基于递归自编码器的广告短语相关性

发布时间：2018-01-15 15:04

本文关键词：基于递归自编码器的广告短语相关性　出处：《计算机应用》2016年01期 　论文类型：期刊论文

【摘要】：针对现有广告短语相关性研究成果多采用字面匹配,忽略了短语所包含的深层语义信息,限制了任务的性能等问题,提出了采用深度学习算法研究广告短语的相关性,采用递归自编码器(RAE)对短语进行深层结构分析,使得短语向量包含深层的语义信息,以此来构建广告语境下的短语相关性计算方法。具体地,给定一个包含若干词的序列,序列中所有相邻的两个元素尝试合并产生一个重构误差,遍历将重构误差最小的元素两两合并,形成类似哈夫曼树结构的短语树。采用梯度下降法最小化短语树的重构误差,采用余弦距离度量短语之间的相关性。实验结果显示,通过引入词语权重信息,加大了重要词语在最终短语向量表示中贡献的信息量,使得RAE更适合短语计算;比起传统LDA和BM25算法,在50%召回率的条件下,提出的算法的准确率分别提高了4.59个百分点和3.21个百分点,这证明了所提算法的有效性。
[Abstract]:In view of the existing research on the relevance of advertising phrases, most of them use literal matching, ignoring the deep semantic information contained in the phrase, limiting the performance of the task and so on. A depth learning algorithm is proposed to study the correlation of advertising phrases, and a recursive self-encoder is used to analyze the deep structure of phrases, which makes phrase vectors contain deep semantic information. Specifically, given a sequence containing a number of words, all the adjacent elements in the sequence try to merge to produce a reconstruction error. Traversing the elements with the minimum reconstruction error to form a phrase tree similar to the Huffman tree, the gradient descent method is used to minimize the reconstruction error of the phrase tree. CoSine distance is used to measure the correlation between phrases. The experimental results show that by introducing word weight information, the amount of information contributed by important words in the final phrase vector representation is increased. Make RAE more suitable for phrase calculation; Compared with the traditional LDA and BM25 algorithms, the accuracy of the proposed algorithm is increased by 4.59% and 3.21% respectively under the condition of 50% recall rate. This proves the validity of the proposed algorithm.
【作者单位】：武汉大学计算机学院;桂林航天工业学院广西高校机器人与焊接技术重点实验室培育基地;
【基金】：国家自然科学基金资助项目(11301106) 广西自然科学基金资助项目(2014GXNSFAA1183105) 广西高校科研资助项目(ZD2014147,YB2014431)~~
【分类号】：TP18;TP391.3
【正文快照】： 0引言计算广告指的是基于用户给定的搜索信息,通过计算得到与之最匹配的广告,然后进行精准定向投放的广告机制。采用该机制可以大幅度提高广告主投放广告的点击率,加大广告投放所在网站的访问量,从而帮助用户获得优质的信息。计算广告是支撑互联网广告产业链发展的基础,其涉及

【参考文献】