CN102945230A - Natural language knowledge acquisition method based on semantic matching driving - Google Patents

Abstract

The utility model discloses a natural language knowledge acquisition method based on semantic matching driving. The natural language knowledge acquisition method comprises the following steps: (1) defining a semantic model for natural language processing; (2) defining a representing method of lexical semantic; (3) defining semantic matching relations among the lexicons; (4) defining a processing method of statements; and (5) transforming analysis results into knowledge points. According to the natural language knowledge acquisition method based on the semantic matching driving, an analysis scheme better accordant with semantic logics can be selected as the final analysis result from multiple grammatical analysis schemes according to semantic matching values by adopting semantic matching information to cooperate with common syntax rules in a small quantity of natural languages. Through the adoption of the method, the natural language statement analysis can be carried out and knowledge contained in the statements of the natural languages can be acquired. Experiments indicate that the method has higher feasibility.

Description

The natural language knowledge acquisition method that a kind of semantic-based coupling drives

Technical field

The invention belongs to computing machine natural language understanding field, particularly the natural language knowledge acquisition method of a kind of semantic-based coupling driving.

Background technology

In knowledge was integrated, a large amount of knowledge was lain in the natural language statement, only had the automatic analysis of having realized the natural language statement, could effectively obtain the knowledge that contains in the statement.Therefore, natural language processing technique just becomes the integrated Key technology of knowledge.

Natural language processing technique mainly contains Rule-based method and based on two kinds of thinkings of method of statistics, but these two kinds of methods are not all utilized semantic information fully, are difficult to obtain high-quality result.Therefore, the researcher more and more payes attention to semantic effect, the method of natural language has appearred analyzing based on lexical semantic knowledge bases such as Wordnet, hownet, framenet, but the description granularity that comprises semantic information in these knowledge bases is too thick, there is not expression to the level of morpheme, also not accurate enough.These shortcomings cause and are difficult to utilize them to form statement disposal route and the model of a cover system.

Aspect the Sentence analysis of semantic-based, scholars have made research to a certain degree: Yao Tianshun and had studied the natural language understanding that semantic-based drives, but the description of lexical semantic is more coarse, and is also abundant not to the utilization of semantic information, and analytical approach is inadequate system also; The HPSG method is based on lexical information and drives, if but the information spinner of vocabulary is used for describing syntax rule, and very few also not accurate enough to the description of semanteme, incompatible with the characteristics of Chinese.Document (Tom OH, Janyce W, Exploiting Semantic Role Resources for Preposition Disambiguation[J] .Computational Linguistics, 2008,35 (2): 151-184.) studied the prepositional phrase disambiguation that the language material treebank is combined with framenet, document (Patwardhan S, Banerjee S, Pedersen T.Using measures of semantic relatedness for word sense disambiguation[C] .Proceedings of the 4th International Conference on Intelligent Text Processing and Computational Linguistics (CICLING-03), Mexico City, 2,003 241 – 257.) having studied the use semantic relation carries out the lexical semantic disambiguation.Although these researchs obtain certain achievement, also do not form a cover system land productivity carries out natural language processing with semantic information model and method.

Form the semantic model of a complete natural language processing, the semanteme that must represent comparatively exactly vocabulary, should may satisfy specific to the level special talent of morpheme the requirement of semantic model, and existing lexical semantic represents that mode is mostly less than this not being carried out deep research.Case grammar uses " lattice " to describe the Deep Semantics relation, but case grammar only is used to carry out grammatical analysis, seldom describes the semanteme of vocabulary with lattice.Mode is to describe one of action executing state key factor, the accurate semanteme of a lot of vocabulary has contained the mode of certain movement concept in itself in the natural language, and existing various lexical semantic represents that mode do not consider mode, therefore is difficult to describe exactly the semanteme of concept.Body can strictly represent Concept Semantic, and body normal operation description logic represents all concepts; But in the description logic semantic relation between the concept is carried out treating of equality, the essential characteristic of the semantic relation of action property concept do not carried out special semantic interpretation and processing.

Summary of the invention

For the deficiencies in the prior art, the object of the present invention is to provide and a kind ofly utilize semantic information to carry out natural language processing and obtain the natural language knowledge acquisition method that the semantic-based coupling of the knowledge that wherein contains drives.

Technical scheme of the present invention is achieved in that for can the automatic analysis natural language and obtain the middle knowledge that contains proposed the semantic model of a natural language processing, this model is judged best grammatical analysis scheme with a hypothesis axiom.In order to satisfy the semantic expressiveness demand of this model, description logic has been carried out becoming privileged processing, enable to represent more suitably that natural language vocabulary is semantic.Based on this model, a kind of natural language analysis method that drives based on the lexical semantic coupling has been proposed; The method is utilized semantic matches information, in conjunction with common syntax rule in a small amount of natural language, according to the semantic matches value, can be in multiple grammatical analytical plan, select a kind of analytical plan that relatively meets semantic logic as final analysis result, can carry out the natural language Sentence analysis and obtain the knowledge that wherein contains by the method.The present invention is based on the natural language knowledge acquisition method that semantic matches drives, comprise following process:

(1) semantic model of definition natural language processing: comprise the semantic matches relation between definition vocabulary; Target modified in the semanteme of definition vocabulary; And the semantic matches value of definition statement;

(2) method for expressing of definition lexical semantic;

(3) the semantic matches relation between definition vocabulary: comprise the semantic matches relation between noun-noun; Semantic matches relation between noun-verb; Noun-adjectival semantic matches relation; Semantic matching relationship arranged side by side; The semantic matches relation of adverbial word; The semantic matches relation of measure word; The semantic matches relation of pronoun; And the grammer coupling is closed;

(4) definition statement disposal route: three levels and the syntax thereof that relate to semantic structure; Comprise the thinking of obtaining best grammatical analysis scheme, obtain best grammatical analysis scheme, and several parts such as best semantic matches value of simple clause;

(5) analysis result is converted into the knowledge point.

The invention has the beneficial effects as follows: the semantic model that the present invention proposes a natural language processing.In this model, the concept of semantic matches has been proposed, judge best grammatical analysis scheme with a hypothesis axiom.Lexical semantic represents mode owing to use this model must have comparatively accurately, and the present invention processes having done based on the semantic expressiveness mode of description logic to become privileged, and its characteristics are:

(1) with " lattice " and " mode " as basic semantic relation;

(2) carry out modification, limited and given special semantic interpretation to description logic, made it to be convenient to represent the lexical semantic of natural language;

Based on this model, the present invention proposes the method for the natural language processing of a kind of semantic-based coupling driving, the method has following characteristics:

(1) the bottom-up simple clause who has proposed the semantic-based coupling sums up method;

(2) purpose of Sentence analysis and result obtain tacit knowledge in the statement.

Description of drawings

Fig. 1 is the vocabulary definition that the present invention is based on semantic-based relation in the natural language knowledge acquisition method that semantic matches drives;

Fig. 2 the present invention is based on grammar rule synoptic diagram in the natural language knowledge acquisition method that semantic matches drives.

Embodiment

The specific implementation process that the present invention is based on the natural language knowledge acquisition method that semantic matches drives is as follows:

Step 1:

(1a) the semantic matches relation between definition vocabulary:

Definition 1: in the lexical semantic knowledge base, any two notional word W _XAnd W _YBetween the inherent semantic relation that has, be called the semantic matches relation.With function match (W _X, W _Y) represent that its level of intimate, the value of function are exactly the semantic matches value.The semantic matches relation is irrelevant with concrete statement.If W _XWith W _YBetween do not have semantic matches relation, match (W then is set _X, W _Y)=MAX, MAX are large constants.

(1b) definition 2: any notional word W in the statement _i(removing predicate head) all semanteme modifies in another one notional word W _Gi, claim W _GiW _iThe semantic target of modifying.

(1c) definition 3: in the situation of specific grammatical analysis option A i, suppose that V is predicate head, S is the actor of V, and O is the receptor that holds of V, Wi be in the statement notional word and! (W _i∈ S, V, O}), W _GiW _iSemanteme modify target, with function match (W _i, W _Gi) represent its semantic matches value, so, the semantic matches value Value of whole statement _AiCan represent with formula (1):

${\mathrm{Vaule}}_{\mathrm{Ai}}=K_{\mathrm{SVO}}*(\mathrm{match}(S,V)+\mathrm{match}(O,v))+K_W\underset{i=1}{\overset{n}{\mathrm{\Σ}}}*\mathrm{match}(W_i,W_{\mathrm{Gi}})---\left(1\right)$

Target modified in the semanteme of S and O is V, and n is the number (not comprising S, V, O) of notional word, K _SVOAnd K _WiBe the weights coefficient.The less expression semantic matches of the value of it should be noted that degree is larger.

(1d) hypothesis axiom 1 (best grammatical analysis axiom): suppose that a statement has m kind grammatical analysis scheme, meets the grammatical analysis option A of semantic logic most _iSatisfy condition: A _i=argmin (Value _Ai), namely the grammatical analysis scheme of semantic matches value minimum is best grammatical analysis scheme.

Step 2:

(2a) definition 4 (key concepts): that set before making up semantic base, that need not to carry out semantical definition and carry out the limited assemble of symbol that special semantic interpretation is processed by system during in Sentence analysis and reasoning, key concept quantity is few, uses C _WB={ W _B1, W _B2W _BkExpression, give tacit consent to each key concept W _BiInherit in the root concept.

(2b) definition 5 (lattice): the special key concept set of the Deep Semantics relation of correlation circumstance between expression action genus and other things, use C _VC={ C ₁, C ₂C _mExpression.For example " reason " is lattice of an expression action reason.

(2c) definition 6 (mode): the special key concept set of the executing state of expression action genus, use C _VM={ M ₁, M ₂M _nExpression.For example " finishing " is the mode that the expression action has been finished.

(2d) definition 7 (basic semantic relations): that before making up semantic base, set, unmodifiable and need to carry out the semantic relation that special semantic interpretation is processed during at Sentence analysis, comprise following semantic relation:

R _C(V, C, W) lattice semantic relation: have the Deep Semantics relation that lattice are C, wherein C ∈ C in order to expression action genus V and concept W _VC

R _M(V, M) modal semantic relation: the executing state in order to expression action genus V is M, wherein C ∈ C _VM

R _AP(W, P) attribute semantic relation: be the semantic relation of the attribute of noun genus W in order to represent concept P;

R _AS(W, S) state semantic relation: be the semantic relation of the state of noun genus W in order to represent concept S;

R _D(W) inherit semantic relation: can only use once in the semantical definition formula of concept, for example W ₁Inherit in W ₂

R _P(W ₁, W ₂) finite aggregate that part of semantic concerns is described: concept W is described ₂Semanteme be W ₁The part of semanteme;

R _W(W ₁, W ₂) finite aggregate of whole relation is described: concept W is described ₁Semantic subsume W ₂

R _VSThe actor of (V, W) expression action V is the semantic relation of W, R _VO(V, W) represents the semantic relation that the receptor is W that holds of action V.

(2e) definition 8 (expansion semantic relations): the semantic relation of the new definition of institute is gathered in making up the semantic base process, and quantity is not limit.When Sentence analysis, all expansion semantic relations have unified disposal route, do not carry out special processing.Use R _R(W ₁, W ₂) expression, illustrate that the concept W that is defined is W ₁And W ₂Between a kind of semantic relation.In non-definition, use R _R(W, W ₁, W ₂) expression W ₁And W ₂Between have the expansion semantic relation of W by name.

(2f) based on basic description logic, its concept definition is limited and conversion by regular 1-7, become the Concept Semantic define method of this paper.

Rule 1 (concept definition rule):

1) if W _B1, W _B2Key concept, R ₁, R ₂Semantic relation, new symbol then

$W=\⫬W_{B1}|\⫬R_1\left(W_{B1}\right)|\⫬W_{B1}|\⫬R_1\left(W_{B1}\right)|R_1\left(W_{B1}\right)\∪R_2\left(W_{B2}\right)|R_1\left(W_{B1}\right)\∩$

$R_2\left(W_{B2}\right)$ It is concept;

2) if W _B1, W _B2Concept, R ₁, R ₂Semantic relation, new symbol then

$W=\⫬W_{B1}|\⫬R_1\left(W_{B1}\right)|\⫬W_{B1}|\⫬R_1\left(W_{B1}\right)|R_1\left(W_{B1}\right)\∪R_2\left(W_{B2}\right)|R_1\left(W_{B1}\right)\∩$

$R_2\left(W_{B2}\right)$ It is concept;

Rule do not have in 1 concept also, concept hands over, the service regeulations of measure word, their processing mode is seen regular 2-2.Rule 2-3 is for the definition of conception mode being converted into one group of semantic relation, and requires noun to satisfy the principle of single succession.

Rule 2 (concept processing rule also): if W ₁, W ₂Concept, W ₁, W ₂The most recent co mmon ancestor concept be W _P, W has W=W when new ideas ₁∪ W ₂Semantic the time because single inheritance rules, W can be defined as W=R _D(W _P) ∩ (R _P(W _P, W ₁) ∪ R _P(W _P, W ₂)).For example, parents=R _D(people) ∩ (R _P(people, father) ∪ R _P(people, mother)).

Rule 3 (processing rule that concept is handed over): if W ₁, W ₂Be concept, W has W=W when new ideas ₁∩ W ₂Semantic the time because single inheritance rules, W can be defined as W=R _D(W ₁) ∩ R _W(W ₁, W ₂) or W=R _D(W ₂) ∩ R _W(W ₂, W ₁).

Rule 2 and rule 3 only with concept also, concept hands over a kind of expression mode of having changed, the certain semantic explanation that only need stipulate these two kinds expression modes gets final product, do like this is in order to solve a kind of flexible means of many succession issues and bonding inheritance principles, effect is similar to the interface among the java, with the retrieval of quickening concept and the speed of coupling.

In natural language, noun all can semantically directly or indirectly inherited in key concept, for the more clear semanteme that represents accurately noun, adopts single principle of inheriting when requiring semantic nouns to represent, for having the semantic concepts of inheriting, adopt rule 2 and rule 3 to process more.Transitive verb is representing that semantically noun does an action that applies to another noun; Intransitive verb then represents a kind of variation of noun self; Adjective is at the state or the attribute that semantically all represent between noun or noun; Adverbial word is at semantically all implementation status (mode) and the correlation circumstance (lattice) of expression action.The semanteme that therefore, can represent with the mode that satisfies rule 4 all kinds of vocabulary in the natural language.

Rule 4 (concept classification definition rules): concept is classified by character and is expressed as noun, verb, adjective, adverbial word in the natural language.Suppose that Num (R, W) is the occurrence number of semantic relation R in the definition with the definition of Def (W) expression concept W, every class vocabulary definition should be satisfied following rule:

The single of noun inherits: satisfy condition $(R_D\left(W_P\right)\⋐\mathrm{Def}\left(W\right))\∩(\mathrm{Num}(R_D,W)=1)$ Concept W.

Verb: satisfy condition $(R_{\mathrm{VS}}\left(W_1\right)\∪\left(R_{\mathrm{VO}}\left(W_2\right)\right)\⋐\mathrm{Def}\left(W\right)\∩(\mathrm{Num}(R_{\mathrm{VS}},W_1)=1)$

$\∩(\mathrm{Num}(R_{\mathrm{VO}},W_2)=1).$

Adjective: satisfy condition $(R_{\mathrm{AP}}(W,W_1)\∪R_{\mathrm{AS}}(W,W_1))\⋐\mathrm{Def}\left(W)\right)\∩$

$(\mathrm{Num}(R_{\mathrm{AP}},W)=1)\∩(\mathrm{Num}(R_{\mathrm{AS}},W)=1).$

Adverbial word: satisfy condition $(R_M(W_1,W)\⋐\mathrm{Def}\left(W\right))\∩(\mathrm{Num}(R_M,W)=1).$

Rule 5 (processing rules of measure word): measure word (

Generality quantifier and existential quantifier

) do not obtain special treating, be used as " number of times " number of times of value representation action of lattice semantic relation, the number of noun represented as the value of " quantity " attribute semantic relation.Because this does not affect the syntactic structure analysis of natural language, when reasoning, make a concrete analysis of according to " number of times " lattice of action and the value of noun " quantity " attribute.

Rule 6 (example arranges rule): when defined notion W, if the concept W in the definition _iOccur m time, and the n that occurs referring to for this m time semantic { S ₁, S ₂S _n, then available { W, W#1 ... W#n-1}, n of distinguishing W is semantic, and W#i may be interpreted as example when reasoning.

Rule 7 (polysemant is disposed rule): a lot of polysemants are arranged, if polysemant W has n semantic { S in the natural language ₁, S ₂S _n, then for concept of each concrete semantical definition, define altogether n concept { W@1, W@2 ... it is different semantic that W@n} distinguishes this n of expression.

(2g) hypothesis axiom 2: inherit semantic relation and have unidirectional delivery, lower floor's concept is inherited semantic relation that Upper Concept has.

Theorem 1.According to inheritance R _D, all nouns consist of one tree.

According to regular 1-7 and definition 4-8, can define the semanteme of vocabulary in the natural language.Suppose to represent semantic relation with a directed line segment that according to the definition of theorem 1 and vocabulary and rule as can be known, the semanteme of vocabulary W can be represented by one group of directed line segment in the noun tree.

Step 3:

(3a) the semantic matches relation between definition noun-noun:

9 (conjunctive word compiles) of definition: the set of all vocabulary that comprise in the nominal definition formula, use C _RWExpression.For example the conjunctive word of the noun W among accompanying drawing Fig. 1 compiles

C _RW={W，W _P，W _r1，W _r2，W _r3，W _r4，W _r5，W _v，W _vc}

In the analysis hereinafter with ∝ for representing to inherit semantic relation, W _X∝ W _YExpression W _XInherit in W _Y, and regulation W ∝ W.

(1) basic semantic matching relationship

Definition 10 (direct semantics matching relationships): if vocabulary W _X, W _YSatisfy following condition, use symbol

Expression:

Condition: suppose W _YConjunctive word to compile be C _WY, then

$\∃W_Z\∩(W_Z\∈C_{\mathrm{WY}})\∩(W_X\∝W_Z).$

When

The time, match (W _X, W _Y)=K _T* d (W _X, W _Z).

K _TBe the matching relationship coefficient, the type that concerns R according to mating is set to different constants, and 1≤K is generally arranged _T≤ 3.

For example: { W among accompanying drawing Fig. 1 _Dr1, W _Dvc, W _Dr2, W _D2, W _Dr3, W _Dr4, W _Dr5In each vocabulary have the direct semantics matching relationship with W.

Definition 11 (inheriting the semantic matches relation): if vocabulary W _X, W _YSatisfy following condition, use symbol

Expression:

Condition:

When The time, match (W _X, W _Y)=match (W _X, W _Z)+d (W _Y, W _Z).

For example: { W among accompanying drawing Fig. 1 _Dr1, W _Dvc, W _Dr2, W _D2, W _Dr3, W _Dr4, W _Dr5And W _D1, W _D2Has the semantic matches of succession relation.

Definition semantic distance function d (W _X, W _Y): two vocabulary W that expression has inheritance _X, W _YBetween the succession number of times.

(2) comprise the semantic matches relation

Definition 12 (explicit semantic subsume relations): if vocabulary W _X, W _YSatisfy following condition, use symbol W _X⊙ W _YExpression.

Condition: have concept W _Z, satisfy condition

$\∃W_Z(R_W(W_Y,W_Z)\⋐\mathrm{Def}\left(W_Y\right))\∩(W_X\∝W_Z)$

Work as W _X⊙ W _YThe time, match (W is arranged _X, W _Y)=K _P* (d (W _X, W _Z).K _PFor comprising the matching relationship coefficient.

Definition 13 (implicit semantic subsume relation): if vocabulary W _X, W _YSatisfy following condition, use symbol W _XZero W _YExpression.

Condition: have concept W _Z, satisfy condition

$\∃W_Z(R_P(W_Z,W_X)\⋐\mathrm{Def}\left(W_X\right))\∩(W_Z\∝W_Y)$

Work as W _XZero W _YThe time, match (W _X, W _Y)=K _P* (d (W _Z, W _Y)).

Definition 14 (comprising the semantic matches relation): if vocabulary W _X, W _YSatisfy following condition, use symbol W _X◎ W _YExpression:

Condition:

Work as W _X◎ W _YThe time, match (W _X, W _Y)=

min{match(W _X,W _Z)+match(W _Z,W _Y),match(W _X,W _Y)}

Theorem 3: when vocabulary WX, WY satisfied WX ◎ WY, WY had the semantic relation of WX.

(3b) relation of the semantic matches between noun-verb

The semantic matches relation of noun-verb can be divided into two classes:

1) SVO semantic matches relation: the receptor that holds of actor or action may be done in noun

2) lattice semantic matches relation: noun and verb have lattice semantic matches relation

Suppose that verb is V, the gerund of executing in the definition of V is S ₀, being subjected to gerund is O ₀Since when definition with S ₀Be set to implement the top noun of V, O ₀Be set to bear the top noun of this action, so only have same S ₀Or O ₀Having the noun S of certain relation and noun O just might execution action V, namely consists of the semantic matches of SVO.The SVO semantic matches has 6 kinds of situations, and it is worth available Value _SVOExpression, computing formula is as follows:

Value _SVO=match(S，S ₀)+match(O，O ₀)

Definition 15 (conventional SVO semantic matches relation): (S ∝ S satisfies condition ₀) ∩ (O ∝ O ₀).

Definition 16 (heavily loaded SVO semantic matches relation): satisfy condition:

$((S\∝S_0)\∩(\∃R_{\mathrm{VO}}(R_{\mathrm{VO}}(V,W)\⋐\mathrm{Def}\left(S\right))\∩(O\∝W))\∪((O\∝O_0)\∩$

$(\∃R_{\mathrm{VS}}(R_{\mathrm{VS}}(V,W)\⋐\mathrm{Def}\left(O\right))\∩(S\∝W))$

For noun S and O and verb V, when not satisfying SVO when coupling in the definition of V, and the defined declaration of S, O they satisfy the SVO coupling.

Example: ring=R _D(ornaments) ∩ R _VS(wearing the people) ∩ R _VO(wearing ornaments) ∩ R _C(wearing position, hand) is owing to comprised R in " ring " _VS(wearing the people) is so { people wears, ring } consists of heavily loaded SVO semantic matches relation.

Definition 17 (comprising SVO semantic matches relation): ((S ◎ S satisfies condition ₀) ∩ (O ∝ O ₀)) ∪ ((S ∝ S ₀) ∩ (O ◎ O ₀)).

Example: class=R _D(set) ∩ R _W(set, student), because " student " can " eat " " meal ", " student " is the part of " class ", so { class eats, meal } consists of whole SVO semantic matches relation.

Definition 18 (similar SVO semantic matches relation): ((S ∽ S satisfies condition ₀) ∩ (O ∝ O ₀)) ∪ ((S ∝ S ₀) ∩ (O ∽ O ₀)).

Definition 19 (metaphor SVO semantic matches relation): under the following conditions, may have metaphor in the conjecture statement:

Condition 1: in whole statement, can satisfy front four kinds of SVO coupling without any noun.

A condition 2: have noun S or O in the statement, satisfy! (S ∝ S ₀) ∩ (O ∝ O ₀), conjecture is likened S to S ₀

Or condition 3: have noun S or O in the statement, satisfy (S ∝ S ₀) ∩! (O ∝ O ₀), conjecture is likened O to O ₀

For metaphor SVO semantic matches relation, Value _SVO=K _F* (match (S, W _P)+match (O, W _P))

K _FBe weights coefficient, W _PS and S ₀Most recent co mmon ancestor.Because belong to conjecture character, K _FValue should be larger, to avert evil influence.

Definition 20 (lattice semantic matches relations): for noun W and verb V, satisfy $\∃R_C(V,C,$ $W_C)\⋐\mathrm{Def}\left(V\right)\∩(W\∝W_C).$

Match (W, V)=K _C* d (W, W _C), K _CBe the weights coefficient.

(3c) noun-adjectival semantic matches relation

For adjective W _VAWith noun W _N, satisfy

$\∃W((R_{\mathrm{AS}}(W,W_1)\∪\left(R_{\mathrm{AP}}(W,W_1)\right)\⋐\mathrm{Def}\left(W_{\mathrm{VA}}\right)\∩(W_N\∝W)),$

Match (W _VA, W _N)=K _A* d (W _N, W), K _ABe weights coefficient (general K _A=1).

(3d) semantic matching relationship arranged side by side

Semantic matching relationship arranged side by side only is used for the judgement of statement parallel construction, to determine the scope of conjunction.

Definition 21 (semantic similar): because nominal definition has adopted the method for single succession, two noun W _X, W _YAlthough in definition, do not have inheritance, at W semantically _XBut may be W _YA kind of, the concept that is equivalent in the description logic contains, and uses symbol W _X∽ W _YExpression.Can the Tableau algorithm in the description logic be improved, to judge the Concept Semantic similarity relation.

Definition 22 (noun semantic matching relationships arranged side by side): for two noun W _X, W _YAvailable match (W _X, W _Y)=K _T* (d (W _X, W _E)+d (W _Y, W _E)) calculate a numerical value, as heuristic information, W _EW _X, W _YNearest common ancestor's node.When satisfying W _X∽ W _YThe time, also may be coordination.

Definition 23 (verb semantic matching relationships arranged side by side): for two verb V _X, V _YAvailable match (W _X, W _Y)=K _T* (d (S _X0, S _Y0)+d (O _X0, O _Y0)) calculate a numerical value, as heuristic information, { S _X0, S _Y0, O _X0, O _Y0W _X, W _YActor in the definition and be subjected to the person.

(3e) relation of the semantic matches between other class vocabulary

The semantic matches of adverbial word relation: about adverbs modify adjective and adverbial word, very complicated situation is arranged also, this paper wouldn't discuss, suppose adverbial word can semantic matches in verb, adjective and adverbial word, stipulate match (W ₁, W ₂)=0.The semantic matches relation of measure word: lexicon should be preserved the incidence relation of measure word and noun.If measure word W can modification noun W _N, then stipulate match (W, W _N)=0; Otherwise match (W, W _N)=MAX.The semantic matches relation of pronoun: according to the relation that refers to of pronoun, pronoun is replaced to corresponding noun process, such as " I " am processed by " people ".

(3f) grammer matching relationship

Pay special attention to: the various semantic matches relations of front are inherent, and are irrelevant with concrete statement.In concrete statement, the vocabulary of possible some type is modified mutually, but do not have inherent semantic relation between the vocabulary itself, a kind of grammatical phenomenon (being phraseological modified relationship) of semantic modified relationship just may be arranged in this statement, mainly comprise following two kinds of situations:

(1) modified relationship between uncommon part of speech: between verb-verb; Between adverbial word-noun; Between adjective-verb etc.Such as " like swimming " " to be frank " etc.; These all belong to the grammer matching relationship, do not have inherent semantic matches relation between the vocabulary itself, just have phraseological modified relationship in statement.Its semantic matches is worth available match (W in the Sentence analysis process _X, W _Y)=MAX/K _GCalculate K _GType weights (K generally speaking _G=1, or satisfy K _G＜1.5).

(2) the flexible conversion of parts of speech, for example adjective often can be applied flexibly and is adverbial word, and this situation this paper does not consider.

Step 4:

(4a) three levels and the syntax thereof of definition semantic structure

To carry out Sentence analysis according to the semantic model of this paper, must have the statement abstract representation method of suitable semantic model.Any statement is all formed through iteration by statement relatively simple for structure, and phrase is seen as an ingredient in the statement.In order to satisfy the semantic analysis needs of semantic model, the semantic structure of statement can be divided into three levels according to complexity and the characteristics of semantic structure: simple sentence, special simple sentence, complex sentence.

Definition 24 (simple sentences): only have a verb or adjective to make the statement C of predicate _S, available grammar G ₁Come abstractdesription.

Thought design grammar G with case grammar ₁, mentality of designing: suppose that V is predicate, S is the actor of V; O is the receptor that holds of V, A _BIt is preposition attribute; A _AIt is postpositive attributive; P _DBe the adverbial modifier or complement, be equivalent to one group of lattice in the case grammar; P _CIt is one lattice content; N is noun; N _PBe noun phrase.

Grammar G ₁In fuzzy rules more (detailed fuzzy rules is more, slightly), mentality of designing of its crucial rule is as follows:

1) C _S→ P _DA _BSA _AP _DVP _DA _BOA _AP _D(appearance of SVO sequentially has 10 kinds, and accompanying drawing Fig. 2 is wherein a kind of)

2) S → n|SA _AA _BS (actor done in a plurality of vocabulary, such as the S among Fig. 2)

3)P _D→P _C|P _DP _C

S, O, A _B, A _A, P _CIn the service regeulations of the vocabulary such as preposition, conjunction, auxiliary word, number, measure word can write out easily.

The concrete expression mode of the G1 of the syntax is as follows:

Definition 25 (special simple sentences): have a plurality of verbs or adjective, but semantically do not comprising the statement of subordinate clause, available grammar G ₂Come abstractdesription.

Grammar G ₂Mentality of designing: can not produce on the basis of subordinate clause, to grammar G guaranteeing ₁The few rule of middle interpolation can generating grammar G ₂, mainly contain following 2 kinds of situations:

1) situation of predicate made in a plurality of verbs or adjective

2) S, O, A made in verb or adjective _B, A _A, P _CSituation

Grammar G ₂Key be verb phrase V _VNoun phrase N can not directly be followed in front and back _P, namely N can not appear _P+ V _VOr V _V+ N _P

The concrete expression mode of the G2 of the syntax is as follows:

Definition 26 (complex sentences): in grammar G ₂The regular N of middle interpolation _P→ C _S, form grammar G ₃Because regular N _P→ C _SIllustrate that a simple sentence or special simple sentence can make in the complex sentence arbitrarily composition, realized the simple sentence recurrence, so grammar G ₃Complex sentence can be described.

(4b) obtain the thinking of best grammatical analysis scheme

(1) lexical ambiguity digestion procedure

Suppose W ₁W ₂W _kThe lexical semantic number be respectively n ₁, n ₂N _k, entirely arranging for each semanteme, the result is { L ₁, L ₂L _M, M=n then _1*n _{2* *}n _k, suppose one of them

W _mN the meaning of a word, L then _iC _SA sequence of words without lexical ambiguity.Exhaustive each { L in the grammatical analysis process ₁, L ₂L _MAnalysis result, select best L _iJust can clear up lexical ambiguity.

(2) analyze thinking

According to axiom 1, obtain all grammatical analysis schemes, for each grammatical analysis option A _i, calculate A according to formula 1 _iCorresponding semantic matches value, and select best grammatical analysis scheme.

Definition 27 (simple clause): satisfy grammar G in the statement ₁Or G ₂Substring be simple clause.

Suppose axiom 4 (the semantic target axiom of modifying): the semanteme modification target of supposing notional word W is W _Gi, then for the simple clause C that meets semantic logic in the statement _S, satisfy (W ∈ C _S) → (W _Gi∈ C _S).For attribute A _B(supposing next-door neighbour S) satisfies (W ∈ A _B) → (W _Gi∈ (A _B∪ S)), other situation of attribute is similar.For the adverbial modifier or complement (P _D), satisfy (W ∈ P _D) → (W _Gi∈ (P _D∪ V)).

According to semantic characteristics of modifying target, all grammatical analysis schemes can be divided into 2 layers:

1) simple clause's level grammatical analysis scheme;

2) the grammatical analysis scheme of simple clause inside.

(4c) obtain best grammatical analysis scheme

(1) can sum up simple clause's Rule of judgment

For statement C _S, carry out grammar G ₁, G ₂, G ₃The CYK Algorithm Analysis, satisfy table 1 conditional substring s (i, j) be the simple clause that can sum up.

Table 1 can be summed up simple clause's Rule of judgment

(2) bottom-up simple clause sums up method

Available bottom-up simple clause's end method is asked for best subordinate clause level grammatical analysis scheme, sees algorithm 4:

Algorithm 1 (simple clause sums up method):

5) for statement C _S, according to the Rule of judgment of table 1, find out and to sum up the corresponding substring set of simple clause { s ₁, s ₂S _m;

6) for each clause s _i, calculate (algorithm 2) simple clause s _iBest semantic matches value, with s _iBe summed up as N _P, N is set _PEnd semantic;

7) make C _SEqual to sum up the result, with the simple clause s in the recursive procedure _iThe summation of best semantic matches value, carry out the recurrence of step 1-3;

8) have best full sentence semantic matches and be worth corresponding simple clause s _iScope and end sequentially be best grammatical analysis scheme.

The key that the best semantic matches value of calculating simple clause is algorithm, concrete grammar are seen next step " simple clause's best semantic matches value ".

In algorithm, when selecting, simple clause adopted the method for exhaustion, can obtain theoretic best grammatical analysis scheme.But the calculated amount of this method is larger, is difficult for realizing.But when the quantity that can return the clause less than 4 the time, also can consider.In the time can returning clause's quantity too much, (the individual semantic matches degree of k＜m) preferably simple clause is carried out recursive search, to ask for suboptimum grammatical analysis scheme can only to select k.

(3) sum up semanteme

In algorithm 1, with simple clause C _SBe summed up as N _PAfter, N _PDo not have semanteme, can't carry out next step semantic matches and calculate, the mode of solution is as follows:

1) regulation is come N by end _PCan be matched with any vocabulary W, the semantic matches value is:

Match (N _P, W)=MAX/K _C(K is arranged generally _C1)

2) if N _PMake S or the O of new goal clause, then can be with N _PSemanteme be set to former C _SIn S or O.

(4d) obtain simple clause's best semantic matches value

Calculate simple clause's best semantic matches value, according to axiom 1 and axiom 4, there is multiple grammatical analysis scheme simple clause inside, must obtain all grammatical analysis schemes, for every kind of grammatical analysis scheme, the semanteme of its notional word is modified target and is determined, just can calculate semantic matches value under this grammatical analysis scheme according to formula 1, and the grammatical analysis scheme with minimum semantic matches value is exactly required analysis result.

The grammatical analysis scheme of simple clause inside can be divided into 3 layers: 1) SVO combination level; 2) A _A, P _D, A _BLevel; 3) A _A, P _D, A _BInner grammatical analysis scheme.Can select wherein best grammatical analysis scheme by algorithm 2.

Algorithm 2 (simple clause's best semantic matches value):

8) if simple clause is special simple sentence, find all methods that it is summed up as simple sentence

9) for every kind of end method, special simple sentence is summed up as simple sentence

10) for this simple sentence, find out all possible SVO combined method

11) for every kind of SVO combined method, with simple sentence C _SSegmentation if S or O are phrase, is then carried out algorithm 3; Suppose C _SBe divided into { L ₁, L ₂..L _n}

12) each segmentation L _iIn can comprise A at most _A, P _D, A _BThree partial contents are found out L _iIn all A _A, P _D, A _BDivision methods

13) for every kind of A _A, P _D, A _BDivision methods by the means that the syntax and semantics the matching analysis combines, determines that target modified in the semanteme of each notional word, makes A _A, P _D, A _BThe semantic matches value minimum

14) ask for the semantic matches value of full sentence, select the minimum corresponding analytic process of semantic matches value as the grammatical analysis scheme of the best

Suppose

For simple clause C _SCarry out grammar G ₁The operation result of CYK algorithm, expression can generate the grammar symbol collection of substring s (i, j).

(1) the grammatical analysis scheme of SVO combination level

In simple sentence, suppose noun W ₁And W ₂Satisfy SVO coupling, then { W with verb V ₁, V, W ₂It is a SVO combination.But S (O is similar) may be a phrase, has { W in sentence ₁, V, W ₃And { W ₂, V, W ₃The SVO coupling, and V, W ₃Not at W ₁And W ₂Centre, and W ₁And W ₂Middle substring s (m, n) satisfies:

The time, at W ₁+ s (m, n)+W ₃The phrase that forms is S, the S that in like manner can obtain more to grow or O.

Algorithm 3 (S or O segmentation):

3) obtain phrase S (or O), according to the described principle of this step, find out the institute that satisfies the SVO coupling among the S

Noun is arranged, be assumed to be { n ₁, n ₂..n _m}

4) according to { n ₁, n ₂..n _mPhrase S is divided into the m-1 section, according to regular S → n|SA _AA _BEach may not comprise A for empty segmentation to S as can be known _AA _B

(2) A _A, P _D, A _BThe division methods of level

Suppose segmentation L _iSubstring be s (m, n), then satisfy

P, q is grammatical A _A, P _D, A _BDivision methods, segmentation result is: A _A=s (m, p), P _D=s (p, q) A _B=s (q, n).

(3) A _A, P _D, A _BInner best grammatical analysis scheme

Theorem 2:A _A, P _D, A _BInner best grammatical analysis scheme is to work as A _A, P _D, A _BEach interior notional word has the best semantic corresponding grammatical analysis scheme under the target conditions of modifying.

Definition 28 (simple noun phrase): not comprising verb and adjectival noun phrase is exactly simple noun phrase.

Theorem 3: the attribute (A of simple sentence _AOr A _B) best grammatical analysis scheme be equivalent to the best grammatical analysis scheme of a simple noun phrase.

Proof: because simple sentence only comprises a verb or predicate V made in adjective, so A _A, A _BIn do not comprise verb and adjective.According to axiom 4, A _A, A _BThe grammatical analysis scheme be equivalent to simple noun phrase N _PThe grammatical analysis scheme, N _P∈ { (A _B+ S), (A _A+ S), (A _B+ O), (A _A+ O) }.

Simple noun phrase N _PDifferent grammatical analysis schemes only be subjected to the impact of { conjunction/preposition/auxiliary word/measure word }.The key of grammatical analysis is scope and their the end order of selected { conjunction/preposition/auxiliary word/measure word }.

Determining of A, scope: in { conjunction/preposition/auxiliary word/measure word }, suppose w _BFor preposition type, w _MFor preposition type, then its scope can be summed up as two kinds of forms 1) ..N _Bn..N _B1... w _M..N _A1..N _Am..; 2) W _B..N _Bn..N _B1... W _M..N _A1..N _Am...; { N wherein _B1, N _B2... N _BnThe noun of first half in the scope, { N _A1, N _A2... N _AmIt is the noun of first half in the scope.

According to the backward semantic custom of modifying of Chinese, can be at { N _A1, N _A2... N _AmIn find out grammatical and N _B1Noun N with best semantic matches value _AjAs boundary after the scope.The scope prezone of form 1 can be determined with similar method.

(2) determining of end order: { conjunction/preposition/auxiliary word/measure word } and scope thereof should be summed up by certain bar grammar rule, and the available method of exhaustion is obtained their the best end order.Generally speaking, after the simple sentence statement has carried out repeatedly segmentation, A _A, A _BIn the number n of { conjunction/preposition/auxiliary word/measure word } that comprise be generally less than 4, have the calculating feasibility.

Definition 29 (noun sequences): not having the simple noun phrase of { conjunction/preposition/auxiliary word/measure word } is noun sequence.

After { conjunction/preposition/auxiliary word/measure word } all summed up, simple noun phrase just had been known as a noun sequence, and may also there be one or two noun sequence in { conjunction/preposition/auxiliary word/measure word } scope inside in addition.Only noun affects semantic modified relationship in noun sequence, according to the backward semantic custom of modifying of Chinese, supposes that noun sequence is L _N=W ₁W ₂W _mThen by semantic definite L _NIn arbitrarily noun to modify target concrete grammar (approximation method) as follows:

Algorithm 4 (the best semantic target of modifying of noun sequence):

Set C is set _WBe sky, for L _NIn each noun W _iIf, match (W _i, W _m)＜MAX is with W _iBe added to C _W, do following operation:

1) supposes C _WElement be W by successively order ₁-W ₂-...-W _n(n〉1), then be done as follows: with L _NBe divided into the n+1 section, the semanteme modification target that they are set is W _m, and each section carried out recurrence.

2) work as C _WIn when only having a noun, carry out step 3-4.

3) the forward direction modified relationship is set: for any segmentation, if there is W _xW _X+1W _Y-1W _y, satisfy condition:

1. any W _X+1W _yBetween noun and W _yAfter the semantic matches value of noun be MAX; 2. match (W _y, W _x)＜MAX; W then is set _ySemanteme to modify target be W _xW then is set _X+1W _Y-1Between the modification target W of noun _y

4) if L _NIn also have noun W _yDo not modify target, it is W that its modification target then is set _Y+1

P _DAnalytical approach be similar to A _A, A _B, key is to carry out boundary line delimitation according to preposition, and the content in the preposition scope also is converted into a simple noun phrase.

(4) disposal route of special simple sentence

Sum up all non-predicate verb/adjectives, statement is converted into simple sentence, select best end scheme.Disposal route is as follows:

1) statement is carried out the CYK algorithm of grammar G 2, find verb (adjective) phrase that to do predicate, may have kinds of schemes.

2) for each scheme, sum up remaining verb (or adjective), choose the analytical plan of semantic matches value minimum.

Also need to arrange when summing up non-predicate verb (adjective) and sum up semanteme, such as " beautiful necklace ", summing up semanteme is " necklace ".

Step 5:

Can simple sentence be converted into a knowledge point according to the grammatical analysis result with best semantic matches value, each simple clause of complex sentence is converted into the knowledge point, whole complicated sentence is converted into one group of knowledge point.

Example: " moulding is that the green bronze ornaments of insect are welcome by Brazilian girl very to statement." knowledge five-tuple (seeing Table 2):

The more educated example of table 2 statement

After statement is converted into the knowledge point of depositing with the structural data form, just can carry out various Intelligent Information Processing to these knowledge datas easily.

Above-described embodiment only is for the invention example clearly is described, and is not the restriction to the invention embodiment.For those of ordinary skill in the field, can also make other changes in different forms on the basis of the above description.Here need not also can't give all embodiments exhaustive.And the apparent variation of being amplified out thus or change still are among the protection domain of the invention claim.

Claims (6)

1. semantic-based mates the natural language knowledge acquisition method that drives, and comprises following process:

(1) semantic model of definition natural language processing: comprise the semantic matches relation between definition vocabulary; Target modified in the semanteme of definition vocabulary; And the semantic matches value of definition statement;

(2) method for expressing of definition lexical semantic;

(3) relation of the semantic matches between vocabulary: comprise the semantic matches relation between noun-noun; Semantic matches relation between noun-verb; Noun-adjectival semantic matches relation; Semantic matching relationship arranged side by side; The semantic matches relation of adverbial word; The semantic matches relation of measure word; The semantic matches relation of pronoun; And the grammer coupling is closed;

(4) definition statement disposal route: three levels and the syntax thereof that relate to semantic structure; Comprise the thinking of obtaining best grammatical analysis scheme, obtain best grammatical analysis scheme, and several parts such as best semantic matches value of simple clause;

(5) analysis result is converted into the knowledge point.

2. semantic-based according to claim 1 mates the natural language knowledge acquisition method that drives, and wherein step (1) is carried out according to the following procedure:

(1a) the semantic matches relation between definition vocabulary:

Definition 1: in the lexical semantic knowledge base, any two notional word W _XAnd W _YBetween the inherent semantic relation that has, be called the semantic matches relation, wherein adverbial word is treated by notional word; With function match (W _X, W _Y) represent that its level of intimate, the value of function are exactly the semantic matches value; The semantic matches relation is irrelevant with concrete statement; If W _XWith W _YBetween do not have semantic matches relation, match (W then is set _X, W _Y)=MAX, MAX are large constants;

(1b) definition 2: remove any notional word W outside the predicate head in the statement _iAll semanteme is modified in another one notional word W _Gi, claim W _GiW _iThe semantic target of modifying;

(1c) definition 3: in arbitrarily grammatical analysis option A _iSituation under, suppose that V is predicate head, S is the actor of V, O is the receptor that holds of V, Wi be in the statement notional word and! (W _i∈ S, V, O}), W _GiW _iSemanteme modify target, with function match (W _i, W _Gi) represent its semantic matches value, so, the semantic matches value Value of whole statement _Ai, can represent with formula (1):

Target modified in the semanteme of S and O is V, and n does not comprise S, the number of the notional word of V and O, K _SVOAnd K _WiBe the weights coefficient, the less expression semantic matches of semantic matches value degree is larger;

(1d) hypothesis axiom 1---best grammatical analysis axiom: suppose that a statement has m kind grammatical analysis scheme, meets the grammatical analysis option A of semantic logic most _iSatisfy condition: A _i=argmin (Value _Ai), namely the grammatical analysis scheme of semantic matches value minimum is best grammatical analysis scheme.

3. semantic-based according to claim 1 mates the natural language knowledge acquisition method that drives, and wherein step (2) is carried out according to the following procedure:

(2a) definition 4---key concept: that set before making up semantic base, that need not to carry out semantical definition and carry out the limited assemble of symbol that special semantic interpretation is processed by system during in Sentence analysis and reasoning, key concept quantity is few, uses C _WB={ W _B1, W _B2W _BkExpression, give tacit consent to each key concept W _BiInherit in the root concept;

(2b) definition 5---lattice: the special key concept set of the Deep Semantics relation of correlation circumstance between expression action genus and other things, use C _VC={ C ₁, C ₂C _mExpression;

(2c) definition 6---mode: the special key concept set of the executing state of expression action genus, use C _VM={ M ₁, M ₂M _nExpression;

(2d) definition 7---basic semantic relation: that before making up semantic base, set, unmodifiable and need to carry out the semantic relation that special semantic interpretation is processed during at Sentence analysis, comprise following semantic relation:

R _C(V, C, W) lattice semantic relation: have the Deep Semantics relation that lattice are C, wherein C ∈ C in order to expression action genus V and concept W _VC

R _M(V, M) modal semantic relation: the executing state in order to expression action genus V is M, wherein C ∈ C _VM

R _AP(W, P) attribute semantic relation: be the semantic relation of the attribute of noun genus W in order to represent concept P;

R _AS(W, S) state semantic relation: be the semantic relation of the state of noun genus W in order to represent concept S;

R _D(W) inherit semantic relation: can only use once in the semantical definition formula of concept;

R _P(W ₁, W ₂) finite aggregate that part of semantic concerns is described: concept W is described ₂Semanteme be W ₁The part of semanteme;

R _W(W ₁, W ₂) finite aggregate of whole relation is described: concept W is described ₁Semantic subsume W ₂

R _VSThe actor of (V, W) expression action V is the semantic relation of W, R _VO(V, W) represents the semantic relation that the receptor is W that holds of action V;

(2e) definition 8---expansion semantic relation: the semantic relation of the new definition of institute is gathered in making up the semantic base process, and quantity is not limit; When Sentence analysis, all expansion semantic relations have unified disposal route, do not carry out special processing; Use R _R(W ₁, W ₂) expression, illustrate that the concept W that is defined is W ₁And W ₂Between a kind of semantic relation; In non-definition, use R _R(W, W ₁, W ₂) expression W ₁And W ₂Between have the expansion semantic relation of W by name;

(2f) based on basic description logic, its concept definition is limited and conversion to rule 7 by rule 1, become the Concept Semantic define method;

Rule 1---concept definition rule:

1) if W _B1, W _B2Key concept, R ₁, R ₂Semantic relation, new symbol then

It is concept;

2) if W _B1, W _B2Concept, R ₁, R ₂Semantic relation, new symbol then

It is concept;

Rule do not have in 1 concept also, concept hands over, the service regeulations of measure word, their processing mode is seen rule 2, rule 3, they are for the definition of conception mode being converted into one group of semantic relation, and require noun to satisfy the principle of single succession;

Rule 2---concept processing rule also: if W ₁, W ₂Concept, W ₁, W ₂The most recent co mmon ancestor concept be W _P, W has W=W when new ideas ₁∪ W ₂Semantic the time because single inheritance rules, W can be defined as W=R _D(W _P) ∩ (R _P(W _P, W ₁) ∪ R _P(W _P, W ₂));

The processing rule that rule 3---concept is handed over: if W ₁, W ₂Be concept, W has W=W when new ideas ₁∩ W ₂Semantic the time because single inheritance rules, W can be defined as W=R _D(W ₁) ∩ R _W(W ₁, W ₂) or W=R _D(W ₂) ∩ R _W(W ₂, W ₁);

Rule 4---the concept classification definition rule: concept is classified by character and is expressed as noun, verb, adjective and adverbial word in the natural language; Suppose that Num (R, W) is the occurrence number of semantic relation R in the definition with the definition of Def (W) expression concept W, every class vocabulary definition should be satisfied following rule:

The single of noun inherits: satisfy condition

Concept W;

Verb: satisfy condition

Adjective: satisfy condition

Adverbial word: satisfy condition

Rule 5---the processing rule of measure word: measure word does not obtain special treating, and is used as " number of times " number of times of value representation action of lattice semantic relation, represent the number of noun as the value of " quantity " attribute semantic relation;

Rule 6---example arranges rule: when defined notion W, if the concept W in the definition _iOccur m time, and the n that occurs referring to for this m time semantic { S ₁, S ₂S _n, then available { W, W#1 ... W#n-1}, n of distinguishing W is semantic, and W#i may be interpreted as example when reasoning;

Rule 7---polysemant is disposed rule: a lot of polysemants are arranged, if polysemant W has n semantic { S in the natural language ₁, S ₂S _n, then for concept of each concrete semantical definition, define altogether n concept { W@1, W@2 ... it is different semantic that W@n} distinguishes this n of expression;

(2g) hypothesis axiom 2: inherit semantic relation and have unidirectional delivery, lower floor's concept is inherited semantic relation that Upper Concept has;

Theorem 1: according to inheritance R _D, all nouns consist of one tree;

Arrive rule 7 and define 4 to defining 8 according to rule 1, can define the semanteme of vocabulary in the natural language; Suppose to represent semantic relation with a directed line segment that according to the definition of theorem 1 and vocabulary and rule as can be known, the semanteme of vocabulary W can be represented by one group of directed line segment in the noun tree.

4. semantic-based according to claim 1 mates the natural language knowledge acquisition method that drives, and wherein step (3) is carried out according to the following procedure:

(3a) the semantic matches relation between definition noun-noun:

Definition 9---conjunctive word compiles: the set of all vocabulary that comprise in the nominal definition formula, use C _RWExpression;

With ∝ for representing to inherit semantic relation, W _X∝ W _YExpression W _XInherit in W _Y, and regulation W ∝ W;

(1) basic semantic matching relationship

Definition 10---direct semantics matching relationship: if vocabulary W _X, W _YSatisfy following condition, use symbol Expression:

Condition: suppose W _YConjunctive word to compile be C _WY, then

When

The time, match (W _X, W _Y)=K _T* d (W _X, W _Z);

K _TBe the matching relationship coefficient, the type that concerns R according to mating is set to different constants, 1≤K _T≤ 3;

Definition 11---inherit the semantic matches relation: if vocabulary W _X, W _YSatisfy following condition, use symbol

Expression:

Condition:

When

The time, match (W _X, W _Y)=match (W _X, W _Z)+d (W _Y, W _Z);

Definition semantic distance function d (W _X, W _Y): two vocabulary W that expression has inheritance _X, W _YBetween the succession number of times;

(2) comprise the semantic matches relation

Definition 12---explicit semantic subsume relation: if vocabulary W _X, W _YSatisfy following condition, use symbol W _X⊙ W _YExpression;

Condition: have concept W _Z, satisfy condition

Work as W _X⊙ W _YThe time, match (W is arranged _X, W _Y)=K _P* (d (W _X, W _Z).K _PFor comprising the matching relationship coefficient;

Definition 13---implicit semantic subsume relation: if vocabulary W _X, W _YSatisfy following condition, use symbol W _XZero W _YExpression;

Condition: have concept W _Z, satisfy condition

Work as W _XZero W _YThe time, match (W _X, W _Y)=K _P* (d (W _Z, W _Y));

Definition 14---comprise the semantic matches relation: if vocabulary W _X, W _YSatisfy following condition, use symbol W _X◎ W _YExpression:

Condition:

Work as W _X◎ W _YThe time, match (W _X, W _Y)=min{match (W _X, W _Z)+match (W _Z, W _Y), match (W _X, W _Y)

Theorem 3: when vocabulary WX, WY satisfied WX ◎ WY, WY had the semantic relation of WX;

(3b) relation of the semantic matches between noun-verb

The semantic matches relation of noun-verb can be divided into two classes:

1) SVO semantic matches relation: the receptor that holds of actor or action may be done in noun;

2) lattice semantic matches relation: noun and verb have lattice semantic matches relation;

Suppose that verb is V, the gerund of executing in the definition of V is S ₀, being subjected to gerund is O ₀Since when definition with S ₀Be set to implement the top noun of V, O ₀Be set to bear the top noun of this action, so only have same S ₀Or O ₀Having the noun S of certain relation and noun O just might execution action V, namely consists of the semantic matches of SVO; The SVO semantic matches has 6 kinds of situations, and it is worth available Value _SVOExpression, computing formula is as follows:

Value _SVO=match(S，S ₀)+match(O，O ₀)

Definition 15 (conventional SVO semantic matches relation): (S ∝ S satisfies condition ₀) ∩ (O ∝ O ₀);

Definition 16 (heavily loaded SVO semantic matches relation): satisfy condition:

For noun S and O and verb V, when not satisfying SVO when coupling in the definition of V, and the defined declaration of S, O they satisfy the SVO coupling;

Definition 17---comprise SVO semantic matches relation:

((S ◎ S satisfies condition ₀) ∩ (O ∝ O ₀)) ∪ ((S ∝ S ₀) ∩ (O ◎ O ₀));

Definition 18---similar SVO semantic matches relation:

((S ∽ S satisfies condition ₀) ∩ (O ∝ O ₀)) ∪ ((S ∝ S ₀) ∩ (O ∽ O ₀));

Definition 19---metaphor SVO semantic matches relation: under the following conditions, may have metaphor in the conjecture statement:

Condition 1: in whole statement, can satisfy front four kinds of SVO coupling without any noun;

A condition 2: have noun S or O in the statement, satisfy! (S ∝ S ₀) ∩ (O ∝ O ₀), conjecture is likened S to S ₀

Or condition 3: have noun S or O in the statement, satisfy (S ∝ S ₀) ∩! (O ∝ O ₀), conjecture is likened O to O ₀

For metaphor SVO semantic matches relation, Value _SVO=K _F* (match (S, W _P)+match (O, W _P));

K _FBe weights coefficient, W _PS and S ₀Most recent co mmon ancestor.Because belong to conjecture character, K _FValue should be larger, to avert evil influence;

Definition 20---lattice semantic matches relation: for noun W and verb V, satisfy

Match (W, V)=K _C* d (W, W _C), K _CBe the weights coefficient;

(3c) noun-adjectival semantic matches relation

For adjective W _VAWith noun W _N, satisfy

Match (W _VA, W _N)=K _A* d (W _N, W), K _ABe weights coefficient (general K _A=1);

(3d) semantic matching relationship arranged side by side

Semantic matching relationship arranged side by side only is used for the judgement of statement parallel construction, to determine the scope of conjunction;

Definition 21---semantic similar: because nominal definition has adopted the method for single succession, two noun W _X, W _YAlthough in definition, do not have inheritance, at W semantically _XBut may be W _YA kind of, the concept that is equivalent in the description logic contains, and uses symbol W _X∽ W _YExpression; Can the Tableau algorithm in the description logic be improved, to judge the Concept Semantic similarity relation;

Definition 22---noun semantic matching relationship arranged side by side: for two noun W _X, W _YAvailable match (W _X, W _Y)=K _T* (d (W _X, W _E)+d (W _Y, W _E)) calculate a numerical value, as heuristic information, W _EW _X, W _YNearest common ancestor's node; When satisfying W _X∽ W _YThe time, the W in the statement _XWith W _YIt may be coordination;

Definition 23---verb semantic matching relationship arranged side by side: for two verb V _X, V _YAvailable match (W _X, W _Y)=K _T* (d (S _X0, S _Y0)+d (O _X0, O _Y0)) calculate a numerical value, as heuristic information, { S _X0, S _Y0, O _X0, O _Y0W _X, W _YActor in the definition and be subjected to the person;

(3e) semantic matches of adverbial word relation: suppose adverbial word can semantic matches in verb, adjective and adverbial word, stipulate match (W ₁, W ₂)=0; The semantic matches relation of measure word: lexicon should be preserved the incidence relation of measure word and noun; If measure word W can modification noun W _N, then stipulate match (W, W _N)=0; Otherwise match (W, W _N)=MAX; The semantic matches relation of pronoun: according to the relation that refers to of pronoun, pronoun is replaced to corresponding noun process;

(3f) grammer matching relationship

In concrete statement, the vocabulary of some type is modified mutually, but not inherent semantic relation between the vocabulary itself, and a kind of grammatical phenomenon of semantic modified relationship just may be arranged in this statement, i.e. phraseological modified relationship comprises following situation:

Modified relationship between uncommon part of speech: between verb-verb, between adverbial word-noun and between adjective-verb; These all belong to the grammer matching relationship, do not have inherent semantic matches relation between the vocabulary itself, just have phraseological modified relationship in statement; Its semantic matches is worth available match (W in the Sentence analysis process _X, W _Y)=MAX/K _GCalculate K _GType weights (K generally speaking _G=1, or satisfy K _G＜1.5).

5. semantic-based according to claim 1 mates the natural language knowledge acquisition method that drives, and wherein step (4) is carried out according to the following procedure:

(4a) three levels and the syntax thereof of definition semantic structure

To carry out Sentence analysis according to the semantic model of this paper, must have the statement abstract representation method of suitable semantic model; Any statement is all formed through iteration by statement relatively simple for structure, and phrase is seen as an ingredient in the statement; In order to satisfy the semantic analysis needs of semantic model, the semantic structure of statement can be divided into three levels according to complexity and the characteristics of semantic structure: simple sentence, special simple sentence, complex sentence;

Definition 24---simple sentence: only have a verb or adjective to make the statement C of predicate _S, available grammar G ₁Come abstractdesription;

Thought design grammar G with case grammar ₁, mentality of designing: suppose that V is predicate, S is the actor of V; O is the receptor that holds of V, A _BIt is preposition attribute; A _AIt is postpositive attributive; P _DBe the adverbial modifier or complement, be equivalent to one group of lattice in the case grammar; P _CIt is one lattice content; N is noun; N _PBe noun phrase;

Grammar G ₁In fuzzy rules more, mentality of designing of its crucial rule is as follows:

1)C _S→P _DA _BSA _AP _DVP _DA _BOA _AP _D

2)S→n|SA _AA _BS

3)P _D→P _C|P _DP _C

S, O, A _B, A _A, P _CIn preposition, conjunction, auxiliary word, number, the service regeulations of measure word can write out easily;

SVO in the simple sentence has 6 kinds of different order: SVO, SOV, VSO, OSV, VOS, OVS altogether; SV, VS; Dispense sometimes S or O in the simple sentence, have 4 kinds of different order: SV, VS, OV, VO; When predicate made in intransitive verb or adjective, this adjective was expressed as V, has 2 kinds of different order: SV, VS; Altogether have 10 kinds of orders; The concrete expression mode of the G1 of the syntax is as follows:

Definition 25---special simple sentence: have a plurality of verbs or adjective, but semantically do not comprising the statement of subordinate clause, available grammar G ₂Come abstractdesription;

Grammar G ₂Mentality of designing: can not produce on the basis of subordinate clause, to grammar G guaranteeing ₁The few rule of middle interpolation can generating grammar G ₂, mainly contain following 2 kinds of situations:

1) situation of predicate made in a plurality of verbs or adjective;

2) S, O, A made in verb or adjective _B, A _A, P _CSituation

Grammar G ₂Key be verb phrase V _VNoun phrase N can not directly be followed in front and back _P, namely N can not appear _P+ V _VOr V _V+ N _P

The concrete expression mode of the G2 of the syntax is as follows:

Definition 26---complex sentence: in grammar G ₂The regular N of middle interpolation _P→ C _S, form grammar G ₃Because regular N _P→ C _SIllustrate that a simple sentence or special simple sentence can make in the complex sentence arbitrarily composition, realized the simple sentence recurrence, so grammar G ₃Complex sentence can be described;

(4b) obtain the thinking of best grammatical analysis scheme

(1) lexical ambiguity digestion procedure

Suppose W ₁W ₂W _kThe lexical semantic number be respectively n ₁, n ₂N _k, entirely arranging for each semanteme, the result is { L ₁, L ₂L _M, M=n then _1*n _{2* *}n _k, suppose one of them

W _mN the meaning of a word, L then _iC _SA sequence of words without lexical ambiguity; Exhaustive each { L in the grammatical analysis process ₁, L ₂L _MAnalysis result, select best L _iJust can clear up lexical ambiguity;

(2) analyze thinking

According to axiom 1, obtain all grammatical analysis schemes, for each grammatical analysis option A _i, calculate A according to formula 1 _iCorresponding semantic matches value, and select best grammatical analysis scheme;

Definition 27---simple clause: satisfy grammar G in the statement ₁Or G ₂Substring be simple clause;

Suppose axiom 4---the semantic target axiom of modifying: the semanteme modification target of supposing notional word W is W _Gi, then for the simple clause C that meets semantic logic in the statement _S, satisfy (W ∈ C _S) → (W _Gi∈ C _S); For attribute A _B(supposing next-door neighbour S) satisfies (W ∈ A _B) → (W _Gi∈ (A _B∪ S)), other situation of attribute is similar; For the adverbial modifier or complement (P _D), satisfy (W ∈ P _D) → (W _Gi∈ (P _D∪ V));

According to semantic characteristics of modifying target, all grammatical analysis schemes can be divided into 2 layers:

1) simple clause's level grammatical analysis scheme;

2) the grammatical analysis scheme of simple clause inside;

(4c) obtain best grammatical analysis scheme

(1) can sum up simple clause's Rule of judgment

For statement C _S, carry out grammar G ₁, G ₂, G ₃The CYK Algorithm Analysis, satisfy table 1 conditional substring s (i, j) be the simple clause that can sum up;

Table 1 can be summed up simple clause's Rule of judgment

(2) bottom-up simple clause sums up method

Available bottom-up simple clause's end method is asked for best subordinate clause level grammatical analysis scheme, sees algorithm 4:

Algorithm 1---simple clause sums up method:

1) for statement C _S, according to the Rule of judgment of table 1, find out and to sum up the corresponding substring set of simple clause { s ₁, s ₂S _m;

2) for each clause s _i, calculate simple clause s according to algorithm 2 _iBest semantic matches value, with s _iBe summed up as N _P, N is set _PEnd semantic;

3) make C _SEqual to sum up the result, with the simple clause s in the recursive procedure _iThe summation of best semantic matches value, carry out the recurrence of step 1-3;

4) have best full sentence semantic matches and be worth corresponding simple clause s _iScope and end sequentially be best grammatical analysis scheme;

The key that the best semantic matches value of calculating simple clause is algorithm, concrete grammar are seen step (4d) " simple clause's best semantic matches value ";

In algorithm, when selecting, simple clause adopted the method for exhaustion, can obtain theoretic best grammatical analysis scheme; But the calculated amount of this method is larger, is difficult for realizing; But when the quantity that can return the clause less than 4 the time, also can consider; In the time can returning clause's quantity too much, (the individual semantic matches degree of k＜m) preferably simple clause is carried out recursive search, to ask for suboptimum grammatical analysis scheme can only to select k;

(3) sum up semanteme

In algorithm 1, with simple clause C _SBe summed up as N _PAfter, N _PDo not have semanteme, can't carry out next step semantic matches and calculate, the mode of solution is as follows:

1) regulation is come N by end _PCan be matched with any vocabulary W, the semantic matches value is:

Match (N _P, W)=MAX/K _C(K is arranged generally _C1)

2) if N _PMake S or the O of new goal clause, then can be with N _PSemanteme be set to former C _SIn S or O;

(4d) obtain simple clause's best semantic matches value

Calculate simple clause's best semantic matches value, according to axiom 1 and axiom 4, there is multiple grammatical analysis scheme simple clause inside, must obtain all grammatical analysis schemes, for every kind of grammatical analysis scheme, the semanteme of its notional word is modified target and is determined, just can calculate semantic matches value under this grammatical analysis scheme according to formula 1, and the grammatical analysis scheme with minimum semantic matches value is exactly required analysis result;

The grammatical analysis scheme of simple clause inside can be divided into 3 layers: 1) SVO combination level; 2) A _A, P _D, A _BLevel; 3) A _A, P _D, A _BInner grammatical analysis scheme; Can select wherein best grammatical analysis scheme by algorithm 2;

Algorithm 2---simple clause's best semantic matches value:

1) if simple clause is special simple sentence, finds all methods that it is summed up as simple sentence

2) for every kind of end method, special simple sentence is summed up as simple sentence

3) for this simple sentence, find out all possible SVO combined method

4) for every kind of SVO combined method, with simple sentence C _SSegmentation if S or O are phrase, is then carried out algorithm 3; Suppose C _SBe divided into { L ₁, L ₂..L _n}

5) each segmentation L _iIn can comprise A at most _A, P _D, A _BThree partial contents are found out L _iIn all A _A, P _D, A _BDivision methods

6) for every kind of A _A, P _D, A _BDivision methods by the means that the syntax and semantics the matching analysis combines, determines that target modified in the semanteme of each notional word, makes A _A, P _D, A _BThe semantic matches value minimum

7) ask for the semantic matches value of full sentence, select the minimum corresponding analytic process of semantic matches value as the grammatical analysis scheme of the best

Suppose

For simple clause C _SCarry out grammar G ₁The operation result of CYK algorithm, expression can generate the grammar symbol collection of substring s (i, j);

(1) the grammatical analysis scheme of SVO combination level

In simple sentence, suppose noun W ₁And W ₂Satisfy SVO coupling, then { W with verb V ₁, V, W ₂It is a SVO combination.But S (O is similar) may be a phrase, has { W in sentence ₁, V, W ₃And { W ₂, V, W ₃The SVO coupling, and V, W ₃Not at W ₁And W ₂Centre, and W ₁And W ₂Middle substring s (m, n) satisfies:

The time, at W ₁+ s (m, n)+W ₃The phrase that forms is S, the S that in like manner can obtain more to grow or O;

Algorithm 3---S or O segmentation:

1) obtains phrase S or O, according to the described principle of this step, find out all nouns that satisfy the SVO coupling among the S, be assumed to be { n ₁, n ₂..n _m}

2) according to { n ₁, n ₂..n _mPhrase S is divided into the m-1 section, according to regular S → n|SA _AA _BEach may not comprise A for empty segmentation to S as can be known _AA _B

(2) A _A, P _D, A _BThe division methods of level

Suppose segmentation L _iSubstring be s (m, n), then satisfy

P, q is grammatical A _A, P _D, A _BDivision methods, segmentation result is: A _A=s (m, p), P _D=s (p, q) A _B=s (q, n);

(3) A _A, P _D, A _BInner best grammatical analysis scheme

Theorem 2:A _A, P _D, A _BInner best grammatical analysis scheme is to work as A _A, P _D, A _BEach interior notional word has the best semantic corresponding grammatical analysis scheme under the target conditions of modifying;

Definition 28---simple noun phrase: not comprising verb and adjectival noun phrase is exactly simple noun phrase;

Theorem 3: the attribute A of simple sentence _AOr A _BBest grammatical analysis scheme be equivalent to the best grammatical analysis scheme of a simple noun phrase;

Proof: because simple sentence only comprises a verb or predicate V made in adjective, so A _A, A _BIn do not comprise verb and adjective.According to axiom 4, A _A, A _BThe grammatical analysis scheme be equivalent to simple noun phrase N _PThe grammatical analysis scheme, N _P∈ { (A _B+ S), (A _A+ S), (A _B+ O), (A _A+ O) };

Simple noun phrase N _PDifferent grammatical analysis schemes only be subjected to the impact of { conjunction/preposition/auxiliary word/measure word }; The key of grammatical analysis is scope and their the end order of selected { conjunction/preposition/auxiliary word/measure word };

Determining of A, scope: in { conjunction/preposition/auxiliary word/measure word }, suppose w _BFor preposition type, w _MFor preposition type, then its scope can be summed up as two kinds of forms 1) ..N _Bn..N _B1... w _M..N _A1..N _Am..; 2) w _B..N _Bn..N _B1... w _M..N _A1..N _Am...; { N wherein _B1, N _B2... N _BnThe noun of first half in the scope, { N _A1, N _A2... N _AmIt is the noun of first half in the scope;

According to the backward semantic custom of modifying of Chinese, can be at { N _A1, N _A2... N _AmIn find out grammatical and N _B1Noun N with best semantic matches value _AjAs boundary after the scope; The scope prezone of form 1 can be determined with similar method;

(2) determining of end order: { conjunction/preposition/auxiliary word/measure word } and scope thereof should be summed up by certain bar grammar rule, and the available method of exhaustion is obtained their the best end order; Generally speaking, after the simple sentence statement has carried out repeatedly segmentation, A _A, A _BIn the number n of { conjunction/preposition/auxiliary word/measure word } that comprise be generally less than 4, have the calculating feasibility;

Definition 29---noun sequence: not having the simple noun phrase of { conjunction/preposition/auxiliary word/measure word } is noun sequence;

After { conjunction/preposition/auxiliary word/measure word } all summed up, simple noun phrase just had been known as a noun sequence, and may also there be one or two noun sequence in { conjunction/preposition/auxiliary word/measure word } scope inside in addition; Only noun affects semantic modified relationship in noun sequence, according to the backward semantic custom of modifying of Chinese, supposes that noun sequence is L _N=W ₁W ₂W _mThen by semantic definite L _NIn arbitrarily noun to modify the target concrete grammar as follows:

The best semantic target of modifying of algorithm 4---noun sequence:

Set C is set _WBe sky, for L _NIn each noun W _iIf, match (W _i, W _m)＜MAX is with W _iBe added to C _W, do following operation:

1) supposes C _WElement be W by successively order ₁-W ₂-...-W _n(n〉1), then be done as follows: with L _NBe divided into the n+1 section, the semanteme modification target that they are set is W _m, and each section carried out recurrence;

2) work as C _WIn when only having a noun, carry out step 3-4;

3) the forward direction modified relationship is set: for any segmentation, if there is W _xW _X+1W _Y-1W _y, satisfy condition: 1. any W _X+1W _yBetween noun and W _yAfter the semantic matches value of noun be MAX; 2. match (W _y, W _x)＜MAX; W then is set _ySemanteme to modify target be W _xW then is set _X+1W _Y-1Between the modification target W of noun _y

4) if L _NIn also have noun W _yDo not modify target, it is W that its modification target then is set _Y+1

P _DAnalytical approach be similar to A _A, A _B, key is to carry out boundary line delimitation according to preposition, and the content in the preposition scope also is converted into a simple noun phrase; Since comparatively loaded down with trivial details, do not discuss in detail here;

(4) disposal route of special simple sentence

Sum up all non-predicate verb/adjectives, statement is converted into simple sentence, select best end scheme; Disposal route is as follows:

1) statement is carried out the CYK algorithm of grammar G 2, find verb (adjective) phrase that to do predicate, may have kinds of schemes;

2) for each scheme, sum up remaining verb (or adjective), choose the analytical plan of semantic matches value minimum;

It is semantic also to need to arrange end when summing up non-predicate verb (adjective).

6. semantic-based according to claim 1 mates the natural language knowledge acquisition method that drives, and wherein step (5) is carried out according to the following procedure:

According to the grammatical analysis result with best semantic matches value, simple sentence is converted into a knowledge point, each simple clause of complex sentence is converted into the knowledge point, whole complicated sentence is converted into one group of knowledge point;

After statement is converted into the knowledge point of depositing with the structural data form, just can carry out various Intelligent Information Processing to these knowledge datas easily.

Images