, user age, information energy, information value, and then constructs interaction rules. Secondly, a mathematical model of information diffusion is established to simulate the evolution of public opinion in social networks based on these characteristics. Finally, in order to verify the effectiveness of the proposed model, empirical analysis and comparison experiments with real online social network events are carried out, revealing a similarity between the simulation structure and real data beyond 0.97.
-
Sentiment Analysis and Social Computing2019, 33(9): 123-128.At present, deep learning models have achieved good results in text sentiment analysis. Following this thread, we propose a method based on recurrent entity networks for fine-grained sentiment analysis. In this method, predefined evaluation attribute categories information is embedded in networks, and the sentiment features about each attribute categories are extracted through the expanded internal memory chains. The dynamic memory unit controls the distant sentiment dependence about attribute categories. Finally, for a given attribute category, attention mechanism is applied to extract sentiment features from internal memory chains. In our experiment, the proposed method achieves a nearly one percentage point improvement in Sentihood data compared to the current highest accuracy method.
-
Sentiment Analysis and Social Computing2019, 33(9): 129-140.A new method for emotion analysis by automatic extraction of sentimental syntactic patterns is proposed. Based on the syntax dependency tree of the sentence to be analyzed, four basic operations are defined in this paper, i.e., branching, grafting, pruning and decomposing, according to Chinese grammatical characteristics. These operations can simultaneously compress the feature space of dependency tree and transform sentences into a set of subtrees representing syntactic relations. Finally, the genetic algorithm is used to obtain the optimal emotional subtrees. The experiments conducted on the evaluation data from NLPCC 2014 show that this method performs very well in judging whether a sentence expresses emotion or not. When combined with dictionary-based sentiment analysis, it can improve the performance of the dictionary-based sentiment analysis method by reducing its deficiency of high false positive rate on objective sentences.
