CN101414310A - Method and apparatus for searching natural language - Google Patents

Abstract

The invention discloses a searching method and a device for natural languages. The method is as follows: building a Chinese framework knowledge base CFN and a professional field knowledge body base; then utilizing the Chinese framework knowledge base CFN to carry out Chinese framework meaning character automatic marking on a searching sentence inputted in natural language searching and extracting a triad with meaning information from the searching sentence according to the marking, finally taking the triad as the searching input and utilizing the body base to generate the searching answer. The invention can be applied to identify the searching sentence inputted in the natural languages by the user. When in answer extraction, a large quantity of answer bases are not needed for matching.

Description

A kind of method and apparatus of Natural Language Search

Technical field

The present invention relates to the Natural Language Search technical field, particularly about a kind of searching method and device of natural language.

Background technology

Chang Yong search technique mainly is based on keyword matching or subject classification in the prior art, but owing to lack semantic information, lack knowledge understanding and inferential capability, exist the search return message to comprise a large amount of irrelevant informations, and return message also exists quality to hang down problems such as the precision that reaches information dropout, precision ratio is not enough, its main cause is the defective that the Internet exists aspect information representation and retrieval, do not offer the enough readable informations of computing machine, limited the automatic analysis ability of computing machine in retrieval.

The method of available technology adopting is, at first semantic analysis carried out in user's search input, cooperates part-of-speech tagging, finds out the significant keyword of search engine.And then the index file of business datum retrieved.

As in search the input " how going to Wutai Mountain? " from packet header can carry out semantic fractionation to sentence according to Chinese dictionary, be that participle becomes: " packet header ", " how going ", " Wutai Mountain " these semantic just main bodys also are the keywords that Natural Language Search needs.Because for search, searched content has been carried out the establishment of index in advance by the entry in the dictionary.So the answer of returning when search may be relevant information, the relevant information in Wutai Mountain, the while in packet header also to comprise the information of how to go to Wutai Mountain from packet header that the user need inquire about.As seen in the prior art because the semanteme of user input is not understood accurately, so when the information of returning, the needed information of feedback user that can not be promptly and accurately.

Summary of the invention

The invention provides a kind of searching method and device of natural language, when being used to solve prior art and carrying out Natural Language Search, just return the problem of a large amount of related web pages for inquiring user.

A kind of method of Natural Language Search comprises:

A plurality of lemmas, framework with identical semanteme and the framework element that constitutes framework are preserved in A, structure Chinese framework knowledge base CFN and professional domain ontologies storehouse in the described Chinese framework knowledge base, wherein said framework is used to explain described identical semanteme;

B, at the search statement of inquiring user input, lemma at least one verb in the described search statement and the Chinese framework knowledge base is mated, find the affiliated framework of described verb, and described search statement is marked according to the framework element that comprises in the described framework;

C, select in the described verb one, and generate tlv triple according to described mark extracts described semantic predicate and this semantic predicate from described search statement main body and/or object as semantic predicate;

D, with described tlv triple as inquiry input, utilize described professional domain ontologies storehouse to generate the candidate answers collection.

Wherein, the content in the described Chinese framework knowledge base is described by the Semantic Web SGML.

Described Chinese language knowledge framework storehouse comprises framework storehouse, sentence storehouse and lemma storehouse:

Described framework storehouse is to be unit with the framework, preserves the definition of framework, the framework element that constitutes framework and the relation between framework and the framework;

Described sentence storehouse record has the sentence of framework semantic tagger information, and the described sentence that has framework semantic tagger information is the framework that provided according to the framework storehouse and the framework semantic information and the syntactic information of framework element mark sentence;

The involved lemma of each framework is preserved in described lemma storehouse.

Wherein, make up professional domain ontologies storehouse, concrete steps comprise:

Make up the ontology model in this field with reference to the taxonomic hierarchies standard relevant with professional domain;

By the body edit tool relation of the notion of each knowledge entry in the ontology library, each knowledge entry and example are represented with the Semantic Web SGML, and be stored as computer-readable document format.

After the described step B, further comprise:

When in the search statement a plurality of verb being arranged, entry relationship in each verb and the ontology library compared obtain the semantic index of described verb, and according to the semantic predicate of described semantic index selection verb as described statement, described semantic index is used to weigh the importance of verb.

Wherein, described step D comprises:

From described search statement, extract tlv triple according to described mark with semantic information;

According to described tlv triple generated query statement, in ontology library, search related content with this tlv triple coupling;

If search success then generate the candidate answers collection; If search failure, then utilize corresponding rule searching to create inference machine and carry out reasoning, and generate corresponding data model and inquire about, generate corresponding candidate answers collection after the successful inquiring.

After the described generation candidate answers collection, further comprise:

The answer that candidate answers is concentrated is sorted, and the answer after will sorting returns to inquiring user.

Further, when the search statement of user's input is question sentence, after generating tlv triple, comprising:

Carry out the question sentence analysis, extract the interrogative and the query purpose speech of described question sentence, obtain the inquiry message of this question sentence;

Described inquiry message and tlv triple are imported as inquiry, utilized described professional domain ontology library to generate the candidate answers collection.

According to said method, the present invention also provides a kind of Natural Language Search device, comprising:

Memory module, be used to store Chinese framework knowledge base CFN and professional domain ontologies storehouse, preserve a plurality of lemmas, framework with identical semanteme and the framework element that constitutes framework in the described Chinese framework knowledge base, wherein said framework is used to explain described identical semanteme;

Analysis module, be used for when inquiring user inputted search statement, lemma at least one verb in the described search statement and the Chinese framework knowledge base is mated, find the affiliated framework of described verb, and described search statement is marked according to the framework element that comprises in the described framework;

The semantic predicate module, one that is used for selecting described verb as semantic predicate, and generates tlv triple according to described mark extracts described semantic predicate and this semantic predicate from described search statement main body and/or object;

The answer generation module is used for described tlv triple is imported as inquiry, utilizes described professional domain ontologies storehouse to generate the candidate answers collection.

Wherein, described memory module also is used for utilizing the Semantic Web SGML to describe the content of Chinese framework knowledge base.

Further, described analysis module comprises:

The framework determining unit is used for when inquiring user inputted search statement the lemma in verb in the search statement and the Chinese framework knowledge base being mated, and finds the affiliated framework of described verb;

The mark unit, the framework element that is used for comprising according to described framework marks described search statement.

Described semantic predicate module comprises:

Selected cell is used for selecting a verb as semantic predicate from the verb of search statement;

Extraction unit is used for and extracts the main body of described semantic predicate and this semantic predicate and/or object generates tlv triple according to described mark from described search statement.

Described answer generation module comprises:

Query unit is used for described tlv triple is imported as query search, utilizes described professional domain ontologies storehouse to generate the candidate answers collection;

Reasoning element is used for searching when failure when enquiry module, utilizes corresponding rule searching to create inference machine and carries out reasoning, and generate corresponding data model and inquire about and generate the candidate answers collection.

Sequencing unit is used for the answer that candidate answers is concentrated is sorted, and according to this ordering answer is returned to the user.

Further, described selected cell also is used for when search statement has a plurality of verb, entry relationship in each verb and the ontology library compared obtain the semantic index of described verb, and according to the semantic predicate of verb of described semantic index selection as described statement, described semantic index is used to weigh the importance of verb.

This device also comprises:

The question sentence module is used for carrying out the question sentence analysis when the search statement of user's input is question sentence, extracts the interrogative and the query purpose speech of described question sentence, obtains the inquiry message of this question sentence;

Then described answer generation module also is used for described inquiry message and tlv triple are imported as inquiry, utilizes described professional domain ontology library to generate the candidate answers collection.

The present invention utilizes CFN that the natural search statement of inquiring user input is marked automatically, extracts the tlv triple with semantic information then, described tlv triple is carried out the search of answer in ontology library as the inquiry input.So can be fast and search definite answer efficiently because before carrying out ontology library search, carried out semantic analysis and mark.

Description of drawings

Fig. 1 is the process flow diagram of the method for a kind of Natural Language Search of the embodiment of the invention;

Fig. 2 is the annexation figure of each word bank in the Chinese framework semantic knowledge-base in the embodiment of the invention;

Fig. 2 A is the frame network figure that each framework constitutes in the Chinese framework knowledge base in the embodiment of the invention;

Fig. 3 is the embodiment of the invention is extracted tlv triple from search statement a process flow diagram;

Fig. 3 A is that the embodiment of the invention utilizes Chinese framework knowledge base query statement to be carried out the process flow diagram of semantic character labeling;

Fig. 4 utilizes ontology library for the embodiment of the invention and carries out the process flow diagram of the extraction of answer;

Fig. 4 A is the fundamental diagram of inference machine;

Fig. 5 is a kind of querying method process flow diagram at the simple search statement of the embodiment of the invention;

Fig. 6 utilizes the inventive method to be applied to the process flow diagram of tour field;

Fig. 6 A is sight spot, lodging, the vehicles, amusement, food and drink and the relation model figure between 6 classes (notion) of doing shopping;

Fig. 7 carries out the process flow diagram that tlv triple is extracted for the embodiment of the invention to question sentence;

Fig. 8 is the installation drawing of a kind of Natural Language Search device of the embodiment of the invention;

Fig. 9 is an analysis module installation drawing in a kind of Natural Language Search device of the embodiment of the invention;

Figure 10 is semantic predicate modular device figure in a kind of Natural Language Search device language of the embodiment of the invention;

Figure 11 is an answer generation module installation drawing in a kind of Natural Language Search device language of the embodiment of the invention.

Embodiment

In the embodiment of the invention, make up Chinese framework knowledge base CFN and professional domain ontologies storehouse, utilize Chinese framework knowledge base that the query statement of Natural Language Search input is marked then, and according to described mark the extraction from query statement has the tlv triple of semantic information, at last described tlv triple is imported as inquiry, utilized the answer of described ontology library generated query.

Below in conjunction with Figure of description the specific embodiment of the present invention is elaborated, as shown in Figure 1, the method for a kind of Natural Language Search of the embodiment of the invention comprises step:

Step 101, make up Chinese framework knowledge base (Chinese Frame Net, CFN).

It is the Chinese framework knowledge base of description object with limited set of words that the embodiment of the invention has at first made up one, and with Semantic Web SGML (extend markup language (XML, Extensible MarkupLanguage), resource description framework (RDF, Resource Description Framework), Web body SGML (OWL, Web Ontology Language)) the various resources of this semantic knowledge-base have been represented.

(1) Chinese framework knowledge base mainly is made up of framework storehouse, sentence storehouse and lemma storehouse, and particular content comprises:

Lemma is mainly deposited in A, lemma storehouse, and described lemma is the class word with identical semanteme, and wherein said identical semanteme is a framework.

For example explain " statement " semantic lemma, as shown in table 1 comprising:

Speech v	Confess v	Assert n	Declare v	Announce v	Declaration n	Assert v	Claim v	Statement n	Statement v
										V sings one's own praises	Warning v	Comment v	Comment n	Note v	Complaint v	Honest v	Hysteria is said v	Pass on v	Leak v
Call out v	Explain orally v	V is described	Express v	Emphasize v	Mention v	Advertise v	Expression v	Propose v	Statement v
										Reaffirm v	Narration v	Say v	Talk about v	Report v	Report n	Disclose v	Say v	V speaks	Comment v
State v	Talk v	Talk v	Speech v	State v	Report v	Report n	Say frankly v	Tell v	V states outright
										Scatter v	Mention v	Mention v	Disclose v	Declaration n	Read out v	Declare v	Advocate v	Explain and publicise v	Make noise v
Comment v	Comment v	Comment n	Divulge v	Refer to v	V tries to justify oneself	Accusation v	Smooth dew v	State frankly v	Speech n
										Report v	Appraise v through discussion

Table 1

B, framework storehouse are to be unit with the framework, clearly provide the definition of framework and the framework element (being also referred to as semantic role) of framework, and describe the conceptual relation between this framework and other frameworks.

Framework is mainly deposited in the storehouse: the 1. definition of framework; 2. (the different components that constitute framework serve as different roles to the framework element, are referred to as semantic role and are also referred to as the framework element.Comprising core frame element and non-core framework element); 3. the relation of framework.

The related content that below is " statement " framework comprises: the definition, as shown in table 1, the non-core framework element of core frame element (core semantic role) (non-core semantic role) that comprise framework are shown in table 2 and table 3.

The framework definition of " statement ": what this framework was expressed is the behavior that the speaker conveys a message to obedient person with language.

Table 2

Table 3

C, sentence storehouse record have the sentence of framework semantic tagger information, and the principle of mark is according to the mark of the sentence under the framework of framework storehouse example, and is the branch framework, divides lemma to deposit.

CFN provides the sentence that has framework semantic tagger information for each senses of a dictionary entry of each lemma, and these sentences are from real database for natural language, rather than create by linguist or dictionary editor.Choosing on the sentence, making every effort to demonstrate as much as possible all possible syntactic-semantic combination of this lemma.This makes the data of CFN provide abundant material for the syntactic-semantic combinatorial property of summarizing word, for automatic semantic tagger Study on Technology provides training data.

A sentence example of " statement " framework:

Britain side's face measure in requital announces that also the diplomat of 4 Russian embassies is the person non grata.

＜spkr-np-subj English jn aspect n〉announces v as v revenge v measure n also d＜tgt＜the diplomat u n of the msg-dj-obj 4m name q Russia nsy n of embassy be v not d be subjected to the u people n of v welcome v.

(2) contact between each element in the Chinese framework knowledge base:

As shown in Figure 2, lemma storehouse, sentence storehouse and framework storehouse three's relation comprises in embodiments of the present invention: the lemma storehouse depends on the framework storehouse, be that specific word is under the jurisdiction of specific framework (though the phenomenon of one-to-many is arranged, promptly a lemma can be under the jurisdiction of several frameworks), because same lemma is under different frameworks, its semantic collocation pattern is different with the sentence structure way of realization, so the sentence storehouse depends on lemma storehouse and framework storehouse again.

Also have multiple contact between framework and the framework, constitute a knowledge network shown in Fig. 2 A, wherein the contact between each framework comprises: inheritance, total points relation, total territory/minute territory relation, reference relation, cause-effect relationship, follow-up relation.Simultaneously a framework relates to a plurality of lemmas, marks with the framework element set of same framework; Conversely, a polysemant is represented a plurality of lemmas, belongs to several different frameworks, promptly represents that with different framework elements such information has been arranged, and an application system just might be distinguished the different meanings of same morphology in different environments for use.

Step 102, structure professional domain ontologies storehouse specifically comprise:

At first determine the field and the scope of body with reference to the taxonomic hierarchies standard, and list the important terms in the body, described term roughly shows all things that relate to of modeling process, and the relation between the attribute that these things had and these attributes etc.Define the relation between the support, attribute, attribute of class and class, the restriction of attribute then, obtain the ontology model of this body at last.

By body edit tool (wherein comparatively common body edit tool comprises Ontolingua, OntoEdit, Ontosaurus and Prot é g é etc.) notion of each knowledge entry in the ontology model, relation and example (being tlv triple) are showed with the Web SGML, and be stored as computer-readable document form.

The reverse-power (Inverse Of) of having set up strict difinition between the class of body, transitive relation (TranstiveProperty), funtcional relationship (Functional Property), symmetric relation (Symmetric Property), inverse function relation (Inverse Functional Property) and to the restriction of attribute.

The tlv triple of step 103, search statement extracts.

Behind the search statement that receives user's input, at first carry out pre-service, promptly carry out the part of speech of participle and all words of mark.Extract all verbs in the described search statement then, and the lemma in each verb and the Chinese framework knowledge base is mated, find the affiliated framework of each verb, described search statement is marked according to the framework element in the described framework.Select in the verb one to generate tlv triple at last with semantic information as the semantic predicate of described search statement and the subject and object that extracts this semantic predicate, described subject and object is a previous noun and a back noun adjacent with semantic predicate in the query statement, and described tlv triple has been expressed the semantic information of INQUIRE statement and the annexation between each framework element.

Wherein, can lack main body or object in the tlv triple, promptly tlv triple be by semantic predicate add the above semantic predicate main body and/object forms.

Further, if there is not verb in the search statement, then described semantic predicate then is the word that can represent this statement search intention.If described search statement is not for comprising the question sentence of verb, then described semantic predicate is an interrogative, and subject and object then is the noun adjacent with interrogative.

As shown in Figure 3, be example with the verb tlv triple, the extraction of tlv triple is further detailed, specifically comprise step:

S301, query statement is carried out semantic character labeling according to Chinese framework knowledge base.As shown in Figure 3A, specifically comprise step:

S3A01, the search statement that inquiring user is imported carry out pre-service, extract all verbs in this search statement.

S3A02, the lemma in described verb and the Chinese framework knowledge base is mated, thereby obtain framework under this verb.

S3A03, described search statement is marked according to the framework element that is comprised in this framework.Specifically comprise three layers:

Ground floor framework element mark, the framework element is the various participants in the framework, the framework element is divided into core frame element and non-core framework element.The core frame element is the mandatory component of a framework on conceptual understanding, and their types in different frameworks are different with quantity, demonstrates the individual character of framework.Non-core framework element is the individual character of display frame not, peripheral semantic components such as expression time, space, environmental baseline, reason, purpose.Second layer phrase type mark.The 3rd layer of syntactic function mark.

The semantic predicate of S302, the described search statement of selection, described semantic predicate are can explain the word of the topmost search purpose of search statement really.

1, when having only a verb in the search statement of user's input, then this verb is the semantic predicate of described search statement.

When if a plurality of verb is arranged in the search statement of 2 users input, then (entry relationship comprises: the notion in the model with the entry relationship in each verb and the ontology library, and the example of relation between the notion and notion, for example: automotive-type is arranged in the vehicles, between automobile and the sight spot class relation is arranged, and sight spot and automobile all have instantiation separately) comparing obtains the semantic index of each verb, and described semantic index is used to weigh the importance of semantic predicate.Select the semantic predicate of a verb according to semantic index then as this search statement.

The extraction of S303, tlv triple.

Generate the tlv triple of the semantic information that can express this search statement according to the subject and object of the described semantic predicate of mark extraction.The main body or the object that can lack semantic predicate in the described tlv triple.

Owing to have a large amount of spoken languages in the statement of user's input, so the semanteme of query statement is understood according to the spoken vocabulary dictionary of vocabulary correspondence in the restricted domain.

Step 104 utilizes ontology library to carry out the extraction of answer.As shown in Figure 4, specifically comprise step:

S401, utilize described tlv triple generated query statement in ontology library, to search the relevant information that is complementary with this tlv triple.

If S402 searches successfully, change S405 over to after then generating the candidate answers collection, if search failure, then change S403 over to.

S403, utilize corresponding rule searching (the whole relations in the ontology library of depositing in custom rule in the inference machine and the inference machine) to create inference machine, carry out reasoning, and generate corresponding data model, inquire about once more.

If the S404 successful inquiring then generates corresponding answer set, and changes S405 over to; If inquiry is failure once more, then go to S406.

S405, the concentrated answer that checks on one's answers are sorted, and the answer after will sorting returns to the user.

S406, looked into content and can not be found by inquiring user returns.

In embodiments of the present invention, when the tlv triple extracted and the relevant information in the ontology library were mated, the kit-Jena of rdf model can be resolved and inquire about to employing.Jena body resolver can be resolved RDF, and the inquiry of RDQL is supported and to the parsing of OWL.Jena provides the RBR machine simultaneously.

Jena provides RBR machine (as RDFS Reasoner, OWL Reasoner etc.), and the user can also self-defined as required inference rule in addition, also can register and use third party's inference engine.Shown in Fig. 4 A, the principle of work of inference machine is: the inference machine login mechanism is created out inference machine according to basic RDF vector description (information resources) and Ontology, inference machine can generate the model object (InferenceGraph that comprises inference mechanism thus, InfGraph), in Jena, figure (Graph) is also referred to as model (Model), and the form of expression is model interface (ModelInterface), the application programming interface that can use a model then (Model API, ModelApplication Programming Interface) and body application programming interface (Ontology API, OntologyApplication Programming Interface) this model is operated and handled, thus the information retrieval of realization semantic level.

As shown in Figure 5, the embodiment of the invention also provides a kind of querying method at the simple search statement, specifically comprises step:

Step 501, structure professional domain ontologies storehouse.

Make up the professional domain ontologies towards the restricted domain, with reference to " (Chinese classification scheme vocabulary ", professional domain relevant criterion, and, make up the ontology model in this field according to all information relevant as can be known of relation between each component or the like in the basic term of professional domain and the professional domain with this professional domain.Adopt OWL that ontology model is encoded then, by the body edit tool Prot é g é of Stanford Univ USA, the notion of each clauses and subclauses in the ontology library, relation and example are showed with OWL and RDF, and be stored as the OWL document at last.

The reverse-power (Inverse Of) of having set up strict difinition between the class of body, transitive relation (TranstiveProperty), funtcional relationship (Functional Property), symmetric relation (Symmetric Property), inverse function relation (Inverse Functional Property) and to the restriction of attribute.

Step 502, at first search statement is carried out pre-service, extract the tlv triple in the search statement, utilize described tlv triple to generate the SPARQL query statement then, in ontology library, search the relevant information that is complementary with this tlv triple, if search successfully, then change step 504 over to,, then changestep 503 over to if search failure.

Step 503, utilize corresponding rule searching to create inference machine, carry out reasoning, and generate corresponding data model, inquire about once more, if successful inquiring then changes step 504 over to; If inquiry is failure once more, then return institute's query contents and can not find.

Step 504, candidate answers is sorted, and the answer after will sorting returns to inquiring user.

As shown in Figure 6, the relevant question sentence below in conjunction with the tour field inquiring user is proposed is described further the embodiment of the invention.Because user's major part in the time of the inquiry related content all is the form input with question sentence, so done the processing of optimizing at the inquiry question sentence especially in the present embodiment, concrete steps comprise:

Step 601, structure Chinese framework knowledge base (CFN).

Step 602, structure tour field ontologies storehouse.

Towards the travel information in somewhere, choose distinctive tourist attractions, all set up corpus at each sight spot, make up the ontology library of tour field.On the basis of sight spot corpus, promptly swim according to tourism six key elements, purchase, joy, food, live, OK, document has been carried out the extraction of term, and with reference to " Chinese classification scheme vocabulary " and " tourist service basis term " (gb/t 16766-1997), " tourism planning general rule " (gb/t 18971-2003), each subject of tourist industry is affiliated classification in the Chinese Library classification, " tourist industry standards system table ", " travel agency's domestic travel quality of service requirement " (lb/t004-1997), " guide service quality " (gb/15971-1995), CNS net (www.chinagb.org), tourism planning general rule (gb/t 18971-2003), tourist resources Classification Count and evaluation (gb/t 18972-2003), tourist service basis term (gb/t 16766-1997) etc. has carried out the Primary Construction of tourism ontology model.Fig. 6 A is sight spot, lodging, the vehicles, amusement, food and drink and the relation model figure between 6 classes (notion) of doing shopping.

System adopts OWL Lite to carry out the coding of ontology model, and has used the body edit tool Prot é g é of Stanford Univ USA.The reverse-power (Inverse Of) of having set up strict difinition between the class of body, transitive relation (Transtive Property), funtcional relationship (Functional Property), symmetric relation (Symmetric Property), inverse function relation (Inverse Functional Property) and to the restriction of attribute.

By Prot é g é, the notion relevant with database, relation and example are showed with OWL and RDF, be stored as the OWL document.

Step 603, the query statement that user search is imported carry out the problem classification.

Problem is carried out the branch time-like, different problem classification can be arranged from different angles.Native system has been taked multi-angle classification form, on the basis of TREC (Text Retrieval Conference) classification, utilizes the thought of body, and problem is classified.

According to the statistics in question sentence storehouse, the question sentence type of being carried for the tour field inquiring user at present is divided into following three classes:

(1) simply asks the main body of body, object.Comprise and refer in particular to interrogative sentence and be non-yet inquiry personage, time, numeral, entity.

As: how is the weather in Wutai Mountain? is there there the hotel near the Wutai Mountain?

(2) method for inquiring belongs to description.

As: is how driving left for Wutai Mountain, to be gone from Beijing?

(3) problem of reason, definition class.

Step 604, the query statement that utilizes Chinese framework knowledge base that user search is imported extract the tlv triple with semantic information, and concrete steps comprise as shown in Figure 7:

S701, utilize Chinese framework knowledge base that query statement is carried out semantic character labeling.

Mark has three layers, and ground floor is the framework element, and the framework element is divided into core frame element and non-core frame

Table 4

The frame element.The core frame element is the mandatory component of a framework on conceptual understanding, and their types in different frameworks are different with quantity, demonstrates the individual character of framework.Non-core framework element is the individual character of display frame not, peripheral semantic components such as expression time, space, environmental baseline, reason, purpose.The second layer is phrase type mark, and the 3rd layer is the syntactic function mark.Provided the frame description of " arrival " framework in the table 4.

Example sentence: " driving from Taiyuan to Wutai Mountain, how to walk recently? "

Carry out behind the CFN mark be:

＜mot-vp-va drives〉＜src-pp-adva is from Taiyuan〉＜tgt=reaches〉＜goal-sp-obj Wutai Mountain〉how to walk recently?

S702, question sentence analysis.

Obtain interrogative and query purpose speech.Because question sentence can be determined the search purposes of inquiring user by interrogative and query purpose speech.

The extraction of S703, tlv triple.

At first from the verb of question sentence, obtain semantic predicate, and the semantic predicate that gets access to and the entry relationship in the ontology library are compared.Weigh the subject and object of important, the rule-based scoring back extraction semantic predicate of semantic predicate by semantic index.

Example sentence: drive from Taiyuan to Wutai Mountain, how to walk recently?

At first pass through pre-service, directly extracting framework element＜mot-vp-va by the information of CFN mark drives 〉,＜src-pp-adva is from Taiyuan 〉,＜tgt=reaches 〉,＜goal-sp-obj Wutai Mountain 〉, discern, judge the second largest class that belongs in the TREC classification through problem types: the method class in the description, analyzing the comparison composition simultaneously is the route property value.Do you satisfy＜self driving?, starting point, Taiyuan 〉,＜self driving?, destination, Wutai Mountain〉the example of automobile subclass self driving, the route property value to all examples compares then.

For example: frameworks such as embodiment of the invention utilization " arrival ", " passing through ", " setting out ", " displacement ", " existence ", question sentence to the inquiry traffic route or the vehicles carries out the question sentence analysis, utilizes the lemma in the framework that verb has been carried out the synonym expansion simultaneously.

The CFN ground floor can be the vehicles and very fast the identifying of starting point and destination.Table 5 is the part question sentence mark example in tourist communications field.

Table 5

The extraction ofstep 605, answer.

The search purposes of described tlv triple and inquiring user is imported as inquiry, generated SPARQL query language and Jena inference machine and carry out searching of answer in described tour field ontologies storehouse, concrete querying flow comprises:

When the user import an inquiry " how going to Wutai Mountain? " from packet header, then system therefrom extracts starting point, verb and destination＜packet header by above-mentioned steps, goes to Wutai Mountain 〉, and the search purposes that the question sentence analysis obtains the user is: interrogation link is how to get to.Generate the SPARQL query statement according to tlv triple and search purposes information, in ontology library, search and inquire the relevant information that content is complementary.

If search successfully, then directly generate the candidate answers collection; If search failure, then generate corresponding rule searching, and create inference machine, carry out reasoning, generate corresponding data model then, inquire about once more, search success and then generate corresponding candidate answers collection, and the answer that candidate answers is concentrated is sorted.Result after will sorting at last returns to the user.If still fail after generating corresponding rule searching, then return the sky answer to inquiring user.

The answer of returning of example is:

1, train 1674/1675: packet header---Xinzhou train 2462/2463: packet header---Xinzhou big bus: Xinzhou---Wutai Mountain

2, aircraft MU5690: airport, packet header---Taiyuan Wu Su airport limousine: Taiyuan---Wutai Mountain

3, big bus: packet header---Taiyuan big bus: Taiyuan---Wutai Mountain.

As shown in Figure 8, the embodiment of the invention also provides a kind of Natural Language Search device to comprise memory module 801, analysis module 802, question sentence module 803, semantic predicate module 804, answer generation module 805:

Memory module 801, be used to make up Chinese framework knowledge base CFN and professional domain ontologies storehouse, preserve a plurality of lemmas, framework with identical semanteme and the framework element that constitutes framework in the described Chinese framework knowledge base, wherein said framework is used to explain described identical semanteme, and all the elements in the wherein said Chinese framework knowledge base are all described by the Semantic Web SGML.

Analysis module 802, be used for when inquiring user inputted search statement, lemma at least one verb in the described search statement and the Chinese framework knowledge base is mated, find the affiliated framework of described verb, and described search statement is marked according to the framework element that comprises in the described framework.

As shown in Figure 9, described analysis module comprises framework determining unit and mark unit:

Framework determining unit 901 is used for when inquiring user inputted search statement the lemma in verb in the search statement and the Chinese framework knowledge base being mated, and finds the affiliated framework of described verb.

Mark unit 902, the framework element that is used for comprising according to described framework marks described search statement.

Question sentence module 803 is used for carrying out the question sentence analysis when the search statement of user's input is question sentence, extracts the interrogative and the query purpose speech of described question sentence, obtains the inquiry message of this question sentence;

Semantic predicate module 804, one that is used for selecting described verb as semantic predicate, and generates tlv triple according to described mark extracts described semantic predicate and this semantic predicate from described search statement main body and/or object.

As shown in figure 10, described semantic predicate module comprises selectedcell 1001 andextraction unit 1002, wherein:

Described selectedcell 1001 is used for when the search statement of user's input has only a verb, and then this verb is the semantic predicate of described search statement.If when a plurality of verb was arranged in the search statement of user input, then the entry relationship in each verb and the ontology library (being attribute) being compared obtained the semantic index of each verb, described semantic index is used to weigh the importance of semantic predicate.Select the semantic predicate of a verb according to semantic index then as this search statement.

Describedextraction unit 1002 is used for and extracts the main body of described semantic predicate and this semantic predicate and/or object generates tlv triple according to described mark from described search statement.

Answer generation module 805, be used for extracting tlv triple from described search statement with semantic information according to described mark, described tlv triple comprises the main body and/or the object of verb and verb, and described tlv triple imported as query search, utilize described professional domain ontologies storehouse to generate the candidate answers collection.When described search statement was question sentence, then this answer generation module also was used for described inquiry message and tlv triple are imported as inquiry, utilizes described professional domain ontology library to generate the candidate answers collection.

As shown in figure 11, described answer generation module comprisesquery unit 1101,reasoning element 1102, sequencing unit 1103:

Query unit 1101 is used for described tlv triple is imported as query search, utilizes described professional domain ontologies storehouse to generate the candidate answers collection.

Reasoning element 1102 is used for searching when failure when enquiry module, utilizes corresponding rule searching to create inference machine and carries out reasoning, and generate corresponding data model and inquire about and generate the candidate answers collection.

Sequencing unit 1103 is used for the answer that candidate answers is concentrated is sorted, and according to this ordering answer is returned to the user.

Because all the elements in the Chinese framework knowledge base all are described with Semantic Web, institute thinks readable, the intelligible semantic dictionary of computer utility, for the semantic knowledge in the realization Semantic Web is shared and intelligent, personalized Web service provides basic resource.

And, corresponding relation between sentence storehouse record semantic role in the Chinese framework knowledge base and phrase type, the syntactic function, replaced from the description of intuition role's selectional restriction, than the result of artificial description more specifically, more accurate, also more with practical value.

Method of the present invention is not limited to the embodiment described in the embodiment, and those skilled in the art's technical scheme according to the present invention draws other embodiment, belongs to technological innovation scope of the present invention equally.

Obviously, those skilled in the art can carry out various changes and modification to the present invention and not break away from the spirit and scope of the present invention.Like this, if of the present invention these are revised and modification belongs within the scope of claim of the present invention and equivalent technologies thereof, then the present invention also is intended to comprise these changes and modification interior.

Claims (17)

Translated fromChinese

1、一种自然语言搜索的方法，其特征在于，包括：1. A method for natural language search, comprising:A、构建汉语框架知识库CFN和专业领域知识本体库，所述汉语框架知识库中保存具有相同语义的多个词元、框架以及构成框架的框架元素，其中所述框架用于表述所述相同语义；A. Construct the Chinese frame knowledge base CFN and professional field knowledge ontology base, in which a plurality of lexical units with the same semantics, frames and frame elements constituting the frame are stored in the Chinese frame knowledge base, wherein the frame is used to express the same Semantics;B、针对查询用户输入的搜索语句，将所述搜索语句中的至少一个动词与汉语框架知识库中的词元进行匹配，找到所述动词所属的框架，并根据所述框架中包含的框架元素对所述搜索语句进行标注；B. Match at least one verb in the search sentence with the lemma in the Chinese frame knowledge base for the search sentence input by the querying user, find the frame to which the verb belongs, and according to the frame elements contained in the frame Marking the search statement;C、选择所述动词中的一个作为语义谓词，并根据所述标注从所述搜索语句中提取出所述语义谓词以及该语义谓词的主体和/或客体生成三元组；C. Select one of the verbs as a semantic predicate, and extract the semantic predicate and the subject and/or object of the semantic predicate from the search sentence according to the annotation to generate a triple;D、将所述三元组作为查询输入，利用所述专业领域知识本体库生成候选答案集。D. Using the triples as query input, using the professional field knowledge ontology database to generate a candidate answer set.2、如权利要求1所述的方法，其特征在于，所述汉语框架知识库中的内容由语义Web标记语言描述。2. The method according to claim 1, characterized in that the content in the Chinese framework knowledge base is described by Semantic Web Markup Language.3、如权利要求2所述的方法，其特征在于，所述语义Web标记语言包括可扩展标记语言XML、资源描述框架RDF、本体标记语言OWL。3. The method according to claim 2, wherein the semantic Web markup language includes Extensible Markup Language XML, Resource Description Framework RDF, and Ontology Markup Language OWL.4、如权利要求1所述的方法，其特征在于，所述汉语知识框架库包括框架库、句子库和词元库：4. The method according to claim 1, characterized in that, said Chinese knowledge framework database comprises a framework database, a sentence database and a word unit database:所述框架库是以框架为单位，保存框架的定义、构成框架的框架元素以及框架和框架之间的关系；The frame library is based on the frame as a unit, and saves the definition of the frame, the frame elements constituting the frame, and the relationship between the frame and the frame;所述句子库记录带有框架语义标注信息的句子，所述带有框架语义标注信息的句子是按照框架库所提供的框架和框架元素标注句子的框架语义信息和句法信息；The sentence library records sentences with frame semantic annotation information, and the sentence with frame semantic annotation information is the frame semantic information and syntactic information of the sentence marked according to the frame and frame elements provided by the frame library;所述词元库保存每个框架所涉及到的词元。The lexical library stores the lexical elements involved in each frame.5、如权利要求1所述的方法，其特征在于，构建专业领域知识本体库，包括：5. The method according to claim 1, characterized in that building a professional field knowledge ontology database includes:参照与专业领域相关的分类体系标准构建该领域的本体模型；Construct the ontology model of this field with reference to the classification system standards related to the professional field;通过本体编辑工具把本体库内各知识条目的概念、各知识条目的关系以及实例用语义Web标记语言表示，并存储为计算机可读的文档格式。The concept of each knowledge item in the ontology library, the relationship of each knowledge item, and the instance are represented by the semantic Web markup language through the ontology editing tool, and stored in a computer-readable document format.6、如权利要求1所述的方法，其特征在于，所述步骤B之后，进一步包括：6. The method according to claim 1, characterized in that, after step B, further comprising:当搜索语句中有多个动词时，将每个动词与本体库中的条目关系进行比对得到所述动词的语义指数，并根据所述语义指数选择动词作为所述语句的语义谓词，所述语义指数用于衡量动词的重要性。When there are multiple verbs in the search sentence, each verb is compared with the entry relationship in the ontology database to obtain the semantic index of the verb, and the verb is selected as the semantic predicate of the sentence according to the semantic index, the The semantic index is used to measure the importance of verbs.7、如权利要求1所述的方法，其特征在于，所述步骤D，包括：7. The method according to claim 1, characterized in that said step D comprises:根据所述标注从所述搜索语句中提取具有语义信息的三元组；extracting triples with semantic information from the search statement according to the annotation;根据所述三元组生成查询语句，在本体库中查找与该三元组匹配的相关内容；Generate a query statement according to the triple, and search for relevant content matching the triple in the ontology database;如果查找成功则生成候选答案集；如果查找失败，则利用相应的查询规则创建推理机进行推理，并生成相应的数据模型进行查询，查询成功后生成相应的候选答案集。If the search is successful, the candidate answer set will be generated; if the search fails, the corresponding query rules will be used to create an inference engine for inference, and the corresponding data model will be generated for query, and the corresponding candidate answer set will be generated after the query is successful.8、如权利要求1或7所述的方法，其特征在于，所述生成候选答案集之后，进一步包括：8. The method according to claim 1 or 7, characterized in that, after generating the candidate answer set, further comprising:对候选答案集中的答案进行排序，并将排序后的答案返回给查询用户。Sorts the answers in the candidate answer set and returns the sorted answers to the querying user.9、如权利要求1所述的方法，其特征在于，当用户输入的搜索语句为问句时，在生成三元组之后，进一步包括：9. The method according to claim 1, wherein when the search sentence input by the user is a question sentence, after generating triples, further comprising:进行问句分析，提取所述问句的疑问词和疑问意向词，得到该问句的询问信息；Performing question analysis, extracting interrogative words and interrogative intention words of the question sentence, and obtaining inquiry information of the question sentence;将所述询问信息和三元组作为查询输入，利用所述专业领域本体库生成候选答案集。The query information and triples are used as query input, and the professional field ontology database is used to generate a candidate answer set.10、一种自然语言搜索装置，其特征在于，包括：10. A natural language search device, comprising:存储模块，用于存储汉语框架知识库CFN和专业领域知识本体库，所述汉语框架知识库中保存具有相同语义的多个词元、框架以及构成框架的框架元素，其中所述框架用于表述所述相同语义；The storage module is used to store the Chinese frame knowledge base CFN and the professional field knowledge ontology base, in which a plurality of lexical units with the same semantics, frames and frame elements constituting the frame are stored in the Chinese frame knowledge base, wherein the frame is used to express said same semantics;分析模块，用于当查询用户输入搜索语句时，将所述搜索语句中的至少一个动词与汉语框架知识库中的词元进行匹配，找到所述动词所属的框架，并根据所述框架中包含的框架元素对所述搜索语句进行标注；The analysis module is used to match at least one verb in the search sentence with the lemma in the Chinese frame knowledge base when the query user inputs the search sentence, find the frame to which the verb belongs, and according to the frame contains The frame elements of mark the search statement;语义谓词模块，用于选择所述动词中的一个作为语义谓词，并根据所述标注从所述搜索语句中提取出所述语义谓词以及该语义谓词的主体和/或客体生成三元组；A semantic predicate module, configured to select one of the verbs as a semantic predicate, and extract the semantic predicate and the subject and/or object of the semantic predicate from the search sentence according to the annotation to generate a triple;答案生成模块，用于将所述三元组作为查询输入，利用所述专业领域知识本体库生成候选答案集。An answer generating module, configured to use the triplet as query input, and generate a candidate answer set by using the professional field knowledge ontology database.11、如权利要求10所述的装置，其特征在于，所述存储模块还用于利用语义Web标记语言描述汉语框架知识库中的内容。11. The device according to claim 10, wherein the storage module is further configured to use Semantic Web Markup Language to describe the content in the Chinese framework knowledge base.12、如权利要求10所述的装置，其特征在于，所述分析模块包括：12. The device according to claim 10, wherein the analysis module comprises:框架确定单元，用于当查询用户输入搜索语句时，将搜索语句中的动词与汉语框架知识库中的词元进行匹配，找到所述动词所属的框架；A frame determination unit, configured to match the verbs in the search sentence with the lemmas in the Chinese frame knowledge base when the querying user inputs the search sentence, to find the frame to which the verb belongs;标注单元，用于根据所述框架中包含的框架元素对所述搜索语句进行标注。A labeling unit, configured to label the search statement according to the frame elements included in the frame.13、如权利要求10所述的装置，其特征在于，所述语义谓词模块包括：13. The device according to claim 10, wherein the semantic predicate module comprises:选择单元，用于从搜索语句的动词中选择一个动词作为语义谓词；A selection unit for selecting a verb from among the verbs of the search statement as a semantic predicate;提取单元，用于并根据所述标注从所述搜索语句中提取出所述语义谓词以及该语义谓词的主体和/或客体生成三元组。The extraction unit is configured to extract the semantic predicate and the subject and/or object of the semantic predicate from the search sentence according to the annotation to generate triples.14、如权利10所述的装置，其特征在于，所述答案生成模块包括：14. The device according to claim 10, wherein the answer generation module includes:查询单元，用于将所述三元组作为查询搜索输入，利用所述专业领域知识本体库生成候选答案集；a query unit, configured to use the triplet as a query search input, and use the professional field knowledge ontology database to generate a candidate answer set;推理单元，用于当查询模块查找失败时，利用相应的查询规则创建推理机进行推理，并生成相应的数据模型进行查询生成候选答案集。The reasoning unit is used to create a reasoning machine using corresponding query rules to perform reasoning when the query module fails to find, and generate a corresponding data model for querying to generate a candidate answer set.15、如权利要求14所述的装置，其特征在于，所述答案生成模块还包括：15. The device according to claim 14, wherein the answer generating module further comprises:排序单元，用于对候选答案集中的答案进行排序，并根据该排序将答案返回给用户。The sorting unit is used for sorting the answers in the candidate answer set, and returning the answers to the user according to the sorting.16、如权利要求13所述的装置，其特征在于，所述选择单元还用于当搜索语句中有多个动词时，将每个动词与本体库中的条目关系进行比对得到所述动词的语义指数，并根据所述语义指数选择一个动词作为所述语句的语义谓词，所述语义指数用于衡量动词的重要性。16. The device according to claim 13, wherein the selection unit is further configured to compare each verb with the entry relationship in the ontology database to obtain the verb when there are multiple verbs in the search sentence and select a verb as the semantic predicate of the sentence according to the semantic index, and the semantic index is used to measure the importance of the verb.17、如权利要求10所述的装置，其特征在于，该装置还包括：17. The device of claim 10, further comprising:问句模块，用于当用户输入的搜索语句为问句时，进行问句分析，提取所述问句的疑问词和疑问意向词，得到该问句的询问信息；Question module, for when the search sentence input by the user is a question sentence, analyze the question sentence, extract the interrogative word and interrogative intention word of the question sentence, and obtain the inquiry information of the question sentence;则所述答案生成模块还用于将所述询问信息和三元组作为查询输入，利用所述专业领域本体库生成候选答案集。Then the answer generation module is further configured to use the query information and triples as query input, and use the professional field ontology database to generate candidate answer sets.

Images