Deep learning is the new frontier of machine learning re- search, which has led to many recent breakthroughs in English natu- ral language processing. However, there are inherent differences between Chinese and English, and little work has been done to apply deep learning techniques to Chinese natural language processing. In this paper, we pro- pose a deep neural network model: text window denoising autoencoder, as well as a complete pre-training solution as a new way to solve clas- sical Chinese natural language processing problems. This method does not require any linguistic knowledge or manual feature design, and can be applied to various Chinese natural language processing tasks, such as Chinese word segmentation. On the PKU dataset of Chinese word segmentation bakeoff 2005, applying this method decreases the F1 error rate by 11.9\% for deep neural network based models. We are the first to apply deep learning methods to Chinese word segmentation to our best knowledge.

Web queries are time sensitive which implies that user’s intent for information changes over time. How to recognize temporal intents behind user queries is crucial towards improving the performance of search engines. However, to the best of our knowledge, this problem has not been studied in existing work. In this paper, we propose a timebased query classification approach to understand user’s temporal intent automatically. We first analyzed the shared features of queries’ temporal intent distributions. Then, we present a query taxonomy which group queries according to their temporal intents. Finally, for a new given query, we propose a machine learning method to decide its class in terms of its search frequency over time recorded in Web query logs. Experiments demonstrate that our approach can understand users’ temporal intents effectively.

With the tremendous increasing of information, the demands of in-formation from people advanced the development of Nature Language Process-ing (NLP). As a consequent, Sentence compression, which is an important part of automatic summarization, draws much more attention. Sentence compression has been widely used in automatic title generation, Searching Engine, Topic de-tection and Summarization. Under the framework of discriminative model, this paper presents a decoding method based on Integer Linear Programming (ILP), which considers sentence compression as the selection of the optimal com-pressed target sentence. Experiment results show that the ILP-based system maintains a good compression ratio while remaining the main information of source sentence. Compared to other decoding method, this method has the ad-vantage of speed and using fewer features in the case of similar results obtained.

In the process of developing a domain-specific Chinese-English ma-chine translation system, the accuracy of Chinese word segmentation on large amounts of training text often decreases because of unknown words. The lack of domain-specific annotated corpus makes supervised learning approaches un-able to adapt to a target domain. This problem results in many errors in transla-tion knowledge extraction and therefore seriously lowers translation quality. To solve the domain adaptation problem, we implement Chinese word segmenta-tion by exploring n-gram statistical features in large Chinese raw corpus and bi-lingually motivated Chinese word segmentation, respectively. Moreover, we propose a method of combining multiple Chinese word segmentation results based on linear model to augment domain adaptation. For evaluation, we con-duct experiments of Chinese word segmentation and Chinese-English machine translation using the data of NTCIR-10 Chinese-English patent task. The expe-rimental results showed that the proposed method achieves improvements in both F-measure of the Chinese word segmentation and BLEU score of the Chi-nese-English statistical machine translation system.

Unlike English, Chinese sentences do not have a strict syntactic structure and ellipsis is a common phenomenon, which weaken the effective-ness of syntactic structure in argument extraction. In Chinese event extraction, lots of arguments cannot be extracted from the sentence successfully, because of the loose connection between the nominal trigger and its arguments. This paper brings forward a novel argument extraction approach based on trigger mapping. It maps the nominal trigger to its predicate and uses it as a key to extract syntac-tic features for classification. Experimental results on the ACE 2005 Chinese Corpus show that, in terms of F1-measure in argument identification and role determination, our approach achieves an obvious improvement.

事件关系识别是一项面向文本信息流进行事件关系判定的自然语言处理技术。事件关系识别的核心任务是以事件为基本语义单元，通过分析事件的篇章结构信息及语义特征，实现事件逻辑关系的浅层检测（即判定任意事件之间是否存在语义关系）。本文通过利用同一话题下的事件元素（事件元素指事件核心词和实体）在话题内的分布特性以及事件元素在话题演化过程中的语义依存规律，提出基于推理线索构建的事件关系识别方法。实验结果显示，相比于基于核心词和实体推理的事件关系识别方法，本文方法在F值上获得了9.57\%的性能提升。

Weighted alignment hypergraph [4] is potentially useful for statistical machine translation, because it is the first study to simultane- ously exploit the compact representation and fertility model of word alignment. Since estimating the probabilities of rules extracted from hypergraphs is an NP-complete problem, they propose a divide-and- conquer strategy by decomposing a hypergraph into a set of independent subhypergraphs. However, they employ a Bull’s algorithm to enumerate all consistent alignments for each rule in each subhypergraph, which is very time-consuming especially for the rules that contain non-terminals. This limits the applicability of this method to the syntax translation models, the rules of which contain many non-terminals (e.g. SCFG rules). In response to this problem, we propose an inside-outside algorithm to ef- ficiently enumerate the consistent alignments. Experimental results show that our method is twice as fast as the Bull’s algorithm. In addition, the efficient dynamic programming algorithm makes our approach applicable to syntax-based translation models.

Since the quality of statistical machine translation (SMT) is heavily dependent upon the size and quality of training data, many approaches have been proposed for automatically mining bilingual text from comparable corpo-ra. However, the existing solutions are restricted to extract either bilingual sen-tences or sub-sentential fragments. Instead, we present an efficient framework to extract both sentential and sub-sentential units. At sentential level, we con-sider the parallel sentence identification as a classification problem and extract more representative and effective features. At sub-sentential level, we refer to the idea of phrase table’s acquisition in SMT to extract parallel fragments. A novel word alignment model is specially designed for comparable sentence pairs and parallel fragments can be extracted based on such word alignment. We integrate the two levels’ extraction task into a united framework. Experi-mental results on SMT show that the baseline SMT system can achieve signifi-cant improvement by adding those extra-mined knowledge.

Data quality is one of the key factors in Statistical Machine Translation (SMT). Previous research addressed the data quality problem in SMT by corpus weighting or phrase scoring, but these two types of methods were often investigated independently. To leverage the dependencies between them, we propose an intuitive approach to improve translation modeling by collective corpus weighting and phrase scoring. The method uses the mutual reinforcement between the sentence pairs and the extracted phrase pairs, based on the observation that better sentence pairs often lead to better phrase extraction and vice versa. An effective graph-based random walk is designed to estimate the quality of sentence pairs and phrase pairs simultaneously. Extensive experimental results show that our method improves performance significantly and consistently in several Chinese-to-English translation tasks.

在目前的统计机器翻译翻译方法中，所利用的翻译模型的规模一般非常庞大，因此抬高了机器翻译服务的运行成本并限制了机器翻译的应用场景。在本文中，我们提出了一种基于翻译日志的统计机器翻译模型的剪枝方法。本方法利用翻译规则在翻译日志中的命中频数对机器翻译规则进行过滤，保留当前机器翻译模型所需的最小的规则表。实验表明，本方法能够在仅保留原有模型1~3\%翻译规则的前提下达到原有模型的翻译效果。

Populating existing knowledge base with new facts is important to keep the knowledge base fresh and most updated. Before importing new knowledge into the knowledge base, entity linking is required so that the enti-ties in the new knowledge can be linked to the entities in the knowledge base. During this process, entity disambiguation is the most challenging task. There have been many studies on leveraging name ambiguity problem via a variety of algorithms. In this paper, we propose an entity linking method based on Chi-nese Semantic Knowledge where entity disambiguation can be addressed by re-trieving a variety of semantic relations and analyzing the corresponding docu-ments with similarity measurement. Based on the proposed method, we devel-oped CASIA_EL, a system for linking entities with knowledge bases. We vali-date the proposed method by linking 1232 entities mined from Sina Weibo to a Chinese Semantic knowledge base, resulting in an accuracy of 88.5\%. The re-sults show that the CASIA_EL system and the proposed algorithm are poten-tially effective.

在评价对象抽取与极性判定研究中，语义特征一直处于“缺位”状态。本文提出与评 价对象抽取相关的七类语义特征，评价触发词、评价消解词、评价对象绝缘词、后指动词、 前指动词、心理动词、指向定语的评价名词；与极性判定相关的五类语义特征，褒义性名词、 贬义性名词、语义偏移型名词、度量衡形容词、语义构式，从引入语义特征的必要性以及如 何使用这些特征两方面进行了阐述。实验证明，语义特征的引入有助于提高评价对象抽取及 极性判断的准确率。

情绪因子是用来表达情绪、态度的语言形式,其包含情绪词、情绪短语、情绪表达式、标点符号、表情符号五种表达手段。本研究以情绪因子中的常用情绪词和情绪短语为基础构建情绪词典, 并针对特殊的情绪表达形式, 结合标点符号和表情符号在情绪分析中的功 能建立情绪规则库。系统通过对情绪词典和情绪规则的匹配和计算, 实现了对中文微博情绪的识别和分类,并在2013年 CCF 第二届自然语言处理与中文计算会议中文微博情绪分析评测中取得较好成绩。测试结果证明了该方法的有效性。

The performances of machine learning based opinion analysis sys-tems are always puzzled by the insufficient training opinion corpus. Such prob-lem becomes more serious for the resource-poor languages. Thus, the cross-lingual opinion analysis (CLOA) technique, which leverages opinion resources on one (source) language to another (target) language for improving the opinion analysis on target language, attracts more research interests. Currently, the transfer learning based CLOA approach sometimes falls to over fitting on single language resource, while the performance of the co-training based CLOA ap-proach always achieves limited improvement during bi-lingual decision. Target to these problems, in this study, we propose a mixed CLOA model, which esti-mates the confidence of each monolingual opinion analysis system by using their training errors through bilingual transfer self-training and co-training, re-spectively. By using the weighted average distances between samples and clas-sification hyper-planes as the confidence, the opinion polarity of testing sam-ples are classified. The evaluations on NLP&CC 2013 CLOA bakeoff dataset show that this approach achieves the best performance, which outperforms transfer learning and co-training based approaches.

Entity Linking (EL) is a fundamental technology in Natural Language Processing and Knowledge Engineering. Previous works mainly focus on linking mentioned names recognized in news or articles to knowledge base. However, in social network, user-generated content is quite different from typical news text. Users sometimes use words more informally, even create new words. One entity may have different aliases mentioned by web users, so identifying these aliases calls for more attention than before. Several methods are proposed to mine aliases and a learning-to-rank framework is applied to combine different types of feature together. A binary classifier based on SVM is trained to judge whether the top one candidate given by ranking algorithm is accepted. The evaluation results of NLP\&CC 20131 Entity Linking Track shows the effectiveness of this framework.

In text categorization (TC), labeled data is often limited while unlabeled data is ample. This motivates semi-supervised learning for TC to improve the performance by exploring the knowledge in both labeled and unlabeled data. In this paper, we propose a novel bootstrapping approach to semi-supervised TC. First of all, we give two basic preferences, i.e., sufficiency and diversity for a possibly successful bootstrapping. After carefully considering the diversity preference, we modify the traditional bootstrapping algorithm by training the involved classifiers with random feature subspaces instead of the whole feature space. Moreover, we further improve the random feature subspace-based bootstrapping with some constraints on the subspace generation to better satisfy the diversity preference. Experimental evaluation shows the effectiveness of our modified bootstrapping approach in both topic and sentiment-based TC tasks.

News articles express information by concentrating on named entities like who, when, and where in news. Whereas, extracting the relationships among entities, words and topics through a large amount of news articles is nontrivial. Topic modeling like Latent Dirichlet Allocation has been applied a lot to mine hidden topics in text analysis, which have achieved considerable performance. However, it cannot explicitly show relationship between words and entities. In this paper, we propose a generative model, Entity-Centered Topic Model(ECTM) to summarize the correlation among entities, words and topics by taking entity topic as a mixture of word topics. Experiments on real news data sets show our model of a lower perplexity and better in clustering of entities than state-of-the-art entity topic model(CorrLDA2).We also present analysis for results of ECTM and further compare it with CorrLDA2.

指代消解一直是自然语言处理中的核心问题.以往将指代消解转化成分类问题,基于机器学习构建的指代 消解系统,尽管在性能上有了较大提升,但是在消解过程中所采用的传统机器学习方法,大多仅含单层非线性变换 的浅层学习结构,其局限性在于有限样本和计算单元情况下对复杂函数的表示能力有限,面临复杂问题处理时的 泛化能力受到一定制约.本文提出一种利用 DBN(deep belief nets)模型的 Deep learning 学习机制进行基于语义特 征的指代消解方法. DBN模型由多层无监督的RBM(restricted Boltzmann machine)网络和一层有监督的 BP(back-propagation)网络组成,RBM 网络确保特征向量映射达到最优,最后一层 BP 网络分类 RBM 网络的输 出特征向量,从而训练指代消解分类器.在 ACE04 英文语料及 ACE05 中文语料上进行相关测试,实验结果表明, 增加 RBM 训练层数可以提高系统性能.此外,引入对特征集合的抽象分层因素,也对系统性能提升产生积极作用.

This paper presents a discriminative latent variable model (DPLVM) based classifier for improving the translation error detection performance for statistical machine translation (SMT). It uses latent variables to carry additional information which may not be expressed by those original labels and capture more complicated dependencies be- tween translation errors and their corresponding features to improve the classi¯cation performance. Specifically, we firstly detail the mathemati- cal representation of the proposed DPLVM method, and then introduce features, namely word posterior probabilities (WPP), linguistic features, syntactic features. Finally, we compare the proposed method with Max- Ent and SVM classifiers to verify its effectiveness. Experimental results show that the proposed DPLVM-based classifier reduce classification er- ror rate (CER) by relative 1.75\%, 1.69\%, 2.61\% compared to the MaxEnt classifier, and relative 0.17\%, 0.91\%, 2.12\% compared to the SVM clas- sifier over three different feature combinations.

A number of search user behavior studies show that queries with un-clear intents are commonly submitted to search engines. Result diversification is usually adopted to deal with those queries, in which search engine tries to trade-off some relevancy for some diversity to improve user experience. In this work, we aim to improve the performance of search results diversification by generating an intent subtopics list with fusion of multiple resources. We based our approach by thinking that to collect a large panel of intent subtopics, we should consider as well a wide range of resources from which to extract. The resources adopted cover a large panel of sources, such as external resources (Wikipedia, Google Keywords Generator, Google Insights, Search Engines query suggestion and completion), anchor texts, page snippets and more. We selected resources to cover both information seeker (What a user is searching for) and information provider (The websites) aspects. We also proposed an efficient Bayesian optimi-zation approach to maximize resources selection performances, and a new tech-nique to cluster subtopics based on the top results snippet information and Jac-card Similarity coefficient. Experiments based on TREC 2012 web track and NTCIR-10 intent task show that our framework can greatly improve diversity while keeping a good precision. The system developed with the proposed tech-niques also achieved the best English subtopic mining performance in NTCIR-10 intent task.

作为推荐领域主要的方法之一，基于内容的过滤被广泛地应用到新闻推荐中。为了更好地对新闻和用户进行建模，将语义相似度模型引入到基于内容的推荐系统中，挖掘两者之间的语义关联。首先提出了一种基于本体结构的相似度模型(OBSM)，利用在线百科构建的本体结构，计算新闻和用户之间的语义相似度。为了降低本体结构上噪音数据对推荐效果带来的影响，提出X-Ontology聚类算法对本体结构进行清理，并衍生出OBSM的升级模型，X-OBSM。通过中文和英文实验表明，OBSM和X-OBSM比基准模型具有更好的推荐效果，尤其是X-OBSM对本体结构进行清理后，具有比OBSM更高的计算效率。

As the emergence of various social media, short texts, such as weibos and instant messages, are very prevalent on today’s websites. In order to mine semantically similar information from massive data, a fast and efficient matching method for short texts has become an urgent task. However, the conventional matching methods suffer from the data sparsity in short documents. In this paper, we propose a novel matching method, referred as semantically similar hashing (SSHash). The basic idea of SSHash is to directly train a topic model from corpus rather than documents, then project texts into hash codes by using latent features. The major advantages of SSHash are that 1) SSHash alleviates the sparse problem in short texts, because we obtain the latent features from whole corpus regardless of document level; and 2) SSHash can accomplish similar matching in an interactive real time by introducing hash method. We carry out extensive experiments on real-world short texts. The results demonstrate that our method significantly outperforms baseline methods on several evaluation metrics.

本文主要研究了中文微博的观点要素抽取。采用CRFs模型进行抽取实验，深入分析词性，本体特征以及情感特征对抽取结果的影响；并且结合句子的语法结构， 针对微博中评论对象的情感特征设计多组位置属性复合特征模板。实验方法应用于NLP\&CC2013中文微博观点要素抽取评测，评测语料内容包含热点话题、电影、电视剧等评论。评测结果显著，宽松评价下位居21所测评单位第1，严格评价下位居第3。

针对2013年CCF自然语言处理与中文计算会议(NLP\&CC 2013)中文微博实体链接的任务，使用CCF提供的新浪微博数据作为训练数据以及测试数据，使用西南交通大学耶宝智慧中文分词平台作为自然语言预处理工具。提出一种实体链接的方法，该方法包括从知识库和同义词表查找实体，以及应用改进的拼音编辑距离算法和后缀词表匹配法。提出实体聚类消歧与基于百度百科词频的同类实体消歧相结合的消歧方法。在2013 年CCF 中文微博实体链接评测任务中正确结果的准确率为0.8838，在10 个参赛队伍中名列第三位。结果表明该方法的有效性，并可以适应文本中的噪声。

郑州大学参加了NLP\&CC2013的中文微博实体链接评测任务，主要采取从百度百科对命名实体进行抽取，然后利用命名实体所在上下文语境，抽取核心词汇，进而对实体消歧与链接，最终评测结果微平均准确率为0.8172。

随着web2.0的兴起，社交网络的规模和产生的数据量急剧增长，如何利用社交媒体内容进行知识库的构建与扩展成为当今研究热点。本文利用编写的爬虫，结合百科网页的结构跳转规则及网络词语建立知识库。 使用微博内容的上下文关系及其标签信息，采用机器学习和统计学提取微博间的联系等多策略，多层级的匹配算法对上述任务进行研究。评测准确率为84.99\%，改进后准确率有所提高，达到88.38\% 。

For one-sidedness of the various qualitative expansion methods, we propose a query terms selection method based on Grey Relational Analysis (GRA).We called the fusion expansion technique with GRA (FET-GRA). It calculates weight of expansion term by varied qualitative expansions and comprehensive weight by FET-GRA and thus extracts expansion term in terms of the weight. The experiment result of TREC dataset shows the method (FET-GRA) is substantially superior to TFIDF, Mutual Information , Local Context Analysis.

针对2013年CCF自然语言处理与中文计算会议(NLP\&CC 2013)中文微博实体链接的任务，使用CCF提供的新浪微博数据作为训练数据以及测试数据，使用西南交通大学耶宝智慧中文分词平台作为自然语言预处理工具。提出一种实体链接的方法，该方法包括从知识库和同义词表查找实体，以及应用改进的拼音编辑距离算法和后缀词表匹配法。提出实体聚类消歧与基于百度百科词频的同类实体消歧相结合的消歧方法。在2013 年CCF 中文微博实体链接评测任务中正确结果的准确率为0.8838，在10 个参赛队伍中名列第三位。结果表明该方法的有效性，并可以适应文本中的噪声。

术语定义是术语学的重要内容，是研究工作的基础，必须保证其准确性。而在信息人人可创造、编辑的形势下，术语存在多条定义的情况越来越多，定义的规范性和准确性得不到保证。针对一个术语的多条候选定义，首次提出术语定义辨析模型并给出一种基于互联网的求解方法。该方法从百度百科和百度搜索构建参考释义，总结术语及其定义来源语料中的术语定义模板，根据参考释义和定义的模板从待辨析定义中选出最优定义。实验选取中国知网概念知识元库中部分领域的术语在学术文献和工具书中的定义作为待辨析对象，以所提出的模型和方法进行辨析实验，结果表明该方法辨析的准确率为96.1\%，具有很好的辨析效果。

维吾尔语是一种黏着语，单词不太适合于作为维吾尔语大词汇连续语音识别系统识别单元。本文主要研究了维吾尔语大词汇连续语音识别系统中的识别单元选择问题,设计了更适合维吾尔语的子词识别单元，提出了维吾尔语单词和子词相结合的组合识别单元构建方法，评价了单词、子词和组合识别单元的语言模型和语音识别性能。实验结果表明，本文提出的识别单元在单元数量、语言模型复杂度等方面表现出更加优越的性能，并且使识别系统的单词错误率比基于单词的系统相对减少了22\%。

Learning to rank, which can fuse various of features, performs well in microblog retrieval. However, it is still unclear how the features function in microblog ranking. To address this issue, this paper examines the contribution of each single feature together with the contribution of the feature combinations via the ranking SVM for microblog retrieval modeling. The experimental re-sults on the TREC microblog collection show that textual features, i.e. content relevance between a query and a microblog, contribute most to the retrieval per-formance. And the combination of certain non-textual features and textual fea-tures can further enhance the retrieval performance, though non-textual features alone produce rather weak results.

Sentiment analysis of Chinese microblogs is important for scientific research in public opinion supervision, personalized recommendation and social computing. By studying the evaluation task of NLP\&CC 2012, we mainly implement two tasks, namely the extraction of opinion sentence and the determination of sentiment orientation for microblogs. First, we manually label the sample of microblog corpus supplied by the organization, and expand the sentiment lexicon by introducing the Internet sentiment words; second, we construct the different feature sets based on the analysis of the characteristic of Chinese microblogs. Finally, we use SVM classifier to generate a model based on training corpus, and implement the predication of test corpus. Evaluation results show our work has good performance on two tasks.

汉语复句在书面文本中所占比例较大，处理汉语复句关系是中文信息处理的问题之一。复句关系词对汉语复句的类别起到非常重要的标识作用。本文介绍了清华汉语树库的标注情况，根据清华汉语树库的标注方法，利用规则从中提取复句关系词并标注其类别。然后利用自动句法树和手工句法树的句法、词法、位置特征进行复句关系词的识别和分类。实验结果表明复句关系词判断准确率可达95.7\%，复句关系类别判断F1值为77.2\%。

Microblog, as a way of online communication, can generate large amounts of information in a very short period. Therefore, how to retrieve the latest relevant information becomes a hot research area. Different from tradi-tional information retrieval (IR), the microblog retrieval emphasizes fresh con-tents of the information. In order to solve this problem, we extend the tradition-al IR methods by taking into account the posting time. We propose a time-sensitive retrieval model, which takes the time factor as a prior probability. In the retrieval model, we introduce the pseudo relevance feedback technology as a query expansion approach to improve retrieval performance. Furthermore, we introduce a strategy to filter the initial retrieval results, which takes post quality factors into account including entropy and link features. Experiments on Twitter corpus show that our algorithm is effective to improve the retrieval perfor-mance, and the retrieval results can meet the real time retrieval need well.

篇章分析在很多自然语言处理任务中都是一项重要的基础性工作，具有重要的理论研究价值和实用意 义。但自动分析汉语篇章一直是一个难点，鲜有成熟的工具和方法。本文在标有复句逻辑语义关系的清华汉语 树库上，研究了汉语篇章语义片段自动切分以及篇章关系的自动标注方法。我们比较了不同序列标注模型对汉 语篇章语义单元切分的性能，并提出了基于最大熵模型的汉语篇章结构分析方法。实验结果表明，篇章语义单 元自动切分的F 值能达到89.1\%，当篇章语义结构树的高度不超过6 层时，篇章语义关系标注的F 值为63\%。

本文基于隐主题马尔科夫模型，消除LDA主题模型的主题独立假设，使得文摘生成过程中充分利用文章的结构信息，并结合基于内容的多特征方法提高文摘质量。此外本文提出在不破坏文章结构的前提下，从单文档扩展到多文档的自动文摘策略。最终实现完善的自动文摘系统。本文通过在DUC2007标准数据集上的实验结果，证明了隐主题马尔科夫模型和文档特征的优越性，并且所实现的自动文摘系统ROUGE值有明显提高。

类别、属性和实例的属性值是知识库重要组成部分,由相互协作的用户在网络百科产生的数据中包含大 量的属性和属性值知识。以中文网络百科条目文章为数据源,提出一种无监督的属性和属性值抽取方法。将属 性值看作命名实体,利用频繁模式挖掘和关联分析从文本中抽取类别属性;采用自扩展方法为属性建立触发词 表;基于属性触发词和属性值实体标注挖掘属性值抽取模式,利用层次聚类算法获取高质量的模式。在互动百 科中采集的数据集上进行实验,结果表明了所提方法的可行性和有效性。

Text summarization aims at getting the most important content in a condensed form from a given document while retains the semantic information of the text to a large extent. It is considered to be an effective way of tackling information overload. There exist lots of text summarization approaches which are based on Latent Semantic Analysis (LSA). However, none of the previous methods consider the term description of the topic. In this paper, we propose a comprehensive LSA-based text summarization algorithm that combines term description with sentence description for each topic. We also put forward a new way to create the term by sentence matrix. The effectiveness of our method is proved by experimental results. On the summarization performance, our ap-proach obtains higher ROUGE scores than several well known methods.

时间信息在文本分析中占据着非常重要的地位。在军事应用领域，准确识别时间信息能够有效支持军事情报自动获取、以时间为线索的军事文本信息抽取、检索与自动问答等多种应用。该文提出一种基于层叠有限状态自动机（CFSA）的中文军事文本时间表达式识别与规范化算法C-TERN。C-TERN首先利用成熟的分词工具识别出文本中的时间词，然后将从通用语言和军事语言中提取的时间表达式规则分成多层，逐层进行时间信息的精细识别；在规范化过程中，通过四个步骤分别对特殊时间表达式、简单时间表达式、时间段表达式和绝对/相对时间表达式进行推理计算和规范化。算法注意了规则集提取的正确性、规则之间冲突的消解、以及匹配方式的合理性。在多个数据集上的实验结果显示，C-TERN不但能有效地识别出标准时间、偏移时间和不确定性时间表达式，而且能完成对简单、特殊以及隐含的时间点、时间段和偏移时间的推理与规范化，能够满足军事文本时间信息的处理的需要。

篇章关系识别作为篇章结构分析的重要内容之一，可以分为显式关系识别和隐式关系识别。初步探讨了汉语篇章隐式关系的识别，应用词汇特征，上下文特征，依存树特征，在一个手工标注的小规模汉语篇章语料库上，总正确率达到了62.15\%。

Chinese word segmentation dictionary based on the Double-Array Trie Tree has higher efficiency of search, but the dynamic insertion will con-sume a lot of time. This paper presents an improved algorithm-iDAT, which is based on Double-Array Trie Tree for Chinese Word Segmentation Dictionary. After initialization the original dictionary. We implement a Hash process to the empty sequence index values for base array. The final Hash table stores the sum of the empty sequence before the current empty sequence. This algorithm adopt Sunday jumps algorithm of Single Pattern Matching. With slightly and reason-able space cost increasing, iDAT reduces the average time complexity of the dynamic insertion process in Trie Tree. Practical results shows it has a good op-eration performance.

Tibetan word segmentation (TWS) is the basic problem for Tibetan natural language processing. The paper reformulates the segmentation as a syllable tagging problem, and studies the performance of TWS with different sequence labeling models. Experimental results show that, the TWS system with conditional random field achieves the best performance in the condition of current 4-tag set, at the same time, the other models achieve good results too. All the above show that, the segmentation as a syllable tagging problem that is an efficient approach to deal with TWS.

面对日益增长的多语种商品评论信息资源, 跨语言情感分析将为进一步提升多语言的服务能力打下良好的基础,对商家掌握用户喜好变化有重要意义。目前,跨语言情感分析的研究处于发展之中,具有不可忽视的作用。本文提出利用句法分析模型,将语句分成若干组合词,根据组合词的主谓成分、情感词色彩的强弱不同,分别赋予不同的权重,统计分析该语句的情感分布特征,利用得到的特征参数训练分类器,再将训练好的分类器用于测试语料的情感分类。实验结果表明,该方法的情感分类判别准确率较理想,与已有的判别方法相比有一定提高。此研究方法也可用于语句的比较级判别和否定句的极性判断等领域。

Semantic analysis microblog data is a challenging, emerging research area. Unlike news text, microblogs pose several new challenges, due to their short, noisy, contextualized and real-time nature. In this paper, we investigate how to link entities in microblog posts with knowledge base and adopt a cascade linking approach. In particular, we first use a mention expansion model to identify all possible entities in the knowledge base for a mention based on a variety of sources. Then we link the mentions with the corresponding entities in the knowledge base by collectively considering lexical matching, popularity probability and textual similarity.

通过结合 TF-IDF 方法与方差统计方法提出一种实现多分类特征抽取的计算方法。采用先极性判断,后细粒度情感判断的处理方法, 构建细粒度情感分析与判断流程, 并将其应用于微博短文本的细粒度情感判断中。通过 NLPCC2013 评测所提供的训练语料对该方法有效性进行了验证,并在 NLPCC2013 评测任务中运用该方法, 证实了该方法具有较好的抽取效果。

Microblog is one of the most widely used web applications. Weibo, which is a microblog service in China, produces plenty of opinionated messages every second. Sentiment analysis on Chinese Weibo impacts many aspects of business and politics. In this work, we attempt to address the opinion target extraction, which is one of the most important aspects of sentiment analysis. We propose a unified approach that concentrates on phrase-level target extraction. We assume that a target is represented as a subgraph of the sentence’s dependency tree and define the grammatical relations that point to the target word as TARRELs. We conduct the extraction by classifying grammatical relations with a cost-sensitive classifier that enhances performance of unbalanced data and figuring out the target subgraph by connecting and recovering TAR-RELs. Then we prune the noisy targets by empirically summarized rules. The evaluation results indicate that our approach is effective to the phrase-level target extraction on Chinese microblog messages.

朴素贝叶斯(NB)、支持向量机(SVM)、最近邻(KNN)等算法是目前分析微博情绪的常用方法,它们在粗粒度的微博情绪分析上都取得了较好的效果。而在实际应用中, 微博的细粒度情绪往往更能展示用户对于事物的态度。因此, 针对中文微博的用户情绪分析问题, 提出了一种基于多策略融合的细粒度情绪分析方法。我们首先采用朴素贝叶斯算法对微博的有无情绪分类问题进行研究,然后构建有情绪微博的 21 维特征向量,最后分别采用SVM 和 KNN 算法对微博进行细粒度情绪分析。以新浪微博作为实验对象,结果表明多策略集成方法要好于单一分类算法。而在多策略集成方法中,“NB+SVM”方法又略优于“NB+KNN”方法。

情感分类是自然语言处理的一个热点研究问题。本文研究基于半监督学习的情感分类方法，即在很少规模的标注样本的基础上，借助非标注样本提高情感分类性能。实现半监督情感分类的一个重要思想是构建已标注样本和非标注样本之间的联系，将标注样本的标签传播到非标注样本上面。然而，以往的半监督情感分类研究在构建这种联系的时候仅仅关注评论的文本信息。我们发现由于社会关系网络的存在，评论与评论之间还存在一些社会联系。因此，本文将基于样本的社会关系，提出了一种新的半监督学习方法。具体来讲，本文提出了一种基于文档—词及社会关系的二部图模型，并根据标签传播算法将未标注样本加入到分类器的构建中。实验结果表明，加入社会关系网络的半监督情感分类方法明显优于传统的仅利用评论文本信息的半监督情感分类方法。

Passive voice is widely used in English while it is less used in Chinese, which is more prevalent in patent documents. The difference requires us to transform the voice in Chinese-English machine translation in order to make the result more smooth and natural. Previous studies in this field are based on statistics, but the effect is not very good. In this paper we propose a strategy to deal with the Chinese-English passive voice transformation from the perspective of semantic. Through analyzing the sentences, a series of transformation rules are summarized. Then we test them in our system. Experiment results show that the transformation rules can achieve an accuracy of 89.1\% overall.

自动提取漫画图像中的对白不仅有助于在移动设备上清晰地显示漫画中的文字信息，而且可进一步实现对白翻译、有声阅读等新的漫画移动阅读应用。针对中文漫画图像的特点，提出一种无监督的中文漫画图像对白自动定位方法，以满足中文漫画移动阅读的需求。不同于现有基于学习的方法，该方法不需要训练集，且具有较强的鲁棒性，主要包括三个步骤：1）利用包围漫画图像文字的空白区域（称为气泡，下同）的连通性进行气泡检测，并在气泡中检测完整字符对；2）基于字符形状与字符排版规则的一致性，聚类形成字符行或字符列，并提取字体特征；3）联合多页漫画图像字体特征，采用贝叶斯分类器检测多页漫画中的剩余字符。在包含900页漫画图像上的实验结果表明，该方法可以有效定位中文漫画图像中的对白区域，取得了比较满意的实验结果。

To alleviate the problem with Chinese being poor at telling the dif-ference between spoken and written English which is important for learning and using the language, we propose to classify English sentences with bilingual in-formation into the two categories automatically. Based on the text categoriza-tion technology, we explore a variety of features, including words, statistics and their combinations, and find that a classification accuracy nearly 95\% can be achieved in the open test through Chinese characters + sentence length + aver-age syllable number, or other similar combinations.

The authors are supposed to discover the valuable public opinions based on the massive social media data (the comments data, social relation data and location data collected from social website like twitter), so that it could help users to make better decisions during shopping. In order to solve this problem, the authors used micro-blog data collected from Sina as an example and pro-posed an opinion mining system based on the distributed computing system (Hadoop and related projects). This system was designed for a general purpose. It is not only for the restaurant area but also for the other fields like electronic products.

Automatic summarization is an important research branch of natural language processing. The automatic summarization should provide information to users from different point of views for better understanding. Aiming at the characteristics of the news, an automatic summarization system is constructed from two aspects: keywords and key sentences. Then, the location factor is added to optimize the keywords extraction algorithm. Meanwhile, the key sentences extraction algorithm is improved through introducing keywords factors. On this basis, in allusion to the existing problems of RSS, this paper builds a user-interest model. Finally, after the verification in terms of the feasibility and the effectiveness, the result shows it is effective to improve the accuracy and the user experience of the RSS feeds

TBC

TBC

Traditional methods for sentiment analysis mainly focus on the con-struction of emotional resources based on the review corpus of specific areas, and use phrase matching technologies to build a list of product feature words and opinion words. These methods bring about the disadvantages of inadequate model scalability, low matching precision, and high redundancy. Besides, it is particularly difficult to deal with negative words. In this work, we designed a unified framework based on finite state matching machine to deal with the problems of emotional element extraction. The max-matching principal and negative words processing can be integrated into the framework naturally. In addition, the framework leverages rule-based methods to filter out illegitimate feature-opinion pairs. Compared with traditional methods, the framework achieves high accuracy and scalability in emotional element extraction. Exper-imental results show that the extracting accuracy is up to 84\%, which has in-creased by 20\% comparing with traditional phrase matching techniques.

新闻综述报告可以有效帮助阅读者解决面对海量新闻文档的阅读困难。对新闻集合进行深入分析，可以很好地提取新闻集合潜在的热点、观点。本文利用新闻事件的话题、实体及其关联和发展趋势分析，建立对新闻事件分析报告模型。模型可以从多个角度描述新闻事件，并依照新闻综述报告的写作特点，制定出一种计算机自动分析报告生成的框架，对新闻事件在话题和实体上的分析结果进行组合，自动生成一篇观点分析透彻、图形图表信息生动准确的新闻事件分析报告。

Patent prior-art search is an necessary step to ensure that no previous similar disclosures were made before granting an patent. The task is to identify all relevant information which may invalidate the originality of a claim of a patent application. Using the whole patent or extracting high indicative terms to form a query reduces the search burden on the user. To date, There are no large-scale experiments conducted specifically for evaluating query generation techniques used in patent prior-art search in multiple languages. In the following paper, we firstly introduced seven methods for generating patent queries for ranking. Then a large-scale experimental evaluation was carried out on the CLEF-IP 2009 multilingual dataset in English, French and German. A detail comparison of the different methods in terms of performance and efficiency has been performed in addition to the use of full-length documents as queries in the patent search. The results show that some methods, work well in information retrieval in general, fail to achieve the same effectiveness in the patent search. Different methods demonstrated distinct performance w.r.t query and document languages.

Information disclosure in annual reports is a mandatory re- quirement for publicly traded companies in China. The quality of in- formation disclosure will reduce information asymmetry and therefore support market efficiency. Currently, the evaluation of the information disclosure quality in Chinese reports is conducted manually. It remains an untapped field for NLP and text mining community. The goal of this paper is to develop automatic assessment system for information disclo- sure quality in Chinese annual reports. Our assessment system framework incorporates different technologies including Chinese document model- ing, Chinese readability index construction, and multi-class classifica- tion. Our explorative and systematic experiment results show that: 1) our automatic assessment system can produce solid predictive accuracy for disclosure quality, especially in "excellent" and "fail" categories; 2) our system for Chinese annual reports assessment achieves better predic- tive accuracy in certain perspective than the counterparts of the English annual reports prediction; 3) our readability index for Chinese docu- ments, as well as other findings from system performance, may provide enlightenment for a better understanding about the quality features of Chinese company annual reports.

The traditional method of Named Entity Translation Equivalents ex-traction is often based on large-scale parallel or comparable corpora. But the practicability of the research results is constrained by the relatively scarce of the bilingual corpus resources. We combined the features of Chinese and Japanese, and proposed a method to automatically extract the Chinese-Japanese NE trans-lation equivalents based on inductive learning from monolingual corpus. This method uses the Chinese Hanzi and Japanese Kanji comparison table to calcu-late NE instances similarity between Japanese and Chinese. Then, we use in-ductive learning method to obtain partial translation rules of NEs through ex-tracting the differences between Chinese and Japanese high similarity NE in-stances. In the end, the feedback process refreshes the Chinese and Japanese NE similarity and translation rule sets. Experimental results show that the proposed method is simple and efficient, which overcome the shortcoming that the tradi-tional methods have a dependency on bilingual resource.

基于自定义知识库,采用一种知识库强化规则集,以及与统计模型相结合的日语时间表达式识别方法。 按照 Timex2 标准对时间表现的细化分类,结合日语时间词的特点,渐进地扩展重构日语时间表达式知识库, 实现基于知识库获取的规则集的优化更新,旨在不断提高时间表达式的识别精准度。同时,融合 CRF 统计模 型提高日语时间表达式识别的泛化能力。另外,考察基于短语的翻译模型翻译时间词的精度,提出统计机器翻 译(SMT)结合规则翻译日语时间词的必要性。 实验结果显示日语时间表达式识别的开放测试 F1 值达到 0.8987, 以及基于日汉时间词平行字典与规则的翻译精度和召回率都略高于基于统计机器翻译模型。

词汇衔接是篇章的重要特征，本文在已有的文档词汇衔接评价LC方法的基础上，提出了基于权重的LC，即WLC，该方法通过在文档词图上运行PageRank算法获得词汇权重；此外,本文根据词性信息使得PageRank算法偏向特定的词汇，据此提出了PWLC方法。实验表明在文档级别上本文建议的两种方法与人工评价的相关度都优于LC；同时，融合这两种方法之后，BLEU和TER在文档级别上的评价性能有了显著的提高。

This paper is concerned with exploring efficient domain adap- tation for the task of statistical machine translation, which is based on extracting sentence pairs (pseudo in-domain subcorpora, that are most relevant to the in domain corpora) from a large-scale general-domain web bilingual corpus. These sentences are selected by our proposed un- supervised phrase-based data selection model. Compared with the tra- ditional bag-of-words models, our phrase-based data selection model is more effective because it captures contextual information in modeling the selection of phrase as a whole, rather than selection of single words in isolation. These pseudo in-domain subcorpora can then be used to train small domain-adapted spoken language translation system which outperforms the system trained on the entire corpus, with an increase of 1.6 BLEU points. Performance is further improved when we use these pseudo in-domain corpus/models in combination with the true in-domain corpus/model, with increases of 4.5 and 3.9 BLEU points over single in- and general-domain baseline system, respectively.

Negative and speculative expressions are popular in natural lan-guage. Recently, negation and speculation detection has become an important task in computational linguistics community. However, there are few related research on Chinese negation and speculation detection. In this paper, a su-pervised machine learning method with conditional random fields (CRFs) is proposed to detect negative and speculative information in scientific litera-ture. This paper also evaluates the effectiveness of each feature under the character-based and word-based framework, as well as the combination of features. Experimental results show that the single-word feature and the part of speech feature are effective, and the combined features improve the per-formance furthest. Our Chinese negation and speculation detection system in sentence level achieves 94.70\% and 87.10\% of accuracy, respectively.

本文以藏文音节拼写检查、梵音转写藏文检查、接续关系检查、词语检查为研究内容， 提出了藏文文本自动校对框架和接续关系检查算法。根据该框架及算法，设计并实现了藏文 自动校对系统，并通过相应的实验证明了算法和系统的可靠性和有效性。

Traditional Mongolian and Cyrillic Mongolian are both Mongolian languages that are respectively used in china and Mongolia. With similar oral pronunciation, their writing forms are totally different. A large part of Cyrillic Mongolian words have more than one corresponds in Traditional Mongolian. This makes the conversion from Cyrillic Mongolian to Traditional Mongolian a hard problem. To overcome this difficulty, this paper proposed a Language model based approach, which takes the advantage of context information. Experimental results show that, for Cyrillic Mongolian words that have multiple correspondence in Traditional Mongolian, the correct rate of this approach reaches 87.66\%, thereby greatly improve the overall system performance.

Personal Family Network is an important component of social net-works, therefore, it is of great importance of how to extract personal family re-lationships. We propose a novel method to construct personal families based on bootstrapping and coreference resolution on top of a search engine. It begins with seeds of personal relations to discover relational patterns in a bootstrap-ping fashion, then personal relations are further extracted via these learned pat-terns, finally family networks are fused using cross-document coreference reso-lution. The experimental results on a large-scale corpus of Gigaword show that, our method can build accurate family networks, thereby laying the foundation for social network analysis.

This paper studies the emotion classification task on microblogs. Given a message, we classify its emotion as happy, sad, angry or surprise. Existing methods mostly use the bag-of-word representation or manually designed features to train supervised or distant supervision models. However, manufacturing feature engines is time-consuming and not enough to capture the complex linguistic phenomena on microblogs. In this study, to overcome the above problems, we utilize pseudo-labeled data, which is extensively explored for distant supervision learning and training language model in Twitter sentiment analysis, to learn the sentence representation through Deep Belief Network algorithm. Experimental results in the supervised learning framework show that using the pseudo-labeled data, the representation learned by Deep Belief Network outperforms the Principal Components Analysis based and Latent Dirichlet Allocation based representations. By incorporating the Deep Belief Network based representation into basic features, the performance is further improved.

Chinese microblog is a popular Internet social medium where users express their sentiments and opinions. But sentiment analysis on Chinese microblogs is difficult: The lack of labeling on the sentiment polarities restricts many supervised algorithms; out-of-vocabulary words and emoticons enlarge the sentiment expressions, which are beyond traditional sentiment lexicons. In this paper, emoticons in Chinese microblog messages are used as annotations to automatically label noisy corpora and construct sentiment lexicons. Features including microblogspecific and sentiment-related ones are introduced for sentiment classification. These sentiment signals are useful for Chinese microblog sentiment analysis. Evaluations on a balanced dataset are conducted, showing an accuracy of 63.9\% in a three-class sentiment classification of positive, negative and neutral. The features mined from the Chinese microblogs also increase the performances.

Cross-lingual sentiment classification aims to predict the sentiment orientation of a text in a language (named as the target language) with the help of the resources from another language (named as the source language). However, current cross-lingual performance is normally far away from satisfaction due to the huge difference in linguistic expression and social culture. In this paper, we suggest to perform active learning for cross-lingual sentiment classification, where only a small scale of samples are actively selected and manually annotated to achieve reasonable performance in a short time for the target language. The challenge therein is that there are normally much more labeled samples in the source language than those in the target language. This makes the small amount of labeled samples from the target language flooded in the aboundance of labeled samples from the source language, which largely reduces their impact on cross-lingual sentiment classification. To address this issue, we propose a data quality controlling approach in the source language to select high-quality samples from the source language. Specifically, we propose two kinds of data quality measurements, intra- and extra-quality measurements, from the certainty and similarity perspectives. Empirical studies verify the appropriateness of our active learning approach to cross-lingual sentiment classification.

随着网络数据规模的飞速增长，针对大规模结构化数据的获取、组织和利用已经成为信息处理领域的热点问题。特别是对于结构化百科知识库的查询理解，存在着用户提问的自然语言问句与结构化查询（如SPARQL语句）之间的鸿沟。为此我们提出了从自然语言问句到结构化查询的转换框架，该方法从自然语言问句的句法结构入手，提出了一套启发式识别实体与关系的方法，并利用语料库建立了从实体到知识库的映射，对谓词进行消歧，进而转化为计算机可理解的结构化查询语言。我们从百度知道上抽取了人物、地点、组织三类一共42个问题作为标准测试集。实验结果表明，本文提出的框架能够有效地将中文自然语言问句转换为结构化查询，为下一代智能问答系统打下良好的基础。

Ontology is applied to various fields of computer as a conceptual modeling tool, and is used to organize information and manage knowledge. Ontology extension is used to add the new concepts and relationship into the existing ontology, which is a more complex task. In this paper, we propose a hybrid approach for ontology extension from text using semantic relatedness between words, which exploit co-occurrence analysis, word filter and semantic relatedness between words to find the potential concepts from text, denoted as the extended concepts. And we take advantage of extension rules and subsumption analysis to find the relationship between concepts, which is used to add the extended concepts into the existing ontology. The improved recall, precision and F1-Measure have been presented and used to evaluate our method proposed in this paper. Experimental results show that the proposed method is more reasonable and promising. It has a stronger competitiveness and application ability.

Although recommender system has been studied for many years, the research of social recommender system is just beginning. Plenty of information can be used in social networks to improve the performance of recommender system. However, some information is very sparse when used as features. We call this feature sparsity problem. In this paper, we aimed at solving feature sparsity problem. A new strategy was proposed to expand user features by social relationships. Experiments on two real world datasets demonstrated that our method can significantly improve the recommendation performance.

Expanding corpora is very important in designing a spoken dialogue system (SDS). In this big data era, data is expensive to collect and there are rare annotations. Some researchers make much work to expand corpora, most of which is based on rule. This paper presents a probabilistic method to simulate dialogues between human and machine so as to expand a small corpus with more varied simulated dialogue acts. The method employs Input/output HMM with user history (UH-IOHMM) to learn system and user dialogue behavior. In addition, this paper compares with simulation system based on standard IOHMM. We perform experiments using the WDC-ICA corpus, weather do-main corpus with annotation. And the experiment result shows that the method we present in this paper can produce high quality dialogue acts which are simi-lar to real dialogue acts.

区别于传统的基于图像和西文文档的公式定位方法,针对中文电子文档的特点,提出了 一种基于机器学习和规则相结合的独立公式和内嵌公式的定位方法。在其中,设计了适合中文 文档的页面分行策略和词块划分规则;选择了适合中文文档的公式特征和机器学习算法;针对 公式定位中的过分割问题提出了行合并与词块合并等后处理手段。实验结果表明,该方法可以 有效地从中文电子文档中自动定位公式区域。此外,构建了公开可用的中文数据集,以促进不 同数学公式定位方法间的相互比较及性能评估。

通过对甲骨字的构成分析,针对甲骨文字形多变、异体字多等特点,提出了一种基于语义构件的甲骨文 字库自动生成方法。该方法以动态描述库为基础,首先通过算法提取甲骨字的构件特征信息;然后重组笔元生 成语义构件,再给语义构件加上特征描述生成构件知识库。通过仿射变换重复使用语义构件自动生成任意甲骨 字。实验表明,该方法有效地解决甲骨文无字库输入的实现,并且还可以解决甲骨字编码、构件统计、未释字 的考释等。

汉字字形设计是一项艰巨的工作，与英文字形设计不同，英文只要设计好了二十六个字母就能形成一种字体，而汉字每产生一套字库就要设计成千上万个汉字，非艺术家和汉字专家很难完成。本文针对汉字字形设计和开发方面的困难，提出了基于特征点抽象的汉字字形描述方法和汉字字形生成方法，研究了特征点、特征表达式、特征点的权值和权矢量等在汉字字形生成中的应用技术和方法，并以点的生成为例，设计了汉字点笔画的生成算法，并进行了验证实验。实验结果证明了算法的可靠性和实用性，为汉字其它笔画的生成提供了一种解决方案，可切实提高汉字字形设计的效率。

How to preserve and make use of ancient Chinese characters is not only a mission to contemporary scientists but is also a technical challenge. This paper proposes a feasible solution to enable character collection, management, and access on the Internet. Its advantage lies in a unified representation for en-coded and uncoded characters that provide a visual convenient and efficient re-trieval method that does not require new users to have any prior knowledge about ancient Chinese characters. We also design a system suitable for describing the relationships between ancient Chinese characters and contemporary ones. As the implementation result, a website is established for public access to ancient Chi-nese characters.