技术领域technical field
本发明涉及基于会话搜索的信息检索，尤其涉及一种基于马尔科夫决策过程模型的会话搜索方法。The invention relates to information retrieval based on conversational search, in particular to a conversational search method based on a Markov decision process model.
背景技术Background technique
会话搜索是在一种在搜索会话的过程中通过搜索引擎与用户的交互来实现信息检索的搜索技术。在搜索一个主题的相关信息时，用户会在一个会话中不断地变更查询内容，直至找到其需要的信息。会话搜索中用户会根据搜索引擎的返回结果来调整查询内容，搜索引擎则将其作为用户的反馈，用以完善会话搜索的结果。会话搜索中用户查询的变更模式没有统一的规律，搜索引擎难以准确获取用户的意图，因而会话搜索是一项挑战性的信息检索任务。目前利用和学习查询变更的方法主要有：Conversational search is a search technology that realizes information retrieval through the interaction between the search engine and the user during the search session. When searching for information about a topic, users vary their queries throughout a session until they find what they need. In conversational search, users will adjust the query content according to the results returned by the search engine, and the search engine will use it as user feedback to improve the results of conversational search. In conversational search, there is no uniform pattern of user query change patterns, and it is difficult for search engines to accurately obtain user intentions, so conversational search is a challenging information retrieval task. At present, the methods of using and learning query changes mainly include:
(1)基于不同种类的搜索进行分类，比如进行具体化，概括，变换，或者细微变化，然后执行检索过程。(1) Classify based on different kinds of searches, such as specificization, generalization, transformation, or subtle changes, and then execute the retrieval process.
(2)另一种方法是将查询映射为语义图形表示，如本体或查询日志形成的查询流图，然后研究查询在图中的变化规律。(2) Another approach is to map queries into semantic graph representations, such as query flow graphs formed by ontology or query logs, and then study the changing rules of queries in the graph.
然而，第一种方法依赖于完备的查询记录，这样的数据往往难以获得。第二种方法中的本体化映射很有挑战，这有可能引入不精确的中间数据，并且损坏搜索的准确性。所以，尽管这些方法可以应用到信息检索任务中，比如查询的规范化，联想查询等，但是难以直接应用于会话搜索。However, the first approach relies on well-documented queries, which are often difficult to obtain. Ontological mapping in the second approach is challenging, which has the potential to introduce imprecise intermediate data and damage search accuracy. Therefore, although these methods can be applied to information retrieval tasks, such as query normalization, predictive queries, etc., they are difficult to be directly applied to conversational search.
为了能充分利用会话搜索中查询变更的信息，本文发明了一种基于马尔科夫决策过程模型的会话搜索方法，可以有效的加强会话搜索的效果。In order to make full use of the information of query changes in conversational search, this paper invents a conversational search method based on the Markov decision process model, which can effectively enhance the effect of conversational search.
发明内容Contents of the invention
发明目的：本发明提供了一种新颖的基于马尔科夫决策过程模型的信息检索方法。每次查询的语句q可以由多个词语t∈q组成。可以利用相邻查询之间查询语句的变更，以及前次检索的返回结果来加强会话搜索的准确性。该方法首先获取语料库，然后训练数据集，生成训练文档，接着对数据预处理，最后接收查询语句，返回优化查询地结果。Purpose of the invention: The present invention provides a novel information retrieval method based on the Markov decision process model. The statement q of each query can consist of multiple words t∈q. The change of query statements between adjacent queries and the returned results of the previous retrieval can be used to enhance the accuracy of session search. This method first obtains the corpus, then trains the data set, generates training documents, then preprocesses the data, finally receives query statements, and returns optimized query results.
一种基于马尔科夫决策过程模型的会话搜索方法，该方法包括如下步骤：A conversational search method based on a Markov decision process model, the method comprises the following steps:
1)准备阶段1) Preparation stage
a)爬取足够多的网页，获取语料库全集C。C为爬取的网页集合经过筛选后的结果。每个网页内对应文档d，C＝{di}a) Crawl enough web pages to obtain the complete set C of the corpus. C is the filtered result of the crawled webpage collection. Corresponding document d in each webpage, C={di }
b)训练人员进行自发的会话搜索并记录其过程以获取训练数据(训练数据包括涉及的查询，查询的更改，用户在搜索引擎返回的结果中点击的文档及其点击停留时间等)，生成训练文档b) Train personnel to conduct spontaneous conversational search and record its process to obtain training data (training data includes related queries, query changes, documents clicked by users in the results returned by the search engine and their click dwell time, etc.), generate training document
c)结束准备阶段c) Finish the preparation phase
2)训练阶段2) Training stage
a)数据预处理，统计检索阶段中需要使用的Ps(t|d)和Pus(t|d)a) Data preprocessing, Ps (t|d) and Pus (t|d) that need to be used in the statistical retrieval stage
b)解析步骤1-b)中生成的训练文档，其中包括多个会话的信息b) parse the training document generated in step 1-b), which includes information from multiple sessions
c)读取强化学习中的一个会话，一个会话包括一次或多次用户对搜索结果操作的信息c) Read a session in reinforcement learning, a session includes information about one or more user operations on search results
d)读取会话中的一次用户对搜索结果操作的信息(包括涉及的查询，查询的更改，点击的文档，点击的停留时间等)，并由此更新Pus(t|d)的值d) Read the information of a user's operation on the search results in a session (including the query involved, the change of the query, the document clicked, the dwell time of the click, etc.), and update the value of Pus (t|d) accordingly
e)重复步骤d)直到会话结束e) Repeat step d) until the session ends
f)重复步骤c),d),e)直到所有会话都被处理完毕f) Repeat steps c), d), e) until all sessions are processed
g)结束训练阶段g) End the training phase
3)检索阶段3) Retrieval stage
a)接收用户当前输入的查询语句qia) Receive the query statement qi currently input by the user
b)通过公式计算每个文档d与当前查询qi的关联度b) Calculate the degree of association between each document d and the current query qi by formula
c)计算每个文档d与整个会话的关联度c) Calculate the relevance of each document d to the entire session
d)返回关联度高的前N篇文档(本发明取10)d) return the first N documents with high correlation (this invention takes 10)
e)重复a),b),c),d)直到用户结束查询e) Repeat a), b), c), d) until the user ends the query
f)结束检索阶段f) End retrieval phase
其中所述步骤1-b)所述的训练文件Wherein the training file described in step 1-b)
1)定义一组主题(topics)，训练人员根据主题的描述自发地进行特定搜索。搜索引擎采用雅虎提供的Yahoo BOSS APIs(Build your Own Search Service)1) Define a set of topics, and train people to spontaneously conduct specific searches according to the descriptions of the topics. The search engine uses Yahoo BOSS APIs (Build your Own Search Service) provided by Yahoo
2)系统记录用户和检索系统的交互。包括涉及的查询，查询的更改，点击的文档，点击的停留时间等2) The system records the interaction between the user and the retrieval system. Include queries involved, changes to queries, documents clicked, dwell time of clicks, etc.
3)生成训练文件3) Generate training files
4)结束4) end
其中所述步骤2-b)所述的数据预处理：Wherein said step 2-b) described data preprocessing:
1)计算和使用狄利克雷平滑的作为词语与文档关联度的初始值，其中#(t,d)为词语t在文档d中出现的次数，P(t|C)为t出现在全集C中的次数，|d|为该文献的长度，μ为狄利克雷方法的参数，本发明中设置为50001) calculate and smoothed using Dirichlet As the initial value of the degree of association between the word and the document, where #(t,d) is the number of times the word t appears in the document d, P(t|C) is the number of times t appears in the complete set C, and |d| is the number of times the word t appears in the document The length of, μ is the parameter of Dirichlet method, is set to 5000 in the present invention
2)结束2) end
其中所述步骤2-f)所述的更新Pus(t|d)的值的过程：Wherein the process of updating the value of Pus (t|d) described in step 2-f):
1)如果是第一次交互，则不改变Pus(t|d)的值1) If it is the first interaction, the value of Pus (t|d) will not be changed
2)如果不是第一次交互，设当前交互的查询内容为qi,前一次交互的查询内容为qi-1,令qtheme为qi和qi-1的最长公共子序列，则+Δq＝qi-qtheme,-Δq＝qi-1-qtheme。更新Pus(t|d)分为权值不变，降低权值和增加权值的情况2) If it is not the first interaction, let the query content of the current interaction be qi , the query content of the previous interaction be qi-1 , let qtheme be the longest common subsequence of qi and qi-1 , then +Δq=qi -qtheme , -Δq=qi-1 -qtheme . Updating Pus (t|d) is divided into cases where the weight remains unchanged, the weight is reduced and the weight is increased
a)词语t与文档d关联度权值不变的情况。对于且t∈-Δq的情况，搜索引擎不改变其权值a) The case where the weight of the association degree between word t and document d remains unchanged. for And in the case of t∈-Δq, the search engine does not change its weight
b)降低词语t与文档d关联度权值的情况。当查询变更了，不论+Δq还是-Δq，只要出现在上次的搜索返回的结果集Di-1中，都要降低这些词语的权值。Pus(t|d)为词语t默认对当前查询和待评估文档之间的相关性的贡献。而由于词语t已经出现在文档集Di-1中，为了体现新颖度，词语t在文档集Di-1中出现的频率越高，权值就减得越多。因此，对于(t∈+Δqort∈-Δq)and t∈Di-1，有如下公式：b) Reduce the weight of the association degree between word t and document d. When the query is changed, regardless of +Δq or -Δq, as long as they appear in the result set Di-1 returned by the last search, the weights of these words must be reduced. Pus (t|d) is the contribution of term t default to the relevance between the current query and the document to be evaluated. Since the word t has already appeared in the document set Di-1 , in order to reflect the novelty, the higher the frequency of the word t in the document set Di-1 , the more the weight will be reduced. Therefore, for (t∈+Δqort∈-Δq)and t∈Di-1 , there is the following formula:
此处采用对数函数是为了防止数值下溢。The logarithmic function is used here to prevent numerical underflow.
其中确定的过程:which determines the process of:
i.将对qi-1搜索返回结果的前十个片段和满意的点击作为有效的搜索结果，记为所谓满意的点击是指在点击的文档上停留时间超过30si. Take the first ten fragments and satisfactory clicks of the search results returned by qi-1 as valid search results, recorded as The so-called satisfactory click refers to staying on the clicked document for more than 30s
ii.对于所有的搜索结果找出文本关联度与上次查询qi-1最大的即ii. For all search results Find the one with the largest text correlation degree and the last query qi-1 which is
其中in
iii.计算的值；以2-a方法iii. Calculate The value; with 2-a method
iv.结束确定的过程iv. End Confirmation the process of
c)增加词语t与文档d关联度权值的情况。c) The case of increasing the weight of the association degree between word t and document d.
i.当为一个增加的词语并且没有出现在上一次查询的结果集Di-1中，本发明将根据反文档的频率增加该词语的权值。对于有如下公式：i. When it is an added word and does not appear in the result set Di-1 of the previous query, the present invention will increase the weight of the word according to the frequency of the anti-document. for There are the following formulas:
log Pus(t|d)new＝(1+idf(t))log Pus(t|d)log Pus (t|d)new ＝(1+idf(t))log Pus (t|d)
其中：in:
idf(t)是反文档的频率，定义为：其中D是搜索引擎返回的全部文档数目。DW是D中t出现的文档数目。idf(t) is the frequency of inverse documents, defined as: where D is the total number of documents returned by the search engine. DW is the number of documents in D where t occurs.
ii.对于t∈qtheme，也需要增加权值，因为主题词通常是一个会话中的话题类或常用词，并不是整个全集中常用词语。因此，idf(t)并不适用此处。本发明用词语t在先前最大收益文档出现的频率的逆运算，来代替idf(t)。公式如下：ii. For t∈qtheme , it is also necessary to increase the weight, because the topic word is usually a topic class or a common word in a conversation, not a common word in the entire corpus. Therefore, idf(t) does not apply here. The present invention uses the inverse operation of the frequency of occurrence of word t in the previous maximum income document, to replace idf(t). The formula is as follows:
3)结束更新Pus(t|d)的值的过程3) End the process of updating the value of Pus (t|d)
其中所述步骤3-b)所述的文档d与查询qi的关联度计算：The calculation of the degree of association between the document d and the query qi described in step 3-b):
1)若i＝1，即用户在本会话中的第一次查询1) If i=1, it is the user's first query in this session
则关联度Score(q1,d)＝logP(q1|d)Then the degree of association Score(q1 ,d)=logP(q1 |d)
若i>1，则关联度:If i>1, the degree of association:
其中P(qi|d)为即时收益，Ps(t|d)由步骤2-b)求得。Where P(qi |d) is the immediate income, Ps (t|d) is obtained from step 2-b).
α,β,∈,δ为每种类型动作的折扣因子。根据历次实验得到α＝2.15，β＝1.75，∈＝0.07，δ＝0.42α, β, ∈, δ are discount factors for each type of action. According to previous experiments, α=2.15, β=1.75, ∈=0.07, δ=0.42
2)结束计算文档d与查询qi的关联度的过程2) End the process of calculating the degree of association between document d and query qi
其中所述步骤3-c)所述的文档d与会话搜索的关联度计算：Wherein the step 3-c) described document d and the association degree calculation of conversational search:
其中Score(q1,d)＝logP(q1|d)where Score(q1 ,d)=logP(q1 |d)
MDP模型中的γ为折扣因子，本发明取γ＝0.8。考虑到用户必然不满意重复查询之间的查询和搜索结果，本发明将重复查询之间的折扣因子设为0。公式如下：γ in the MDP model is a discount factor, and the present invention takes γ=0.8. Considering that the user must be dissatisfied with the query and search results between repeated queries, the present invention sets the discount factor between repeated queries to 0. The formula is as follows:
2)结束2) end
马尔科夫决策过程(MDP)一种广泛适用的决策模型。MDP为决策过程相关的所有代理建立状态空间S和动作空间A。代理的动作影响状态环境，使得状态作不确定的转换，动作的反馈也会影响代理的动作选择。本发明中，查询q建模为状态S，搜索引擎对词语和文档关联度权值的调整作为动作A，搜索引擎将累积收益值高的文档作为搜索结果返回给用户。用户之前的查询影响搜索结果，同时搜索结果会影响用户下一次查询的决策。这个过程不停迭代，直至查询结束。Markov decision process (MDP) is a widely applicable decision model. MDP establishes state space S and action space A for all agents involved in the decision process. The action of the agent affects the state environment, making the state transition indefinitely, and the feedback of the action will also affect the action choice of the agent. In the present invention, the query q is modeled as a state S, the search engine adjusts the word and document relevance weights as an action A, and the search engine returns documents with high cumulative income values to the user as search results. The user's previous query affects the search results, and the search results affect the user's next query decision. This process iterates continuously until the query ends.
有益效果，本发明提供了一种新颖的基于马尔科夫决策过程模型的信息检索方法。每次查询的语句q可以由多个词语t∈q组成。可以利用相邻查询之间查询语句的变更，以及前次检索的返回结果来加强会话搜索的准确性。该方法首先获取语料库，然后训练数据集，生成训练文档，接着对数据预处理，最后接收查询语句，返回优化查询地结果。本发明方法并不依赖于完备的查询记录，可直接应用于会话搜索。能充分利用会话搜索中查询变更的信息，可以有效的加强会话搜索的效果。Beneficial effects, the present invention provides a novel information retrieval method based on the Markov decision process model. The statement q of each query can consist of multiple words t∈q. The change of query statements between adjacent queries and the returned results of the previous retrieval can be used to enhance the accuracy of session search. This method first obtains the corpus, then trains the data set, generates training documents, then preprocesses the data, finally receives query statements, and returns optimized query results. The method of the invention does not depend on complete query records, and can be directly applied to session search. The information of the query change in the session search can be fully utilized, and the effect of the session search can be effectively enhanced.
附图说明Description of drawings
图1总述本发明的工作流程图；Fig. 1 sums up the work flowchart of the present invention;
图2本发明的基于马尔科夫决策过程模型的会话搜索方法工作流程图；Fig. 2 is a flow chart of the conversation search method based on the Markov decision process model of the present invention;
图3生成训练文件的工作流程图；Fig. 3 generates the workflow diagram of training file;
图4训练阶段工作流程图；Figure 4 is a workflow diagram of the training phase;
图5根据交互内容更新Pus(t|d)的值工作流程图；Fig. 5 is a flow chart of updating the value of Pus (t|d) according to the interactive content;
图6信息检索阶段工作流程图。Figure 6 Workflow diagram of the information retrieval phase.
具体实施方式Detailed ways
下面结合附图对本发明进行详细说明。The present invention will be described in detail below in conjunction with the accompanying drawings.
本发明是基于马尔科夫决策过程模型的会话搜索方法，意在提高信息检索时的准确性，为用户提供有用且满意的信息。如图1所示，描述了本发明的处理过程。本发明先获取语料库和训练数据，然后对数据进行预处理，接着执行具体查询检索，最后输出结果。The invention is a conversation search method based on a Markov decision process model, and aims to improve the accuracy of information retrieval and provide users with useful and satisfactory information. As shown in FIG. 1, the process of the present invention is described. The invention first acquires corpus and training data, then preprocesses the data, then executes specific query and retrieval, and finally outputs the result.
本发明中，将过程分为三个阶段，准备阶段，训练阶段和检索阶段，如图2所示。本发明的关键之处在于根据训练数据优化词语和文档关联度的权值，并在查询过程中实时调整词语和文档关联度的权值，以及计算文档和会话查询关联度的方法。In the present invention, the process is divided into three stages, preparation stage, training stage and retrieval stage, as shown in FIG. 2 . The key point of the present invention is to optimize the weight of word and document relevance according to the training data, adjust the weight of word and document relevance in real time during the query process, and calculate the method of document and session query relevance.
步骤2-0是本发明的基于马尔科夫决策过程模型的会话搜索方法的起始状态；Step 2-0 is the initial state of the conversation search method based on the Markov decision process model of the present invention;
准备阶段包括步骤2-1，2-2；The preparation phase includes steps 2-1, 2-2;
步骤2-1获取语料库，爬取足够多的网页；Step 2-1 Obtain the corpus and crawl enough web pages;
步骤2-2获取训练数据，生成训练文档，详细步骤见图4；Step 2-2 Obtain training data and generate training documents, see Figure 4 for detailed steps;
训练阶段包括步骤2-3，步骤2-4，步骤2-5；The training phase includes steps 2-3, steps 2-4, and steps 2-5;
步骤2-3数据预处理，计算词语t与文档d之间的关联度值；Step 2-3 data preprocessing, calculating the correlation value between word t and document d;
步骤2-4读取训练文档，读取会话中的交互信息；Steps 2-4 read the training document and read the interaction information in the session;
步骤2-5根据交互内容更新Pus(t|d)的值；Steps 2-5 update the value of Pus (t|d) according to the interactive content;
检索阶段包括步骤2-6，步骤2-7；The retrieval phase includes steps 2-6 and steps 2-7;
步骤2-6接收用户输入的查询语句，计算每个文档与整个会话的关联度；Steps 2-6 receive the query statement input by the user, and calculate the degree of relevance between each document and the entire session;
步骤2-7根据文档与查询的关联度，返回关联度高的前N个文档，本发明取N为10；Step 2-7 returns the first N documents with high correlation according to the correlation between the document and the query, and the present invention takes N as 10;
步骤2-8为结束状态。Steps 2-8 are the end states.
图3是生成训练文件的过程的详细描述。Figure 3 is a detailed description of the process of generating training files.
步骤3-0是生成训练文件的开始状态；Step 3-0 is to generate the starting state of the training file;
步骤3-1根据语料库全集C，定义一组主题，训练人员根据主题的描述进行搜索；Step 3-1 Define a group of topics according to the complete corpus C, and trainers search according to the description of the topics;
步骤3-2记录用户和检索系统的交互。包括涉及的查询，查询的更改，点击的文档，点击的停留时间等；Step 3-2 records the user's interaction with the retrieval system. Including the query involved, the change of the query, the document clicked, the dwell time of the click, etc.;
步骤3-3生成训练文件；Step 3-3 generates training files;
步骤3-4为生成训练文件的过程的结束状态，Step 3-4 is the end state of the process of generating the training file,
图4是针对训练阶段的详细描述。Figure 4 is a detailed description for the training phase.
步骤4-0是开始训练步骤；Step 4-0 is to start the training step;
步骤4-1对数据进行预处理，计算和使用狄利克雷平滑的并保存，#(t,d)为词语t在文档d中出现的次数，P(t|C)为t出现在全集C中的次数，|d|为该文献的长度，μ为狄利克雷方法的参数，本发明中设置为5000；Step 4-1 preprocess the data and calculate and smoothed using Dirichlet And save, #(t,d) is the number of times the word t appears in document d, P(t|C) is the number of times t appears in the complete set C, |d| is the length of the document, μ is Dirichlet The parameters of the method are set to 5000 in the present invention;
步骤4-2读取训练文件；Step 4-2 reads the training file;
步骤4-3读取一个会话；Step 4-3 reads a session;
步骤4-4取出会话中的交互；Step 4-4 takes out the interaction in the session;
步骤4-5判断会话中的交互是否结束，如果结束转到步骤4-6，如果没结束转到步骤4-4；Step 4-5 judges whether the interaction in the session is over, if it is over, go to step 4-6, if not, go to step 4-4;
步骤4-6根据交互内容更新Pus(t|d)的值，详细步骤见图5；Steps 4-6 update the value of Pus (t|d) according to the interactive content, and the detailed steps are shown in Figure 5;
步骤4-7判断训练文件中的会话是否读取结束，如果是全部训练结束则转到步骤4-8，如果否则转到步骤4-3；Step 4-7 judges whether the session in the training file has been read and finished, if all training is finished then go to step 4-8, if otherwise go to step 4-3;
步骤4-8为训练阶段结束状态。Steps 4-8 are the end state of the training phase.
图5是更新Pus(t|d)的值的详细描述。FIG. 5 is a detailed description of updating the value of Pus (t|d).
步骤5-0是更新Pus(t|d)的值的开始状态；Step 5-0 is the start state of updating the value of Pus (t|d);
步骤5-1判断是否属于第一次交互，属于则转到步骤5-2，否则转到步骤5-3；Step 5-1 judges whether it belongs to the first interaction, if yes, go to step 5-2, otherwise go to step 5-3;
步骤5-2属于第一次交互，权值不变，转到步骤5-7；Step 5-2 belongs to the first interaction, the weight value remains unchanged, go to step 5-7;
步骤5-3不属于第一次交互，判断t∈+Δq或者t∈-Δq。属于则转到步骤5-4，否则转到步骤5-5。确定+Δq，-Δq，qtheme方法：设当前交互的查询内容为qi,上一次交互的查询内容为qi-1,令qtheme为qi和qi-1的最长公共子序列，+Δq＝qi-qtheme,-Δq＝qi-1-qtheme；Step 5-3 does not belong to the first interaction, judge t∈+Δq or t∈-Δq. If yes, go to step 5-4, otherwise go to step 5-5. Determine +Δq, -Δq, qtheme method: set the query content of the current interaction as qi , the query content of the last interaction as qi-1 , let qtheme be the longest common subsequence of qi and qi-1 , +Δq=qi -qtheme ,-Δq=qi-1 -qtheme ;
步骤5-4当查询变更了，不论+Δq还是-Δq，只要出现在上次的搜索结果Di-1中，都要降低这些词语的权值，之后直接转到步骤5-7。因此，对于(t∈+Δq or t∈-Δq)and t∈Di-1，有如下公式：Step 5-4 When the query is changed, regardless of +Δq or -Δq, as long as it appears in the last search result Di-1 , the weight of these words must be reduced, and then directly go to step 5-7. Therefore, for (t∈+Δq or t∈-Δq)and t∈Di-1 , there is the following formula:
此处运用对数函数是为了防止数值下溢。P(t|d)通过计算t出现在全集C中的概率求得。其中，其中确定的程:The logarithmic function is used here to prevent numerical underflow. P(t|d) is obtained by calculating the probability that t appears in the complete set C. where, where determined program:
1)将返回结果的前十个片段和满意的点击作为有效的搜索结果，记为满意的点击的定义为点击的文档停留时间超过30s。1) Take the first ten fragments of returned results and satisfactory clicks as effective search results, recorded as Satisfactory clicks are defined as the clicked documents stay longer than 30s.
2)对于所有的搜索结果找出文本关联度与上次查询qi-1最大的即2) For all search results Find the one with the largest text correlation degree and the last query qi-1 which is
其中in
3)根据t和查找的值。3) According to t and look up value.
步骤5-5如果t不满足(t∈+Δq or t∈-Δq)and t∈Di-1，，判断t∈+Δq或则t∈qtheme，如果满足则转到步骤5-6，否则转到步骤5-2；Step 5-5 If t does not satisfy (t∈+Δq or t∈-Δq)and t∈Di-1 , judge t∈+Δq Or t∈qtheme , if satisfied, go to step 5-6, otherwise go to step 5-2;
步骤5-6满足t∈+Δq或则t∈qtheme，增加权值，之后直接转到步骤5-7。Steps 5-6 satisfy t∈+Δq Or t∈qtheme , increase the weight, and then directly go to steps 5-7.
1)对于t∈+Δq有如下公式：1) For t∈+Δq There are the following formulas:
logPus(t|d)new＝(1+idf(t))logPus(t|d)logPus (t|d)new ＝(1+idf(t))logPus (t|d)
2)对于t∈qtheme，也需要增加权值，由于主题词通常是一个会话中的话题类或常用词，并不是整个全集中常用词语。因此，idf(t)并不适用此处。本发明中用词语t在先前最大收益文档出现的频率的逆运算，来代替idf(t)。公式如下：2) For t ∈ qtheme , it is also necessary to increase the weight, because the topic word is usually a topic class or a common word in a conversation, not a common word in the entire corpus. Therefore, idf(t) does not apply here. In the present invention, the inverse operation of the frequency of occurrence of the word t in the previous maximum income document is used, to replace idf(t). The formula is as follows:
其中：idf(t)是逆文本指数，公式为：其中D是搜索引擎返回的全部网页数，DW是t出现的网页数。Among them: idf(t) is the inverse text index, the formula is: Among them, D is the number of all web pages returned by the search engine, and DW is the number of web pages where t appears.
步骤5-7为更新Pus(t|d)的值的结束状态。Steps 5-7 are the end state of updating the value of Pus (t|d).
图6为信息检索阶段工作流程图Figure 6 is a workflow flowchart of the information retrieval stage
步骤6-0是信息检索阶段的开始状态；Step 6-0 is the start state of the information retrieval phase;
步骤6-1接收用户输入的查询语句；Step 6-1 receives the query statement input by the user;
步骤6-2通过公式计算每个文档d与当前查询qi的关联度；Step 6-2 calculates the degree of association between each document d and the current query qi by a formula;
步骤6-3计算每个文档d与整个会话的关联度以考虑用户之前的搜索对搜索结果的影响；Step 6-3 calculates the relevance of each document d to the entire session to consider the impact of the user's previous searches on the search results;
步骤6-4返回与整个会话的关联度高的文档；Step 6-4 returns documents with a high degree of relevance to the entire session;
步骤6-5判断用户查询是否结束，结束则执行步骤6-1，否则执行步骤6-6；Step 6-5 judges whether the user query is over, and then executes step 6-1, otherwise executes step 6-6;
步骤6-6为信息检索阶段的结束状态。Step 6-6 is the end state of the information retrieval phase.
本发明方案所公开的技术手段不仅限于上述实施方式所公开的技术手段，还包括由以上技术特征任意组合所组成的技术方案。应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也视为本发明的保护范围。The technical means disclosed in the solutions of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be pointed out that for those skilled in the art, some improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610237174.9ACN105930400B (en) | 2016-04-15 | 2016-04-15 | A Conversational Search Method Based on Markov Decision Process Model |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610237174.9ACN105930400B (en) | 2016-04-15 | 2016-04-15 | A Conversational Search Method Based on Markov Decision Process Model |
| Publication Number | Publication Date |
|---|---|
| CN105930400A CN105930400A (en) | 2016-09-07 |
| CN105930400Btrue CN105930400B (en) | 2019-10-11 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610237174.9AActiveCN105930400B (en) | 2016-04-15 | 2016-04-15 | A Conversational Search Method Based on Markov Decision Process Model |
| Country | Link |
|---|---|
| CN (1) | CN105930400B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107729512A (en)* | 2017-10-20 | 2018-02-23 | 南京大学 | A kind of session searching method based on part Observable markov decision process model |
| CN109117475B (en)* | 2018-07-02 | 2022-08-16 | 武汉斗鱼网络科技有限公司 | Text rewriting method and related equipment |
| CN109241243B (en)* | 2018-08-30 | 2020-11-24 | 清华大学 | Candidate document sorting method and device |
| CN109783709B (en)* | 2018-12-21 | 2023-03-28 | 昆明理工大学 | Sorting method based on Markov decision process and k-nearest neighbor reinforcement learning |
| CN111241407B (en)* | 2020-01-21 | 2023-07-28 | 中国人民大学 | Personalized search method based on reinforcement learning |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102117321A (en)* | 2010-01-06 | 2011-07-06 | 微软公司 | Automated discovery aggregation and organization of subject area discussions |
| CN102262661A (en)* | 2011-07-18 | 2011-11-30 | 南京大学 | Web page access forecasting method based on k-order hybrid Markov model |
| CN103425710A (en)* | 2012-05-25 | 2013-12-04 | 北京百度网讯科技有限公司 | Subject-based searching method and device |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9076448B2 (en)* | 1999-11-12 | 2015-07-07 | Nuance Communications, Inc. | Distributed real time speech recognition system |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102117321A (en)* | 2010-01-06 | 2011-07-06 | 微软公司 | Automated discovery aggregation and organization of subject area discussions |
| CN102262661A (en)* | 2011-07-18 | 2011-11-30 | 南京大学 | Web page access forecasting method based on k-order hybrid Markov model |
| CN103425710A (en)* | 2012-05-25 | 2013-12-04 | 北京百度网讯科技有限公司 | Subject-based searching method and device |
| Title |
|---|
| 基于MarKov链的Web访问序列挖掘算法研究;肖哲;《中国优秀硕士学位论文全文数据库》;20100415(第2010年第04期);I138-417* |
| Publication number | Publication date |
|---|---|
| CN105930400A (en) | 2016-09-07 |
| Publication | Publication Date | Title |
|---|---|---|
| CN103678576B (en) | The text retrieval system analyzed based on dynamic semantics | |
| CN103605665B (en) | Keyword based evaluation expert intelligent search and recommendation method | |
| CN105930400B (en) | A Conversational Search Method Based on Markov Decision Process Model | |
| CN104408148B (en) | A kind of field encyclopaedia constructing system based on general encyclopaedia website | |
| CN107193803A (en) | A kind of particular task text key word extracting method based on semanteme | |
| CN101968819B (en) | Audio and video intelligent cataloging information acquisition method facing wide area network | |
| CN101630327A (en) | Design method of theme network crawler system | |
| CN107153658A (en) | A method for discovering public opinion hot words based on keyword weighting algorithm | |
| US8229960B2 (en) | Web-scale entity summarization | |
| CN114706972B (en) | An automatic generation method of unsupervised scientific and technological information summaries based on multi-sentence compression | |
| CN103838735A (en) | Data retrieval method for improving retrieval efficiency and quality | |
| CN110134799B (en) | BM25 algorithm-based text corpus construction and optimization method | |
| CN109597995A (en) | A kind of document representation method based on BM25 weighted combination term vector | |
| Du et al. | An approach for selecting seed URLs of focused crawler based on user-interest ontology | |
| CN102637179B (en) | Method and device for determining lexical item weighting functions and searching based on functions | |
| CN111460158A (en) | Microblog topic public emotion prediction method based on emotion analysis | |
| CN118626611A (en) | Retrieval method, device, electronic device and readable storage medium | |
| Kanapala et al. | Passage-based text summarization for legal information retrieval | |
| CN108491407B (en) | Code retrieval-oriented query expansion method | |
| CN111680505B (en) | A Markdown feature-aware unsupervised keyword extraction method | |
| JP5427694B2 (en) | Related content presentation apparatus and program | |
| CN112214511A (en) | An API Recommendation Method Based on WTP-WCD Algorithm | |
| Yadav et al. | Large text document summarization based on an enhanced fuzzy logic approach | |
| CN119557500B (en) | A method and system for accurate search of Internet massive data based on AI technology | |
| Armano et al. | Semantic Enrichment of Contextual Advertising by using Concepts. |
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |