2022 Volume 36 Issue 11
Published: 12 January 2023
-
Select all|
-
Survey2022, 36(11): 1-19.In recent years, multi-hop reading comprehension has become a hot topic in natural language understanding. Compared with single-hop reading comprehension, multi-hop reading comprehension is more challenging considering the involvement of multiple clues (e.g. multiple documents reading) the expectation for explicit reasoning paths. This paper summarizes the multi-hop reading comprehension task. This paper first gives the definition of multi-hop reading comprehension task. And then, and according to different reasoning methods, multi-hop reading comprehension models are can be divided into three categories: models based on structured reasoning, models based on evidence extraction, and models based on question decomposition. This paper analyzes the experimental results of these models on the common multi-hop reading comprehension datasets, revealing their advantages and disadvantages. Finally, the future research directions are discussed.
-
Survey2022, 36(11): 20-37.Machine translation quality estimation refers to the estimation of the quality of the outputs by machine translation system without the human reference translations. It is of great value to the research and application of machine translation. In this survey, we firstly introduce the background and significance of machine translation quality estimation. Then we introduce in detail the specific task objectives and evaluation indicators of word-level QE, sentence-level QE, and document-level QE. We further summarize the development of QE methods to three main stage: methods based on feature engineering and machine learning, methods based on deep learning, and methods integrated with pre-training model. Representative research works in each stage are introduced, and the current research status and shortcomings are analyzed. Finally, we outline the outlook for the future research and development of QE.
-
Language Analysis and Calculation2022, 36(11): 38-49.Implicit discourse relation recognition automatically identifies the semantic relation between arguments. The key to this task involves two issues: one is to represent the argument semantics, the other is to recognize the relation between arguments. Focusing on better representation of the arguments, this paper introduces the contrast learning into the process of argument representation learning. We further propose a method generating confused samples based on conditional auto-encoders, so as to enhance the confused data in contrastive learning. Experiments on the Penn Discourse Treebank (PDTB) corpus show that,our method increases F1 score by 4.68%, 4.63%, 3.14% and 12.77% on four top relations (Comparison, Contingency, Expansion, and Temporal), respectively.
-
Language Analysis and Calculation2022, 36(11): 50-59.Discourse analysis is a well-recognized topic in natural language processing. Rapid as the development of discourse analysis based on formal grammar, the function and semantics of discourse as a whole have not been well addressed. This paper proposes a Functional Pragmatics Recognition Model Based on Global and Structure Information (FPRGS). The FPRGS first obtains the interactive information of discourse units and integrates the global information of the article. Then it uses the gated semantic network to combine the structural information of discourse units with semantic information. The experimental results in the Chinese macro discourse tree-bank show that the proposed model can effectively identify the discourse units' functional pragmatics.
-
Language Analysis and Calculation2022, 36(11): 60-67.Relation identification of complex sentences is to distinguish the categories of semantic relations of sub-sentences. Focused on the sentences without explicit connectives, this paper applies the Attention mechanism combined with the GCN to classify the semantic relationship of Chinese complex sentences. The Bert based sentence representation formed by word vector is input into the Bi-LSTM. The sentence position representation is obtained and weighted via attention mechanism. A graph network is then constructed to capture the semantic information between sentences. The experiments on the CCCS and CDTB datasets reveal that the proposed method achieves 76.2% and 74.4% F1 value, respectively, increasing about 2.1% compared with the previous best models.
-
Language Analysis and Calculation2022, 36(11): 68-78.Macro discourse structure analysis aims to facilitate the understanding of the content and purpose in a discourse by revealing its structure. This paper proposes a pointer network model integrating top-down and bottom-up construction strategies. It can use the semantic information of the two construction strategies at the same time, so as to select the appropriate construction method. Experiments on Chinese Macro Chinese Discourse Treebank (MCDTB 2.0) show that the model proposed in this paper can effectively reduce the error propagation in the construction process and achieve the best performance.
-
Information Extraction and Text Mining2022, 36(11): 79-90.Entity and relationship extraction is to extract triples from unstructured natural language text. The existing pipeline method extracts entities first and then relations, without capturing the internal relations and dependencies of the two subtasks. This article proposes a joint extraction model of entities and relations based on CWHC-AM, regarding the multiple relationship issue in extraction as multiple binary relationship tasks. The multi-layer pointer network labeling scheme is adopted to transform the joint extraction task of entities and relations into a sequence labeling problem. And the adversarial training is introduced to improve the robustness of the model. Experiments on the Baidu DuIE2.0 Chinese dataset show that the method in this article can effectively extract multiple relations and binary relations at the same time with better results than the baseline model.
-
Information Extraction and Text Mining2022, 36(11): 91-100.The academic abstract summarizes key points in a research paper, with a series of moves conveying different information. The automatic recognition and extraction of moves could provide a valuable foundation for other tasks related with the academic abstract. This paper proposed a move recognition algorithm for academic abstract based on ERNIE-BiGRU model. Firstly, a multi-move structure splitting method based on dependency structure is proposed, identifying the multiple single-move structure in the academic abstract. Secondly, a single-move structure corpus is constructed, and a pre-training model of knowledge-enhanced semantic representation is employed to train sentence-level word vectors. Finally, the trained word vector with single move structure information is input into BiGRU for automatic recognition of moves. The experimental results show that the algorithm has good robustness and high recognition accuracy, achieving 96.57% and 93.75% recognition accuracy for structured and unstructured abstracts, respectively.
-
Information Extraction and Text Mining2022, 36(11): 101-109.To deal with the semantic sparsity caused by same entities in different forms in the culture of cross-border ethnic groups, this paper proposes a cross-border ethnic culture retrieval method based on entity semantic expansion. It uses the cross-border ethnic cultural knowledge map to associate the entities between various culture texts in the form of knowledge triples with addtional entity category tags. The TransH model is applied to represent entities and their extended semantic information, which is integrated into the query as kind of semantic enhancement. Experimental results show that the proposed method is 5.4% higher than the baseline model.
-
Question Answering and Dialogue System2022, 36(11): 110-120,130.The automated short answer grading(ASAG) system reduces the time-consuming manual scoring for educators with Natural Language Processing technology. It is worth noting that most ASAG system has shortcomings that students have intentionally fraud the model to get high scores by copying or slightly rewriting the standard solution. This paper explored a rule-based data augmentation approach to investigate the robustness of the ASAG system. However, natural languages have a discrete factor that limits the diversity of samples synthesized by rule-based data augmentation. In this paper, a knowledge distillation-based data augmentation strategy is proposed to stack different individual data augmentation methods in a parallel manner. In addition, the paper proposes a supervised contrast learning-based ASAG system that enables the model to learn effective sentence representations. We evaluate our model on two datasets from the University of North Texas and SemEval-2013. The performances our model are substantially improved compared to the baselines.
-
Question Answering and Dialogue System2022, 36(11): 121-130.Open-domain Question Answering (OpenQA) is an important task in natural language processing. However, OpenQA models tend to match texts on a superficial level between questions and documents and often make stupid errors on some easy questions. Part of the reason for these errors is that reading comprehension datasets lack some common patterns in the actual scenes. To eliminate the effects of these patterns, we propose several methods to improve the robustness of OpenQA models. Besides, we build a new dataset to evaluate the performance of models in the real world. The experimental results show that the proposed methods can improve the performance of OpenQA models on this dataset.
-
Information Retrieval2022, 36(11): 131-139.In the E-Commence scenario, it is a crucial issue to construct an effective Session-Based Recommendation (SBR) model, helping users quickly find products of interest based on their short-term interests. Most current models fails to capture the interactive relationship of item embeddings on the session graph and global graph. This paper proposes to utilize the interactive attention and improved parameter self-adaptive strategy to enhance the GNN-based commodity recommendation model. The interactive attention layer is applied to amend the commodity representation in the global graph and the session graph through strong correlation extraction, and the parameter adaptive layer is to obtain the final representation of the items for prediction. Experimental results show that our model is significantly superior to other off-the-shelf models on the public dataset of Tmall.
-
Natural Language Understanding and Generation2022, 36(11): 140-147.Simiar Case matching is an important task in intelligent justice, especially for case retrieval and same-case same-judgment. Owing to the long text and the subtle difference between legal documents, existing deep matching models are difficult to achieve ideal results. To address this issue, this paper proposes a method of integrating case elements to improve the matching of similar cases with a focus on the private lending cases. First, six types of private lending case elements are formulated and extracted by regular expressions, represented in the form of one -hot word vectors. Then the legal text is filtered and formed in reverse order, represented by BERT capture the long-distance dependence. The legal text representation and the case element representation is fused by the linear network and then encoded by BiLSTM for high-dimensional representation. Finally, the vector representation similarity matrix is constructed through the twin network framework, and the final similarity is decided by semantic interaction and vector pooling. The experimental results show that the proposed model is better than the baseline model on the CAIL2019-SCM public data set.
-
Natural Language Understanding and Generation2022, 36(11): 148-155,168.Focused on the Cail2020 multi-hop machine reading comprehension data set, this paper presents TransformerG, a multi-hop reading comprehension model based on the integration of paragraph graph structure and attention mechanism.This model combines the feature of question node, question entity node, sentence node and sentence entity node in the text to predict the answer span. In addition, a sentence level sliding window method is designed to substitute the truncation of long text in the pre-training model. The proposed TransformerG model ranks Top 2 in the machine reading comprehension setting of Cail2020 Competition.
-
Natural Language Understanding and Generation2022, 36(11): 156-168.Dense video captioning can automatically generate sentence sequence corresponding to video content, involving issues in both computer vision and natural language processing. To capture the audio information as well as the temporal structure and semantic relationship between events, this paper proposes a dense video captioning method based on multi-modal features. Firstly, Timeception layer is used as basic module in action proposal generation stage to better adapt various time span of action segments. Secondly, audio features are used to enhance the effect of proposal and description generation stages. Finally, the temporal semantic relation module models the temporal structure and semantic information between events to further enhance the accuracy of description generation. In addition, this paper also constructs a dataset named SDVC to explore the effectiveness of this method in application of real learning scene. The experimental results on ActivityNet Captions and SDVC datasets show that the AUC of action proposal generation increases by 0.8% and 6.7%, respectively; and in turn, using generated action proposals for description generation, BLEU_3 and BLEU_4 of SDVC dataset increased by 2.3% and 2.2%, respectively.
-
Speech Processing2022, 36(11): 169-176.The problem of long-tail distributed data is common in NLP practice. Taking the polyphone disambiguation task in text-to-speech (TTS) as an example, the extreme data imbalance and the lack of tail data affect industrial online TTS applications. Observging that the Chinese polyphone is long-tail distributed on both “character” and “pronunciation” dimensions, this paper proposes a double-weighted (DW) algorithm, which can be combined with the other two long-tail algorithms: MARC and Decouple-cRT. Given the perspectives of both open-source data and industrial data, DW demonstrates improvement in accuracy compared to the baseline model and the two original algorithms.
