2021 Volume 35 Issue 10
Published: 15 October 2021
-
Select all|
-
Survey2021, 35(10): 1-20.Recently, the artificial intelligence-based dialogue system has been widely applied in, human-computer interaction, intelligent assistant, smart customer service, Q&A consulting, and so on. This paper proposes a definition about the task-oriented dialogue system, which is to satisfy the user’s requirements and certain tasks with the least response turns in dialogue between human and machine. Furthermore, three critical technical problems and challenges are summarized: the user’s intent detection in the complex context, the limitation annotated data, and the personalized response under the multi-modal situation. The research progress in these three challenges are discussed in the paper. Finally, we outline the research directions in the future and the key issues in the next generation of task-oriented dialogue system.
-
Language Analysis and Calculation2021, 35(10): 21-31.Topic Continuity and Topic Shift are important pragmatic functions in discourse. From the perspective of topic sharing at the beginning of a punctuation clause (P-Clause), this paper classifies the Topic Continuity and Topic Shift into five categories: Topic Continuity at the beginning of a P-Clause, Sub-topic Continuity among a P-Clause, Completely Topic Shift, Pivotal Topic Shift and New Branch Topic Shift. Then focusing on the New Branch topic, we conduct a statistic analysis of the syntactic constituents and semantic roles of the New Branch topic based on a corpus of Generalized Topic Structure with 330, 000 characters. As for syntactic constituents, the subject of an object clause or a complement clause, the small subject of a sentence with S-V structured predicate, the subject of an adverbial starting sentence, the object at the end of a sentence, the object of a co-verbal sentence, the pivot of a pivotal sentence, the prepositional object, and even the adverbial can serve as New Branch topics, which introduce one or more New Branch clauses. Among them, the object at the end of a sentence is the most frequently used as a New Branch topic, but no indirect object was found to be a New Branch topic. As for semantic roles, most of subject arguments (agent, sentiment, experiencer, theme) and object arguments (patient, relative, result, target, dative) are found to be New Branch topics. As for the supporting arguments (manner) and environment arguments (location, goal), a few of them can function as New Branch topics and lead to New Branch clauses. Among them, relative and patient are the most typical New Branch topics, followed by agent, result and target, but cause and aim can hardly function as New Branch topics. Our study reveals a possible path for syntactic and semantic constraints on Topic Shift.
-
Language Analysis and Calculation2021, 35(10): 32-38.160 paraphrase texts of native Chinese speakers are compared with the original text, and 6, 484 pairs of paraphrase sentences are extracted. From generation methods, two generation methods can be categorized: the change of words and the recasting of whole sentences. From the perspective of pragmatic principles, it is found that the paraphrase differs from previous studies in that sentence pairs often do not have the same logical semantic truth value, while they convey the same pragmatic meaning and have equivalent pragmatic functions in a specific context. This shows that the recognition of paraphrase sentences in real communication depends not only on the knowledge base of grammar and semantics, but also on the knowledge base containing pragmatic knowledge and contextual information.
-
Language Resources Construction2021, 35(10): 39-47,55.The construction of the moral dictionary is a fundamental resource for artificial intelligence ethical computing. Moral behavior is complex and varied, and the taxonomy of English moral dictionary is still under development. As an endeavor to build the first Chinese dictionary resources in this aspect, this paper proposes the task of constructing a Chinese moral dictionary for artificial intelligence ethics calculation. Four polar labels and four type labels are designed to for the finally achieved Chinese moral dictionary of 25, 012 words. Experimental results show that the dictionary resource be applied for moral knowledge learning and moral word tagging, thus providing data support for the analysis of moral text at the sentence level.
-
Knowledge Representation and Acquisition2021, 35(10): 48-55.TransC is an efficient method for embedding knowledge graphs. It establishes the embedding of concepts, instances, and relations by distinguishing concepts and instances. TransC encodes the concept as a sphere, and the radius of the sphere is randomly initialized and updated iteratively during training.This leads to two problems in the model. First, part of the sphere radius obtained from training does not match the model training target. Second, the semantic information provided by the concept itself is ignored. This paper proposes a model named TransIC to deal with the two issues above. TransIC adopts a novel concept sphere radius solution method based on IC parameters, so that the obtained radius meets the TransC goal, and enriches the semantic information of the concept embedding vector. Then it is based on TransC and introduces a concept sphere radius based on IC parameters during the concept coding phase.Finally, the two tasks of link prediction and triple classification are completed on the public data set YAGO39K, and the experimental performance of the method in this paper is compared with the performance of TransC and other models. The results show that TransIC has achieved a significant improvement in most indicators.
-
Knowledge Representation and Acquisition2021, 35(10): 56-63.There are a large amount of logical knowledge that portray the logical evolutionary relationships between things in the open texts. Logical knowledge bases are an important foundation to advancing the knowledge reasoning, the development of large-scale logical reasoning knowledge bases can help support conduction-driven decision-making tasks for entities or events. This paper presents an overview of the logical knowledge base, including categories and basic compositions. It also proposes a method for entity description and event causal logical knowledge extraction from the large-scale open text. Finally, a reliable interpretable path reasoning algorithm and financial entity influence generation system based on the logical reasoning knowledge base is explored for the financial domain. The algorithm model and system have achieved good results.
-
Ethnic Language Processing and Cross Language Processing2021, 35(10): 64-72.The bilingual sentence similarity aims to calculate the semantic similarity between different language sentences, which are of substantial application in the fields of information retrieval, parallel corpus construction, and machine translation. Challenged by the lack of parallel corpora and the obvious semantic and syntactic differences between Lao and Chinese, this paper proposes a model of bilingual sentence similarity with textual features for Chinese and Lao. Firstly, text features including part of speech and word co-occurrence in Chinese and Lao are fused with GloVe pretrained word vectors. Secondly, long-distance context features and deep-level semantic information are distinguished based on a multi-layered siamese network, which is composed of bidirectional long-term and short-term memory self-attention networks. Finally, the method of transfer learning is used to initialize the model by its parameters, and different strategies of fine-tuning are used to enhance the generalization ability of the model. Experimental results indicate that the recall rate, precision and F1 value of the proposed method reach 82.5%, 85.78% and 84.00%, respectively.
-
Information Extraction and Text Mining2021, 35(10): 73-80.As a subtask of information extraction, relation extraction aims to extract the structured knowledge from unstructured text, which is very important for the downstream tasks such as automatic question answering and knowledge graph construction. Focused on document-level relation extraction, this paper proposed a graph attention convolution model to deal with long-distance dependence issue. The model uses a multi-head attention mechanism to construct a dynamic topological graph for coreference, syntax and other information. Then it uses the graph convolution model and dynamic graph to capture global and local dependency information between entities. Experiments on the DocRED corpus and the self-expanding ACE 2005 corpus comfirm improvements on F1 values by 2.03 and 3.93, respectively.
-
Information Extraction and Text Mining2021, 35(10): 81-89.Entity relation extraction is an important research issue in the field of information extraction, which plays an important role in the automatic construction of knowledge base. This paper proposes a new entity relationship extraction model based on hybrid neural network, named BiGRU-Att-PCNN. Firstly, a Bi-directional Gated Recurrent Unit (BiGRU) is constructed to better obtain the relevant information of context word order in text sequence. Then, the attention mechanism is used to automatically capture the sequence features with high influence on the relationship. Finally, with Piecewise Convolution Neural Network (PCNN), the relevant environmental feature information is better learned from the adjusted sequence for relation extraction. Experiment on SemEval 2010 Task 8 proves that the proposed method achieves 86.71% in F1 value.
-
Information Extraction and Text Mining2021, 35(10): 90-100.Current aspect extraction methods adopt global or local self-attention mechanism without discriminating the sentiment words in the sentence. This paper proposes an aspect extraction method via interactive attention with sentiment words. We arrange all the sentiment word in the text in order, and apply Bi-directional Long-Short Term Memory network to encode sentence, and adopt fully connected neural network and High Way network to encode sentiment words. Then, we model the interactive attention with the sentence and sentiment words representations, which can precisely locate aspects with the help of sentiment words. Experimental results on three datasets from SemEval 2014-2015 show that our method can effectively improve the performance of aspect extraction by 5.53%、2.90% and 5.76% respectively.
-
Information Extraction and Text Mining2021, 35(10): 101-109.The event detection task is aimed at automatically retrieving structured event information from text. Existing neural event detection methods based on representation learning is restricted the limited manual labeling data sets. To capture more semantic knowledge in different task scenarios and large-scale language resources, this paper proposes a BERT-based multitask event detection model. This method takes advantage of the semantic knowledge which has already been contained by BERT as the foundation to further improve the representation learning and semantic perception ability of the multitask model. Experiments show that this method can effectively improve the comprehensive performance of event detection, achieving 76.7% F1 value in event classification of ACE2005 corpus. In addition, the paper explains the training process of the multitask model in detail in the experimental part, analyzes the effects of the multitask architecture on the event detection process from the perspective of interpretability.
-
Information Extraction and Text Mining2021, 35(10): 110-118.Extreme multi-label text classification (XMTC) is a challenging for large-scale label sets, complex interclass relationship and unbalanced data distribution. To better employ the label semantic information, this paper proposes an XMTC strategy by using semantic guidance from hierarchical labels, which offers weakly supervised semantic information for models to restrict the boundary of multi-label semantics during training and predicting. The experimental results over benchmark datasets show that the strategy can effectively improve performance of existing models, especially for short-text datasets in which the top macro-P of 21.23% is observed.
-
Sentiment Analysis and Social Computing2021, 35(10): 119-127.Rumor stance detection is to determine if each user’s stance on the rumor is supporting, denying or others by analyzing the posts on social media. Existing works models the conversations as unidirectional trees, focusing only on partial structure and semantic information of them. To address this issue, an enhanced bidirectional tree neural networks model is proposed. Firstly, a gate mechanism is designed to learn the representations jointly from the bottom-up and top-down propagation, which effectively extracted the global context of the conversation. Then, a local inference module is incorporated into the model to strengthen the semantic relations between the rumor and responsive posts. Results on the RumourEval 2017 Twitter dataset demonstrate that the proposed model achieves the best performance of 52.5% in macro-averaged F1 scores (1.6% improvements), especially good at detecting the denying stance which is the most challenging for stance detection.
-
Sentiment Analysis and Social Computing2021, 35(10): 128-136.As a fine-grained task, aspect-level sentiment analysis is aimed at identifying the sentiment polarity corresponding to the specific aspect in a sentence. Most existing works are focused on the design of attention network to highlight the different contributions of words in context and associate the context and the aspect properly. In this paper, we put forward to combine the hierarchical attention and gate networks to process aspect-level sentiment analysis task. we obtain the representation of the context through the attention between the context and a new representation of the aspect weighted by the context. At the same time, the useful information in the context is selected through the gate networks to enrich the representation of the context. The design of hierarchical attention gate networks is to make the representation of the context and the aspect more accurate. The experimental results on the Sem-Eval 2014 Task4 and Twitter show the validity of the model.
