Table of Content

20 January 2024, Volume 60 Issue 1

Previous Issue    Next Issue

Enhanced Prompt Learning for Few-shot Text Classification Method

LI Ruifan, WEI Zhiyu, FAN Yuantao, YE Shuqin, ZHANG Guangwei

2024, 60(1):  1-12.  DOI: 10.13209/j.0479-8023.2023.071

Asbtract ( )   HTML   PDF (723KB) ( )  

Related Articles | Metrics

An enhanced prompt learning method (EPL4FTC) for few-shot text classification task is proposed. This algorithm first converts the text classification task into the form of prompt learning based on natural language inference. Thus, the implicit data enhancement is achieved based on the prior knowledge of pre-training language models and the algorithm is optimized by two losses with different granularities. Moreover, to capture the category information of specific downstream tasks, the triple loss is used for joint optimization. The masked-language model is incorporated as a regularizer to improve the generalization ability. Through the evaluation on four Chinese and three English text classification datasets, the experimental results show that the classification accuracy of the proposed EPL4FTC is significantly better than the other compared baselines.

A Low-Resource Named Entity Recognition Method for Cultural Heritage Field Incorporating Knowledge Fusion

LI Chao, HOU Xia, QIAO Xiuming

2024, 60(1):  13-22.  DOI: 10.13209/j.0479-8023.2023.070

Asbtract ( )   HTML   PDF (1003KB) ( )  

Related Articles | Metrics

In cultural heritage field, entity nesting of cultural relics data is obvious, the entity boundary is not unique, and the marked data in the field of cultural relics is extremely lacking. All the problems above can lead to the low recognition performance of named entities in the field of cultural relics. To address these issues, we construct a dataset called FewRlicsData for NER in the field of cultural heritage and propose a knowledge-enhanced, low-resource NER method RelicsNER. This method integrates the semantic knowledge of category description information into the cultural relics text, employs the span-based method to decode and solve the entity nesting problem, and uses the boundary smoothing method to alleviate the overconfidence problem of span recognition model. Compared with the baseline model, the proposed method achieves higher F1 scores on the FewRlicsData dataset and demonstrates good performance in named entity recognition tasks in the cultural heritage field. Experimental results on the public dataset OntoNotes 4.0 indicate that the proposed method has good generalization ability. Additionally, small-scale data experiments on OntoNotes 4.0 and MSRA datasets show that the performance of the proposed method surpasses that of the baseline model, demonstrating its applicability in low-resource scenarios.

Reducing Multi-model Biases for Robust Visual Question Answering

ZHANG Fengshuo, LI Yu, LI Xiangqian, XU Jin’an, CHEN Yufeng

2024, 60(1):  23-33.  DOI: 10.13209/j.0479-8023.2023.072

Asbtract ( )   HTML   PDF (5709KB) ( )  

Related Articles | Metrics

In order to enhance the robustness of the visual question answering model, a bias reduction method is proposed. Based on this, the influence of language and visual information on bias effect is explored. Furthermore, two bias learning branches are constructed to capture the language bias, and the bias caused by both language and images. Then, more robust prediction results are obtained by using the bias reduction method. Finally, based on the difference in prediction probabilities between standard visual question answering and bias branches, samples are dynamically weighted, allowing the model to adjust learning levels for samples with different levels of bias. Experiments on VQA-CP v2.0 and other data sets demonstrate the effectiveness of the proposed method and alleviate the influence of bias on the model.

A Context-Aware Query Suggestion Method Based on Multi-source Data Augmentation through Cross-Attention

ZHANG Naizhou, CAO Wei

2024, 60(1):  34-42.  DOI: 10.13209/j.0479-8023.2023.074

Asbtract ( )   HTML   PDF (617KB) ( )  

Related Articles | Metrics

Most existing neural network-based approaches for query suggestion use solely query sequences in query logs as training data. However, these methods cannot fully mine and infer all kinds of semantic relationships among words or concepts from query sequences because queries in query sequences inherently suffer from a lack of syntactic relation, even a loss of semantics. To solve this problem, this paper proposes a new neural network model based on multi-source data augmentation through cross-attention (MDACA) for generating context-aware query suggestions. Proposed model adopts a Transformer-based encoder-decoder model that incorporates document-level semantics and global query suggestions into query-level information through cross-attention. The experimental results show that in contrast to the current suggestion models, the proposed model can generate context-aware query suggestions with higher relevance.

Can ChatGPT Be Served as the Sentiment Expert? An Evaluation of ChatGPT on Sentiment and Metaphor Analysis

ZHANG Yazhou, WANG Mengyao, RONG Lu, YU Yang, ZHAO Dongming, QIN Jing

2024, 60(1):  43-52.  DOI: 10.13209/j.0479-8023.2023.075

Asbtract ( )   HTML   PDF (772KB) ( )  

Related Articles | Metrics

To explore the potential for subjective understanding, the subjectivity and metaphorical nature of ChatGPT, this paper evaluates ChatGPT on five sentiment, humor, and metaphor benchmark datasets and discusses its strengths and limitations on different tasks by comparing it with the most cutting-edge models in the field. In addition, this paper also compares the performance of ChatGPT and humans in sentiment analysis, with gaps of 9.52%, 16.64% and 6.69% in human results on sentiment, humor and metaphor tasks. The results suggest that although ChatGPT achieves the best performance in dialogue generation, it still has potential for improvement in sentiment understanding. Finally, this paper investigates ChatGPT’s sensitivity to cueing templates in an emotion understanding scenario by improving the cueing templates.

A Long Dialogue Summary Model Integrating Salience Discourse Context Window Sampling Methods

WU Jie, WANG Pengming, XIONG Zhengkun

2024, 60(1):  53-61.  DOI: 10.13209/j.0479-8023.2023.078

Asbtract ( )   HTML   PDF (856KB) ( )  

Related Articles | Metrics

A long dialogue summary generation model with integrated salience discourse context window sampling method (SDCWS) is proposed according to the characteristics of dialogue corpus. The model is divided into two modules. 1) The salience discourse context window sampling module (CWS) evaluates the dialogue discourse for salience, uses the salient discourse as the sampling anchor point, and then sets the sampling window to extract the discourse adjacent to the left and right of the sampling anchor point together as fragments, containing richer discourse relations. 2) The inter-fragment information fusion summary generation module (IF) uses the transformer block to fuse information from mutually independent fragments, enhancing the semantic relationships between fragments and assigning blended weights to fragments during summary generation. The loss-of-consistency mechanism is used to encourage the salience discourse context window sampling module to determine better sampling anchors. Experimental results on the publicly available query-based long conversation summary dataset QMSum show that scores of the proposed model are significantly higher than the best existing model on the ROUGE evaluation metric.

Interpretable Biomedical Reasoning via Deep Fusion of Knowledge Graph and Pre-trained Language Models

XU Yinxin, YANG Zongbao, LIN Yuchen, HU Jinlong, DONG Shoubin

2024, 60(1):  62-70.  DOI: 10.13209/j.0479-8023.2023.073

Asbtract ( )   HTML   PDF (1332KB) ( )  

Related Articles | Metrics

Joint inference based on pre-trained language model (LM) and knowledge graph (KG) has not achieved better results in the biomedical domain due to its diverse terminology representation, semantic ambiguity and the presence of large amount of noise in the knowledge graph. This paper proposes an interpretable inference method DF-GNN for biomedical field, which unifies the entity representation of text and knowledge graph, denoises the subgraph constructed by a large biomedical knowledge base, and further improves the information interaction mode of text and subgraph entities by increasing the direct interaction between corresponding text and subgraph nodes, so that the information of the two modes can be deeply integrated. At the same time, the path information of the knowledge graph is used to provide interpretability for the model reasoning process. The test results on the public dataset MedQA-USMLE and MedMCQA show that DF-GNN can more reliably leverage structured knowledge for reasoning and provide explanatory properties than existing biomedical domain joint inference models.

Radiology Report Generation Method Based on Multi-scale Feature Parsing

WANG Rui, LIANG Jianguo, HUA Rong

2024, 60(1):  71-78.  DOI: 10.13209/j.0479-8023.2023.076

Asbtract ( )   HTML   PDF (1926KB) ( )  

Related Articles | Metrics

When using deep learning models to automatically generate radiology reports, due to the extreme imbalance of data, it is difficult for current models to identify abnormal regional features, which leads to misjudgment and missed judgment of the disease. In order to improve the model’s ability to identify diseases and improve the quality of reports, the authors use a multi-scale feature parsing Transformer (MFPT) model to generate radiology reports. Among them, a key feature enhanced attention (KFEA) module is constructed to strengthen the utilization of key features. A multi-modal feature fusion (MFF) module is designed to promote the feature fusion of semantic features and visual features and alleviate the impact caused by feature differences. This paper explores the role of stage-aware (SA) module in optimizing primary features in radiology reporting tasks. Finally, compared with the current mainstream models on the popular radiology report dataset IU X-Ray, the results show that the proposed model has achieved the current best effect.

Bi-Attention Text-Keyword Matching for Law Recommendation

DING Na, LIU Peng, SHAO Huipeng, WANG Xuekui

2024, 60(1):  79-88.  DOI: 10.13209/j.0479-8023.2023.077

Asbtract ( )   HTML   PDF (686KB) ( )  

Related Articles | Metrics

This paper proposed a bi-directional attention based text-keyword matching model for law recommendation (BiAKLaw). In this model, BERT is utilized as a basic matching model, bi-directional attention mechanism is implemented to extract token-level alignment features and keyword-level differential features, and these features are fused with keyword attentive semantic representations for a better matching effect. The experimental results on the traffic accident and intentional injury datasets demonstrate that, compared with BERT, the proposed model increases F1 evaluation metric by 3.74% and 3.43% respectively.

MPI Solver of Particle-Laden Compressible High Enthalpy Flow: Numerical Method and Validation

LI Qing, YU Zhaosheng, LIU Pengxin, LI Tingting, CHEN Jianqiang, YUAN Xianxu

2024, 60(1):  89-108.  DOI: 10.13209/j.0479-8023.2023.022

Asbtract ( )   HTML   PDF (2831KB) ( )  

Related Articles | Metrics

Based on the theoretical equations of compressible particle-laden flow in multi-physical conditions, a MPI-Particle Solver based on dynamic linked list array is developed, which is coupled with a compressible flow solver of carried phase. The dispersed particle phase is different from the carried fluid phase, the previous one is based on the Lagrange coordinate system, while the later one is based on the Euler coordinate system. The traditional way is to use the global array to assign the memory to the dispersed particle phase to achieve the transfer between two-phase, but at expense of low computational efficiency. This paper uses dynamic linked list array to allocate memory to dispersed particle phase, achieve both problems. A series of DNS validations with references case have been done, in terms of multi-physical effects and two canonical cases at incompressible limit.

Intelligent Diagnosis on Anterior Cruciate Ligament Deficiency Based on Plantar Pressure and ConvLSTM Neural Network

LI Dai, WANG Tianmu, ZHANG Si, QIN Yue, XIE Fugui, LIU Xinjun, NIE Zhenguo, HUANG Hongshi

2024, 60(1):  109-117.  DOI: 10.13209/j.0479-8023.2023.089

Asbtract ( )   HTML   PDF (1993KB) ( )  

Related Articles | Metrics

Based on Convolutional Long-Short Term Memory Neural Network, the authors proposed a deep learning method PressureConvLSTM to extract features during walking in both spatial and temporal dimensions. Classification based on plantar pressure of anterior cruciate ligament deficiency (ACLD) was applied to distinguish walking gait information. Experiment results combined with clinical data showed that PressureConvLSTM model obtained 95% test accuracy when diagnosing ACLD, which was well performed over other traditional deep learning models.

Evaluating the Characters of Songbirds’ Vocalizations during Mobbing Event in Dark Coniferous Forest Using Bio-acoustic Indicators

WANG Jiangyue, TIAN Jia, ZHOU Zhengyang, MA Xiaoyun, LONG Yu, WANG Rongjiang, LI Sheng

2024, 60(1):  118-126.  DOI: 10.13209/j.0479-8023.2023.099

Asbtract ( )   HTML   PDF (15525KB) ( )  

Related Articles | Metrics

This study investigated the acoustic variation of songbirds’ vocalization during mobbing event in the sub-alpine forests using playback experiment and automatic acoustic recording (AAR). We played Glaucidium brodiei vocalization and Phylloscopus trochiloides alarm call to induce songbirds’ mobbing behavior and recorded their response vocalizations using automatic acoustic recorders in the Abies-Picea dark coniferous forest in Wanglang National Nature Reserve, Sichuan Province, during the summer of 2021. Altogether 24 experiments were performed on 12 experiment sites, with a total length of 1057 min acoustic recordings, including 12 G. brodiei vocalization experiments, and 12 P. trochiloides alarm call experiments. A total of 28 species of birds belonging to 17 families and 3 orders were observed to participate in mobbing behavior. By analyzing the acoustic indexes of recordings, it was concluded that 1) both G. brodiei vocalization and P. trochiloides alarm call could trigger mobbing event, during which Phylloscopidae and Paridae species were the majority being involved; 2) during playback period, songbirds’ vocalizations are more intense (G. brodiei experiments: SPL_avg=−31.02±4.87 dB vs −42.74±4.68 dB, p = 0.001; P. trochiloides experiments: SPL_avg = −33.26±4.05 dB vs −46.38±4.54 dB, p = 0.001) and less complicated (G. brodiei experiments: H = 0.76±0.02 vs 0.80±0.03, p = 0.001; P. trochiloides experiments: H = 0.77±0.02 vs 0.82±0.02, p = 0.001) than that of pre-playback period; 3) the variability of acoustic indices from pre-playback to during-playback didn’t change significantly between G. brodiei experiments and P. trochiloides experiments. The results provide a new insight into avian behavior studies in the acoustic aspect, and novel behavioral application scenarios for large-scale soundscape monitoring data. 

Research and Development of Living Plants in-vivo Fluorescence Imaging System for Experimental Teaching

WANG Fanlin, HE Xinqiang, WANG Donghui

2024, 60(1):  127-132.  DOI: 10.13209/j.0479-8023.2023.019

Asbtract ( )   HTML   PDF (17058KB) ( )  

Related Articles | Metrics

Through the development of plant culture and imaging system, a low-cost plant living fluorescence imaging system was developed. The system adopts a double-layer box structure design, which is divided into two parts: the constant culture system and the fluorescence recording system. It can provide the necessary life support function for plant growth, and can also obtain the key images in the process of plant growth in real time, including the time-space mode of monitoring the fluorescence signal in the development of the transgenic plant. The EGFP transgenic tobacco and Arabidopsis were successfully cultured and observed in real time using this system. The low cost greatly expanded the application of the system in experimental teaching, and also laid a foundation for the further study of the observation and detection of follow-up life activities.

Precipitation Trend in Warm Seasons during 1981–2015 over the Tibetan Plateau: A Perspective of Circulation Change

SUN Yawei, Chan-Pang NG, LI Liye, ZHANG Qinghong

2024, 60(1):  133-144.  DOI: 10.13209/j.0479-8023.2023.082

Asbtract ( )   HTML   PDF (2729KB) ( )  

Related Articles | Metrics

The long-term precipitation trend over the Tibetan Plateau (TP) was studied using a high spatiotemporal resolution precipitation dataset. It was found that the warm season (May–Sept.) precipitation over TP increased significantly during 1981–2015 (grid-mean: 0.9 mm/a), which was most significant in the northeastern, central, and western TP. The geopotential height field of 500 hPa over TP was divided into 9 circulation patterns (T1–T9) by using the obliquely rotated Principal Components in the T-mode (PCT) method, in which T2 and T4 were the dominant circulation patterns (DT) for the increase of precipitation. The geopotential height field of DT was low over the west but high over the east. DT dominated the increase of precipitation was reflected in the increase of precipitation days and daily precipitation. The increase in the number of DTs led to increased precipitation in the central and western TP. Another dominant mechanism was the optimization of precipitation conditions: T2 was dominated by the optimization of dynamic conditions driven by the larger gradient of geopotential height, and T4 was dominated by the optimization of thermal conditions driven by more water vapor.

Comprehensive Detection Payload Technology for Space Environment of FY-3E Satellite

SHEN Guohong, HUANG Cong, ZHANG Pengfei, ZHANG Xiaoxin, WANG Jinhua, LI Jiawei, ZONG Weiguo, ZHANG Shenyi, ZHANG Xianguo, SUN Yueqiang, YANG Yong, ZHANG Huanxin, ZOU Hong, WANG Jindong, SUN Ying, BAI Chaoping, TIAN Zheng

2024, 60(1):  145-156.  DOI: 10.13209/j.0479-8023.2023.096

Asbtract ( )   HTML   PDF (19964KB) ( )  

Related Articles | Metrics

To monitor the space environment and its effects in the low-Earth sun-synchronous orbit of China’s FY-3 satellite, a comprehensive detection technology based on the type Ⅱ loads of the space environment monitor is proposed. In the process of ground development, various technical indicators of the space environment comprehensive detection payload have been calibrated and experimentally verified by different methods such as standard radiation source, equivalent signal source, particle accelerator and standard magnetic field. The results show that the multi-direction full-spectrum particle detection achieves an energy range of 30 keV–300 MeV, with the accuracy of ≤10%. The magnetic field detection realizes the measurement range of −65023–+65023 nT, with the accuracy of ≤0.73 nT. The potential detection realizes the measurement range of −32.4–+23.7 kV, with the sensitivity of ≤10V. The detection of radiation dose realizes the measurement range of 0–3×104 rad (Si), with the sensitivity of ≤8.3 rad (Si). Through comprehensive observation of particle radiation environment, change of in-situ magnetic field vector, radiation dose accumulation and change of satellite surface potential in satellite operation orbit, the space environment monitor provides necessary data support for space activities, satellite design, space science research and space weather early warning and prediction. 

Investigating Fish Diversity in Huangjinxia Section of the Upper Hanjiang River Based on Environmental DNA Metabarcoding

DING Yang, LI Yanyan, ZHAO Jinyong, PENG Wenqi, ZHANG Jing, REN Jinhao

2024, 60(1):  157-164.  DOI: 10.13209/j.0479-8023.2023.080

Asbtract ( )   HTML   PDF (15908KB) ( )  

Related Articles | Metrics

 In 2020, Environmental DNA Metabarcoding was used to study the fish diversity in the Huangjinxia section of the upper Hanjiang River, and the historical survey data were compared to construct a fish list in the Huangjinxia section of the upper Hanjiang River. A total of 20 fish species were detected from 9 sampling sites, accounting for 45.45% of the list; Cyprinus_carpio, Carassius_auratus, Distoechodon_tumirostris and Squalidus_argentatus are the dominant species. The Shannon index of fish in the upper and lower reaches of the Huangjinxia was significantly different (P < 0.01, n=9), and the Shannon index of fish in the upper reaches was significantly larger than that in the lower reaches. The main reason for the spatial differences in fish diversity is the Huangjinxia hydraulic hub. The fish composition detected by environmental DNA was similar to that obtained by traditional methods. As an emerging biomonitoring tool, environmental DNA macrobarcoding technology can complement traditional fish monitoring methods to rapidly detect fish diversity and its spatial distribution in the upper reaches of the Hanjiang River. 

Habitat Creation Modes Based on Balance between Water and Soil and Its Application in Ecological Restoration Engineering

WANG Zhiyong, LI Jianing, PENG Xiao, HONG Min, ZHANG Kai

2024, 60(1):  165-174.  DOI: 10.13209/j.0479-8023.2023.092

Asbtract ( )   HTML   PDF (53797KB) ( )  

Related Articles | Metrics

Methods for regulating the relationship between water and soil were summed up in 23 traditional waterscape and farmland design wisdom. Creating a reciprocal and mutually reinforcing relationship between water and soil in local area is the core of creating habitats through collection and diversion of water by cutting and filling earthwork. Moreover, two habitat creation modes, named collection mode and diversion mode, were put forward. Collection mode means building water catchment by cutting earthwork in land slope. Diversion mode means building dividing ridge by filling earthwork in inundation area. At the same time, spatial distribution pattern of two modes in river basin scale was proposed based on dynamic spatial relationship between slope and runoff inundation area. It presented a multiscale mosaic structure among diverse dividing ridges and water catchments. Potential in regulating the balance between water and soil and creating habitats being suitable for planting were shown clearly by the applications in ecological engineering practices. It could provide method and technology reference for ecological engineering practices in China.

Mechanism of Operation Mode Effects on Phosphorus Removal and Recovery Performance of PAOs-biofilm

BI Zhen, OUYANG Zhikang, QIAN Mengmeng, LIU Yuqing, HUANG Yong

2024, 60(1):  175-182.  DOI: 10.13209/j.0479-8023.2023.039

Asbtract ( )   HTML   PDF (682KB) ( )  

Related Articles | Metrics

In order to investigate effects of operation mode on the phosphorus removal and recovery performance of PAOs-biofilm, the continuous operating PAOs-biofilm reactor was carried out to recovery phosphorus from synthetic sewage wastewater with different operation modes. It showed that the operating modes resulted in significant differences of the environment for phosphorus uptake and release (phosphorus concentration in water). Under the “anaerobic-aerobic alternating” mode, the high phosphorus loading in aerobic phase strengthened the phosphorus uptake ability of the PAOs-biofilm, and the phosphorus recovery performance depended on the limitation of phosphorus uptake. In contrast, under the “aerobic-anaerobic alternating” mode, the high concentration of phosphorus in anaerobic phase depressed the phosphorus release ability of the PAOs-biofilm, which could be the main reason for the limited phosphorus recovery performance. Such differences in environment resulted from the operation modes led to different aerobic phosphorus uptake, anaerobic phosphorus release and carbon source utilization rate of biofilm. The average Prel/Cupt under “anaerobic-aerobic alternating mode” was 0.213 mmol P /mmol C, which was much higher than that of 0.046 mmol P/mmol C under “aerobic-anaerobic alternating” mode. In summary, the PAOs-biofilm showed higher efficiency in phosphorus removal under “anaerobic-aerobic alternating” mode, while it exhibited a more superior performance in phosphorus recovery under “aerobic-anaerobic alternating” mode, with phosphorus recovery liquid of 99.90 mg/L.

Optimization of Reaction Conditions for Formaldehyde Treatment of Semi-coking Wastewater

WANG Yali, BAI Xubo, SUN Juanjuan

2024, 60(1):  183-187.  DOI: 10.13209/j.0479-8023.2023.024

Asbtract ( )   HTML   PDF (3185KB) ( )  

Related Articles | Metrics

Phenolic substances in semi-coking wastewater were converted into phenolic resin by formaldehyde, therefore the resources were utilized. Volatile phenol, COD, ammonia nitrogen and oil in semi-coking wastewater before and after phenolic resin formation were detected to optimize the reaction time, temperature and raw material ratio, eventually the optimal reaction conditions for phenolic resin preparation were determined. Meanwhile, the physicochemical analysis of phenolic resin materials was carried out by XRD and SEM. The results showed that a volume ratio of 1:40 (formaldehyde vs. semi-coking wastewater), a reaction temperature of 90ºC, and a reaction time of 4 hours were the optimum reaction conditions for semi-coking wastewater treated by formaldehyde.

Translation of Chinese Version of Three Dimensional Meaning in Life Scale and Examinations of Reliability and Validity

MIAO Miao, ZHOU Zhiwei, ZHONG Shaoting, HE Qinghuan, QI Wei

2024, 60(1):  188-194.  DOI: 10.13209/j.0479-8023.2023.100

Asbtract ( )   HTML   PDF (399KB) ( )  

Related Articles | Metrics

To translate and validate the Chinese Version of the Three Dimensional Meaning in Life Scale (3DM), which consists of purpose, significance, and coherence, its reliability, validity, and measurement invariance are examined. Sample 1, which consisted of 544 undergraduate students, was used for exploratory factor analysis. Sample 2, which consisted of 559 adults, was used for confirmatory factor analysis and test of measurement invariance. The Meaning in Life Questionnaire was used as criterion validity; the Positive Affect and Negative Affect Scale and the Satisfaction with Life Scale were used as convergent–discriminant validity. One month later, a follow-up survey was conducted to assess test-retest reliability. The results showed that the three-dimensional structure was supported. The Cronbach’s α of the total scale and the three subscales ranged between 0.82 and 0.95, and their test-retest reliability ranged between 0.71 and 0.83. The total scale and the three subscales were positively associated with presence of meaning, search for meaning, positive affect, and life satisfaction, while negatively associated with negative affect. In addition, results showed acceptable measurement invariance across gender and time. 3DM scale displayed satisfactory reliability and validity, and could be used to assess different components of meaning in life.