基于逐步优化分类模型的跨领域文本情感分类

跨领域文本情感分类已成为自然语言处理领域的一个研究热点。针对传统主动学习不能利用领域间的相关信息以及词袋模型不能过滤与情感分类无关的词语,提出了一种基于逐步优化分类模型的跨领域文本情感分类方法。首先选择源领域和目标领域的公共情感词作为特征,在源领域上训练分类模型,再对目标领域进行初始类别标注,选择高置信度的文本作为分类模型的初始种子样本。为了加快目标领域的分类模型的优化速度,在每次迭代时,选取低置信度的文本供专家标注,将标注的结果与高置信度文本共同加入训练集,再根据情感词典、评价词搭配抽取规则以及辅助特征词从训练集中动态抽取特征集。实验结果表明,该方法不仅有效地改善了跨领域情感分类效果,而且在一定程度上降低了人工标注样本的代价。     

改进的语义分割模型及其应用

训练语义分割网络模型需要较为繁琐的人工标注作为训练标签,同时语义分割模型在构建和运行过程中也存在超参数较难确定以及模型过于庞大等问题。为解决这类问题,提出了一种基于标注框生成热点图的标签生成方法,简化了语义分割训练标签的人工标注过程。以及在可微分神经网络结构搜索方法的基础上提出了一种对硬件要求更低的神经网络结构搜索方法,并基于此种方法改进了特征金字塔结构,构建了一个改进的语义分割模型,并在安全帽与口罩检测数据集上进行了试验。与U-Net、FPN等模型比较,新的模型在参数量、计算速度以及精确度上都更有优势。     

一种结合众包的排序学习算法

针对有监督排序学习所需带标记训练数据集不易获得的情况,引入众包这种新型大众网络聚集模式来完成标注工作,为解决排序学习所需大量训练数据集标注工作耗时耗力的难题提供了新的思路。首先介绍了众包标注方法,着重提出两种个人分类器模型来解决众包结果质量控制问题,同时考虑标注者能力和众包任务的难度这两个影响众包质量的因素。再基于得到的训练集使用RankingSVM进行排序学习并在微软OHSUMED数据集上衡量了该方法在NDCG@n评价准则下的性能。实验结果表明该众包标注方法能够达到95%以上的正确率,所得排序模型的性能基本和RankingSVM算法持平,从而验证了众包应用于排序学习的可行性和优越性。     

一种自动分类的网页搜索排序算法

针对传统网页排序算法Okapi BM25通常会出现网页与查询关键词领域无关的领域漂移现象,以及改进算法需要人工建立领域向量的问题,提出了一种基于BM25和Softmax回归分类模型的网页搜索排序算法。该方法首先对网页文本进行数据预处理并利用词袋模型进行网页文本的向量表示,之后通过少量的网页数据来训练Softmax回归分类模型,来预测测试网页数据的类别分数,并与BM25信息检索的分数结合在一起,得到最终的网页排序结果。实验结果显示该检索算法无须人工建立领域向量,即可达到很好的网页排序结果。     

一种结合LPA半监督学习的排序学习算法

针对有监督排序学习所需训练集的大量标注数据不易获得的情况,引入基于图的标签传播半监督学习。利用有限的已标记数据和大量未标记数据来完成训练数据的自动标注工作,解决大量训练数据集标注工作耗时耗力的难题。首先以训练数据为节点建立εNN图模型实现标签传播算法进行训练数据的自动标注,再基于得到的训练集使用Ranking SVM实现排序学习,在OHSUMED数据集上衡量该方法在MAP和NDCG@n评价准则下的性能。实验结果表明,该方法的性能优于普通pointwise排序学习方法,略低于普通pairwise排序学习方法,能够在达到可用性要求的前提下节省接近60%的训练集标注工作量。     

基于Active Learning的中文分词领域自适应

在新闻领域标注语料上训练的中文分词系统在跨领域时性能会有明显下降。针对目标领域的大规模标注语料难以获取的问题,该文提出Active learning算法与n-gram统计特征相结合的领域自适应方法。该方法通过对目标领域文本与已有标注语料的差异进行统计分析,选择含有最多未标记过的语言现象的小规模语料优先进行人工标注,然后再结合大规模文本中的n-gram统计特征训练目标领域的分词系统。该文采用了CRF训练模型,并在100万句的科技文献领域上,验证了所提方法的有效性,评测数据为人工标注的300句科技文献语料。实验结果显示,在科技文献测试语料上,基于Active Learning训练的分词系统在各项评测指标上均有提高。
     

适应文档检索的半监督多样本排序学习算法

针对标记训练集不足的问题,提出了一种协同训练的多样本排序学习算法,从无标签数据挖掘隐含的排序信息。算法使用了两类多样本排序学习机,从当前已有的标记数据集分别构造两个不同的排序函数。相应地,每一个无标签查询都有两个不同的文档排列,由似然损失来计算这两个排列的相似性,为那些文档排列相似度低的查询贴上标签,使两个多样本排序学习机新增了训练数据。在排序学习公开数据集LETOR上的实验结果证实,协同训练的排序算法很有效。另外,还讨论了标注比例对算法的影响。     

判别式藏语文本词性标注研究

该文在分析了现有藏文词性标注方法的基础上,提出感知机训练模型的判别式藏语词性标注方法,重点研究了符合藏语词法特性的模型训练特征模板、模型训练和词性标注方法。并且在人工标注的测试集上获得了98.26%的词性标注精确率,可以实际应用到藏语自然语言处理中。     

面向少量标注数据的中文命名实体识别

训练数据的缺乏是目前命名实体识别存在的一个典型问题。实体触发器可以提高模型的成本效益，但这种触发器需要大量的人工标注，并且只适用于英文文本，缺少对其他语言的研究。为了解决现有TMN模型实体触发器高成本和适用局限性的问题，提出了一种新的触发器自动标注方法及其标注模型GLDM-TMN。该模型不仅能够免去人工标注，而且引入了Mogrifier LSTM结构、Dice损失函数及多种注意力机制增强触发器匹配准确率及实体标注准确率。在两个公开数据集上的仿真实验表明：与TMN模型相比，在相同的训练数据下，GLDM-TMN模型的F₁值在Resume NER数据集和Weibo NER数据集上分别超出TMN模型0.0133和0.034。同时，该模型仅使用20%训练数据比例的性能就可以优于使用40%训练数据比例的BiLSTM-CRF模型性能。     

用于提高谷歌图像搜索结果的二分类器在线学习方法

对于基于关键词的图像检索,利用检索结果的视觉相似性学习二分类器有望成为改善检索结果的最有效途径之一. 为改善搜索引擎的搜索结果,本文提出一种算法框架并且基于此框架着重研究训练数据选择这一关键问题. 训练数据选择过程由两个阶段组成：1）训练数据初始化以开始分类器学习过程;2）分类器迭代学习过程中的动态数据选择. 对于初始训练数据的选择,我们探讨了基于聚类和基于排序两种方法,并且对比了自动训练数据选择与人工标注的结果. 对于动态数据选择,我们比较了支持向量机和基于最大最小后验伪概率的贝叶斯分类器的分类效果. 组合上述两个阶段的不同方法,我们得到了8种不同的算法,并将其用于谷歌搜索引擎进行基于关键词的图像检索. 实验结果证明,如何从含有噪声的搜索结果中选择训练数据是搜索结果改善的关键问题. 实验显示我们的方法能够有效的改善谷歌搜索的结果,尤其是排序在前的结果. 尽早为用户提供更相关的结果能够更大程度的减少用户逐个翻页查看结果的工作. 另外,如何使自动训练数据选择与人工标注媲美仍是需要继续研究的一个问题.     

Active evaluation of ranking functions based on graded relevance

Evaluating the quality of ranking functions is a core task in web search and other information retrieval domains. Because query distributions and item relevance change over time, ranking models often cannot be evaluated accurately on held-out training data. Instead, considerable effort is spent on manually labeling the relevance of query results for test queries in order to track ranking performance. We address the problem of estimating ranking performance as accurately as possible on a fixed labeling budget. Estimates are based on a set of most informative test queries selected by an active sampling distribution. Query labeling costs depend on the number of result items as well as item-specific attributes such as document length. We derive cost-optimal sampling distributions for the commonly used performance measures Discounted Cumulative Gain and Expected Reciprocal Rank. Experiments on web search engine data illustrate significant reductions in labeling costs.     

Exploiting location information for Web search

Most Web pages contain location information, which are usually neglected by traditional search engines. Queries combining location and textual terms are called as spatial textual Web queries. Based on the fact that traditional search engines pay little attention in the location information in Web pages, in this paper we study a framework to utilize location information for Web search. The proposed framework consists of an offline stage to extract focused locations for crawled Web pages, as well as an online ranking stage to perform location-aware ranking for search results. The focused locations of a Web page refer to the most appropriate locations associated with the Web page. In the offline stage, we extract the focused locations and keywords from Web pages and map each keyword with specific focused locations, which forms a set of <keyword, location> pairs. In the second online query processing stage, we extract keywords from the query, and computer the ranking scores based on location relevance and the location-constrained scores for each querying keyword. The experiments on various real datasets crawled from nj.gov, BBC and New York Time show that the performance of our algorithm on focused location extraction is superior to previous methods and the proposed ranking algorithm has the best performance w.r.t different spatial textual queries.     

Retrieving keyworded subgraphs with graph ranking score

Keyword queries have long been popular to search engines and to the information retrieval community and have recently gained momentum for its usage in the expert systems community. The conventional semantics for processing a user query is to find a set of top-k web pages such that each page contains all user keywords. Recently, this semantics has been extended to find a set of cohesively interconnected pages, each of which contains one of the query keywords scattered across these pages. The keyword query having the extended semantics (i.e., more than a list of keywords hyperlinked with each other) is referred to the graph query. In case of the graph query, all the query keywords may not be present on a single Web page. Thus, a set of Web pages with the corresponding hyperlinks need to be presented as the search result. The existing search systems reveal serious performance problem due to their failure to integrate information from multiple connected resources so that an efficient algorithm for keyword query over graph-structured data is proposed. It integrates information from multiple connected nodes of the graph and generates result trees with the occurrence of all the query keywords. We also investigate a ranking measure called graph ranking score (GRS) to evaluate the relevant graph results so that the score can generate a scalar value for keywords as well as for the topology.     

Regularized query classification using search click information

Hundreds of millions of users each day submit queries to the Web search engine. The user queries are typically very short which makes query understanding a challenging problem. In this paper, we propose a novel approach for query representation and classification. By submitting the query to a web search engine, the query can be represented as a set of terms found on the web pages returned by search engine. In this way, each query can be considered as a point in high-dimensional space and standard classification algorithms such as regression can be applied. However, traditional regression is too flexible in situations with large numbers of highly correlated predictor variables. It may suffer from the overfitting problem. By using search click information, the semantic relationship between queries can be incorporated into the learning system as a regularizer. Specifically, from all the functions which minimize the empirical loss on the labeled queries, we select the one which best preserves the semantic relationship between queries. We present experimental evidence suggesting that the regularized regression algorithm is able to use search click information effectively for query classification.     

Topic-sensitive PageRank: a context-sensitive ranking algorithm for Web search

The original PageRank algorithm for improving the ranking of search-query results computes a single vector, using the link structure of the Web, to capture the relative "importance" of Web pages, independent of any particular search query. To yield more accurate search results, we propose computing a set of PageRank vectors, biased using a set of representative topics, to capture more accurately the notion of importance with respect to a particular topic. For ordinary keyword search queries, we compute the topic-sensitive PageRank scores for pages satisfying the query using the topic of the query keywords. For searches done in context (e.g., when the search query is performed by highlighting words in a Web page), we compute the topic-sensitive PageRank scores using the topic of the context in which the query appeared. By using linear combinations of these (precomputed) biased PageRank vectors to generate context-specific importance scores for pages at query time, we show that we can generate more accurate rankings than with a single, generic PageRank vector. We describe techniques for efficiently implementing a large-scale search system based on the topic-sensitive PageRank scheme.     

Web数据集成系统基于QC模型的物化视图选择

在Web数据集成系统中,物化视图能够有效地减少网络传输代价,提高系统的查询效率．如何选择查询进行物化,使得选中的查询满足集成层的空间限制,同时获取最大物化收益,成为集成系统中一个迫切需要解决的问题．传统方法没有考虑到海量XML查询之间的包含关系,其选择的物化视图中可能包含冗余的信息．针对上述问题,提出了①Web数据集成系统中海量查询集合的QC(query containment)模型,该模型能够捕捉查询之间最常见的包含关系;②基于QC模型的物化视图选择算法,算法考虑了物化视图选择相关的主要因素,包括查询提交的频率、空间代价、查询重写能力和查询结果的完备性,提出了查询位图的物化视图组织方式,从而获取更加合理的物化视图选择方案．实验结果证明了该方法的有效性．     

Improving large-scale search engines with semantic annotations

Traditional search engines have become the most useful tools to search the World Wide Web. Even though they are good for certain search tasks, they may be less effective for others, such as satisfying ambiguous or synonym queries. In this paper, we propose an algorithm that, with the help of Wikipedia and collaborative semantic annotations, improves the quality of web search engines in the ranking of returned results. Our work is supported by (1) the logs generated after query searching, (2) semantic annotations of queries and (3) semantic annotations of web pages. The algorithm makes use of this information to elaborate an appropriate ranking. To validate our approach we have implemented a system that can apply the algorithm to a particular search engine. Evaluation results show that the number of relevant web resources obtained after executing a query with the algorithm is higher than the one obtained without it.     

混合P2P环境下有效的查询扩展及其搜索算法

查询扩展是解决信息获取领域中用词歧义性问题的关键技术,并被广泛应用于搜索引擎中,获得了巨大的成功.然而,由于P2P(peer-to-peer)系统是一个分散的、动态的系统,在P2P环境下进行有效的查询扩展具有一定的挑战性.首先,利用查询与文档的关联关系构建了LEM(local expansion method)查询扩展方法;然后,基于查询与文档用词的直接关联,提出了HEM(history_based expansion method)查询扩展方法.在此基础上,提出了一种基于查询扩展的混合P2P环境下的搜索算法.实验及分析结果表明,查询扩展及其搜索算法能够极大地提高搜索的效果.     

Using Local Popularity of Web Resources for Geo-Ranking of Search Engine Results

Search engines retrieve and rank Web pages which are not only relevant to a query but also important or popular for the users. This popularity has been studied by analysis of the links between Web resources. Link-based page ranking models such as PageRank and HITS assign a global weight to each page regardless of its location. This popularity measurement has shown successful on general search engines. However unlike general search engines, location-based search engines should retrieve and rank higher the pages which are more popular locally. The best results for a location-based query are those which are not only relevant to the topic but also popular with or cited by local users. Current ranking models are often less effective for these queries since they are unable to estimate the local popularity. We offer a model for calculating the local popularity of Web resources using back link locations. Our model automatically assigns correct locations to the links and content and uses them to calculate new geo-rank scores for each page. The experiments show more accurate geo-ranking of search engine results when this model is used for processing location-based queries.     

Semplore: A scalable IR approach to search the Web of Data

The Web of Data keeps growing rapidly. However, the full exploitation of this large amount of structured data faces numerous challenges like usability, scalability, imprecise information needs and data change. We present Semplore, an IR-based system that aims at addressing these issues. Semplore supports intuitive faceted search and complex queries both on text and structured data. It combines imprecise keyword search and precise structured query in a unified ranking scheme. Scalable query processing is supported by leveraging inverted indexes traditionally used in IR systems. This is combined with a novel block-based index structure to support efficient index update when data changes. The experimental results show that Semplore is an efficient and effective system for searching the Web of Data and can be used as a basic infrastructure for Web-scale Semantic Web search engines.