作者导言：以下是我2018-2019年发表的关于深度学习在问答系统和对话系统的应用的paper，感兴趣的读者可以查阅。大家如果有关于这些paper的问题，欢迎发邮件到我的email (yangliuyx@gmail.com)咨询讨论，我会尽量回复。本来想对每一篇论文做细致的介绍，但是发现CSDN最近的博客编辑器不好用，所以先把近两年来发表的论文贴在这儿，以后有时间我会再更新，添加更多细节文字介绍。... ...继续阅读 (26)

博主导言：苦于Deep Learning的baseline太多实现困难？苦于没有好的基于深度学习处理NLP, IR, QA任务的开源工具？苦于没有发布自己研究的深度文本匹配模型的交流平台？强烈推荐MatchZoo，用深度学习做自然语言处理，信息检索，智能问答等任务的小伙伴看过来。MatchZoo提供了基准数据集（TREC MQ系列数据、WiKiQA数据等）进行开发与测试，整合了当前最流行的深度文本 ...继续阅读 (24)

中文简介：本文对企业邮件系统中的用户行为进行了建模分析，首先分析了影响用户邮件回复行为的几类因素，然后基于分析结果建立了预测用户邮件回复行为和邮件回复时间的机器学习模型。基于Avocado邮件数据的实验结果表明，本文提出的特征和模型对于用户邮件回复行为的预测准确度大幅度超过了以往的基准方法。论文出处：SIGIR'17英文摘要：Email is still among the most popula ...继续阅读 (13)

1. 概要Gradient Tree Boosting (别名 GBM, GBRT, GBDT, MART)是一类很常用的集成学习算法，在KDD Cup, Kaggle组织的很多数据挖掘竞赛中多次表现出在分类和回归任务上面最好的performance。同时在2010年Yahoo Learning to Rank Challenge中, 夺得冠军的LambdaMART算法也属于这一类算法。因此Tree Boosting算法和深度学习算法DNN/CNN/RNN等等一样在工业界和学术界中得到了非常广泛的应用。最近研读了UW Tianqi Chen博士写的关于Gradient Tree Boosting 的Slide和Notes, 牛人就是牛人，可以把算法和模型讲的如此清楚，深入浅出，感觉对Tree Boosting算法的理解进一步加深了一些。本来打算写一篇比较详细的算法解析的文章，后来一想不如记录一些阅读心得和关键点，感兴趣的读者可以直接看英文原版资料如下:A. Introduction to Boosted Tree. https://xgboost.readthedocs.io/en/latest/model.htmlB. Introduction to Boosted Trees. By Tianqi Chen. http://homes.cs.washington.edu/~tq ...继续阅读 (30)

1. 概要 Gradient Tree Boosting (别名 GBM, GBRT, GBDT, MART)是一类很常用的集成学习算法，在KDD Cup, Kaggle组织的很多数据挖掘竞赛中多次表现出在分类和回归任务上面最好的performance。同时在2010年Yahoo Learning to Rank Challenge中, 夺得冠军的LambdaMART算法也属于这一类算法。因此Tree Boosting算法和深度学习算法DNN/CNN/RNN等等一样在工业界和学术界中得到了非常广泛的应用。 ...继续阅读 (13)

中文简介：本文针对当前深度学习模型包括基于CNN或者LSTM的模型适用于Answer Sentence Selection这个task时必须额外combine 传统的text matching feature的问题，提出了一个attention based neural matching model。该模型提出使用value-shared weighting scheme以及基于attentionscheme学习问题词的重要性。基于标准Benchmark TREC QA data的实验结果证明，该模型不需要额外combine任何text matching的feature也可以得到和之前的深度学习模型以及基于特征工程的方法相当或者更好的排序效果。如果额外combine一个简单的QL的feature，该模型的排序效果可以超过当前Answer Sentence Selection 这个任务state-of-the-art的方法的效果。论文出处：CIKM'16英文摘要：As an alternative to question answering methods based on feature engineering, deep learning approaches such as convolutional neural networks (CNNs) ...继续阅读 (14)

中文简介：本文针对当前深度学习模型包括基于CNN或者LSTM的模型适用于Answer Sentence Selection这个task时必须额外combine 传统的text matching feature的问题，提出了一个attention based neural matching model。该模型提出使用value-shared weighting scheme以及基于attention ...继续阅读 (12)

中文简介：本文是对前面的SIGIR‘16工作的拓展, 主要是对PV model适用于IR的task时的三方面的问题进行了更加深入的分析，并且提出了针对这三个问题的相应改进。论文出处：ICTIR' 16英文摘要：Previous studies have shown that semantically meaningful representations of words and text can be acquired through neural embedding models. In particular, paragraph vector (PV) models have shown impressive performance in some natural language processing tasks by estimating a document (topic) level language model. Integrating the PV models with traditional language model approaches to retrieval, however, produces unstable performance and limited improvement ...继续阅读 (56)

中文简介：本文是对前面的SIGIR‘16工作的拓展, 主要是对PV model适用于IR的task时的三方面的问题进行了更加深入的分析，并且提出了针对这三个问题的相应改进。论文出处：ICTIR' 16英文摘要：Previous studies have shown that semantically meaningful representations of words and text can ...继续阅读 (18)

中文简介：本文对如何基于Paragraph Vector model改进Ad-hoc Retrieval task进行了分析，主要针对IR的场景提出了对PV model的三方面的改进。实验表明，改进后的模型进行检索的效果超过了基于topic model增强的LM的效果。论文出处：SIGIR'16英文摘要：Incorporating  topic  level  estimation  into  language  models has  been  shown  to  be beneficial for  information  retrieval(IR) models such as cluster-based retrieval and LDA-based document representation.  Neural embedding models,  such as paragraph vector (PV) models, on the other hand have shown their eeffectiveness and efficiency in learning semantic representations ...继续阅读 (18)

中文简介：本文对如何基于Paragraph Vector model改进Ad-hoc Retrieval task进行了分析，主要针对IR的场景提出了对PV model的三方面的改进。实验表明，改进后的模型进行检索的效果超过了基于topic model增强的LM的效果。论文出处：SIGIR'16英文摘要：Incorporating  topic  level  estimation  into ...继续阅读 (22)

中文简介： 本文提出了概率生成模型 Shell Topic Model (STM)对社交论坛文本中的组织性短语(Organizational Phrases)和主题词(topical contents)进行建模分析，主要的应用有组织性短语的挖掘和文档建模。论文出处：ICDM‘14.英文摘要：Threaded debate forums have become one of the major social media platforms. Usually people argue with one another using not only claims and evidences about the topic under discussion but also language used to organize them, which we refer to as shell. In this paper, we study how to separate shell from topical contents using unsupervised methods. Along this line, we develop a latent variable model named Shell To ...继续阅读 (53)

中文简介： 本文提出了概率生成模型 Shell Topic Model (STM)对社交论坛文本中的组织性短语(Organizational Phrases)和主题词(topical contents)进行建模分析，主要的应用有组织性短语的挖掘和文档建模。论文出处：ICDM‘14.英文摘要：Threaded debate forums have become one of the major so ...继续阅读 (9)

中文简介：本文对智能个人助理（如Google Now，Microsoft Cortana）中的信息卡片排序进行了研究，从user modeling的角度提出了三组排序特征：implicit feedback features, entity based user interests features以及user demographic features. 其中entity features的提取用到了word embedding和knowledge base信息进行语义相似度的计算。基于大规模产品日志数据的实验表明，提出的排序特征可以提升信息卡片的检索质量。论文出处：ECIR'16.英文摘要： Proactive  search  systems  like  Google  Now  and  Microsoft  Cortana have gained increasing popularity with the growth of mobile Internet. Unlike traditional reactive search systems where search engines return results in response to queries issued by t ...继续阅读 (56)

中文简介：本文对智能个人助理（如Google Now，Microsoft Cortana）中的信息卡片排序进行了研究，从user modeling的角度提出了三组排序特征：implicit feedback features, entity based user interests features以及user demographic features. 其中entity features的提取用 ...继续阅读 (13)

中文简介：本文对non-factoid 问题的答案句子检索进行了研究，基于learning to ranking的框架，在传统文本匹配特征的基础上提出了给予语义匹配和上下文信息的特征，并且证明了这些特征对于答案句子检索的有效性。本文使用TREC GOV2数据集，并且开源了code和标注数据集，下载链接参见论文脚注。论文出处：ECIR'16.英文摘要： Retrieving finer grained text units such as passages or sentences as answers for non-factoid Web queries is becoming increasingly important for applications such as mobile Web search. In this work, we introduce the answer sentence retrieval task for non-factoid Web queries, and investigate how this task can be effectively solved under a learning to rank framework. We design two types of features ...继续阅读 (58)

中文简介：本文对non-factoid 问题的答案句子检索进行了研究，基于learning to ranking的框架，在传统文本匹配特征的基础上提出了给予语义匹配和上下文信息的特征，并且证明了这些特征对于答案句子检索的有效性。本文使用TREC GOV2数据集，并且开源了code和标注数据集，下载链接参见论文脚注。论文出处：ECIR'16.英文摘要： Retrieving finer graine ...继续阅读 (16)

中文简介：本文对如何对网上论坛讨论中用户交互关系进行统计建模分析进行了研究。论文出处：CIKM‘13.英文摘要： Online discussion forums are popular social media platforms for users to express their opinions and discuss controversial issues with each other. To automatically identify the sides/stances of posts or users from textual content in forums is an important task to help mine online opinions. To tackle the task, it is important to exploit user posts that implicitly contain support and dispute (interaction) information. The challenge we face is how to mine such interaction information from the content of posts and how to use them to help ...继续阅读 (38)

中文简介：本文对如何对网上论坛讨论中用户交互关系进行统计建模分析进行了研究。论文出处：CIKM‘13.英文摘要： Online discussion forums are popular social media platforms for users to express their opinions and discuss controversial issues with each othe ...继续阅读 (20)

中文简介：本文对如何基于情感分析和概率矩阵分解从网络论坛讨论中挖掘用户关系进行了深入研究。论文出处：NAACL'13.英文摘要： Advances in sentiment analysis have enabled extraction of user relations implied in online textual exchanges such as forum posts. However,recent studies in this direction only consider direct relation extraction from text. As user interactions can be sparse in online discussions,we propose to apply collaborative filtering through probabilistic matrix factorization to generalize and improve the opinion matrices extracted from forum posts. Experiments with two tasks show that the learned latent factor representation can give g ...继续阅读 (44)

中文简介：本文对如何基于情感分析和概率矩阵分解从网络论坛讨论中挖掘用户关系进行了深入研究。论文出处：NAACL'13.英文摘要： Advances in sentiment analysis have enabled extraction of user relations implied in online textual exchanges such as forum posts. Howev ...继续阅读 (15)

中文简介：想知道如何基于雅虎问答社区帖生成旅行指南吗？本文介绍了相关统计模型和技术。论文出处：COLING’14英文摘要： In this paper we study how to summarize travel-related information in forum threads to generate supplementary travel guides. Such summaries presumably can provide additional and moreup-to-date information to tourists. Existing multi-document summarization methods have limitations for this task because (1) they do not generate structured summaries but travel guides usually follow a certain template, and (2) they do not put emphasis on named entities but travel guides often recommend points of interest to travelers. To overcome t ...继续阅读 (34)

中文简介：想知道如何基于雅虎问答社区帖生成旅行指南吗？本文介绍了相关统计模型和技术。论文出处：COLING’14英文摘要： In this paper we study how to summarize travel-related information in forum threads to generate supplementary travel guides. Such summarie ...继续阅读 (12)

中文简介： 本文对如何在问答社区对用户主题兴趣及专业度建模分析进行了研究，并且提出了针对此问题的统计图模型Topics Expertise Model.论文出处：CIKM‘13.英文摘要： Community Question Answering (CQA) websites, where people share expertise on open platforms, have become large repositories of valuable knowledge. To bring the best value out of these knowledge repositories, it is critically important for CQA services to know how to find the right experts, retrieve archived similar questions and recommend best answers to new questions. To tackle this cluster of closely related problems in a principled approach, we proposed Topic Expertise Model (TEM), a novel pr ...继续阅读 (80)

中文简介： 本文对如何在问答社区对用户主题兴趣及专业度建模分析进行了研究，并且提出了针对此问题的统计图模型Topics Expertise Model.论文出处：CIKM‘13.英文摘要： Community Question Answering (CQA) websites, where people share expertise on open platforms, have become ...继续阅读 (14)

Given n, how many structurally unique BST's (binary search trees) that store values 1...n?For example,Given n = 3, there are a total of 5 unique BST's.1 3 3 2 1 \ / / / \ \ 3 2 1 1 3 2 / / \ \ 2 1 2 3思路分析：这题关键在于把握BST的结构特征，从i=0,1,2...逐渐找出递推规律，用count[i]表示有i个节点是BST的数量，那么i = 0 时 count[i] = 1// 空树i = 1 时 count[i] = 1//以1为rooti = 2 时 count[i] = count[0]* count[1]//以1为root                      ...继续阅读 (43)

Given n, how many structurally unique BST's (binary search trees) that store values 1...n?For example,Given n = 3, there are a total of 5 unique BST's. 1 3 3 2 1 \ / ...继续阅读 (11)

Implement the following operations of a stack using queues.push(x) -- Push element x onto stack.pop() -- Removes the element on top of the stack.top() -- Get the top element.empty() -- Return whether the stack is empty.Notes:You must use only standard operations of a queue -- which means only push to back, peek/pop from front, size, and is empty operations are valid.Depending on your language, queue may not be supported natively. You may simulate a queue by using a list or deque (double-ended queue), as long as you use only standard operations of a queue.You may assume that all operations are ...继续阅读 (37)

Implement the following operations of a stack using queues.push(x) -- Push element x onto stack.pop() -- Removes the element on top of the stack.top() -- Get the top element.empty() -- Return whether ...继续阅读 (9)

Follow up for "Find Minimum in Rotated Sorted Array":What if duplicates are allowed?Would this affect the run-time complexity? How and why?Suppose a sorted array is rotated at some pivot unknown to you beforehand.(i.e., 0 1 2 4 5 6 7 might become 4 5 6 7 0 1 2).Find the minimum element.The array may contain duplicates.思路分析：这题类似于LeetCode Find Minimum in Rotated Sorted Array ，具体解析可以参考前一篇，这题的特殊之处是数组中可以有重复元素，那么当A[m]=A[l]时，执行l++即可。此时，就只可以排除掉一个元素，而不是一半元素，最坏情况下算法时间复杂度是O（n），不再是O（logN）。AC Code：public class Solution { public int findMin(int[] nums) { int n = nums.length; int l = 0; ...继续阅读 (31)