2020 Volume 34 Issue 7
Published: 10 August 2020
-
Select all|
-
Survey2020, 34(7): 1-18.Machine translation is a task which translates a source language into a target language of the equivalent meaning via a computer, which has become an important research direction in the field of natural language processing. Neural machine translation models, as the main stream in the reasearch community, can perform end-to-end translation from source language to target language. In this paper, we select several main research directions of neural machine translation, including model training, simultaneous translation, multi-modal translation, non-autoregressive translation, document-level translation, domain adaptation, multilingual translation, and briefly introduce the research progresses in these directions.
-
Survey2020, 34(7): 19-29.In recent years, increasing attention has been attracted to the research field related to cross-modality, especially vision and language. This survey focuses on the task of image captioning and summarizes literatures from four aspects, including the overall architecture, some key questions for cross-modality research, the evaluation of image captioning and the state-of-the-art approaches to image captioning. In conclusion, we suggest three directions for future research, i.e., cross-modality representation, automatic evaluation metrics and diverse text generation.
-
Survey2020, 34(7): 30-41.Syntactic parsing aims to analyze an input sentence for its syntactic structure. It is one of the most classic tasks in natural language processing. Current researches of syntactic parsing are focused on improving the accuracy of syntactic parsers via automatic learning from data. This paper surveys recent developments in syntactic parsing, classifies and introduces the new approaches and new discoveries over the past year in three subareas (supervised parsing, unsupervised parsing, and cross-domain/cross-language parsing), and finally discusses the future perspective of syntactic parsing research.
-
Knowledge Representation and Acquisition2020, 34(7): 42-49,78.Entity disambiguation is the process of linking recognized entity mentions to its corresponding entry in a particular knowledge base. This paper combines the mainstream deep learning-based entity disambiguation method and the entity knowledge description. Experiments demonstrate that the proposed method obtains competitive or state-of-the-art F1 at public datasets.
-
Knowledge Representation and Acquisition2020, 34(7): 50-59.Representation learning of knowledge graph aims to project entities and relations into continuous low-dimensional vector space. Most existing translation-based representation methods, such as TransE、TransH and TransR, usually utilize only triples of knowledge graph, and fail to deal with complex relationships such as one-to-many, many-to-one, and many-to-many. To address this issue, this paper proposes a representation learning model of knowledge graph integrating entity description and type, which is called TDT model. Firstly, the Doc2Vec model is used to obtain the embedding of all entity descriptions. Secondly, treating the hierarchical types as projection matrices for entities, the embedding of entity type information can be obtained via multiplying the projection matrix with triple embedding. Finally, TDT model integrates the information of triple(T), entity description(D), and entity type information(T) in a low-dimensional vector space. This paper evaluates TDT model via the experiments of link prediction and triple classification on the real-world datasets. The results show that new method significantly outperforms other baselines, such as TransE, TransR, DKRL and SimplE etc.
-
Machine Translation2020, 34(7): 60-67.For sentence-level neural machine translation, the problem of incomplete semantic representation is noticeable since the context information of the current sentence is not considered. We extract effective information from each sentence in a document by dependency parsing, and then complement the extracted information into the source sentences, making the semantic representation of the sentences more complete. We conduct experiments on Chinese-English language pair, and propose a training method on large-scale parallel language pairs for the scarcity of document-level parallel corpus. Compared with the baseline model, our approach improves 1.47 BLEU significantly. Experiments show that the document-level neural machine translation based on context recovery can effectively solve the problem of incomplete semantic representation of sentence-level neural machine translation.
-
Machine Translation2020, 34(7): 68-78.Visual semantic information can improve the performance of machine translation. However, most of the existing work incorporate the overall visual semantic information of the image into the translation model, ignoring the possible different local semantic object features. To deal with this issue, a multimodal machine translation model incorporating image attention is proposed in this paper. We incorporate the interaction information between local and global image visual information with the words of source language as an image attention into the traditional textual attention, for better alignment from hidden states of the decoder to the source words. We carry several experiments on Multi30k dataset, the results on English-German and Indonesian-Chinese tasks (the latter is annotated by human manually) show that our model has a good improvement compared with the existing recurrent neural network based multimodal machine translation model.
-
Information Extraction and Text Mining2020, 34(7): 79-88.As a controllable text generation task,text style transfer has attracted more and more attention in recent years. Based on the variational auto-encoder model, the content and style of source sentences are separated in the latent space through the adversarial training between the discriminator and the variational auto-encoder. Due to the defect in the method using fixed binary vector for the style representation, we proposed a more fine-grained joint representation method which combines the latent variable extracted from an independent encoder with a style label to improve the accuracy of style transferation. Experimental results show that the joint representation method achieves higher accuracy compared with two baseline models on Yelp, a common dataset in the style transfer field.
-
Information Retrieval and Question AnsweringQ2SM: Dialogue State Tracking Algorithm Based on BERT for Multi-Domain Task-Oriented Dialogue System2020, 34(7): 89-95.Dialogue State Tracking(DST) is the core task in task-oriented dialogue system. For multi-domain task dialogue system, this paper introduces a novel model named Q2SM(Query to State Model) based on BERT. The upstream of Q2SM is a model based on sentence representation with BERT and similarity, while the downstream is a new type of RNN named DST-RNN. Experiments on WOZ2.0 and MultiWOZ2.0 datasets show that the model outperforms the existing state-of-the-art model by 1.09% in joint-accuracy and 2.38% in F1 score. And the model ablation study show that the DST-RNN could also speed up model convergence.
-
Sentiment Analysis and Social Computing2020, 34(7): 96-104.In the sentiment analysis in micro-blogs, most existing topic sentiment models do not fully consider the users personality characteristics. Based on the JST model, this paper proposes a time-based personality modeling method to incorporate users personality features into the topic sentiment model. Since the microblog data contains a lot of unique information such as emoticons, we also introduce emoticons into the JST model. As a result, an probabilistic model named UC-JST(Joint Sentimet/Topic model based on User Character)is proposed. Tested on the real Sina Weibo dataset, the results show that UC-JST performs better than JST, TUS-LDA ,JUST and TSMMF in terms of sentiment classification accuracy.
-
NLP Application2020, 34(7): 105-112.The traditional lexicography was mainly subject to manual compilation, which is inefficient and costs a lot of resources. This paper proposes a gated context-aware network for definition generation. It utilizes GRU to model the definitions of words and generates the textual definition for the target word automatically. The model is based on the encoder-decoder architecture. Firstly, the context of the target word is encoded by bidirectional GRU. Then, different matching strategies are used to interact the target word with context and the context information is incorporated into the target word embedding from two aspects of coarse-grained and fine-grained by the attention mechanism to obtain the meaning of the target word in a specific context. The decoding process based on the contextual and semantic information to generate context-dependent definition of the target word. In addition, the quality of generated definitions is further improved by providing the character level information of target words. The experimental results show that the proposed model improves the perplexity of definition modeling and the BLEU score of definition generation on the English Oxford dictionary dataset by 4.45 and 2.19 respectively, and can generate readable and understandable definitions.
