CN103793491B - Chinese news story segmentation method based on flexible semantic similarity measurement - Google Patents

Abstract

The invention discloses a Chinese news story segmentation method based on flexible semantic similarity measurement. The method includes the following steps that a target text set is input, and word segmentation is conducted on news storyboards Ti in the text set; a context relation diagram is built; iteration spreading is conducted on context relevancy between words through the context relation diagram and a rapid sorting algorithm to acquire a flexible semantic relevant matrix; flexible semantic similarity between sentences is defined through the flexible semantic relevant matrix; a Chinese news story is segmented through the flexible semantic similarity. By the adoption of the flexible measurement method, the semantic similarity between the words and between word sets can be more reasonably expressed. Experiments show that compared with a traditional similarity measurement method, in a Chinese news story segmentation technology and based on same segmentation principles, the flexible semantic similarity measurement method can improve segmentation accuracy to 3% to 10%.

Description

A kind of Chinese News Stories dividing method based on flexible Semantic Similarity tolerance

Technical field

The present invention relates to Chinese News Stories segmentation field, particularly in a kind of tolerance based on flexible Semantic Similarity Civilian News Stories dividing method.

Background technology

Popularization with network and development, for example: in the multimedia of Broadcast Journalism, minutes, online open class etc Holding rapidly increases, and is badly in need of now a kind of effective method and is automatically organized this kind of multi-medium data, for base Text retrieval and analysis in theme.One multimedia document, the broadcasts-based network of such as a hour, generally by multiple events Thing (story) form, in order to carry out efficient semantic retrieval, instruct user go to find their subject of interest beginning and End is critically important, and meanwhile, a multimedia document split is by topic tracking^[1], classification and summarize^[2]Contour level The important prerequisite condition that secondary semanteme browses.The purpose of News Stories cutting techniques is that and for News Stories script to be divided into master Inscribe consistent story.Technically, the efficiency of News Stories cutting techniques is related to two factors: one is the phase between word Measure like the similitude between property and this sentence set；Two is the criterion of segmentation News Stories script.

Many work before are all conceived to segmentation criterion reasonable in design, for example: texttiling^[3][4]Minimum normalizing Change segmentation criterion (minimum ncuts)^[5][6], maximum vocabulary connect criterion^[7]Deng.Compared with widely studied segmentation criterion, Great majority work at this stage is all using simple based on the rigid similarity measurement mode repeating, i.e. phase between identical word It is 1 like property, the similitude between different terms is 0.Clearly this based on repeat rigid method for measuring similarity have ignored Potential semantic dependency between different terms so that semantic relation tolerance is inaccurate, tie by the Chinese News Stories segmentation obtaining Really inaccurate.It is thus desirable to proposing a kind of more rational Semantic Similarity metric form to help the efficiency improving segmentation and essence Degree.

Content of the invention

The invention provides a kind of Chinese News Stories dividing method based on flexible Semantic Similarity tolerance, energy of the present invention Enough rational Semantic Similarity representing between word, and the precision of Chinese News Stories cutting techniques can be significantly improved, Described below:

A kind of Chinese News Stories dividing method based on flexible Semantic Similarity tolerance, methods described includes following step Rapid:

(1) target collected works are inputtedTo each News Stories script t in collected works_iCarry out participle；

(2) set up context relation figure；

(3) by described context relation figure and quick sorting algorithm, the context dependence between word is iterated Propagate and obtain flexible semantic dependency matrix；

(4) it is defined by described flexibility flexible Semantic Similarity between sentence for the semantic dependency matrix；

(5) using described flexibility Semantic Similarity, Chinese News Stories are split.

The described step setting up context relation figure particularly as follows:

1) read in each News Stories script successively, word frequency statisticses are carried out to the word being comprised；

2) according to the word frequency threshold value defining, frequent words and low frequency word are deleted；

3) using the word retaining as context relation in figure node, its set be v；

4) whether any two word in judging to gather simultaneously appears in a certain News Stories script, and this two words The distance between language is less than or equal to distance threshold, if it is sets up side between this two words, the set on side is e； Rejudge other any two words if not, until the word in whole set is all traversed；

5) the weights s on side_cBy the weights sim between word_cThe weights sim of (a, b), word itself_c(a a) represents；

6) described context diagram is shown as g=v, e, s_c.

Weights sim between described word_c(a, b) particularly as follows:

${\mathrm{sim}}_c(a,b)=\frac{\mathrm{freq}(a,b)}{{\mathrm{freq}}_{\max }+\mathrm{\ϵ}}$

Wherein, freq (a, b) represents the number of times that word a and word b occurs simultaneously, freq_max=max(_i,j){freq(i, J) } represent the frequency maxima to (i, j) for the word, ε is a constant in order to guarantee 0≤sim_c(a,b)≤1.

The weights sim of described word itself_c(a,a)=1.

Described by described context relation figure and quick sorting algorithm, the context dependence between word is changed Generation propagate the step obtaining flexible semantic dependency matrix particularly as follows:

1) defining the Semantic Similarity between context relation in figure word is sim_s(a, b), following three criterions of satisfaction:

Word is 1 with the similitude of itself, i.e. sim_s(a,a)=1；sim_s(a, b) and sim_c(a, b) positive correlation；sim_s Similitude between (a, b) and their neighbours is directly proportional；

2) the iterative diffusion process of definition Semantic Similarity:

${\mathrm{sim}}_s^{\left(0\right)}(a,b)={\mathrm{sim}}_c(a,b);$ ${\mathrm{sim}}_s^{\left(t\right)}(a,b)=\frac{c}{z}\underset{u~a,v~b}{\mathrm{\σ}}\mathrm{si}{m}_s^{(t-1)}(a,b);$ ${\mathrm{sim}}_s(a,b)=\underset{t\→\∞}{\lim }{\mathrm{sim}}_s^{\left(t\right)}(a,b);$

Wherein, u～a, v～b represent that u and v is word a and the neighbor node of word b in context relation in figure respectively, z It is normalization factor, c is controlling elements,Represent the Semantic Similarity of word a and word b during the t time iteration,Represent the Semantic Similarity of word a and word b during the t-1 time iteration,Represent initialization；

3) use quick sorting algorithm solve 2) defined in relational expression, obtain semantic dependency, to each two word all Ask for semantic dependency, several semantic dependencies constitute flexible semantic dependency matrix, this correlation matrix is designated as s_s.

The described step being defined by described flexibility flexible Semantic Similarity between sentence for the semantic dependency matrix Particularly as follows:

$\mathrm{sim}(s_i,s_j|s_s)=\frac{{{f_i}^{t}s}_s{f}_i}{\left|\right|f_i\left|\right|\left|\right|f_j\left|\right|}$

Wherein s_iAnd s_jRepresent sentence respectively, | | f_i| | and | | f_j| | represent two norms of two sentence word frequency vectors respectively, T is transposition.

The beneficial effect of the technical scheme that the present invention provides is: the present invention is proposed a kind of non-supervisory by quick sorting algorithm The Semantic Similarity measure of formula, improves potential between word to enable to incorporate to traditional cosine similarity Semantic relation, and improve Chinese News Stories cutting techniques using this flexible Semantic Similarity.Flexible tolerance proposed by the present invention Method can more reasonably represent the Semantic Similarity between word and between set of words.Test result indicate that, in In civilian News Stories cutting techniques, criterion is split based on identical, compared with traditional method for measuring similarity, using this flexibility Segmentation precision can be brought up to 3%-10% by Semantic Similarity measure.

Brief description

Fig. 1 is the flow chart of the Chinese News Stories cutting techniques based on flexible Semantic Similarity；

Fig. 2 is the schematic diagram of context relation figure；

Fig. 3 is between story on standard data set cctv and tdt2 and the contrast of the internal sentence similitude ratio of story Figure；

Fig. 4 is that Chinese News Stories partitioning algorithm uses on 100 groups of random parameters on standard data set cctv-75-s The Comparative result figure of three kinds of different similarity measurement modes.

Specific embodiment

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with accompanying drawing to embodiment party of the present invention Formula is described in further detail.

The tolerance of Semantic Similarity is an extremely challenging research topic in natural language processing.Existing method master It is divided into two classes: supervised and non-supervisory formula.The method of supervised mainly includes wordnet^[8][9]And disco.Wordnet uses Similitude between tolerance any two English word.The collected works that wordnet dependence has identified, by ranking, verb, adjective Carry out distinguishing hierarchy with adverbial word, the foundation of division is the semantical definition to these words for the language specialist.Terseness due to wordnet And validity, wordnet has been widely applied in natural language processing task.Similar with wordnet, disco is as another A kind of conventional supervised method, for retrieving the similitude between any given two word.Compared with wordnet, disco Support more rich languages, for example: English, German, French, Spanish etc..The method of supervised can be used directly to carry Before the language space that defines it is not necessary to any extra calculating, meanwhile, the method for supervised also almost covers whole normal Word.But, the method for supervised depends on the knowledge of linguist, and between word, the tolerance of similitude is generally by subjective consciousness Defined, meanwhile, the method for supervised is not suitable for the application based on specific collected works.The method of non-supervisory formula mainly include pmi, Lsa and plsa.The statistics that pmi is got by query web search engine, two words of statistics simultaneously appear in same Number of times in webpage, number of times is more, then the pmi score of this two words is higher.Lsa is also a kind of semantic phase of non-supervisory formula Like property measure, the mechanism that it has incorporated mankind's learning knowledge goes to obtain the similitude between word or text fragment.lsa Committed step be that dimensionality reduction is carried out by singular value decomposition.Lsa can also solve synonymy in natural language processing simultaneously Problem.Plsa is to lsa innovatory algorithm.Different from the lsa algorithm coming from linear algebra, in the base inheriting lsa advantage On plinth, plsa is analyzed to the correlation between paired word using the method for probability theory, and can be good at processing Synonymy and the problem of ambiguity.Compared with lsa, plsa has more generality.

In recent years, the development of graph theory causes the attention of natural language processing scholars, and widdows et al. proposes one kind It is used for obtaining Semantic Similarity based on the non-supervisory formula method of graph model.In this graph model, node represents word, while being word Relation between language.Additionally, this graph model is based on specific collected works, the ambiguousness of word can be processed.Ambwani et al. proposes The graph model of another kind of tolerance phrase semantic similitude, each word is represented as a series of node, and each node corresponds to One sentence in this word coverage, the weights on side represent the correlation between word.This model is by between word Influence each other and incorporate wherein, to determine the correlation between word according to context simultaneously.Semantic Similarity discussed above The non-supervisory formula method calculating is all based on specific collected works, is more suitable for specifically applying compared with supervised method.Non- at these Inside the method for supervised, because terseness and the high efficiency of graph model is so that Semantic Similarity calculating side based on graph model Method causes the concern of more and more natural language processing researchers.

Semantic Similarity tolerance between set of words (such as paragraph, text etc.) is also a problem demanding prompt solution. The conventional method of set of words Semantic Similarity tolerance is cosine similarity.Based on bag of words it is assumed that each set of words is expressed Vectorial for word frequency, cosine similarity is used for measuring the angle between word frequency vector, and angle is bigger, and similitude is less；Conversely, more Greatly.Because the use very simple of cosine similarity is with effectively, therefore cosine similarity is widely used and set of words is semantic The tolerance of similitude, but cosine similarity only considered the relation between identical word, have ignored in set of words word it Between correlation, this make set of words similitude tolerance inaccurate.In order that the tolerance of set of words Semantic Similarity More accurately and meaningful, when measuring similarity between set of words, the correlation between word is taken into account more For meaningful.Therefore, need proposition badly and a kind of correlation between word is dissolved into similarity measurement between set of words Method.

In order to the rational Semantic Similarity representing between word, and Chinese News Stories can be significantly improved divide Cut the precision of technology, embodiments provide a kind of Chinese News Stories segmentation side based on flexible Semantic Similarity tolerance Method, referring to Fig. 1, this method, in carrying out Semantic Similarity calculating and News Stories cutting procedure, is all for a certain specific Data set.Meanwhile, in order to embody the reasonability of flexible Semantic Similarity measure, redesigned checking criterion reasonable to this Property is verified, described below:

101: input target collected worksParticiple is carried out to each News Stories script ti in collected works；

Pass through this step and every words in News Stories script are split several words, this step is art technology Well known to personnel, the embodiment of the present invention does not repeat to this.

102: set up context relation figure；

1) read in News Stories script successively, word frequency statisticses are carried out to the word included in it；

2) according to the word frequency threshold value defining, frequent words and low frequency word are deleted.

3) using the word retaining as context relation in figure node, its set be v.

4) whether any two word in judging to gather simultaneously appears in a certain News Stories script, and this two words The distance between language is less than or equal to distance threshold, if it is sets up side between this two words, the set on side is e； Rejudge other any two words if not, until the word in whole set is all traversed.

5) the weights s on side_cBy the weights sim between word_cThe weights sim of (a, b), word itself_c(a a) represents；

Weights sim between word_c(a, b) is defined by below equation:

${\mathrm{sim}}_c(a,b)=\frac{\mathrm{freq}(a,b)}{{\mathrm{freq}}_{\max }+\mathrm{\ϵ}}$

Wherein freq (a, b) represents the number of times that word a and word b occurs simultaneously, freq_max=max(_i,j){freq(i,j)} Represent the frequency maxima to (i, j) for the word, ε is a constant in order to guarantee 0≤sim_c(a,b)≤1(a≠b).

Meanwhile, the weights sim of word itself_c(a,a)=1.

6) therefore context relation figure can be expressed as g=v, e, s_c, Fig. 2 is the upper and lower of a certain document in structure data set The schematic diagram of civilian graph of a relation, wherein, w_uRepresent t_iIn u-th word, lines represent the relation between word.

103: by context relation figure and quick sorting algorithm^[10][11]Context dependence between word is changed In generation, propagates and obtains flexible semantic dependency matrix；

1) defining the Semantic Similarity between context relation in figure word is sim_s(a, b), it meets following three standards Then:

Word is 1 with the similitude of itself, i.e. sim_s(a,a)=1；

sim_s(a, b) and sim_c(a, b) positive correlation；

sim_sSimilitude between (a, b) and their neighbours is directly proportional.

2) the iterative diffusion process defining Semantic Similarity is defined as follows by relationship below:

${\mathrm{sim}}_s^{\left(0\right)}(a,b)={\mathrm{sim}}_c(a,b);$

${\mathrm{sim}}_s^{\left(t\right)}(a,b)=\frac{c}{z}\underset{u~a,v~b}{\mathrm{\σ}}\mathrm{si}{m}_s^{(t-1)}(a,b);$

${\mathrm{sim}}_s(a,b)=\underset{t\→\∞}{\lim }{\mathrm{sim}}_s^{\left(t\right)}(a,b)$

Wherein, u～a, v～b represent that u and v is word a and the neighbor node of word b in context relation in figure respectively, z It is normalization factor, c is controlling elements,Represent the Semantic Similarity of word a and word b during the t time iteration,Represent the Semantic Similarity of word a and word b during the t-1 time iteration,Represent initialization.

3) use quick sorting algorithm solve 2) defined in relational expression, obtain semantic dependency, to each two word all Ask for semantic dependency, several semantic dependencies constitute flexible semantic dependency matrix, this correlation matrix is designated as s_s.With Similar, traditional rigid Semantic Similarity is defined as s_h=i, wherein i represent unit matrix.

The hypothesis that quick sorting algorithm is based on: if neighbours' word of two words more similar (correlation be more than or Equal to 0.5), then this two words are also more similar；

Quick sorting algorithm is using context graph of a relation as input；Algorithm complex is o (k | v |²), wherein, k is in g The average number of degrees, | v | represents the nodes of context relation in figure, and o is algorithm complex.

The full point based on gpu for the quick sorting algorithm is achieved to parallel algorithm in this invention.Found by Germicidal efficacy, with Same context relation figure as input, the speed of the quick sorting algorithm realized based on gpu than tradition based on cpu realize fast The speed of fast sort algorithm improves about 1000 times.

Quick sorting algorithm is output as being iterated the flexible semantic dependency matrix s after propagating_s={sim_s(a, b)}_a,b∈c；

104: by flexible semantic dependency matrix to sentence (the one section of word continuously occurring as in News Stories script Set) between flexible Semantic Similarity be defined；

In News Stories cutting techniques, it is also desirable to measure set of words i.e. sentence in addition to phrase semantic similitude Similitude between son.In story, each sentence can be identified as word frequency vector, be used for recording each word and go out in sentence Existing number of times.For given flexible semantic dependency matrix, the flexible Semantic Similarity between sentence is defined as follows:

$\mathrm{sim}(s_i,s_j|s_s)=\frac{{f_i}^t{s}_s{f}_i}{\left|\right|f_i\left|\right|\left|\right|f_j\left|\right|}$

Wherein s_iAnd s_jRepresent sentence respectively, | | f_i| | and | | f_j| | represent two norms of two sentence word frequency vectors respectively, T is transposition.This definition is the improvement to traditional cosine similarity, and semantic dependency potential between different terms is considered by it Enter, therefore, it is possible to more reasonably represent the Semantic Similarity between sentence.

105: using flexible Semantic Similarity, Chinese News Stories are split.

1) Chinese News Stories cutting techniques based on criterion be Normalization norm^[5][6]；

Wherein, Normalization norm is particularly as follows: this criterion is based on graph model；Sentence is identified as the node in graph model, sentence Relation between son is expressed as the side in graph model；Similitude between sentence is expressed as the weights on side；News Stories are split Problem is converted into graph model segmentation problem.

2) between using sentence, flexible Semantic Similarity carries out Chinese to the News Stories script included in input data set News Stories are split.

A kind of Chinese news event based on flexible Semantic Similarity tolerance with specific test, the present invention being provided below The feasibility of thing dividing method is verified:

Standard data set is tested:

For verifying the validity of this method, this method is tested on two standard data set cctv and tdt2. Cctv data set contains 71 Chinese news storyboards altogether, can say cctv according to News Stories length and identification error rate Data set is divided into 8 subsets, is designated as cctv_59_f/s, cctv_66_f/s, cctv_75_f/s and cctv_ref_f/s respectively, Wherein f represents long story set, and s represents short story set, and ref represents reference data set.Tdt2 data set comprises 177 Chinese news According to identification error rate, storyboard, can say that this 177 scripts are divided into two subsets, be designated as tdt2_ref and tdt2_ respectively rcg.Respectively using rigid Semantic Similarity s_h, context Semantic Similarity s_cWith flexible Semantic Similarity s_sTo cctv and tdt2 News Stories in data set are split, and compare the quality of its segmentation precision.Wherein segmentation precision scores to represent by f1. Table 1 lists using different segmentation precisions on cctv and tdt2 data set for the similarity measurement mode.

Table 1

It is observed that compared with traditional rigid Semantic Similarity, can be made using flexible Semantic Similarity from table 1 Segmentation precision is significantly improved, and increase rate is about 3% to 10%.Simultaneously it has also been found that context Semantic Similarity is than hard The Semantic Similarity of property is good, and context Semantic Similarity can be made to get a promotion by quick sorting algorithm.In order to show The robustness of this method, this method implements another more strict experiment on cctv_75_s data set, by than less Use the segmentation precision of 100 groups of random parameters with method.Fig. 4 shows this experimental result.In news therefore cutting techniques, sentence Between the quality of similitude can be measured with the ratio of the similitude between story and the similitude within story, this ratio pair Answer the ga s safety degree of sentence, wherein, ratio is more little then to prove that similitude is better, and this ratio to be defined by below equation:

$r\left(c\right|s_s)=\frac{\exp \left({\mathrm{mean}}_{\mathrm{lab}\left(s_i\right)\&notequal;\mathrm{lab}\left(s_j\right)}\mathrm{sim}(s_i,s_j|s_s)\right)}{\exp \left({\mathrm{mean}}_{\mathrm{lab}\left(s_i\right)=\mathrm{lan}\left(s_j\right)}\mathrm{sim}(s_i,s_j|s_s)\right)}$

Wherein lab (s_i) represent sentence s_iThe label of affiliated story, lab (s_j) represent sentence s_jThe label of affiliated story, Mean represents and averages.

This method is by rigid Semantic Similarity s_h, context Semantic Similarity s_cWith flexible Semantic Similarity s_sIn criterion numeral Contrasted according on collection, comparing result represents in the diagram.It is found through experiments, using flexible Semantic Similarity s_sObtained R ratio lower than other two kinds of similitudes, and context Semantic Similarity s_cS lower than rigid Semantic Similarity_h.Should Experiment shows flexible Semantic Similarity s_sThan traditional rigid Semantic Similarity s_hMore reasonable, and calculated by quicksort The flexible Semantic Similarity (i.e. Semantic Similarity after iterative diffusion) that method solves is more reasonable.During this method is applied to Segmentation precision can be made to increase significantly in civilian News Stories cutting techniques.

Bibliography:

[1].j.allan,ed.,topic detection and tracking:event-based information organization,kluwer academic publishers,2002.

[2].l.-s.lee and b.chen,“spoken document understanding and organization,”vol.22,no.5,pp.42–60,2005.

[3].s.banerjee and i.a.rudnicky,“a texttiling based approach to topic boundary detection in meetings,”in interspeech,2006.

[4].l.xie,j.zeng,and w.feng,“multi-scale texttiling for automatic story segmentation in chinese broadcast news,”in airs,2008.

[5].i.malioutov and r.barzilay,“minimum cut model for spoken lecture segmentation,”in acl,2006.

[6].j.zhang,l.xie,w.feng,and y.zhang,“a subword normalized cut approach to automatic story segmentation of chinese broadcast news,”in airs, 2009.

[7].z.liu,l.xie,and w.feng,“maximum lexical cohesion for fine-grained news story segmentation,”in interspeech,2010.

[8].t.pedersen,s.patwardhan,and j.michelizzi,“wordnet::similarity- measuring the relatedness of concepts,”in aaai(intelligent systems demonstration),2004.

[9].christiane fellbaum,ed.,wordnet:an electronic lexical database, mit press,1998.

[10].g.jeh and j.widom,“simrank:a measure of structural-context similarity,”in acm sigkdd,2002.

[11].g.he,h.feng,c.li,and h.chen,“parallel simrank computation on large graphs with iterative aggregation,”in acm sigkdd,2010.

It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention Sequence number is for illustration only, does not represent the quality of embodiment.

The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all spirit in the present invention and Within principle, any modification, equivalent substitution and improvement made etc., should be included within the scope of the present invention.

Claims (5)

1. a kind of Chinese News Stories dividing method based on flexible Semantic Similarity tolerance is it is characterised in that methods described bag Include following steps:

(1) target collected works are inputtedTo each News Stories script t in collected works_iCarry out participle；

(2) set up context relation figure；

(3) context dependence between word is iterated propagate by described context relation figure and quick sorting algorithm Obtain flexible semantic dependency matrix；

(4) it is defined by described flexibility flexible Semantic Similarity between sentence for the semantic dependency matrix；

(5) using described flexibility Semantic Similarity, Chinese News Stories are split；

The described step setting up context relation figure particularly as follows:

1) read in each News Stories script successively, word frequency statisticses are carried out to the word being comprised；

2) according to the word frequency threshold value defining, frequent words and low frequency word are deleted；

3) using the word retaining as context relation in figure node, its set be v；

4) judge gather in any two word whether simultaneously appear in a certain News Stories script, and this two words it Between distance be less than or equal to distance threshold, if it is between this two words, set up side, the set on side is e；If No rejudge other any two words, until whole gather in word be all traversed；

5) the weights s on side_cBy the weights sim between word_cThe weights sim of (a, b), word itself_c(a a) represents；

6) described context diagram is shown as g=< v, e, s_c>.

2. method according to claim 1 is it is characterised in that weights sim between described word_c(a, b) particularly as follows:

${\mathrm{sim}}_c(a,b)=\frac{freq(a,b)}{{\mathrm{freq}}_{max}+\mathrm{\&epsiv;}}$

Wherein, freq (a, b) represents the number of times that word a and word b occurs simultaneously, freq_max=max_(i,j){ freq (i, j) } table Show the frequency maxima to (i, j) for the word, ε is a constant in order to guarantee 0≤sim_c(a,b)≤1.

3. method according to claim 1 is it is characterised in that the weights sim of described word itself_c(a, a)=1.

4. method according to claim 1 is it is characterised in that described calculated by described context relation figure and quicksort Method the context dependence between word is iterated propagate the step obtaining flexible semantic dependency matrix particularly as follows:

1) defining the Semantic Similarity between context relation in figure word is sim_s(a, b), following three criterions of satisfaction:

Word is 1 with the similitude of itself, that is,

sim_s(a, a)=1；sim_s(a, b) and sim_c(a, b) positive correlation；sim_sSimilitude between (a, b) and their neighbours just becomes Than；

2) the iterative diffusion process of definition Semantic Similarity:

${\mathrm{sim}}_s^{\left(0\right)}(a,b)={\mathrm{sim}}_c(a,b);{\mathrm{sim}}_s^{\left(t\right)}(a,b)=\frac{c}{z}\underset{u~a,v~b}{\mathrm{\&sigma;}}{\mathrm{sim}}_s^{(t-1)}(a,b);{\mathrm{sim}}_s(a,b)=\underset{t\&rightarrow;\mathrm{\&infin;}}{\lim }{\mathrm{sim}}_s^{\left(t\right)}(a,b);$

Wherein, u～a, v～b represent that u and v is word a and the neighbor node of word b in context relation in figure respectively, and z is to return The one change factor, c is controlling elements,Represent the Semantic Similarity of word a and word b during the t time iteration,Represent the Semantic Similarity of word a and word b during the t-1 time iteration,Represent initialization；

3) use quick sorting algorithm solve 2) defined in relational expression, obtain semantic dependency, each two word is all asked for Semantic dependency, several semantic dependencies constitute flexible semantic dependency matrix, and this correlation matrix is designated as s_s.

5. method according to claim 1 it is characterised in that described by described flexibility semantic dependency matrix to sentence Between the step that is defined of flexible Semantic Similarity particularly as follows:

$sim(s_i,s_j|s_s)=\frac{f_i^t{s}_s{f}_j}{\left|\right|f_i\left|\right|\left|\right|f_j\left|\right|}$

Wherein s_iAnd s_jRepresent sentence respectively, | | f_i| | and | | f_j| | represent two norms of two sentence word frequency vectors respectively, t is to turn Put.