施工组织设计文档智慧辅助审查中的文本分类问题研究

施工组织设计是指导工程建设全过程活动的技术、经济和组织的综合性文件,随着自然语言处理(NLP)等人工智能技术的发展,本文针对施工组织设计文档智慧辅助审查中基础性工作之一-文本分类问题开展研究。为实现施工组织设计文本的自动分类,本文运用Word2vec词嵌入技术对文本进行向量化表示,基于Bi-LSTM捕捉文本上下文序列信息,融入Attention机制,提取文本有效信息,采用softmax激活函数分类。Attention Bi-LSTM在房建数据集上达到了0.97的准确率、召回率以及F1值,整体分类效果在正确率、宏平均、加权平均上均优于其他模型。融入Attention机制的Bi-LSTM文本分类模型通过双向捕获文本的特征并利用Attention机制提取有效信息,达到了联合优化的作用,提高了模型的分类性能。     

一种基于词覆盖的新闻事件脉络链构建方法

针对目前构建新闻脉络链只关注新闻脉络链的主题相似性和文档重要性, 而忽略新闻脉络链逻辑连贯性和可解释性的不足, 以及新闻数据集合指数级增长带来的算法复杂度问题, 从词覆盖的角度提出一种新闻脉络链构建方法, 利用新闻的评论信息来定位新闻事件转折点, 用主题相似与稀疏差异的思想以及RPCA 方法对文档进行逻辑建模, 利用随机游走以及图遍历的方法, 量化并生成可解释且具有很好逻辑连贯性的脉络链。双盲实验表明, 与其他算法相比, 该方法取得较好的效果。     

中国古诗统计建模与宏观分析

利用自然语言处理技术处理文学文本是计算语言学领域近年来的热门话题.该文结合点态互信息量与频率阈值,自动发现中国古诗词汇.基于构建的诗歌词典,利用启发式的正向最大匹配算法,对中国古诗作分词处理.采用主题模型对分词后的诗歌文本进行统计建模,并在此基础上进行了主题演变和诗人群体风格网络的探索性分析.基于全唐诗语料的实验结果表明:主题模型可以给出具有较好解释力的中国古诗统计模型,验证已有的文学史研究,并在传统的文本细读的研究范式之外,对中国诗学提供了全新视角的宏观刻画、描述与阐释.     

基于多特征融合的中文评论情感分类算法

为解决情感分类中词间的语义关系难以表达和分析的问题,提出了一种基于词向量(word representation)和支持向量机(support vector machine)的情感分类算法,对电子商务在线评论的情感分类问题进行研究.首先使用word2vec聚类相似特征,然后使用word2vec和SVM对情感数据进行训练和分类,并分别使用基于词特征和基于词性标注的方法进行特征选择.在京东评论数据上进行的实验结果表明,与现有方法相比,分类准确率和召回率得到了提高.     

ProSeg: a database of local structures of protein segments

Integration of knowledge on the sequence-structure correlation of proteins provides a basis for the structural design of artificial novel proteins. As one of strategies, it is effective to consider a short segment, whose size is in between an amino acid and a domain, as a correlation unit for exploring the structure-to-sequence relationship. Here we report the development of a database called ProSeg, which consists of two sub-databases, Segment DB and Cluster DB. Segment DB contains tens of thousands of segments that were prepared by dividing the primary sequences of 370 proteins using a sliding L-residue window (L = 5, 9, 11, 15). These segments were classified into several thousands of clusters according to their three-dimensional structural resemblance. Cluster DB contains much cluster-related information, which includes image, rank, frequency, secondary structure assignment, sequence profile, etc. Users can search for a suitable cluster by inputting an appropriate parameter (i.e., PDB ID, dihedral angles, or DSSP symbols), which identifies the backbone structure of a query segment. Analogous to a language, ProSeg could be regarded as a ‘structure-sequence dictionary’ that contains over 10,000 ‘protein words’. ProSeg is freely accessible through the Internet ().     

High sensitive determination of trace amount of cobalt by catalytic adsorptive stripping voltammetry

A very sensitive and selective catalytic adsorptive cathodic stripping procedure for trace measurements of cobalt is presented. The method is based on adsorptive accumulation of cobalt-CCA (calcon carboxylic acid) complex onto a hanging mercury drop electrode followed by reduction of the adsorbed species by voltammetric scan using differential pulse modulation. The reduction current is enhanced catalytically by nitrite. The effect of various parameters such as pH, concentration of CCA, concentration of nitrite, accumulation potential and accumulation time on the selectivity and sensitivity were studied. The optimum condition for the analysis of cobalt, include pH 5.2 (Acetate buffer), 2.1 μM clacon carboxylic acid, 0.032 M sodium nitrite and an accumulation potential of 0.05 V (versus Ag/AgCl). Under these optimum conditions and for an accumulation time of 60 s, the measured peak current at −0.480 V is proportional to the concentration of cobalt over the entire concentration range tested 0.003–2.0 ng ml⁻¹ with a detection limit of 1 pg ml⁻¹ for an accumulation time of 60 s and 2.0–10.0 ng ml⁻¹ for an accumulation time of 40 s. The relative standard deviations for ten replicate measurement of 0.5 ng ml⁻¹ of cobalt were 3.1%. The main advantage of this new system is the microtrace Co(II) determination by ASV. The method was applied to determination of cobalt in a water sample and some analytical grade salts with satisfactory results. Published in Elektrokhimiya in Russian, 2009, Vol. 45, No. 2, pp. 221–228. The article is published in the original.     

基于DF算法改进的文本聚类特征选择算法

通过研究文本特征选取中权重的计算问题,提出了一种利用特征词的熵函数加权的权值的计算方法,不但考察了特征词的文档频数,而且考察了它们在文档中出现的次数,使选出的特征子集更具有较好的代表性.实验表明,改进后的算法对聚类结果有了一定的改进.     

中国特色社会主义道路论析

中国特色主义道路是科学社会主义基本原则与中国实际和时代特征相结合的产物,是中国近现代历史发展的必然选择,是我国实现现代化和民族伟大复兴的必由之路。在探索中国特色社会主义道路过程中,以毛泽东、邓小平、江泽民和胡锦涛为核心的党中央领导集体都做出了不懈努力,取得了丰硕成果。中国特色社会主义道路是马克思主义中国化发展进程中的又一次历史性飞跃,为当代世界社会主义运动和发展中国家的建设发展提供了重要经验借鉴。     

基于综合信息的汛期库水位实时动态控制方法

以提高洪水资源利用率为目标,研究利用实时水雨工情及短期雨洪预报等综合信息的汛期库水位实时动态控制方法,即改进的预蓄预泄法.应用实例表明,在不降低水库上下游防洪标准的前提条件下,可使桓仁水库及梯级库群的发电效益明显增加.该综合信息的汛期库水位实时动态控制方法,对北方水资源短缺地区调节性能较高的大型水库(水电站)汛期实时调度,有较好的借鉴意义.