Table of Content

20 January 2020, Volume 56 Issue 1

Previous Issue    Next Issue

Automated ICD Coding Based on Word Embedding with Entry Embedding and Attention Mechanism

ZHANG Hongke, FU Zhenxin, REN Qianping, XU Hui, ZHAO Dongyan, YAN Rui

2020, 56(1):  1-8.  DOI: 10.13209/j.0479-8023.2019.095

Asbtract ( )   HTML   PDF (725KB) ( )  

Related Articles | Metrics

The authors propose a neural model based on word embedding with entry embedding and attention mechanism, which can make full use of the unstructured text in the electronic medical record to achieve automated ICD coding for the main diagnosis of the medical record home page. This method first embeds the words which contain the medical record entries into word embeddings, and enriches word-level representation based on keyword attention. Then, the word attention is used to highlight the role of key words and enhance the text representation. Finally, ICD codes are output by a fully connected neural network classifier. Ablation study on a Chinese electronic medical record data set shows that word embedding with entry embedding, keyword attention and word attention is effective. The proposed model gets the best results for 81 diseases classification compared with baselines and can effectively improve the quality of automated ICD coding.

Discourse-Level Text Generation Method Based on Topical Constraint

HUANG Yan, SUN Haili, XU Ke, YU Xiaoyang, WANG Tongyang, ZHANG Xinfang, LU Songfeng

2020, 56(1):  9-15.  DOI: 10.13209/j.0479-8023.2019.103

Asbtract ( )   HTML   PDF (993KB) ( )  

Related Articles | Metrics

To solve the topic missing problem of text generated by computers, this paper proposed a new discourse-level text generation method based on topical constraint. Providing a short topic description, the approach extracted several topic words from the text, then extended and clustered the keywords to form topical planning which were used to restrain the generation of each paragraphs. The model improved the attention based recurrent neural network form three aspects including topic distribution, attention scoring function and topic coverage generation. In experiments, the proposed method was compared with benchmark models such as Char-RNN, SC-LSTM and MTA-LSTM on three real datasets, three improvement strategies were verified and analysed independently. Experimental results show that proposed model is more efficient than benchmark models on human and BLEU metrics, and the generated text can catch the topic more effectively.

An Interactive Stance Classification Method Incorporating Background Knowledge

LIU Changjian, DU Jiachen, LENG Jia, CHEN Di, MAO Ruibin, ZHANG Jun, XU Ruifeng

2020, 56(1):  16-22.  DOI: 10.13209/j.0479-8023.2019.096

Asbtract ( )   HTML   PDF (682KB) ( )  

Related Articles | Metrics

This paper proposes a stance classification method on interactive text by incorporating background knowledge. This method retrieves relevant background knowledge texts from Wikipedia by using the interactive text as query. The retrieved background knowledge texts are encoded and then ultilized to learn the representation of relavent background knowledge through deep memory network for improving the representation learning of interactive text. The experimental results on three English online debate datasets show that the performance of interactive stance classification can be effectively improved by incorporating background knowledge through choosing the appropriate number of background knowledge embedding layers and the connection method of background knowledge embedding layer.

Representation and Recognition of Clauses Relevance Structure in Chinese Text

FENG Wenhe, CHEN Yilin, REN Yafeng, REN Han

2020, 56(1):  23-30.  DOI: 10.13209/j.0479-8023.2019.094

Asbtract ( )   HTML   PDF (491KB) ( )  

Related Articles | Metrics

The discourse structure is represented as the clause relevance structure, which can effectively describe the direct semantic association between discontinuous and cross-level clauses in a text, compared with the hierarchical discourse structure pattern such as rhetorical structure theory. Firstly, the scheme of clause relevance structure, its judgment criteria and formal constraints. The manual annotation experiments are conducted. Then, the automatic recognition of Chinese clause relevance structure is studied. On the corpus of Chinese discourse clause relevance structure we built, the best recognition accuracy is 92.70% based on the classification model, with the connectives, vocabulary and other classification features. The experimental results show that the ring-removing effect obtained by the overall sampling of corpus is better than that of independent sampling, and the features of vocabulary, clause distance and clause domain contribute greatly to the recognition. Long distance and cross sentences of clause pair are the difficulties of clause relevance recognition, but adjacent clauses and clauses in the same sentence are especially difficult to recognize as uncorrelated clauses.

Analysis of Bi-directional Reranking Model for Uyghur-Chinese Neural Machine Translation

ZHANG Xinlu, LI Xiao, YANG Yating, WANG Lei, DONG Rui

2020, 56(1):  31-38.  DOI: 10.13209/j.0479-8023.2019.093

Asbtract ( )   HTML   PDF (899KB) ( )  

Related Articles | Metrics

The fitting training of neural machine translation is easy to fall into a local optimal solution on a lowresource corpus such as Uyghur to Chinese, resulting in the translation result of a single model may not be a global optimal solution. In order to solve this problem, the probability distribution predicted by multiple models is effectively integrated through the ensemble strategy, and multiple translation models are taken as a whole. At the same time, the translation models with opposite decoding directions are integrated by the reordering method based on cross entropy, and the candidate translation with the highest comprehensive score is selected as the output. The experiment on CWMT2015 Uighur-Chinese parallel corpus shows that proposed method has 4.82 BLEU values improvement compared with a single transformer model.

Distant Supervision for Relation Extraction with Gate Mechanism

LI Xingya, CHEN Yufeng, XU Jin’an, ZHANG Yujie

2020, 56(1):  39-44.  DOI: 10.13209/j.0479-8023.2019.101

Asbtract ( )   HTML   PDF (699KB) ( )  

Related Articles | Metrics

A piecewise convolutional neural network with gating mechanism is proposed, which would automatically filter positive correlation features at word-level. Moreover, the idea of soft-label is introduced to the gating mechanism to weaken the impact of hard labels on noise filtering. Combined with sentence-level noise filtering, the overall performance of the model is improved. The experimental results on the public dataset show that the proposed model has a significant improvement compared to the sentence-level noise filtering methods. 

A Neural Paraphrase Identification Model Based on Syntactic Structure

LIU Mingtong, ZHANG Yujie, XU Jin’an, CHEN Yufeng

2020, 56(1):  45-52.  DOI: 10.13209/j.0479-8023.2019.092

Asbtract ( )   HTML   PDF (1016KB) ( )  

Related Articles | Metrics

Paraphrase identification involves natural language semantic understanding. Most previous methods regarded sentences as sequential structures, and used sequential neural network for semantic composition. These methods do not consider the influence of syntactic structure on semantic computation. In this paper, we proposed a neural paraphrase identification model based on syntactic structure, and designed a tree-based neural network model for semantic composition, which extended the semantic representation from word level to phrase level. Furthermore, this paper proposed a syntactic tree alignment mechanism based on phrase-level semantic representation, and extracted features by using cross-sentence attention mechanism. Finally, a self-attention mechanism was used to enhance semantic representation, which could effectively model context information based on syntactic structure. Experiments on Quora paraphrase dataset show that the performance of paraphrase identification has been improved to 89.3% accuracy. The results further prove that the proposed semantic composition method based on syntactic structure, phrase-level cross sentence attention and self-attention are effective in improving paraphrase identification.

A Multi-Mechanism Fused Paraphrase Generation Model with Joint Auto-Encoding Learning

LIU Mingtong, ZHANG Yujie, ZHANG Shu, MENG Yao, XU Jin’an, CHEN Yufeng

2020, 56(1):  53-60.  DOI: 10.13209/j.0479-8023.2019.104

Asbtract ( )   HTML   PDF (828KB) ( )  

Related Articles | Metrics

Neural network encoder-decoder framework has become the popular method for paraphrase generation, but there are still two problems. On the one hand, there are such issues as inaccurate entity words, unknown words and word repetition in the generated paraphrase sentences. To solve the first problem, we proposed a multimechanism fused paraphrase generation model to improve the decoder. The copy mechanism was used to copy words form input sentence for improving the generation of entity and unknown words. The coverage mechanism was used to model historical attention information to avoid word repetition. On the other hand, the limited-scale parallel paraphrase corpus limits the learning ability of the encoder. We proposed to jointly learn auto-encoding task, which shares one encoder with paraphrase generation task. The joint auto-encoding task enhances the learning ability of the encoder. Experimental results on Quora paraphrase dataset show that the multi-mechanism fused paraphrase generation model with joint auto-encoding task can effectively improve the performance of paraphrase generation.

An Abstractive Summarization Method Based on Encoder-Sharing and Gated Network

TIAN Keke, ZHOU Ruiying, DONG Haoye, YIN Jian

2020, 56(1):  61-67.  DOI: 10.13209/j.0479-8023.2019.100

Asbtract ( )   HTML   PDF (642KB) ( )  

Related Articles | Metrics

This paper proposed an abstractive summarization method based on self-attention based Transformer model, which regarded encoder as part of decoder, and used gated network to control the information flow from encoder to decoder. Compared with the existing methods, proposed method improves the training and inference speed of text summarization task, and improves the accuracy and fluency of generating summary. Experiments on English summarization dataset Gigaword and DUC2004 demonstrate that proposed model outperforms the baseline models on both the quality of summarization and time efficiency.

Consumption Intent Recognition Algorithms for Weibo Users

JIA Yunlong, HAN Donghong, LIN Haiyuan, WANG Guoren, XIA Li

2020, 56(1):  68-74.  DOI: 10.13209/j.0479-8023.2019.102

Asbtract ( )   HTML   PDF (717KB) ( )  

Related Articles | Metrics

The data set is constructed by the data of Jingdong Question Answer Platform and Weibo based on transfer learning method and a bi-directional long-term and short-term memory neural network model based on attention mechanism is proposed to identify users’ implicit consumption intention. For the problem of explicit intention recognition, a new algorithm for extracting consumer intention objects is proposed, which combines TFIDF (term frequency-inverse document frequency) with the verb-object relationship (VOB) in parsing. The experimental results show that the training set can be effectively expanded by merging the data of Jingdong Question Answer Platform and Weibo. The classification model has high accuracy and recall rate, and the method of extracting explicit consumer intent objects by fusing VOB and TF-IDF achieves 78.8% accuracy.

Multimodal Emotion Recognition with Auxiliary Sentiment Information

WU Liangqing, LIU Qiyuan, ZHANG Dong, WANG Jiancheng, LI Shoushan, ZHOU Guodong

2020, 56(1):  75-81.  DOI: 10.13209/j.0479-8023.2019.105

Asbtract ( )   HTML   PDF (1064KB) ( )  

Related Articles | Metrics

Different from the previous studies with only text, this paper focuses on multimodal data (text and audio) to perform emotion recognition. To simultaneously address the characteristics of multimodal data, we propose a novel joint learning framework, which allows auxiliary task (multimodal sentiment classification) to help the main task (multimodal emotion classification). Specifically, private neural layers are designed for text and audio modalities from the main task to learn the uni-modal independent dynamics. Secondly, with the shared neural layers from auxiliary task, we obtain the uni-modal representations of the auxiliary task and the auxiliary representations of the main task. The uni-modal independent dynamics is combined with the auxiliary representations for each modality to acquire the uni-modal representations of the main task. Finally, in order to capture multimodal interactive dynamics, we fuse the text and audio modalities’ representations for the main and auxiliary tasks separately to obtain the final multimodal emotion and sentiment representations with the self attention mechanism. Empirical results demonstrate the effectiveness of our approach to multimodal emotion classification task as well as the sentiment classification task.

Event Coreference Resolution with Document Representation

WU Ruiying, KONG Fang

2020, 56(1):  82-88.  DOI: 10.13209/j.0479-8023.2019.091

Asbtract ( )   HTML   PDF (711KB) ( )  

Related Articles | Metrics

Event coreference resolution is more difficult than entity coreference resolution. The main reason is that the event mentions in the unstructured texts are sparse, and most of them do not have the coreference relationship, at the same time, the semantic information carried by the event itself is richer than entity. In order to accurately extract the coreferential events in the text, for the above characteristics of event coreference resolution, an event coreference resolution platform with text representation is proposed. This platform effectively distinguishes non-event mention, single-chain and coreference event mention through CRF, and uses hierarchical attention mechanism to capture important information at sentence level and text level. Experiments on KBP2015 and 2016 datasets verify the validity of the model, and the CoNLL evaluation standard reaches 43.07% of the F1 value.

Syntax-Enhanced UCCA Semantic Parsing

JIANG Wei, LI Zhenghua, ZHANG Min

2020, 56(1):  89-96.  DOI: 10.13209/j.0479-8023.2019.099

Asbtract ( )   HTML   PDF (644KB) ( )  

Related Articles | Metrics

Considering the close correlation between syntactic and semantic structures, this paper attempts to add syntactic information into the universal conceptual cognitive annotation (UCCA) semantic parsing model to enhance the performance of semantic parsing. Based on the state-of-the-art graph-based UCCA semantic parser, we propose and compare four different approaches for incorporating syntactic information. Experiments are conducted on the English benchmark dataset for the semantic parsing shared task of the SemEval-2019 conference. The results on both the in-domain and out-domain evaluation data show that syntax-enhanced methods can achieve significant improvements of UCCA parsing. After utilizing BERT, syntactic information is still beneficial to some extent.

Neural Network Coupled Model for Conversion and Exploitation of Heterogeneous Lexical Annotations

HUANG Depeng, LI Zhenghua, GONG Chen, ZHANG Min

2020, 56(1):  97-104.  DOI: 10.13209/j.0479-8023.2019.098

Asbtract ( )   HTML   PDF (673KB) ( )  

Related Articles | Metrics

In order to expand the scale of manual annotated data and thereby improve model performance, we attempt to make full use of existing heterogeneous annotations to learn model parameters. We extend coupled sequence labeling model proposed by Li et al. (2015) under the BiLSTM-based deep learning framework. The neural coupled model learn its parameters directly on two heterogeneous training data, and predicts two optimal sequences simultaneously during the test phase. A lot of experiments have been conducted on the part-of-speech (POS) tagging task and the joint word segmentation and POS (WS&POS) tagging task. The results show that neural coupled approach is superior to other methods for exploiting heterogeneous lexical data, including the multi-task learning method and the traditional discrete-feature coupled model. Neural coupled model achieves higher performance on both scenarios, i.e., annotation conversion and boost the final target-side tagging accuracy by exploiting heterogeneous data.

User Profiling Based on Multimodal Fusion Technology

ZHANG Zhuang, FENG Xiaonian, QIAN Tieyun

2020, 56(1):  105-111.  DOI: 10.13209/j.0479-8023.2019.097

Asbtract ( )   HTML   PDF (802KB) ( )  

Related Articles | Metrics

Existing studies in user profiling are unable to fully utilize the multimodal information. This paper presents a cross-modal joint representation learning network, and develop a multi-modal fusion model. Firstly, a stacking method is adopted to learn the joint representation network which fuse the cross-modal information. Then, attention mechanism is introduced to automatically learn the contribution of different modal to the prediction task. Proposed model has a well defined loss function and network structure, which enables combining the related features in various models by learning the joint representations after feature-level and decision-level fusion. The extensive experiments on real data sets show that proposed model outperforms the baselines.

A Parallel Algorithm to Answer Shortest Distance on Dynamic Graph

HAN Shuo, ZOU Lei

2020, 56(1):  112-122.  DOI: 10.13209/j.0479-8023.2019.113

Asbtract ( )   HTML   PDF (1188KB) ( )  

Related Articles | Metrics

The paper presents a parallel algorithm framework to answer shortest distance queries on dynamic graphs. Based on maintaining a delta graph, multiple queries within a batch are executed in parallel over different versions of data graph by multi-threading. For each query, bidirectional breath-first search (BFS) is utilized to reduce search space. During the search process, an exploration direction decision-making function is proposed. Furthermore, adjacency-lists of data graph are encoded by BSR layout, combined with SIMD instructions and graph reordering algorithm, higher degree of data-level parallelism is achieved. Extensive experiments on real graph datasets confirm the superior efficiency of the proposed solution.

Meridional Patterns of Surface Warming and the Underlying Mechanisms in Dry and Moist AGCMs

LI Juan, XIA Yan, YANG Jun

2020, 56(1):  123-134.  DOI: 10.13209/j.0479-8023.2019.120

Asbtract ( )   HTML   PDF (1068KB) ( )  

Related Articles | Metrics

Using an atmospheric general circulation model (AGCM) coupled to a slab ocean, the mechanisms for producing the PWA (polar warming amplification) in idealized conditions are investigated. In the simulations, both ice albedo feedback and cloud radiative effects are turned off, but realistic radiative transfer of greenhouse gases (such as CO₂) and atmospheric heat transport are included. Surface albedo is fixed and meridional oceanic heat transport is set to zero everywhere. Through turning on or off surface evaporation, the model is a moist AGCM or becomes a dry AGCM. Results show that under doubling atmospheric CO₂ concentration, PWA occurs in the moist AGCM but not in the dry AGCM. In the dry AGCM, the increases of surface temperatures are nearly uniform from the equator to the poles. The radiative forcing of increased CO₂ and water vapor feedback are stronger in the tropics than those in the polar region, so that the only mechanism for driving the PWA in the moist AGCM is an enhanced meridional heat transport. In the dry AGCM, the poleward heat transport also increases with a much smaller magnitude, so that it is not able to support a PWA. This study emphasizes that water vapor and its associated meridional heat transport are necessary for the PWA on Earth, and PWA may not occur in a dry atmosphere such as Martian atmosphere.

Discussion on the Seismogenic Structure of Qianguo Earthquake of 1119 AD

SHAO Bo, SHEN Jun, HOU Guiting, YU Xiaohui, DAI Xunye, YU Yang

2020, 56(1):  135-142.  DOI: 10.13209/j.0479-8023.2019.111

Asbtract ( )   HTML   PDF (22409KB) ( )  

Related Articles | Metrics

Based on the full collection and textual research on the basis of predecessors’ research results, as well as analysis in-depth of the implicit constraint conditions of the historical earthquake, we study the epicenter of the earthquake location and seismogenic structure, through three-dimensional petrolic geophysical data, the shallow geophysical prospecting and the joint drilling detection, as well as the seismic geological and geomorphic survey. At the same time in the process of detection, within the scope of influence of Qianguo Earthquake of 1119 AD, we found the greatest Late Pleistocene active fault in the zone, named Gudian Fault. The fault is about 66 km long on the seismic profiles from petroleum exploration and consists of two continuous arcs. The buried depth of the fault breakpoint is shallower than 24 m. Through earthquake risk assessment, we judge that the Gudian Fault is most likely the seismogenic structure of the Qianguo earthquake of 1119 AD.

Random Forest Model for the Estimation of Fractional Vegetation Coverage Based on a UAV-Ground Co-Sampling Strategy

CHENG Junyi, ZHANG Xianfeng, SUN Min, LUO Peng, YANG Wanting

2020, 56(1):  143-154.  DOI: 10.13209/j.0479-8023.2019.110

Asbtract ( )   HTML   PDF (23545KB) ( )  

Related Articles | Metrics

A nonparametric regression — random forest model for the estimation of fractional vegetation coverage (FVC) in a complex topographic area is presented based on low-altitude unmanned aerial vehicle (UAV) hyperspectral imagery. In order to collect a large number of sufficient training samples required for random forest algorithm, the UAV equipped with an optical camera was used to vertically capture the images of land covers in several inaccessible areas such as high mountains, water body and densely forested areas, to increase the density of the ground sampling. The RGBVI (red-green-blue vegetation index) was calculated first and then the Otsu method was adopted to extract the FVC values of the samples from the UAV optical images and ground photos. After that, the hyperspectral images captured by the UAV GaiaSky-mini2 hyperspectral imaging system in the Youlougou Mining area, Chayouzhong County, Inner Mongolia on August 16?18, 2018 were used to extract feature variables, and this feature set was filtered by recursive feature elimination algorithm based on the importance of the variables. On the basis of the optimized feature set and extended training samples using the proposed UAV-ground cosampling approach, the random forest estimation model was constructed to estimate the FVC in the study area. Results indicated that the model achieved a determinant coefficient (R²) of 0.923 and a RMSE of 0.087 on the testing sample set and outperformed the commonly used Pixel Dichotomy method. It can be used in the fast and accurate monitoring of vegetation dynamics in mining areas.

Improving One-Class Classification of Remote Sensing Data by Using Active Learning: A Case Study of Positive and Unlabeled Learning

SUN Yi, LI Peijun

2020, 56(1):  155-163.  DOI: 10.13209/j.0479-8023.2019.035

Asbtract ( )   HTML   PDF (11078KB) ( )  

Related Articles | Metrics

To address the problem that quality and quantity of training samples directly affect accuracy of oneclass classification (OCC) methods, this paper investigates the use of active learning in selection of training samples of target class (positive samples) for improving the performance of OCC, by taking positive and unlabeled learning (PUL) as an example. PUL is first trained with sufficient training samples selected randomly until a stable accuracy is reached. Most informative positive and negative training samples collected by using active learning strategy are then added in PUL classification. The experimental results show that after sufficient samples are used for classification, the use of positive samples selected by using active learning still outperformed that using sufficient positive samples selected randomly. PUL classification by adding both positive and negative samples outperformed that by adding positive samples only. Furthermore, PUL classification using positive samples after removal of redundant positive samples from those directly selected by active learning obtains accuracy comparable to that using more positive samples directly selected by active learning, whereas less samples are needed in the case that redundant samples are removed. This study demonstrates that selecting and adding samples by active learning provides a more effective way of improving accuracy for OCC.

Research on Moving Target Indication Based on Along Track Interferometry of TerraSAR-X Data

JIAO Jian, TIAN Chongrui, HUANG Jianghui, ZENG Qiming

2020, 56(1):  164-172.  DOI: 10.13209/j.0479-8023.2019.112

Asbtract ( )   HTML   PDF (6266KB) ( )  

Related Articles | Metrics

The method to combine along track interferometry (ATI) and constant false alarm rate (CFAR) detection, referred to as ATI-CFAR, is considered to be a promising method for ground moving target indication (GMTI). In order to evaluate the ability of the method used for GMTI with TerraSAR-X data, a set of experimental schemes are presented, which include a synchronous in-situ experiment on a section of Beijing’s North Fifth Ring Road, an improved ATI-CFAR method to estimate ATI amplitude and phase, etc. Experimental results suggest that TerraSAR-X data can be used in GMTI, however ATI phase is easily interfered. A published ATI-CFAR method might overestimate the threshold of interferometric phases, and result in missing detection of moving targets. The proposed method based on the priori knowledge of vehicle velocity is able to effectively improve the detection performance, and it improves the detection rate of moving targets in the study area up to 70%, and has an accuracy of 87.5%. The work of this paper validates the availability and potential of TerraSAR-X data in GMTI application.

Effects of Air Temperature and Soil Moisture on Common Brown-Down Date of Taraxacum mongolicum in Eastern China’s Temperate Zone

XUE Tingting, ZHAO Yuan, CHEN Xiaoqiu, JIANG Mengdi, LIANG Boyi

2020, 56(1):  173-183.  DOI: 10.13209/j.0479-8023.2019.127

Asbtract ( )   HTML   PDF (1051KB) ( )  

Related Articles | Metrics

In order to reveal temporal variations of autumn phenology of herbaceous plants and their climatic attributions, we analyzed changing trend of the common brown-down date of Taraxacum mongolicum and simulated interannual variation of the common brown-down date using plant phenological and meteorological data at 47 phenological stations in eastern China’s temperate zone from 1992 to 2012 and statistical methods. Results show that 1) the common brown-down date of Taraxacum mongolicum was delayed at 34 stations from 1992 to 2012, of which significant delaying trends were observed at 22 stations. By contrast, the common brown-down date of Taraxacum mongolicum advanced at 13 stations, of which significant advancing trends were detected at five stations. 2) The common brown-down date of Taraxacum mongolicum at individual stations correlates mainly negative with average temperature during the growing season (the period from leaf unfolding to common browndown) but positive with relative soil moisture and daily minimum temperature in autumn. 3) Among the 30 effective optimal models at individual stations (p<0.05), the common brown-down date of Taraxacum mongolicum was influenced by relative soil moisture at 22 stations, by daily minimum temperature at 19 stations and by average temperature during the growing season at 21 stations. Moreover, simulation accuracy of the model was significantly affected by interannual variation of the common brown-down date of Taraxacum mongolicum, namely, the smaller the interannual variation of the common brown-down date at a station, the higher the simulation accuracy of the model.

Seasonal and Spatial Variations of Phytoplankton Communities and Correlations with Environmental Factors in Lake Dianchi

FENG Qiuyuan, WANG Shuran, LIU Xueqin, LIU Yong

2020, 56(1):  184-192.  DOI: 10.13209/j.0479-8023.2019.128

Asbtract ( )   HTML   PDF (744KB) ( )  

Related Articles | Metrics

The spatial and seasonal variations of environmental factors and phytoplankton community in Lake Dianchi was investigated. The statistical analysis found that a total of 84 phytoplankton taxon, belonging to 49 genera 6 phyla were identified, of which Chlorophyte were the most abundant, accounting for 59.2%, followed by Cyanophyta, accounting for 16.67% and diatom accounting for 5.95%. The trend of seasonal variation of algae density and biomass was not the same, which was caused by the biomass differences of various species. The Shannon Wiener index (H) was very small throughout the year, and seasonal variation was not significant. Algae density and biomass were positively correlated with total phosphorus (TP), and negatively correlated with nitrate (NO₃^-) and nitrogen phosphorus ratio (N:P). H was positively correlated with NO₃^- and N:P, while negatively correlated with TP. Nutrients had great influences on the density and biomass distributions of various taxa of phytoplankton community. The density and biomass of the most dominant species Microcystis sp. had the same relationships with the environmental factors, which were positively correlated with pH, TP and ammonia (NH₄⁺), and negatively correlated with NO₃^-, N:P, total nitrogen (TN), dissolved organic carbon (DOC) and total organic carbon (TOC). Some biological factors may cause stronger effects on density and biomass distribution of chlorella and diatom, such as interspecific competition and predation, covering the influences of environmental factors.