CN105930400A - Markov decision process model based session search method - Google Patents

Abstract

The invention discloses a Markov decision process model based session search method. The method comprises the following steps of 1) in a preparation stage, crawling sufficient web pages and obtaining a corpus complete set C, wherein C is a result after screening crawled web page sets, and each web page correspondingly contains a document d; and performing spontaneous session search and recording a search process by a trainer to obtain training data, and generating a training document; 2) in a training stage, preprocessing the data, and making statistics on a degree of association between words t required to be used in a retrieval stage and the document d, wherein the degree of association includes Ps(t/d) using Dirichlet smoothing and Pus(t/d) not using the Dirichlet smoothing; and 3) in the retrieval stage, receiving a query statement qi currently input by a user; calculating a degree of association between each document d in the corpus complete set C and the current query statement qi through a formula; calculating a degree of association between each document d and search of the whole session; and returning first N documents with high degree of association.

Description

A kind of session searching method based on markov decision process model

Technical field

The present invention relates to the information retrieval of dialogue-based search, particularly relate to a kind of based on Markovian decision The session searching method of process model.

Background technology

Session search is a kind of realizing alternately by search engine and user during search sessions The search technique of information retrieval.When searching for the relevant information of a theme, user can be the most continuous Ground change inquiry content, until finding its information needed.In session search, user can returning according to search engine Returning result and adjust inquiry content, search engine is then as the feedback of user, in order to improve session search Result.The rule that in session search, the change pattern of user's inquiry is the most unified, search engine is difficult to accurately obtain The intention of user, thus session search is a challenging information retrieval task.Inquire about currently with study The method of change mainly has:

(1) classify based on different types of search, such as embody, summarize, conversion, or Slight change, then performs retrieving.

(2) another kind of method is inquiry to be mapped as semantic figure represent, as body or inquiry log are formed Inquiry flow graph, then research inquiry Changing Pattern in the drawings.

But, first method depends on complete inquiry record, and such data are often difficult to obtain. Ontological in second method maps very challenging, and this likely introduces coarse intermediate data, and damages The accuracy of bad search.So, although these methods may apply to, in information retrieval task, such as inquire about Standardization, Associate Inquire etc., it can be difficult to directly apply to session search.

In order to make full use of, session search be inquired about the information of change, invent a kind of based on Ma Er herein The session searching method of section husband decision process model, can effectively strengthen the effect of session search.

Summary of the invention

Goal of the invention: the invention provides the information retrieval based on markov decision process model of a kind of novelty Method.The statement q of inquiry can be made up of multiple word t ∈ q every time.Can utilize and inquire about between adjacent inquiry The change of statement, and the return result of previous retrieval strengthen session search accuracy.First the method obtains Take corpus, then training dataset, generate Training document, then to data prediction, finally receive inquiry Statement, returns Optimizing Queries ground result.

A kind of session searching method based on markov decision process model, the method comprises the steps:

1) preparatory stage

A) crawl abundant webpage, obtain corpus complete or collected works C.C is that the collections of web pages crawled is after screening Result.Corresponding document d, C={d in each webpage_i}

B) trainer carries out spontaneous session and searches for and record its process to obtain training data (training data Including the inquiry related to, the change of inquiry, document that user clicks in the result that search engine returns and point thereof Hit the time of staying etc.), generate Training document

C) preparatory stage is terminated

2) training stage

A) data prediction, needs the P used in statistics retrieval phase_s(t | d) and P_us(t|d)

B) analyzing step 1-b) the middle Training document generated, including the information of multiple sessions

C) reading a session in intensified learning, a session includes that one or many user is to Search Results The information of operation

D) read information that Search Results operates by a user in session (include the inquiry related to, inquiry Change, the document of click, the time of staying etc. of click), and thus update P_usThe value of (t | d)

E) step d) is repeated until conversation end

F) step c) is repeated, d), e) until all sessions are all processed complete

G) training stage is terminated

3) retrieval phase

A) the query statement q that user is currently entered is received_i

B) each document d and current queries q is calculated by formula_iThe degree of association

C) degree of association of each document d and whole session is calculated

D) front N piece document (present invention takes 10) that the degree of association is high is returned

E) repeat a), b), c), d) until user terminates inquiry

F) retrieval phase is terminated

Wherein said step 1-b) described in training file

1) defining one group of theme (topics), trainer spontaneously carries out specific search according to the description of theme. Search engine uses the Yahoo BOSS APIs (Build your Own Search Service) that Yahoo provides

2) system records user and searching system is mutual.Including the inquiry related to, the change of inquiry, click on Document, the time of staying etc. of click

3) training file is generated

4) terminate

Wherein said step 2-b) described in data prediction:

1) calculateSmooth with using Di Li CrayAs word The initial value of language and document associations degree, wherein (t, d) is the number of times that occurs in document d of word t to #, and P (t | C) is T occurs in the number of times in complete or collected works C, the length that | d | is the document, and μ is the parameter of Di Li Cray method, this 5000 it are set in bright

2) terminate

Wherein said step 2-f) described in renewal P_usThe process of the value of (t | d):

1) if the most mutual, P is not the most changed_usThe value of (t | d)

2) if not the most mutual, if the most mutual inquiry content is q_i, in the most mutual front inquiry Hold for q_i-1, make q_themeFor q_iAnd q_i-1Longest common subsequence, then+Δ q=q_i-q_theme,-Δ q= q_i-1-q_theme.Update P_us(t | d) it is divided into weights constant, reduce weights and increase the situation of weights

A) situation constant with document d degree of association weights for word t.ForAnd the situation of t ∈-Δ q, Search engine does not change its weights

B) situation of word t and document d degree of association weights is reduced.When inquiry changes, no matter+Δ q or -Δ q, if the result set D that the search occurring in last time returns_i-1In, the weights of these words will be reduced. P_us(t | d) it is the contribution to the correlation between current queries and document to be assessed of the word t acquiescence.And due to word Language t is already present in document sets D_i-1In, in order to embody novelty degree, word t is in document sets D_i-1Middle appearance Frequency is the highest, and weights just subtract the most.Therefore, for (t ∈+Δ q ort ∈-Δ q) and t ∈ D_i-1, have Equation below:

$\log P_{us}{\left(t\right|d)}_{new}=(1-P_{us}(t|d_{i-1}^{*}\left)\right)\log P_{us}\left(t\right|d)$

Use logarithmic function is to prevent numerical underflows herein.

Wherein determineProcess:

I. will be to q_i-1Search returns front ten fragments of result with satisfied click as effective Search Results, It is designated asSo-called satisfied click refers to that the time of staying is more than 30s on the document clicked on

Ii. for all of Search ResultsFind out textual association degree and inquire about q with last time_i-1Maximum 'sI.e.

$d_{i-1}^{*}=\underset{d_k\∈D_{i-1}^e}{\mathrm{argmax}}\left(q_{i-1}\right|d_k)$

Wherein

Iii. calculateValue；With 2-a method

Iv. terminate to determineProcess

C) situation of word t and document d degree of association weights is increased.

I. when for a word increased and the result set D that do not appears in last inquiry_i-1In, this Bright will increase the weights of this word according to the frequency of anti-document.ForThere are following public affairs Formula:

log P_us(t|d)_new=(1+idf (t)) log P_us(t|d)

Wherein:

Idf (t) is the frequency of anti-document, is defined as:Wherein D is that search engine returns All number of documents.D_WIt it is the number of documents that in D, t occurs.

Ii. for t ∈ q_theme, it is also desirable to increase weights, because descriptor is typically the topic in a session Class or everyday words, be not everyday expressions in whole complete or collected works.Therefore, idf (t) and inapplicable herein.The present invention The inverse operation of frequency occurred with word t previously maximum return document,Replace idf (t). Formula is as follows:

$\log P_{us}{\left(t\right|d)}_{new}=(1+(1-P_{us}\left(t|d_{i-1}^{*}\right)\left)\right)\log P_{us}\left(t\right|d)$

3) terminate to update P_usThe process of the value of (t | d)

Wherein said step 3-b) described in document d with inquiry q_iCalculation of relationship degree:

1) if i=1, the i.e. user first time in this session inquires about

Then degree of association Score (q₁, d)=logP (q₁|d)

If i > 1, the then degree of association:

$\begin{array}{c}Score(q_i,d)=\log P\left(q_i\right|d)+\mathrm{\α}\underset{t\∈q_{theme}}{\mathrm{\Σ}}\[1-P\left(t\right|d_{i-1}^{*})\]\log P_{us}\left(t\right|d)\\ -\mathrm{\β}\underset{t\∈d_{i-1}^{*}}{\underset{t\∈+\mathrm{\Δ}q}{\mathrm{\Σ}}}{P}_{us}\left(t\right|d_{i-1}^{*})\log P_{us}\left(t\right|d)+\∈\underset{t\∉d_{i-1}^{*}}{\underset{t\∈+\mathrm{\Δ}q}{\mathrm{\Σ}}}idf\left(t\right)\log P_{us}\left(t\right|d)\\ -\mathrm{\δ}\underset{t\∈-\mathrm{\Δ}q}{\mathrm{\Σ}}{P}_{us}\left(t\right|d_{i-1}^{*})\log P_{us}\left(t\right|d)\end{array}$

Wherein P (q_i| d) it is instant income,P_s(t | d) by step 2-b) try to achieve.

α, β, ∈, δ are the discount factor of each type action.α=2.15 are obtained according to all previous experiment, β=1.75, ∈=0.07, δ=0.42

2) process calculating document d with the degree of association of inquiry qi is terminated

Wherein said step 3-c) described in the calculation of relationship degree of document d and session search:

$1)---{\mathrm{Score}}_{session}(q_n,d)={\mathrm{\Σ}}_{i=1}^n{\mathrm{\γ}}^{n-i}Score(q_i,d)$

Wherein Score (q₁, d)=logP (q₁|d)

γ in MDP model is discount factor, and the present invention takes γ=0.8.Repeat in view of user is necessarily dissatisfied Inquiry between inquiry and Search Results, the discount factor repeated between inquiry is set to 0 by the present invention.Formula is such as Under:

2) terminate

A kind of generally applicable decision model of markov decision process (MDP).MDP is that decision process is correlated with All agencies set up state space S and motion space A.The action of agency affects conditions Ambient so that state Making uncertain conversion, the feedback of action also can affect the Action Selection of agency.In the present invention, inquiry q modeling For state S, the adjustment of word and document associations degree weights will be tired out by search engine as action A, search engine The document that long-pending financial value is high returns to user as Search Results.Inquiry before user affects Search Results, with Time Search Results can affect the decision-making that user inquires about next time.This process does not stop iteration, until poll-final.

Beneficial effect, the invention provides the information retrieval based on markov decision process model of a kind of novelty Method.The statement q of inquiry can be made up of multiple word t ∈ q every time.Can utilize and inquire about between adjacent inquiry The change of statement, and the return result of previous retrieval strengthen session search accuracy.First the method obtains Take corpus, then training dataset, generate Training document, then to data prediction, finally receive inquiry Statement, returns Optimizing Queries ground result.The inventive method is not rely on complete inquiry record, can directly answer Search for for session.The information inquiring about change in session search can be made full use of, can effectively strengthen session and search The effect of rope.

Accompanying drawing explanation

The workflow diagram of Fig. 1 summation present invention；

The session searching method workflow diagram based on markov decision process model of Fig. 2 present invention；

Fig. 3 generates the workflow diagram of training file；

Fig. 4 training stage workflow diagram；

Fig. 5 updates P according to interaction content_usThe value workflow diagram of (t | d)；

Fig. 6 information retrieval stage workflow diagram.

Detailed description of the invention

The present invention is described in detail below in conjunction with the accompanying drawings.

The present invention is session searching method based on markov decision process model, it is intended that when improving information retrieval Accuracy, provide the user useful and satisfied information.As it is shown in figure 1, describe the process of the present invention Journey.The present invention first obtains corpus and training data, then pre-processes data, then performs specifically to look into Ask retrieval, finally export result.

In the present invention, process is divided into three phases, preparatory stage, training stage and retrieval phase, such as Fig. 2 Shown in.The present invention's it is critical that optimizes word and the weights of document associations degree according to training data, and Query script adjusts word and the weights of document associations degree in real time, and calculates document and the session inquiry degree of association Method.

Step 2-0 is the initial state of the session searching method based on markov decision process model of the present invention；

Preparatory stage includes step 2-1,2-2；

Step 2-1 obtains corpus, crawls abundant webpage；

Step 2-2 obtains training data, generates Training document, and detailed step is shown in Fig. 4；

Training stage includes step 2-3, step 2-4, step 2-5；

Step 2-3 data prediction, calculates and associates angle value between word t and document d；

Step 2-4 reads Training document, reads the interactive information in session；

Step 2-5 updates P according to interaction content_usThe value of (t | d)；

Retrieval phase includes step 2-6, step 2-7；

Step 2-6 receives the query statement of user's input, calculates the degree of association of each document and whole session；

Step 2-7, according to the degree of association of document Yu inquiry, returns the top n document that the degree of association is high, and the present invention takes N is 10；

Step 2-8 is done state.

Fig. 3 is the detailed description of the process generating training file.

Step 3-0 is the beginning state generating training file；

Step 3-1, according to corpus complete or collected works C, defines one group of theme, and trainer is carried out according to the description of theme Search；

Step 3-2 record user and searching system mutual.Including the inquiry related to, the change of inquiry, click on Document, the time of staying etc. of click；

Step 3-3 generates training file；

Step 3-4 is the done state of the process generating training file,

Fig. 4 is the detailed description for the training stage.

Step 4-0 is to start training step；

Data are pre-processed by step 4-1, calculateSmooth with using Di Li CrayAnd preserve, # (t, d) is the number of times that occurs in document d of word t, P (t | C) be T occurs in the number of times in complete or collected works C, the length that | d | is the document, and μ is the parameter of Di Li Cray method, this 5000 it are set in bright；

Step 4-2 reads training file；

Step 4-3 reads a session；

It is mutual that step 4-4 is taken out in session；

Step 4-5 judges whether terminating alternately in session, if terminating to forward step 4-6 to, if do not terminated Forward step 4-4 to；

Step 4-6 updates P according to interaction content_usThe value of (t | d), detailed step is shown in Fig. 5；

Whether the session in step 4-7 training of judgement file is read and is terminated, if all training terminates then to forward to Step 4-8, if otherwise forwarding step 4-3 to；

Step 4-8 is training stage done state.

Fig. 5 is to update P_usThe detailed description of the value of (t | d).

Step 5-0 is to update P_usThe beginning state of the value of (t | d)；

Step 5-1 judges whether to belong to the most mutual, belongs to, forwards step 5-2 to, otherwise forwards step 5-3 to；

Step 5-2 belongs to the most mutual, and weights are constant, forward step 5-7 to；

Step 5-3 is not belonging to the most mutual, it is judged that t ∈+Δ q or t ∈-Δ q.Belong to, forward step 5-4 to, Otherwise forward step 5-5 to.Determine+Δ q ,-Δ q, q_themeMethod: set the most mutual inquiry content as q_i, Last mutual inquiry content is q_i-1, make q_themeFor q_iAnd q_i-1Longest common subsequence, + Δ q=q_i-q_theme,-Δ q=q_i-1-q_theme；

Step 5-4 changes when inquiry, though+Δ q or-Δ q, as long as occurring in the Search Results D of last time_i-1 In, the weights of these words will be reduced, pass directly to step 5-7 afterwards.Therefore, for (t ∈+Δ q or t ∈ -Δq)and t∈D_i-1, there is an equation below:

$\log P_{us}{\left(t\right|d)}_{new}=(1-P_{us}(t|d_{i-1}^{*}\left)\right)\log P_{us}\left(t\right|d)$

Use logarithmic function is to prevent numerical underflows herein.P (t | d) occurred in complete or collected works C by calculating t Probability try to achieve.Wherein, wherein determineJourney:

1) using the click of front ten fragments returning result and satisfaction as effective Search Results, it is designated as The definition of satisfied click is that the document time of staying clicked on is more than 30s.

2) for all of Search ResultsFind out textual association degree and inquire about q with last time_i-1? BigI.e.

$d_{i-1}^{*}=\underset{d_k\∈D_{i-1}^e}{\mathrm{argmax}}P\left(q_{i-1}\right|d_k)$

Wherein

3) according to t andSearchValue.

If step 5-5 t is unsatisfactory for (t ∈+Δ q or t ∈-Δ q) and t ∈ D_i-1, it is judged that t ∈+Δ qOr t is ∈ q_themeIf meeting, forwarding step 5-6 to, otherwise forward step 5-2 to；

Step 5-6 meets t ∈+Δ qOr t is ∈ q_theme, increase weights, the most directly turn To step 5-7.

1) for t ∈+Δ qThere is an equation below:

logP_us(t|d)_new=(1+idf (t)) logP_us(t|d)

2) for t ∈ q_theme, it is also desirable to increase weights, owing to descriptor is typically the topic in a session Class or everyday words, be not everyday expressions in whole complete or collected works.Therefore, idf (t) and inapplicable herein.The present invention The inverse operation of the frequency that middle word t previously maximum return document occurs,Replace idf (t). Formula is as follows:

$\log P_{us}{\left(t\right|d)}_{new}=(1+(1-P_{us}\left(t|d_{i-1}^{*}\right)\left)\right)\log P_{us}\left(t\right|d)$

Wherein: idf (t) is inverse text index, and formula is:Wherein D is that search engine returns Whole webpage numbers, D_WIt it is the webpage number of t appearance.

Step 5-7 is for updating P_usThe done state of the value of (t | d).

Fig. 6 is information retrieval stage workflow diagram

Step 6-0 is the beginning state in information retrieval stage；

Step 6-1 receives the query statement of user's input；

Step 6-2 calculates each document d and current queries q by formula_iThe degree of association；

Step 6-3 calculates the search before consideration user of the degree of association of each document d and whole session to search The impact of result；

Step 6-4 returns the document high with the degree of association of whole session；

Step 6-5 judges whether user's inquiry terminates, and terminates then to perform step 6-1, otherwise performs step 6-6；

Step 6-6 is the done state in information retrieval stage.

Technological means disclosed in the present invention program is not limited only to the technological means disclosed in above-mentioned embodiment, also Including the technical scheme being made up of above technical characteristic.It should be pointed out that, for the art For those of ordinary skill, under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, These improvements and modifications are also considered as protection scope of the present invention.

Claims (3)

1. a session searching method based on markov decision process model, it is characterised in that include as follows Step:

1) preparatory stage

A) crawl abundant webpage, obtain corpus complete or collected works C；C is that the collections of web pages crawled is after screening Result；To having document d, then C={d in each webpage_i}；

B) trainer carries out spontaneous session and searches for and record its process to obtain training data (training data Including the inquiry related to, the change of inquiry, document that user clicks in the result that search engine returns and point thereof Hit the time of staying etc.), generate Training document；

C) preparatory stage is terminated；

2) training stage

A) data prediction, needs the degree of association between word t and the document d used in statistics retrieval phase, P is smoothed including using Di Li Cray_s(t | d) and do not use the smooth P of Di Li Cray_us(t|d)；

B) analyzing step 1-b) the middle Training document generated, including the information of session；

C) reading a session in Training document, a session includes that one or many user is to Search Results The information of operation；

D) (including the inquiry related to, inquiry is more to read in session the information that Search Results operates by user Change, the document of click, the time of staying etc. of click), and thus update P_usThe value of (t | d)；

E) step d) is repeated until conversation end；

F) step c) is repeated, d), e) until all sessions are all processed complete；

G) training stage is terminated；

3) retrieval phase

A) the query statement q that user is currently entered is received_i；

B) each document d and current queries q in corpus complete or collected works C is calculated by formula_iThe degree of association；

C) degree of association of each document d and whole session search is calculated；

D) front N piece document (present invention takes 10) that the degree of association is high is returned；

E) step a) is repeated, b), c), d) until user terminates inquiry；

F) retrieval phase is terminated；

Wherein said step 2-a) described in data prediction:

1) calculateSmooth with using Di Li CrayAs word The initial value of language t and the document d degree of association, wherein # (t, d) is the number of times that occurs in document d of word t, P (t | C) is the number of times that t occurs in corpus complete or collected works C, and | d | is the length of document d, and μ is Di Li Cray side The parameter of method, is set to 5000 in the present invention；

2) terminate；

Wherein said step 2-d) described in more neologism t and document d between the degree of association, i.e. update P_us(t|d) The process of value:

1) if inquiry is mutual for the first time, P is not the most changed_usThe value of (t | d)；

2) mutual, if the most mutual inquiry content is q if not inquiry for the first time_i, the most mutual front looking into Inquiry content is q_i-1, make q_themeFor q_iAnd q_i-1Longest common subsequence ,+Δ q=q_i-q_theme,-Δ q= q_i-1-q_theme；To P_usThe renewal of (t | d) is divided into weights constant, reduces weights and increases the situation of weights；

A) situation constant with the weights of the document d degree of association for word t；ForAnd the feelings of t ∈-Δ q Condition, search engine does not change its weights；

B) situation of word t and document d degree of association weights is reduced；When inquiry changes, whether+Δ q is also It is-Δ q, as long as occurring in document sets D of the Search Results of last time_i-1In, the weights of these words will be reduced； P_us(t | d) it is the word t contribution to the correlation between current queries and document to be assessed；And due to word t It is already present in document sets D_i-1In, in order to embody novelty degree, word t is in document sets D_i-1The frequency of middle appearance The highest, weights just subtract the most；Therefore, for (t ∈+Δ q or t ∈-Δ q) and t ∈ D_i-1, just like Lower formula:

$\log P_{us}{\left(t\right|d)}_{new}=(1-P_{us}(t|d_{i-1}^{*}\left)\right)\log P_{us}\left(t\right|d)$

Use logarithmic function is to prevent numerical underflows herein；

Wherein determineProcess:

I. will be to q_i-1Search returns front ten fragments of result with satisfied click as effective Search Results, It is designated asSo-called satisfied click refers to that the time of staying is more than 30s on the document clicked on；

Ii. for all of document searching resultFind out textual association degree and inquire about q with last time_i-1 MaximumI.e.

$d_{i-1}^{*}=\underset{d_k\∈D_{i-1}^e}{\mathrm{argmax}}P\left(q_{i-1}\right|d_k)$

Wherein

Iii. calculateValue；With 2-a method

Iv. terminate to determineProcess；

C) situation of word t and document d degree of association weights is increased；

I. when for a word increased and the result set D that do not appears in last inquiry_i-1In, this The bright middle weights by proportional for the frequency according to anti-document these words of increase；If one in a lot of documents Common word, in order to ensure increasing the word of a preference when, it is to avoid increase too much；For t∈+Δq andThere is an equation below:

log P_us(t|d)_new=(1+idf (t)) log P_us(t|d)

Wherein:

Idf (t) is the frequency of anti-document, is defined as:Wherein D is that search engine returns All numbers of document；D_WIt it is the number of documents that t occurs in D；

Ii. for t ∈ q_theme, also increase weights, the topic class that is typically in a session due to descriptor or Everyday words, is not everyday expressions in whole complete or collected works；Therefore, idf (t) and inapplicable herein；The present invention uses The inverse operation of the frequency that word t previously maximum return document occurs,Replace idf (t)；Public Formula is as follows:

$\log P_{us}{\left(t\right|d)}_{new}=(1+(1-P_{us}\left(t|d_{i-1}^{*}\right)\left)\right)\log P_{us}\left(t\right|d)$

3) terminate to update P_usThe process of the value of (t | d).

Session searching method the most according to claim 1, it is characterised in that wherein said step 3-b) described Document d with inquiry q_iCalculation of relationship degree:

1) if i=1, the i.e. user first time in this session inquires about

Then degree of association Score (q₁, d)=logP (q₁|d)

If i > 1, the then degree of association:

$\begin{array}{c}Score\left(q_i,d\right)=\log P\left(q_i|d\right)+\mathrm{\α}\underset{t\∈q_{theme}}{\mathrm{\Σ}}\[1-P\left(t|d_{i-1}^{*}\right)\]\log P_{us}\left(t|d\right)\\ -\mathrm{\β}\underset{t\∈d_{i-1}^{*}}{\underset{t\∈+\mathrm{\Δ}q}{\mathrm{\Σ}}}{P}_{us}\left(t|d_{i-1}^{*}\right)\log P_{us}\left(t|d\right)+\∈\underset{t\∉d_{i-1}^{*}}{\underset{t\∈+\mathrm{\Δ}q}{\mathrm{\Σ}}}idf\left(t\right)\log P_{us}\left(t|d\right)\\ -\mathrm{\δ}\underset{t\∈-\mathrm{\Δ}q}{\mathrm{\Σ}}{P}_{us}\left(t|d_{i-1}^{*}\right)\log P_{us}\left(t|d\right)\end{array}$

Wherein P (q_i| d) it is instant income,P_s(t | d) by step 2-a) try to achieve；

α, β, ∈, δ are the discount factor of each type action；α=2.15 are obtained according to all previous experiment, β=1.75, ∈=0.07, δ=0.42；

2) process calculating document d with the degree of association of inquiry qi is terminated.

Session searching method the most according to claim 1, it is characterised in that wherein said step 3-c) described The calculating of document d and the session search degree of association:

1)

Wherein Score (q₁, d)=logP (q₁|d)

γ in MDP model is discount factor, and the present invention takes γ=0.8；Repeat in view of user is necessarily dissatisfied Inquiry between inquiry and Search Results, the discount factor repeated between inquiry is set to 0 by the present invention；Formula is such as Under:

2) terminate.