2022 Volume 36 Issue 1
Published: 28 February 2022
-
Select all|
-
Survey2022, 36(1): 1-11,20.Natural Language Generation in a task-oriented dialogue system (ToDNLG) aims to generate natural language responses given the corresponding dialogue acts, which has attracted increasing research interest. With the development of deep neural networks and pre-trained language models, great success has been witnessed in the research of ToDNLG field. We present a comprehensive survey of the research field, including: (1) a systematical review on the development of NLG in the past decade, covering the traditional methods and deep learning-based methods; (2) new frontiers in emerging areas of complex ToDNLG as well as the corresponding challenges; (3) rich open-source resources, including the related papers, baseline codes and the leaderboards on a public website. We hope the survey can promote future research in ToDNLG.
-
Language Analysis and Calculation2022, 36(1): 12-20.The language models trained on large scale corpus has achieved significant performance in text generation. However, these language models may generate uncertain aggressive texts when perturbed. In this paper, we proposed a method that automatic evaluates aggressiveness of language model, which is divided into two stages: induction and evaluation. In induction stage, based on controllable text generation technology, we update parameters in activation layers of language model along the gradient of trained text classification model to increase the probability of generating aggressive texts. In evaluation stage, we estimate proportion of induced aggressive generated text by utilizing text classification model to evaluate aggressiveness of language model. Experiment result shows that this approach can effectively evaluate aggressiveness of language model in different experiment settings, and analyze the relationship between aggressiveness and scale of model parameters, training corpus and prefix.
-
Language Analysis and Calculation2022, 36(1): 21-30,38.Abstract Meaning Representation is a sentence-level meaning representation, which abstracts a sentence’s meaning into a rooted acyclic directed graph. With the continuous expansion of Chinese AMR corpus, more and more scholars have developed parsing systems to automatically parse sentences into Chinese AMR. To make up for the vacancy of Chinese AMR parsing evaluation methods, we have improved the Smatch algorithm of generating triples to make it compatible with concept alignment and relation alignment. We finally complete a new integrity metric Align-Smatch for paring evaluation. Compared on 100 manually annotated AMR and gold AMR, it is revealed that Align-Smatch works well in alignments and more robust in evaluating arcs. We also put forward some fine-grained metric for evaluating concept alignment, relation alignment and implicit concepts, in order to further measure parsers’ performance in subtasks.
-
Language Analysis and Calculation2022, 36(1): 31-38.The task of AMR-to-Text generation is to generate text with the same semantic representation given an AMR graph. This task can be viewed as a translation task from the source AMR graph to the target sentence. To capture the syntactic information hidden within the sentence, this paper proposes a direct and effective method of integrating syntactic information in the AMR-to-Text task. Experiments on Transformer and the top-performed model the task show that, on the two existing standard English data sets LDC2015E86 and LDC2017T10, both have achieved significant improvements.
-
Language Resources Construction2022, 36(1): 39-46.A dependency treebank of Learner Chinese provides dependency parses for non-native sentences, which could promote the teaching and research on Chinese as a second language, and support related researches such as syntactic analysis of learner language and grammatical error correction. However, few dependency treebanks of learner Chinese are available, and there are still some problems in annotation guidelines. In this paper, we develop the annotation guideline, establish an online annotation platform, and build the Treebank of Learner Chinese. This paper also describes the details in data selection and annotation workflow, evaluates the quality of annotation, and explores the impact of errors on annotation quality and syntactic analysis.
-
Language Resources Construction2022, 36(1): 47-55.Machine translation error analysis, including error classes and error distribution etc. Error analysis of machine translaution output, plays an important role in the research and application of machine translation. In this paper, post-editing is introduced into error analysis to annotate error labels. Automatic error annotation and manual annotation are applied to build a Fine-grained Error Analysis Corpus of English-Chinese Machine Translation (ErrAC), in which every annotated sample includes a source sentence, MT output, reference, post-edit, WER and error type. The annotated error types include addition, omission, lexical error, word order error, untranslated word, named entity translation error etc. Annotator agreement analysis shows the effectiveness of the annotation. The statistics and analysis based on the corpus provide effective guidance for the development of machine translation system and post-editing practice.
-
Information Extraction and Text Mining2022, 36(1): 56-64.Attribute extraction is a key step of constructing a knowledge graph. In this paper, the task of attribute extraction is converted into a sequence labeling problem. Due to a lack of labeling data in product attribute extraction, we use the distant supervision to automatically label multiple source texts related to e-commerce. In order to accurately evaluate the performance of the system, we construct a manually annotated test set, and finally obtain a new data set for product attribute extraction in multi-domains. Based on the newly constructed data set, we carried out intra-domain and cross-domain attribute extraction for a variety of pre-trained language models. The experimental results show that the pre-trained language models can better improve the extraction performance. Among them, ELECTRA performs the best in attribute extraction in in-domain experiments, and BERT performs the best in cross-domain experiments. we also find that adding a small amount of target domain annotation data can effectively improve the performance cross-domain attribute extraction and enhance the domain adaptability of the model.
-
Information Extraction and Text Mining2022, 36(1): 65-74.In recent years, Capsule Neural Networks (Capsnets) have been successfully applied to text classification. In existing studies, all n-gram features play equal roles in text classification, without capturing the importance of each n-gram feature in the specific context. To address this issue, this paper proposes Partially-connected Routings Capsnets with Multi-scale Feature Attention(MulPart-Capsnets) by incorporating multi-scale feature attention into Capsnets. Multi-scale feature attention can automatically select n-gram features from different scales, and capture accurately rich n-gram features for each word by weighted sum rules. In addition, in order to reduce the redundant information transferring between child and parent capsules, dynamic routing algorithm is also improved. In order to verify the effectiveness of the proposed model, our experiments are conducted on seven well-known datasets in text classification. The experimental results demonstrate that the proposed model consistently improves the performance of classification.
-
Information Extraction and Text Mining2022, 36(1): 75-82.Relation extraction aims to extract the relations between entities from unlabeled free text. This paper proposes a relation extraction model that combines the pre-trained language model and label dependency knowledge. Specifically, given a sentence as the input, we first generate a deep contextualized word representation for the sentence and the two target entities using a pre-trained BERT encoder. At the same time, a multi-layer graph convolutional network is applied to model the dependency graph between the relation labels. Finally, we combine the above information to guide the relation classification. The experimental results show that our approach significantly outperforms the baselines.
-
Information Extraction and Text Mining2022, 36(1): 83-91.In text summarization, the mainstream method is to use encoder-decoder architecture to obtain the required context semantic information by using soft attention in the decoding process. Since the encoder sometimes encodes too much information, the generated summary does not always summarize the core content of the source text. To address this issue, this paper proposes a text summarization model based on a dual-attention pointer network. Firstly, in the dual-attention pointer network, the self-attention mechanism collects key information from the encoder, while the soft attention and the pointer network generate more coherent core content through context information. The fusion of both will generate accurate and coherent summaries. Secondly, the improved coverage mechanism is applied to address the repetition problem and improve the quality of the generated summaries. Simultaneously, scheduled sampling and reinforcement learning are combined to generate new training methods to optimize the model. Experiments on the CNN/Daily Mail dataset and the LCSTS dataset show that the proposed model performs as well as many state-of-the-art models.
-
Information Extraction and Text Mining2022, 36(1): 92-103.The current detection method of illegal comments mainly relies on sensitive words screening, incapable of effectively identifying malicious comments without vulgar language. In this paper, a data set of Chinese illegal comments is established by crawler and manual annotation. On the basis of BERT, RCNN combined with attention mechanism is used to further extract the context features of comments, and multi-task joint training is adopted to improve the classification accuracy and generalization ability of the model. The model is independent to sensitive thesaurus. Experimental results show that the proposed model can better understand the semantic information than the traditional model, achieving aprecision of 94.24%, which is 8.42% higher than traditional TextRNN and 6.92% higher than TextRNN combined with attention mechanism.
-
Information Extraction and Text Mining2022, 36(1): 104-116.To address the semantic and context transfer of text information, an improved method of dynamic convolution neural network based on meta learning and attention mechanism is proposed. Firstly, cross category classification is carried out by using the underlying distribution features of the text to make the text information ready for transfer. Secondly, the attention mechanism is used to improve the traditional convolution network to improve the feature extraction ability of the network, and the balanced variables are generated according to the information of the original data set to reduce the impact of the imbalance of data. Finally, the parameters of the model are optimized automatically by using the two-level optimization method. The experimental results on the general text classification THUCNews dataset show that the proposed method has improved the accuracy by 2.27% and 3.26% in the 1-shot and 5-shot experiments, respectively, and on the IMDb dataset, by 3.28% and 3.01%, respectively.
-
Information Retrieval and Question Answering2022, 36(1): 117-126.Sequential recommendation aims at predicting the items to be interacted next time according to the user's historical behavior sequences. Most related studies have shown that items to be interacted depend on different scales of blocks in historical sequences. Implicit multi-scale representation space is often heuristically designed, from which we cannot infer an explicit hierarchy. Therefore, we propose a Dynamic hierarchical Transformer to learn multi-scale implicit representation and explicit hierarchy simultaneously. The Dynamic hierarchical Transformer adopts a multi-layer structure with dynamically generated mask matrices from neighbor block attention per layer in a bottom-up manner. In the derived multi-scale hierarchy, the composition structure per layer is inferred from the block mask matrix and the implicit representation per scale is obtained by the dynamic block mask and self-attention mechanism. Experimental results on two benchmark datasets (MovieLens-100k and Amazon Movies and TV) show that our proposed model improves the precision by 2.09% and 5.43% respectively over the state-of-the-art baselines. Furthermore, the derived multi-scale hierarchy agrees with our intuition through the case study.
-
Information Retrieval and Question Answering2022, 36(1): 127-134.Children's health and disease diagnosis and treatment services are the focus of family and society. With the Chinese medical knowledge graph and medical texts as the data source, this paper develops question answering system for the pediatric diseases and health care knowledge. The system adopts Aho-Corasick automata, regular expressions, syntactic structure and keywords to match user input questions with templates. Cypher sentences are generated by the template to search and retrieve the Chinese medical knowledge graph and medical texts for the candidate answers. The final answers are scored by the authority of the data source and the matching score. According to users’ feedback collected one month after the system was put into trial use in a Grade III Level A hospital, the satisfaction rate reached 85.43%.
-
Sentiment Analysis and Social Computing2022, 36(1): 135-144,172.The COVID-19 rumors published and spread on the online social media have a serious impact on people's livelihood, economy, and social stability. Most existing researches for rumor detection usually assumed that the happened events for modeling and predictions already have enough labeled data. These studies have severe limitations on detecting emergent events such as the COVID-19 which has very few training instances. This article focuses on the problem of few-shot rumor detection, aiming to detect rumors of emergent events with only very few labeled instances. Taking the COVID-19 rumors from Sina Weibo as the target, we construct a Sina Weibo COVID-19 rumor dataset for few-shot rumor detection, and propose a deep neural network based few-shot rumor detection model with meta learning. In the few-shot machine learning scenarios, the experimental results of the proposed model on the COVID-19 rumor dataset and the PHEME public dataset have been significantly improved.
-
Sentiment Analysis and Social Computing2022, 36(1): 145-153.In this paper, a database of more than 110,000 Chinese celebrities' names is constructed. Each entry are annotated with social and cultural labels including gender, birthplace, as well as Chinese character information labels such as Pinyin, strokes and character components. Based on this database, this paper selects names from 1919 to the present and explores Chinese names on character characteristics, gender differences, and diachronic changes. Quantatively, it is found that the female name characters are shorter, more complex but less diversified than those of men. Over the past centry, the use of personal names has become more monotonous and centralized, and the imagery of the characters significantly shifted around the reform and opening-up era. Besides, we also obtained the gender polarity characters table of the names, high-frequency characters table of each stage, characters change trend table, etc.
-
Sentiment Analysis and Social Computing2022, 36(1): 154-162.As a cross-domain research task, depression detection using multimodal information has recently received considerable attention from researchers in several communities, such as natural language processing, computer vision, and mental health analysis. These studies mainly utilize the user-generated contents on social media to perform depression detection. However, existing approaches have difficulty in modeling long-range dependencies(global information). Therefore, how to obtain global user information has become an urgent problem. In addition, considering that social media contains not only textual but also visual information, how to fuse global information in different modalities has become another urgent problem. To overcome the above challenges, we propose a multimodal hierarchical dynamic routing approach for depression detection. We obtain global user information from hierarchical structure and use dynamic routing policy to fuse text and image modalities which can adjust and refine message to detect depression. Empirical results demonstrate the impressive effectiveness of the proposed approach in capturing the global user information and fusing multimodal information to improve the performance of depression detection.
-
Speech Processing2022, 36(1): 163-172.In recent years, speech attributes are often used in computer-aided pronunciation training systems (CAPT). This paper proposes a method for modeling fine-grained speech attributes (FSA), and applies it in cross-language attribute recognition and mispronunciation detection. We achieve an attribute detector group with an optimal average recognition accuracy rate of 95%. As for the mispronunciation detection on the two second language test sets, FSA method achieves an improvement of more than 1% compared to the baselines. In addition, according to the cross-language characteristics of speech attribute, we also set up a comparative experiment to futher test and analyze the methods
