Background technology
Remote information service (Telematics) is the compound word of communication (Telecommunication) and information science (Informatics); So-called Telematics system is promptly through being built in computer system on the automobile, Wireless Telecom Equipment, Satellite Navigation Set, Internet technology etc., the service system that provides information such as literal, voice, image to transmit.TSP platform (Telematics Service Platform) for a kind of be the software platform that the motorist provides Telematics service based on wireless communication technology, satnav (GPS) technology, geographic information system technology, Internet technology and Call Center Platform.Wherein OnStar system and G-BOOK system are the manufacturers of two main successful application Telematics systems, and domesticly are in the starting stage at Telematics,
Along with speech synthesis technique in a large amount of successful Application of navigation field, the application of speech recognition skill also begins to show up prominently in the part navigational system.Speech recognition technology can reduce the number of times of user's operation, improves user experience.Let user experiencing the target of " only need open one's mouth, need not start " through speech recognition technology.Especially get the user for the motorist, in startup procedure, reduce operational motion as far as possible, make things convenient for the user on the one hand, driver's safety guarantee is provided on the one hand.
As Chinese invention patent application " voice control system for vehicle navigation apparatus " (publication number: CN 1841312A) a kind of vehicle navigation apparatus control system is disclosed, comprise one can identify voice messaging sound identification module, judge that voice messaging is the steering order or the instruction discrimination module of map place name.After sound identification module identified the result, Query Result in the phonetic control command storehouse saw that the voice that identify are that steering order still is the map place name.If in the phonetic control command storehouse, find the result, then be steering order; If in the phonetic control command storehouse, do not find the result, then think the map place name.
Can find out that the phonetic entry of this speech control system is necessary for steering order or map place name; And steering order is limited to map steering order, Navigation Control instruction and three kinds of instructions of map inquiry instruction, can't satisfy the demand of vehicle-mounted information service system.
(publication number: CN 101217584A) disclosed sound identification module uses unspecified person Chinese speech recognition technology in Chinese invention patent application " the voice command control method and the system that can be used for automobile "; Utilize microphone input voice command, voice command is discerned through EM220CN.
Therefore, the phonetic entry of this method also is limited on the order phrase.
Along with the development of vehicle-mounted information service system, the use scene of speech recognition on the navigating instrument terminal is at present: the selected earlier type that needs identification, and record button loquiturs then then, and system discerns and returns recognition result automatically afterwards, shown in Fig. 1.
Wherein action type is: search purposes ground, inquire about peripheral facility, inquiry intersection or the like.Though this application can bring certain facility for the user, its limitation is also very obvious.Mainly show as:
1) user needs to limit earlier action type to be identified.
Through limiting action type to be identified, the degree-of-difficulty factor minimizing for speech recognition has increased the query hit rate, but has brought counter productive to be, the user has carried out single stepping more, has reduced the convenience of user experience.
2) user interaction contents.
The content that the user says need be phrase, rather than sentence.Like the action type on the selected search purposes of user ground, the content that the user says is: " railway station, Beijing ", rather than " I will go to railway station, Beijing ", such different design share the mutual requirement of family natural language.
The utility model content
The purpose of the utility model is to provide a kind of voice operating system that uses the vehicle-mounted information service system of natural language.
The voice operating system of the vehicle-mounted information service system of the use natural language of the utility model comprises:
One navigating instrument is established record button and speech input device, in order to receive phonetic entry and to generate voice document;
One vehicle-mounted information service system voice server with the navigating instrument radio communication, receives the voice document that navigating instrument sends;
One voice Cloud Server; Establishing voice Cloud Server network with said vehicle-mounted information service system is connected; The reception voice document also is converted into text-only file and sends to the vehicle-mounted information service system voice server, after resolving through the vehicle-mounted information service system language server recognition result is sent navigating instrument.
Said vehicle-mounted information service system voice server comprises a language processing module; Said speech processing module contains Chinese dictionary and operator scheme storehouse, in order to the text-only file participle, and identifying operation type and operation keyword and operational attribute, and, carry out corresponding operating by it with the operation executing module that recognition result sends navigating instrument.
Said speech processing module also contains colloquial style speech dictionary, in order to the colloquial style speech in the text behind the removal participle.
Said action type comprises: the destination inquiry; The inquiry of periphery facility; The intersection inquiry; Push away under the music; Call.
Said Chinese dictionary adopts tree structure, and ground floor as index, adopts the Hash table storage with the lead-in of Chinese entry; The second layer; Adopt second word of linear precedence table storage entry; Remove identical word and form an orderly linear list; The linear list node to be to extract the interior code value ordering of Chinese character, and whether the pointer and one that store the linear list that the remainder with the word headed by this Chinese character constitutes simultaneously are the sign of speech; At the node of all the other levels of tree, adopt the word storing in order in the entry and the pointer of the linear list that points to its possible follow-up word of institute.
The utility model is also established a user behavior customary rule table, in order to mate to confirm action type and operation keyword and operational attribute with the text of failing to accomplish identification.
Said voice document is through encryption, compression, encoding process, said voice server to said voice document decode earlier, decompress(ion), decryption processing.
The utility model is also established a unidentified knowledge base, and the text in order to storage fails to discern deposits the operator scheme storehouse in after the parsing.
The utility model has realized using the voice operating of the vehicle-mounted information service system of natural language; The user only need be on navigating instrument says with colloquial exchange way and oneself wants the operation carried out; And do not need earlier selected action type, come machine is operated with the interactive mode of phrase again.
The utility model compared with prior art has following advantage:
1) be to have reduced user's operation steps.As shown in Figure 2, the utility model is reduced to the operation of two steps by original three steps operation;
2) use colloquial natural language, replace the interactive mode of original phrase/phrase.
Embodiment
The utility model at first will have been studied applied environment, scene, the flow process that the user uses the natural language recognition technology.Through navigation user being carried out modes such as phone return visit, questionnaire, forum's acquisition of information; Utilize the service sound-recording function of Telematics platform simultaneously, statistical study user's real demand is through analyzing analysis, the research of actual user's operating position; We utilize conclusion, sorting technique; Draw real application demand, confirmed all kinds of user's operation, wherein main action type comprises:
1) destination inquiry;
2) peripheral facility inquiry;
3) intersection inquiry;
4) push away under the music;
5) call.
Certainly, the continuous expansion along with information service also has more action type, but all can adopt the method and system of the utility model to realize voice operating.
As shown in Figure 3, the voice operating system of the utility model comprises three parts: navigating instrument, Telematics speech processes server, voice cloud.The voice operating flow process is following:
The first step: the user starts phonetic entry after pressing record button on the navigating instrument, and the mode navigation system with natural language issues operation information then.Navigational system generates voice document, with recording file encrypt, compression, encoding process, through communication, the recording file after handling is sent to the Telematics voice server;
Second step: the speech processes server is received voice document, decodes, decompress(ion), decryption processing, calls the interface of voice Cloud Server then, voice document is passed to the voice cloud handle.
The 3rd step: the voice Cloud Server is received voice document, voice document is handled generating TXT text (plain text) file, and returns to the natural language processing module of speech processes server.
The 4th step: after the natural language processing module is received the TXT text, carry out natural language processing, parse the operation that the user desires to reach,, recognition result is returned to the operation executing module of navigating instrument like inquiry POI destination operation.
The 5th step: navigating instrument is handled the recognition result of receiving, carries out corresponding operating.If Query Result then directly shows.If call, then directly dial.
Specify the identifying of the natural language text of the utility model below.
Because the natural language processing in vehicle-mounted service system is a specific application area; And be colloquial natural language interaction process flow process; Through research to Problem Areas, draw the just concrete application scenarios of this The Application of Technology, can conclude and sum up main application model; Use the natural language pattern matching algorithm to handle, can solve the application problem of natural language at onboard system.
As shown in Figure 4, identifying mainly comprises: several parts such as text participle, denoising, operation key word recognition, operator scheme are mated, recognition result returns.For the content of text that can not discern, the utility model provides system's self-learning function, can carry out constantly improving with abundant to library and crucial dictionary thereof, colloquial style speech dictionary.
One, text participle
At first to carry out word segmentation processing to mutual natural language processing; Participle technique at present commonly used has forward maximum match participle, reverse maximum match participle, based on the dictionary mechanisms of TRIE index tree, based on two minutes dictionary mechanisms etc. word for word, these participle techniques all respectively have relative merits in efficient, space utilization rate.
The Chinese dictionary of the utility model adopts tree structure.The ground floor of dictionary as index, adopts the Hash table storage with the lead-in of Chinese entry, to improve the seek rate of lead-in.Like this, lead-in becomes root node, and the speech that all lead-ins are identical becomes one group, belongs to same one tree.Because two words are more in Chinese; If the secondary word of entry is still stored with Hash table; Though can improve seek rate, it is very little that the size of this dictionary and the hugest TRIE tree construction are compared improvement, so at the second layer of forest; Adopt the linear precedence table to store second word of entry; Remove identical word and form an orderly linear list, the linear list node to be to extract the interior code value ordering of Chinese character, and whether the pointer and one that store the linear list that the remainder with the word headed by this Chinese character constitutes simultaneously are the sign of speech.At the node of all the other levels of tree, still adopt the word storing in order in the entry and the pointer of the linear list that points to its possible follow-up word of institute.In order to use binary chop to improve matching speed; All linear list below the second layer; But logical organization then is the word number that a Chinese character constitutes; Constitute like this that a support is word for word searched, store with Hash table at the ground floor lead-in, below successively according to the forest structure of linear ordered list storage.In the participle process, utilize above-mentioned data structure to carry out participle matching inquiry successively, solve the participle problem of text.
Two, denoising (removing the colloquial style speech)
Be mingled with the vocabulary of pet phrases such as hesitating, sew language, repeat in the language of spoken words through regular meeting, like " ", " ", " this " etc., the effect of denoising is that the colloquial style speech in the spoken natural language is removed.
One) colloquial style speech dictionary is set up
At first set up everyday spoken english dictionary S1, to commonly used spoken arrangement and statistics in the client's recording file that accumulates in the Telematics operation process, obtain dictionary S2 then.In S2,, the S1 storehouse done with S2 merge processing, obtain new S set 3 according to the different descending sorts of the word frequency of each speech height, i.e. colloquial style speech dictionary, the colloquial style speech in the S3 dictionary is according to occurring arranging from high to low of word frequency in daily life.
Two) denoising process treatment scheme
1) take out each participle Q1 among the text L successively, Q2.。。,Qn;
2) with Qi one by one with the S3 storehouse in each speech Pi match whole word only;
3) if mate successfully, then Qi is a spoken word, then removes, if the coupling failure then continues up to ending;
4) putting the participle phrase that makes new advances in order at last is the text behind the participle after the denoising.
Three, action type, operation keyword and operational attribute identification
One) operator scheme storehouse
Colloquial style language analysis in analysis through user in the Telematics platform being served recording file and the daily life; Conclude and sum up; The utility model has been set up the common natural language operator scheme storehouse of user; Operator scheme under this library storage is all types of, each type operations pattern comprises the operation keyword and the operational attribute of this pattern, and is as shown in table 1:
Table 1
Wherein, for every operator scheme under each action type, all having one or many s' operation keyword and operational attribute, as be numbered in the operator scheme of MA12 in " { } " to the operation key word, is operational attribute in " <>".
Two) user's acquired behavior rule list
The data of user's use habit behavior are through N1 " user is accustomed to collection module " in the car-mounted terminal equipment; Collect all user behaviors; As in a period of time; The time that the number of times that the user makes a phone call is 10 times, make a phone call, listen the song number of times of local storage, song names is listened song time, place or the like; Pass through wireless communication technology then; (like certain free time after the start) general's " user is accustomed to data " is transferred on the car machine in the Telematics speech processes server under certain condition, and by its N2 " user is accustomed to handling " resume module, N2 is from user's (recording user request service related information in the database the service log database on backstage; As ask the number of times 8 times of destination inquiry, to good friend's 3 numbers or the like of making a phone call to transfer) take out existing similar user and be accustomed to data; N2 carries out " POI inquiry use habit storehouse ", " storehouse of making a phone call ", " the inquiry perimeter data storehouse " that the data fusion statistics forms the user with the two according to action type ... Or the like, add up according to certain user according to the data of a plurality of data then, draw the number of times tabulation of certain operation of user; Then regular behavior is divided into from high to low according to the frequency of occurrences and sorts, form user's acquired behavior rule list.As shown in table 2:
Table 2
Three) operation key word recognition
1) take out each participle Qi among the natural language text L one by one, with the keyword MAKm among Qi and each the pattern rules MAj (MAK1, MAK2 ..., MAKn) mate;
2) calculate each keyword matching rate Rm=Qi/MAKm (R1, R2 ..., Rn);
3) calculate average matching rate Ri=(R1+R2+ then ... + Rn)/and n, if Ri, thinks then that the action of text L is the action of Aj bar greater than the matching rate value of agreement.Otherwise, continue coupling and go down;
4) if having no rule to satisfy text L, then use " user's customary rule table " to mate item by item, return to a plurality of selection results of user.Natural language like the user is: " blue and white porcelain ", when mating,, select to inquire about whether the information point of " blue and white porcelain " is arranged earlier according to the height of this user's use habit in user's customary rule table less than concrete rule, if having, then preserve; Whether continue inquiry then has the good friend to be the people of " blue and white porcelain "; If have; Preserve to get up to indicate to make a phone call or the like to this people, a plurality of contents that will preserve then and the related data of action need (like information point title, coordinate, buddy phone number etc.) send to terminal device, and the prompting user selects a certain service content; After the user selected, terminal car machine was carried out corresponding operation.
Four) action type and operational attribute identification
If after confirming that text L belongs to certain action type Ai, verify every operator scheme MAj in the operator scheme storehouse of each action type Ai.The attributes match rate of every MAj operator scheme will reach more than certain threshold value, can think that promptly text L meets this operator scheme MAj, carries out subsequent treatment according to this operator scheme then.
After the operator scheme storehouse was set up, every operator scheme all comprised limited operational attribute information.Like the POI inquiry, mode is expressed as: MA2i={Key}, < POIName>< DistrName >.Basically comprise two generic operation attributes in the POI inquiry, one is the POI title, and one is administrative realm name.System sets up a cover attribute database PDi and a cover matched rule PMi to each operational attribute.For example, set up administrative area attribute database PDi, store the administrative area title in all provinces in the whole nation, city, county, township/town, village for administrative realm name; And matched rule PMi is the matching degree of each speech among all Chinese characters and the PDi in the calculating < DistrName >; When matching degree reaches more than certain threshold value,, just can assert that this attribute is exactly the attribute in administrative area as 90%; And some of the PDi in belonging to indicates and contains this operational attribute information among the text L.
Four, operation is carried out
For the text L that matches operation, carry out corresponding operating and carry out.As inquire about POI, navigating instrument is divided and can be inquired about according to the administrative area, and shows Query Result.
For the text L that does not match any operation, then make a phone call artificial treatment user's operation requests to the user by the person of attending a banquet of speech processes service system meeting notification call central platform.
Should operate text L then, add in the unidentified knowledge base, analyze, resolve to the pattern of certain operation by manual work, as
MAk={key1…keyn},<Property1>,<Property2>,…,<Propertym>。
This operator scheme is joined in the operator scheme storehouse, and system can discern and parse the proper operation demand automatically after running into similar natural language next time.Wherein unidentified knowledge base is used for guaranteeing closed loop and system's self-perfection, learns.
The utility model has provided under the on-vehicle information service platform, utilizes the pattern matching algorithm of natural language to solve user and the free mutual problem of navigating instrument.The natural language speech method of operating of utilizing the utility model to propose can greatly improve the Experience Degree that user and navigating instrument carry out man-machine interaction, increases user's viscosity.