TWI698756B - System for inquiry service and method thereof - Google Patents

Abstract

A system for inquiry service and a method thereof are disclosed, including: using an interaction device to automate the conversation with the client and obtaining the inquiry information of the client, the inquiry information is automatically analyzed by the voice recognition device, the keyword analysis and determination device, and the big data analysis device to obtain a recommended store data corresponding to the inquiry information.

Description

System and method for query service

The present invention relates to a query service technology, especially a system and method for automatic query service.

The traditional telephone enquiry service is a highly labor-intensive and repetitive service. The enquiry service staff asks the customer. After listening to the question, the enquiry service can be started, and new staff may be encountered in the process. If the new staff Insufficient training to quickly find the answer to the question will delay a lot of query service time; secondly, when the online query service staff is busy, it will cause the customer who wants to query to wait for the query service online again Only when the personnel are connected can the inquiry service be performed, and a lot of waiting time is wasted.

Therefore, how to automatically perform telephone inquiry services and provide answers accurately and quickly is one of the urgent problems to be solved.

In order to overcome the deficiencies of the conventional technology, the present invention provides a query service system, which includes: an interactive device for obtaining query information sent by the client requesting the query service, wherein the query data includes voice query data; The phonetic recognition device converts the voice query data in the query information obtained by the interactive device into text query data; the keyword analysis and determination device receives the text query data, and finds the text from the dictionary file database Query data with the same data to obtain keyword features, where the keyword features include signboard names, county names, road names, or industry categories; and a big data analysis device that receives the key from the keyword analysis and determination device Character features, and the big data analysis device includes: a storehouse database, which stores storehouse data, the storehouse data includes signboard names, county names, road names or industry categories; the storehouse comparison module, from Find out the same pronunciation of the sign name in the keyword feature and at least one of the industry category, county and city name and road name in the industrial store database to obtain related industrial store information; the historical query database is stored The historical query record of the query service; and the recommendation module of the first associated store, which uses the inverse document frequency algorithm to calculate the query frequency of the associated store data in the historical query database, so that the query frequency is greater than the first predicted Set the threshold value of the associated store information as the first type of recommended store information, and then recommend the first type of recommended store information to the client.

In one embodiment, the historical query record includes customer satisfaction, and the big data analysis device further includes: a second associated store recommendation module, when the first associated store recommendation module cannot obtain the first type of recommended store When collecting data, find historical query records that are the same as the keyword and have good customer satisfaction from the historical query database to obtain related historical query records, and perform Apriori machine learning algorithm calculations on the related historical query records Obtain a recommended value so that the recommended value is greater than the second preset threshold The associated history query record is the second type of recommended store information, and then the second type of recommended store information is recommended to the client.

In one embodiment, the big data analysis device further includes: a feedback information module, which stores the processing between the client requesting the query service and recommending the recommended store data to the client as the historical query database The historical query record in.

In one embodiment, the query information further includes the county or city where the phone line of the client is located, and when there is no county or city name or road name in the text query data, the keyword analysis and determination device is where the phone line of the client is located Counties and cities are included in this keyword feature.

In one embodiment, the dictionary file database includes a county and city name dictionary file, a road name dictionary file, an industry name dictionary file, a sign name dictionary file, a scenic spot name dictionary file, a common name abbreviation dictionary file, a broken sound word dictionary file, or an industry category Synonym dictionary file.

The present invention also provides a method for query service, which includes the following steps: (1) obtain query information requested by the client for the query service, wherein the query data includes voice query data; (2) the voice query data in the query information Converted into text query data; (3) Find the same data as the text query data from the dictionary file database, and filter the redundant words in the same data to obtain the keyword features, where the keyword features Including the name of the signboard, the name of the county and city, the name of the road or the industry; (4) Find out the same pronunciation of the name of the sign in the keyword feature and at least the name of the industry, county and city from the database of the shop One is the same to obtain the related store information; (5) Use the inverse document frequency algorithm to calculate the related store information in The queried frequency in the historical query database is such that the associated store information whose frequency of query is greater than the first preset threshold is the first type of recommended store information; and (6) the recommended store information is recommended to the client.

In one embodiment, when the step (5) is unable to use the inverse document frequency algorithm to obtain the first type of recommended store information, then find out from the historical query database the same as the keyword characteristics and customer satisfaction It is a good historical query record to obtain the related historical query record, and then the related historical query record is calculated by the Apriori machine learning algorithm to obtain a recommended value, so that the recommended value is greater than the second preset threshold value of the related historical query The record is the second type of recommended store information.

In one embodiment, the method further includes the following steps: (7) The process between the client requesting the query service and recommending the recommended store information to the client is stored as the history in the historical query database Query records.

In one embodiment, the query information further includes the county or city where the phone line of the client is located, and when there is no county or city name or road name in the voice query data, the keyword analysis and determination unit is made to make the phone line of the client The county where you live is included in this keyword feature.

In one embodiment, the dictionary file database includes a county and city name dictionary file, a road name dictionary file, an industry name dictionary file, a sign name dictionary file, a scenic spot name dictionary file, a common name abbreviation dictionary file, a broken sound word dictionary file, or an industry category Synonym dictionary file.

It can be seen from the above that the present invention uses an interactive device to perform automated answering, actively greet the client, and obtain query information from the client, reducing the client If you wait too long for the bad feelings, then use the voice recognition device, keyword analysis and judgment device and big data analysis device to automatically analyze the query data of the client to obtain the recommended store data corresponding to the query information, so as to improve the training level of the customer service staff. 1. It is easy to cause query errors, and the turnover rate of personnel is high, resulting in a significant increase in learning costs, and the feedback information module in the big data analysis device simultaneously returns the processing results of the system and customer service personnel to the historical query database In this way, the accuracy rate of recommended store information generated from the historical query database is improved, and the response to the client is accurately and quickly returned.

1‧‧‧Interactive device

2‧‧‧Voice recognition device

3‧‧‧Keyword analysis and judgment device

4‧‧‧Big data analysis device

5‧‧‧Text-to-speech module

6‧‧‧Client

7‧‧‧Switch

8‧‧‧Customer Service Staff

11‧‧‧Central Control Module

12‧‧‧Interactive Module

21‧‧‧Signal processing module

22‧‧‧Language Classification Module

23‧‧‧Voice to text module

24‧‧‧ pronunciation dictionary file

31‧‧‧Dictionary file database

32‧‧‧Semantic Analysis Module

33‧‧‧Word Segmentation Processing Module

34‧‧‧Keyword Extraction Module

41‧‧‧Workshop Home Database

42‧‧‧Industry shop home comparison module

43‧‧‧History query database

44‧‧‧The first related store recommendation module

45‧‧‧Recommendation module of the second related store

46‧‧‧Feedback Information Module

100‧‧‧System

200‧‧‧Method

S201~S208‧‧‧Step

Figure 1 is a schematic diagram of the query service system of the present invention; and Figure 2 is a flowchart of the steps of the query service method of the present invention.

The following specific examples illustrate the implementation of the present invention. Those familiar with the art can easily understand the other advantages and effects of the present invention from the content disclosed in this specification.

It should be noted that the structures, proportions, sizes, etc. shown in the drawings in this manual are only used to match the contents disclosed in the manual for the understanding and reading of those familiar with the art, and are not intended to limit the implementation of the present invention. Therefore, it does not have any technical significance. Any structural modification, proportional relationship change, or size adjustment should still fall within the scope of the present invention without affecting the effects and objectives that can be achieved. The technical content disclosed by the invention can be covered. At the same time, the terms such as "first", "second" and "one" cited in this manual are only for ease of description. It is not intended to limit the scope of implementation of the present invention, and the change or adjustment of its relative relationship shall be regarded as the scope of implementation of the present invention without substantial changes to the technical content.

Please refer to Figure 1, which is a schematic diagram of the query service system 100 of the present invention. The system includes an interactive device 1, a voice recognition device 2, a keyword analysis and determination device 3, a big data analysis device 4, and a text-to-speech module 5. The interactive device 1, a voice recognition device 2, and a keyword analysis and determination device 3 The big data analysis device 4 can be implemented in the same device to be electrically connected to each other, or implemented in different devices to be electrically connected to each other.

The interactive device 1 interacts with the client 6 when the client 6 requests a query service from the system 100. The interactive device 1 includes a central control module 11 and an interactive module 12.

In one embodiment, the client 6 requests the system 100 to query the service through the switch 7, but it is not limited to this.

In one embodiment, the client 6 is a local telephone or a mobile device, but it is not limited to this.

The central control module 11 receives the inquiry information sent by the client 6 to the system 100 for the inquiry service, and is responsible for the process control of the interaction between the system 100 and the client 6 in the entire inquiry operation.

In one embodiment, the query information includes the incoming line number of the client terminal 6, the county or city where the phone line is located, the incoming line device, the incoming line query time, and the voice query data.

The interactive module 12 is a built-in one-to-one answer sentence automatic generation model, and makes the answering sentence automatic generation model generate a dialogue interactive with the client 6 Character string, the automatic generation model of the answer sentence is built-in the corpus required for natural persons to chat. The content of the corpus contains a large number of contexts and responses. When the voice query data of the client 6 is converted into text query data When inputting the answer sentence automatic generation model, the answer sentence automatic generation model will disassemble the context of the comparison input, and select the best reply from the corpus as the dialogue string. The main function of the interaction module 12 is to maintain interaction with the client 6 and guide the client to answer questions.

In one embodiment, the answer sentence automatic generation model is generated through machine learning algorithm training, wherein the machine learning algorithm includes long-term short-term memory (LSTN) or recurrent neural network (RNN).

The text-to-speech module 5 receives the dialogue string generated by the interactive module 12 and converts the dialogue string into a dialogue voice to transmit the dialogue voice to the client 6.

In one embodiment, the text-to-speech module 5 transmits the reply voice to the client 6 through the switch 7.

When the client 6 makes a service inquiry request to the system 100, the central control module 11 first calls the interactive module 12 to generate a dialogue string, and converts the dialogue string through the 5 TTS text-to-speech component Play the voice to the client 6 to guide the customer to answer questions. During the process, the voice recognition device 2 will be called continuously to analyze the answer of the client 6 and confirm whether there is a recognition result, and the question and answer process will be repeated several times. If the recognition result of the voice recognition device 2 can produce text query data for query , It will continue to be delivered to the background (that is, the keyword analysis and determination device and the big data analysis device) for the query process; if the effective information cannot be identified in the process, or the intention after the semantic analysis is not the query Inquiry, the call will be forwarded to customer service staff 8 for service.

The voice recognition device 2 receives the query information obtained by the interactive device 1 to convert the voice query data in the query information into the text query data. The voice recognition device 2 includes a signal processing module 21 and a voice-to-text model Group 23 and pronunciation dictionary file 24.

The signal processing module 21 uses digital signal processing (DSP) to remove noise in the voice query data.

The pronunciation dictionary file 24 stores the pronunciation combinations of words to provide the application of word-to-phoneme or phoneme-to-character conversion according to the combination. The phoneme is the smallest pronunciation unit in the language, such as the phonetic symbols of Mandarin or the phonetic symbols of English. Not limited to this.

In one embodiment, the pronunciation dictionary file 24 is a pronunciation dictionary file of Mandarin or English, and may also be a pronunciation dictionary file of other languages, but is not limited to this.

The voice-to-text module 23 converts the voice query data from which noise has been removed into the text query data.

In one embodiment, the speech-to-text module 23 builds a first-type neural network based on the trained acoustic model (HMM) and pronunciation dictionary file 24, together with the neural network model BPNN (Backpropagation Neural Network) Lulai converts the voice query data from which the noise has been removed into a single word combination, and then builds a second type of neural network based on the trained language model and the neural network model BPNN to convert the single word combination into a complete And make the complete text sentence as the text query data.

In the first type of neural network, the neural network model BPNN cooperates with the sound The function of the learning model is to replace the voice query data with noise removed into plural phonemes. The function of the neural network model BPNN with the pronunciation dictionary file 24 is to convert the plural phonemes into plural words or words. Single word combination.

In the second type of neural network, the function of the neural network model BPNN in conjunction with the language model is to convert the word or word (for example, the combination of single characters) into the text sentence.

In one embodiment, the speech recognition device 2 further includes a language classification module 22, but it is not limited thereto.

The language classification module 22 performs language classification on the voice query data from which noise has been removed according to the first type of neural network, wherein the language classification is Mandarin or English, but not limited to this.

In one embodiment, the language classification module 22 is integrated in the speech-to-text module 23 for the first type of neural network to perform the language classification on the voice query data from which noise has been removed and generate the word Combination, but not limited to this.

The keyword analysis and determination device 3 receives the text query data obtained by the speech-to-text module 23 to find the keyword characteristics from the text query data. The keyword analysis and determination device 3 includes a dictionary file database 31, The semantic analysis module 32, the word segmentation processing module 33, and the keyword extraction module 34.

The dictionary file database 31 contains the county and city name dictionary file, the road name dictionary file, the industry name dictionary file, the sign name dictionary file, the scenic spot name dictionary file, the broken sound word dictionary file, the common name abbreviation dictionary file (for example: Xingtian Palace=恩The main palace and other information) or industry-specific synonym dictionary files, where the content of the synonym dictionary file is to record the possible conversion of Chinese vocabulary. For example, synonyms for snack bars can also be used. The names are restaurants, eateries, and simple meals, and the synonyms of iPhone are Aifeng, Crazy, Smart Phone, Apple, but not limited to this.

The semantic analysis module 32 identifies the query intent of the client 6 from the text query data. The method of identifying the query intent is to determine the subject and verb, for example: I want to check, please help me check, I want to check , I want to check, help me find, or I want to find, etc., you can determine the intent of the query, but not limited to this.

Since the task of the system 100 is very clear, it is to ask the client 6 to provide the store name or the name of the scenic spot, and then help the client 6 to query related telephone information, so the client 6 questions tend to be short questions, such as: I want to check XXX restaurants The phone number of XX, the phone number of XX hospital for me, the phone number of XXX who has been on the news recently, can you find XXX for me or I want to know the phone number of XX, etc. There are many ways to ask, but the intentions are all The same, therefore, the semantic analysis module 32 only needs to determine whether the text query data is a query. If the query intention cannot be determined, the central control module 11 will control the dialogue interaction with the client 6 to the customer service staff 8 deal with.

The word segmentation processing module 33 performs word segmentation with different numbers of characters on the text query data to obtain word combinations with different numbers of characters, and then compares whether there is a matching word combination in the dictionary file database 31, and Output the matched word combination as a keyword array.

The keyword extraction module 34 deletes redundant words in the keyword array to obtain the keyword characteristics, where the keyword characteristics include signboard names, county names, road names, or industry categories, and the redundant words High-frequency words often appear in general conversations, such as subject words (you, me, him, miss, mr, hello,), Interjections (ㄟ, ah, ah, oh, oh, oh, yo, yo...), verbs (want, want to check, want to know, help me find, please help me, tell me,...) etc., but not Limit this. The purpose of the keyword extraction module 34 is to filter the redundant words used by natural persons in conversation, and only leave keywords that may be the name of the store.

In one embodiment, the keyword analysis and determination device 3 further includes receiving the query information from the voice recognition device 2, and when the word segmentation processing module 33 is unable to obtain from the county and city name dictionary file or the road name dictionary file in the dictionary file database 31 When the matching word combination is found (that is, there is no county or city name or road name in the text query data), the keyword extraction module 34 includes the county and city where the client's phone line in the query information is located in the key Character features, but not limited to this.

The big data analysis device 4 receives the keyword feature obtained by the keyword extraction module 34, and finds recommended store information from the database in the big data analysis device 4 to the client 6 according to the keyword feature. Among them, the big data analysis device 4 includes a shopkeeper database 41, a shopkeeper comparison module 42, a historical query database 43, a first related shop recommendation module 44, a second related shop recommendation module 45, and a feedback information module 46.

The industrial store database 41 stores industrial store information, which includes signboard name, county name, road name, or industry category.

In one embodiment, the shopkeeper database 41 is a specially designed information aggregation table. The web crawler program automatically obtains the registered shopkeeper's name from the Department of Commerce of the Ministry of Economic Affairs every day, and integrates the 104 number check Registered in the station as being in effective use, but not yet dismantled the information of the industrial shop, the professional staff will arrange and classify the pronunciation (that is, the phonetic symbol) and the industry classification, but not Limit this.

The shop comparison module 42 finds the shop shop data from the shop shop database 41 that has the same pronunciation as the name of the sign in the keyword feature to obtain the shop shop data with the same pronunciation of the sign, and compare it to the sign When the homophonic industrial store data has at least one of the industry category, county name, and road name of the keyword feature, the at least one identical industrial store data of the sign with the same sound is the related industrial store data.

The historical query database 43 stores the historical query records of the query service. The historical query records include the text query data queried by the client 6 and the corresponding reply answer, and the call summary record whether the client 6 is correct for the answer Satisfaction.

The first related store recommendation module 44 uses the inverse document frequency (IDF) algorithm to calculate the query frequency of the related store data in the historical query database 43, so that the query frequency is greater than the first preset threshold. The associated store information is the first type of recommended store information, and the first type of recommended store information is then voice-recommended to the client 6 through the text-to-speech module 5.

Inverse document frequency (IDF) calculation formula is as follows, where D represents the total number of query records in the historical query record where the industry is restaurant category, and Dw represents the number of records of the signboard name w appearing in the following formula:

The second associated store recommendation module 45, when the first associated store recommendation module 44 is unable to obtain the first type of recommended store data, finds out from the historical query database 43 the same as the keyword feature and the customer satisfaction is Good calendar Historical query records to obtain related historical query records, and use Apriori machine learning algorithm to calculate the feature value of the related historical query record, and make the related historical query record whose feature value meets the preset recommendation condition as the second type of recommendation Store information to recommend the second type of recommended store information to the client 6 through the text-to-speech module 5 by voice.

The calculation formula of the Apriori machine learning algorithm is as follows:

X is the store X queried by the client 6, Y is the store Y found from the historical query record (that is, the second type of recommended store data), support, confidence, and Lift are the characteristic values, where frq is the calculation The number of times the store has been queried in the historical query record, frq(X) refers to the number of times store X has been queried in the historical query record, frq(X,Y) refers to the historical query that contains both store X and store Y The number of queries in the record, N is all the number of queries in the historical query record, Support is the support degree, which represents the proportion of all the number of queries N, X and Y appear at the same time, Supp(X) is the number of all queries In N, the proportion of store X appearing, Supp(Y) is the proportion of store Y appearing in all the number of queries N, Confidence is the degree of confidence that represents the proportion of all queries where store X appears at the same time as the query store Y , Lift is the gain value, Lift value less than 1 means that the store X and Y are negatively correlated, and equal to 1 means that the store X and Y are irrelevant. Greater than 1 means that both store X and Y are positively correlated. When the Lift value is positively correlated and the support value and confidence value both exceed the second preset threshold, then this association rule is established, and store Y can be recommended To the client 6.

When the modules in the big data analysis device 4 cannot find the first type of recommended store information and the second type of recommended store information based on the keyword characteristics, the central control module 11 controls the communication with the client 6 The dialogue interaction is transferred to the customer service staff 8 for processing, and the client 6 is notified of the third-type recommended store information obtained from the processing.

The feedback information module 46 stores the processing process between the client 6 requesting the query service and recommending each of the recommended store information to the client 6 as the historical query record in the historical query database.

Please refer to Figure 2, which is a flowchart of the steps of the method 200 for querying services of the present invention. The method includes the following steps: in step S201, the query information requested by the client 6 for the query service is obtained, wherein the query data includes the incoming line number of the client 6, the county or city where the phone line is located, the incoming device, and the incoming line Query time and voice query data.

In one embodiment, in step S201, the interactive module 12 is used to generate a dialogue string, and the dialogue string is converted into a dialogue voice through the text-to-speech module 5 and sent to the client 6, thereby guiding the client 6 to answer The question is to obtain the query information from the answer question, but not limited to this.

In step S202, the voice query data in the query information is converted into text query data.

In one embodiment, the step S202 is performed by the voice recognition device 2 Convert the voice query data in the query information into text query data.

In step S203, find the same data as the text query data from the dictionary file database 31, and filter the redundant words in the same data to obtain keyword features, where the keyword features include the name of the sign , County and city name, road name or industry category.

The step S203 further includes the following steps: use the word segmentation processing module 33 to segment the text query data with query intent to obtain word combinations with different numbers of characters, and then compare the dictionary file database Whether the same word combination appears in 31, and output the same word combination as a keyword array; and

The keyword extraction module 34 is used to delete redundant words in the keyword array to obtain the keyword characteristics.

In one embodiment, at the beginning of step S203, the semantic analysis module 32 is used to identify the query intention of the client 6 from the text query data. If so, the step of using the word segmentation processing module 33 is performed. If not, then Return to step S201 to obtain another query information, but not limited to this.

In one embodiment, when the word segmentation processing module 33 cannot compare the word combination from the county and city name dictionary file or the road name dictionary file in the dictionary file database 31 (that is, there is no county or city in the text query data). Name or road name) to make the keyword retrieval module 34 include the county and city where the phone line of the client in the query information is located in the keyword feature, but it is not limited to this.

In step S204, find out from the industrial store database 41 the pronunciation of the sign name in the keyword feature is the same as the name of the industry, county and city, and At least one of the road names is the same in order to obtain the information of the related shops.

In one embodiment, the step S204 is to find from the industrial store database 41 from the store house database 41 through the store house comparison module 42 to obtain the store house data with the same pronunciation of the signboard name in the keyword feature When the store data is compared with the information of the industrial store with the same sound of the signboard and at least one of the industry category, county and city name and road name of the keyword feature is the same, the information of the at least one identical industrial store with the same sound of the signboard is Go to the related store's information.

In step S205, the inverse document frequency algorithm is used to calculate the inquired frequency of the related store information in the historical query database, so that the related store information whose inquired frequency is greater than the first preset threshold is the first type of recommended store data.

In one embodiment, the step S205 is performed through the first associated store recommendation module 44.

In step S206, when the inverse document frequency algorithm cannot be used to obtain the recommended store information of the first type in the step S205, it is found from the historical query database that the feature is the same as the keyword and the customer satisfaction is good Historical query records to obtain associated historical query records, and then perform Apriori machine learning algorithm calculations on the associated historical query records to obtain a recommended value, so that the associated historical query record with the recommended value greater than the second preset threshold is the first The second category recommended store information.

In one embodiment, the step S206 is performed through the second associated store recommendation module 45.

In one embodiment, when the step S206 cannot be based on the keyword feature When the second type of recommended store information is found, the central control module 11 will control the dialogue and interaction with the client 6 to the customer service staff 8 for processing, and recommend the processed third type of recommended store information to the client 6 .

In step S207, the recommended store profile is recommended to the client 6.

In one embodiment, the central control module 11 recommends the first-type recommended store information obtained in step S205 to the client 6, and when the first-type recommended store information cannot be obtained in step S205, the The second type of recommended store information obtained in step S206 is recommended to the client 6. If the second type of recommended store information cannot be obtained in step S206, the customer service staff 8 recommends the third type of recommended store information to the client 6. Client 6, but not limited to this.

In step S208, the process between the client requesting the query service and recommending the recommended store information to the client is stored as the historical query record in the historical query database.

In one embodiment, the step S208 is executed by the central control module 11 or the customer service personnel 8.

The system and method of the query service of the present invention is applicable to the 104 directory. The following provides a scenario example to illustrate the system and method of the query service of the present invention: After the customer (ie client 6) calls the 104 directory, then The service is assigned to the interactive device 1 by the switch 7, and calls the central control module 11 to start a standard dialogue process. First, the interactive module 12 greets the customer and guides the customer to answer questions, and the central control module 11 obtains the customer from the switch 7. Incoming related resources The X set (that is, the query information) obtained is = {X1 customer incoming line number=0910XXXXXX, X2 phone line in the county = Daan District, Taipei, X3=mobile phone incoming line, X4 customer incoming line query time=20XX/04/ 23 11:45, X5=Voice multimedia file of the questions answered by the customer through the phone}.

Hand over X5 to the speech recognition device 2 for processing. After processing through the signal processing module 21, a clearer digital signal X5' without background noise is obtained, and then the language classification module 22 is called to compare the acoustic model and pronunciation dictionary files After 24, after confirming that X5' is the Chinese pronunciation, then X5' will be handed over to the speech-to-text module 23 for processing to obtain the Y set (that is, the text query data) as {Y1 customer question string=Hello, miss hello, I think To check the Ding Tai Fung restaurant on Zhongshan Road, Zhonghe District, please check the reservation phone number for me, thank you}.

Send the Y set to the keyword analysis and determination device 3, first call the semantic analysis module 32 to perform semantic analysis on the string set Y, analyze whether its purpose is a query, and if so, start calling the dictionary file database 31 (road name dictionary, Industry name dictionary, scenic spot name dictionary, sign name dictionary), word segmentation processing module 33 performs Chinese word segmentation comparison operations, cuts Chinese sentences into word combinations with different numbers of characters, and filters the words through the keyword extraction module 34 Subjects, verbs, interjections, auxiliary words and other superfluous words used by natural persons in the combination (for example: hello, miss hello, I want to check, please help me check, thank you), after the above filtering and comparison processing operations After the comparison result of each dictionary file, the keyword features shown in Table 1 can be obtained: "Ding Tai Fung", "Ding Tai Feng", "Ding Tai Feng", "New Taipei City Zhonghe District", "Zhongshan Road", " restaurant".

Keyword feature required to generate query database Y'=(Y'1 store name=Ding Tai Fung, Ding Tai Feng or Ding Tai Feng, Y'2 store in the county and city = New Taipei City Zhonghe District, Y'3 store Road section name=Zhongshan Road, Y'4 store industry type=restaurant}; at this time, the information X set and query obtained by exchange 7 are input into the shop comparison module 42 to compare the shop data The data in library 41, because most Y'1 industrial shop signs are in Chinese, if only the sound-to-word processing is performed through the acoustic model, there will be a lot of candidate combinations due to the factors of homophones and broken sound characters, and the output is unknown. which candidate combination, and therefore the need to advance Y'1 store name into phonetic symbols "ㄉ ㄞ ^{ㄊ a`} ㄈ eng eng ", and this phonetic alphabet pronunciation of the store field than industrial and commercial stores in the database 41, You will find that the stores with the same pronunciation have "Dingtaifeng Mountaineering Supplies Store", "Dingtai Style Thai Cuisine", "Ding Tai Fung Xiao Long Bao" and equivalent signboards, and the industry name that the customer asked is "restaurant". It is a general term for the general catering industry. It is not the "xiaolongbao", "climbing supplies", and "Thai cuisine" of the store industry registered in the storehouse database 41. Therefore, it is necessary to compare the industry-specific data with the dictionary file data The industry-specific synonym dictionary in the database 31 is compared.

If the industry category that the customer asks (ie the restaurant shown in Table 1), through After comparing the industry synonym dictionary, it is found that the industry name registered in the industry database 41 of the same pronunciation store is the same (for example, the synonym of restaurant is equal to Xiaolongbao or cuisine), the confidence level is 60, and the comparison is complete After the synonymous word of the industry category, the county and city and the road name are compared. If the store with the same pronunciation is registered in the county and city in the industrial store’s home database 41 and the county and city inquired by the customer (as shown in Table 1, Zhonghe, New Taipei City) District) the same but different road names, the confidence is 70, and if the same pronunciation of the store in the industrial store home database 41 registered in the county and road name and the customer inquired about the county, city and road name ( When the Zhongshan Road shown in Table 1 is the same, the confidence level is 90, and the data in the industrial store database 41 of the store with the same pronunciation with the highest confidence level (ie, the store’s store data) is the affiliated store’s store. data.

In this case, the customer inquired about Ding Tai Fung Restaurant, but it was recorded in the industrial store’s home database 41 that there was no Ding Tai Fung Restaurant in the customer’s registered information. After comparing the synonym dictionary files, it was found that the synonyms of Yebei Restaurant contained Xiao Long Bao and cuisine, so In the definition of the signature name, Ding Tai Fung Restaurant = Ding Tai Fung Xiao Long Bao and Ding Tai Fung Restaurant = Ding Tai Feng Thai Restaurant.

Generally speaking, the probability of two signs with homophones and different characters appearing on the same road is quite low. This is rarely the case. This example is an extreme case, mainly to show that this method can solve the customer's unclear store name , If you don’t know the name of the street where the store is located, you can also find the relevant store. Even when the customer drives out without knowing his location, the present invention can determine that the customer is inquiring through the phone number through the query parameters. The registration form of the base station can be used to check the information of the base station and indirectly know where the customer is. of Area, such as "Da'an District, Taipei City". If the customer enters the line through the local call, they can know the installation address more accurately, and can more accurately assist the customer to search for stores near the geographic location.

Next, input the name of the signboard of these homophones (that is, the store data of the associated shop) into the first associated store recommendation module 44, and use the built-in IDF algorithm to statistically analyze the signboard names of these homophones in the historical query database 43 The frequency of being queried within, and finally the result Z{store name, store name similarity, store registration phone number} set, and sorted according to IDF score, as shown in Table 2 below.

The output of the above case is {Z1=Ding Tai Fung Xiao Long Bao, Z2=10.38}, {Z'1=Ding Tai Feng Thai Cuisine, Z'2=7.11}; where {Z1=Ding Tai Feng Xiao Long Bao, Z2 is IDF value=10.38} has a higher score. After the comprehensive evaluation, assuming that both Z1 and Z2 exceed the preset threshold, the first level The joint store recommendation module 44 will send the Z1 store name and the phone number registered by this store to the central control module 11, and play a message through the text-to-speech module 5: "Hello, you just asked about the location in {新北市}{Ding Tai Fung Restaurant} in {Zhonghe District}{Zhongshan Road}, the registered phone number is {02-XXXX-XXXX}", and another similar one is in {New Taipei City} {Zhonghe District} {Zhongshan Road} {Top Thai Style Thai Cuisine}, the registration phone number is {02-XXXX-YYYY}.

It can be seen from the above that the present invention uses the interactive device to automate the answering, actively greets the client, and obtains the client's query information, reducing the client's bad feeling of waiting too long, and then using the voice recognition device, the keyword analysis and determination device and the big The data analysis device automatically analyzes the query data of the client to obtain the recommended store data corresponding to the query information, so as to improve the problem of different training levels of customer service personnel, which is easy to cause query errors, and the turnover rate of personnel is high, resulting in a significant increase in learning costs And the feedback information module in the big data analysis device simultaneously feeds back the processing results of the system and customer service personnel to the historical query database, thereby improving the accuracy of recommended store data generated from the historical query database.

1‧‧‧Interactive device

2‧‧‧Voice recognition device

3‧‧‧Keyword analysis and judgment device

4‧‧‧Big data analysis device

5‧‧‧Text-to-speech module

6‧‧‧Client

7‧‧‧Switch

8‧‧‧Customer Service Staff

11‧‧‧Central Control Module

12‧‧‧Interactive Module

21‧‧‧Signal processing module

22‧‧‧Language Classification Module

23‧‧‧Voice to text module

24‧‧‧ pronunciation dictionary file

31‧‧‧Dictionary file database

32‧‧‧Semantic Analysis Module

33‧‧‧Word Segmentation Processing Module

34‧‧‧Keyword Extraction Module

41‧‧‧Workshop Home Database

42‧‧‧Industry shop home comparison module

43‧‧‧History query database

44‧‧‧The first related store recommendation module

45‧‧‧Recommendation module of the second related store

46‧‧‧Feedback Information Module

100‧‧‧System

Claims (10)

A query service system includes: an interactive device for obtaining query information sent by a client requesting a query service, wherein the query data includes voice query data; a voice recognition device is the query information obtained by the interactive device The voice query data is converted into text query data; the keyword analysis and determination device receives the text query data, and finds the same data as the text query data from the dictionary file database to obtain the keyword characteristics, where the Keyword features include signboard names, county and city names, road names, or industry categories; and a big data analysis device that receives the keyword features from the keyword analysis and determination device, and the big data analysis device includes: industrial store information The database is to store the data of the industrial store. The store data includes the name of the sign, the name of the county or city, the road name or the industry category; the industrial store comparison module is to find out the characteristics of the keyword from the database of the industrial store The name of the signboard has the same pronunciation and at least one of the industry category, county and city name, and road name is the same to obtain the information of the related industrial store; the historical query database stores the historical query records of the query service; and the recommended model of the first associated store Group, which calculates the inquired frequency of the related store data in the historical inquiry database through the inverse document frequency algorithm, so that the inquired frequency is greater than the first preset The associated store information of the threshold value is the first type of recommended store information, and the first type of recommended store information is then recommended to the client. For example, the system described in item 1 of the scope of patent application, wherein the historical query record includes customer satisfaction, and the big data analysis device further includes: a second affiliated store recommendation module, which acts as the first affiliated store recommendation module When the first type of recommended store data cannot be obtained, find the historical query records that are the same as the keyword and have good customer satisfaction from the historical query database to obtain the associated historical query records, and the associated historical query The record is calculated by the Apriori machine learning algorithm to obtain the feature value, so that the associated historical query record whose feature value meets the preset recommendation condition is the second type of recommended store information, and then the second type of recommended store information is recommended to the client. For example, the system described in item 1 or 2 of the scope of patent application, wherein the big data analysis department further includes: a feedback information module, which is between the client requesting the query service and recommending the recommended store information to the client The processing process is stored as the historical query record in the historical query database. For example, the system described in item 1 or 2 of the scope of patent application, wherein the query information further includes the county or city where the phone line of the client is located, and the text query data does not include the county or road name, the keyword analysis The determining device incorporates the county and city where the phone line of the client is located into the keyword feature. For example, the system described in item 1 of the scope of patent application, wherein the dictionary file database includes county and city name dictionary files, road name dictionary files, and industry name dictionary files File, sign name dictionary file, scenic spot name dictionary file, common name abbreviation dictionary file, broken sound word dictionary file or industry-specific synonyms dictionary file. A method for querying services includes the following steps: (1) obtaining query information requested by the client for query services, wherein the query data includes voice query data; (2) converting the voice query data in the query information into text query Data; (3) Find the same data as the text query data from the dictionary file database, and filter the redundant words in the same data to obtain the keyword features, where the keyword features include the name of the sign, County and city name, road name or industry category; (4) Find from the industrial store’s home database that the pronunciation of the sign name in the keyword feature is the same and at least one of the industry category, county name and road name is the same, In order to obtain the related store information; (5) Use the inverse document frequency algorithm to calculate the inquired frequency of the related store data in the historical query database, so that the associated store data whose inquired frequency is greater than the first preset threshold is The first type of recommended store information; and (6) recommend the recommended store information to the client. For example, the method described in item 6 of the scope of patent application, wherein, when the step (5) can not use the inverse document frequency algorithm to obtain the information of the first type of recommended store, then find out from the historical query database and the The historical query records with the same keyword characteristics and good customer satisfaction are obtained to obtain the related historical query records, and then the related historical query records are Apriori The machine learning algorithm calculates and obtains the feature value, so that the associated historical query record whose feature value meets the preset recommendation condition is the second type of recommended store information. For example, the method described in item 6 or 7 of the scope of the patent application, wherein the method further includes the following steps: (7) The process between the client requesting the query service and recommending the recommended store information to the client is stored Create the historical query record in the historical query database. For example, the method described in item 6 or 7 of the scope of patent application, wherein the query information further includes the county or city where the phone line of the client is located, and when the voice query data does not contain the name of the county or city, the keyword The analysis and determination unit includes the county and city where the phone line of the client is located in the keyword feature. Such as the method described in item 6 of the scope of patent application, wherein the dictionary file database includes a county and city name dictionary file, a road name dictionary file, an industry name dictionary file, a sign name dictionary file, a scenic spot name dictionary file, and a common name abbreviation dictionary file , Poyin word dictionary file or industry-specific synonyms dictionary file.

Images