TWI869149B - Knowledge recommendation system and knowledge recommendation method - Google Patents

Abstract

A knowledge recommendation method includes: reading the knowledge numbers, the knowledge names and the knowledge keyword groups corresponding to each of the knowledge numbers; calculating the similarities between the voice data and each of the knowledge names, and filtering out multiple candidate knowledge numbers based on the similarities; identifying and marking the voice keywords in the voice data based on each of the keyword groups; adding the time tags to the voice keywords to become the term keywords; comparing at least one knowledge keyword with the term keywords and filtering out multiple knowledge numbers to be recommended from the candidate knowledge numbers; calculating the keyword weight scores corresponding to each of the knowledge numbers to be recommended based on the weights corresponding to each of the knowledge keyword groups; and sorting the knowledge numbers to be recommended based on the keyword weight scores and the similarities corresponding to each of the knowledge numbers to be recommended from high to low to produce knowledge recommendation sorting results.

Description

Knowledge recommendation system and knowledge recommendation method

The present invention relates to a recommendation technology, and in particular to a context-based knowledge recommendation system and a knowledge recommendation method.

Telecom operators, banks, other service industries or counter staff in government agencies face a large number of customer inquiries or requests for industry-related services every day. The complexity of industry-related services will increase the service time for counter staff to handle customer needs. If the counter staff is a new employee who is still in training, they may not be able to provide the most appropriate industry-related services to customers in a timely manner, and the training time of counter staff will also be prolonged. Therefore, according to the content of customer counter conversations, real-time service knowledge recommendations are provided for counter personnel to choose from, helping counter personnel to recall service processes and improve customer satisfaction. In addition to reducing training time and reducing corporate training costs, this is an issue that needs to be solved urgently.

The present invention provides a knowledge recommendation system, including a memory and a processor. The memory stores multiple modules; the processor is coupled to the memory to read and execute multiple modules to generate knowledge recommendation ranking results, wherein the multiple modules include a knowledge data module, a voice similarity calculation module, a keyword recognition module, a keyword deadline module, a keyword comparison module, a keyword weight module and a knowledge recommendation ranking module. The knowledge data module is used to store a plurality of knowledge numbers, a knowledge name corresponding to each knowledge number, and a plurality of knowledge keyword groups, wherein each knowledge keyword group has a corresponding weight and includes at least one knowledge keyword. The voice similarity calculation module is used to calculate the similarity between the voice data and the knowledge name corresponding to each knowledge number, and select a plurality of candidate knowledge numbers from the knowledge numbers according to the similarity. The keyword recognition module is used to recognize and annotate a plurality of voice keywords in the voice data based on the keyword group corresponding to each knowledge number. The keyword expiration module is used to add a time tag to the voice keyword to form a plurality of expiration keywords. The keyword matching module is used to match at least one knowledge keyword with an expiration keyword, and to filter out a knowledge number to be recommended from candidate knowledge numbers based on the expiration keyword that matches the knowledge keyword. The keyword weight module is used to calculate a keyword weight score corresponding to each knowledge number to be recommended according to the weight corresponding to each knowledge keyword group. The knowledge recommendation ranking module is used to sort the knowledge numbers to be recommended from high to low according to the keyword weight score and similarity of each knowledge number to be recommended to generate the knowledge recommendation ranking results.

In one embodiment, the module further includes a dialogue storage module for storing voice data.

In one embodiment, the voice data is composed of multiple sound signals received by two receiving devices.

In one embodiment, the knowledge keyword group includes a category keyword group, a subcategory keyword group, and a general keyword group.

In one embodiment, the category keyword group corresponds to the first weight, the subcategory keyword group corresponds to the second weight, and the general keyword group corresponds to the third weight, wherein the first weight is greater than the second weight, and the second weight is greater than the third weight.

In one embodiment, the voice similarity calculation module is further used to calculate the similarity between the voice data and the knowledge name corresponding to each knowledge number through a string matching algorithm.

In one embodiment, before the keyword weight module calculates the keyword weight score corresponding to each knowledge number to be recommended according to the weight corresponding to each knowledge keyword group, the keyword deadline module is further used to capture part of the knowledge numbers to be recommended corresponding to part of the deadline keywords with time tags within a set time period.

In one embodiment, the knowledge data module is further used to receive at least one newly added knowledge keyword corresponding to at least one of the knowledge keyword groups.

In one embodiment, the knowledge recommendation ranking result includes some recommended knowledge numbers ranked first among the knowledge numbers to be recommended.

In one embodiment, the knowledge recommendation system further includes a display, which is coupled to the processor to display the recommended knowledge description corresponding to each of the recommended knowledge numbers.

The present invention also provides a knowledge recommendation method, comprising: reading a plurality of knowledge numbers, a knowledge name corresponding to each knowledge number, and a plurality of knowledge keyword groups from a knowledge data module, wherein each knowledge keyword group has a corresponding weight and includes at least one knowledge keyword; calculating the similarity between the voice data and the knowledge name corresponding to each knowledge number, and selecting a plurality of candidate knowledge numbers from the knowledge numbers according to the similarity; identifying and annotating a plurality of candidate knowledge numbers in the voice data based on the keyword group corresponding to each knowledge number; Voice keywords; adding time tags to the voice keywords to form multiple time-limit keywords; comparing at least one knowledge keyword with the time-limit keyword, and screening the knowledge number to be recommended from the candidate knowledge numbers based on the time-limit keywords that match the knowledge keyword; calculating the keyword weight score corresponding to each knowledge number to be recommended according to the weight corresponding to each knowledge keyword group; and sorting the knowledge numbers to be recommended from high to low according to the keyword weight score and similarity of each knowledge number to be recommended to generate a knowledge recommendation sorting result.

Based on the above, the knowledge recommendation system and knowledge recommendation method provided by the present invention can provide real-time service knowledge recommendations for counter personnel to select based on the content of customer counter conversations, assist counter personnel to recall service processes, and improve customer satisfaction. In addition to reducing training time, it can also reduce corporate training costs.

1: Knowledge recommendation system

11: Memory

111: Dialogue storage module

112: Knowledge data module

113: Voice similarity calculation module

114:Keyword recognition module

115:Keyword deadline module

116:Keyword matching module

117:Keyword weight module

118: Knowledge recommendation sorting module

12: Processor

13: Display

3: Knowledge recommendation method

S310~S380: Steps

Figure 1 is a diagram of the architecture of a knowledge recommendation system according to an embodiment of the present invention.

Figure 2 is a schematic diagram of a knowledge recommendation system executing a knowledge recommendation method according to an embodiment of the present invention.

Figure 3 is a flow chart of a knowledge recommendation method according to an embodiment of the present invention.

FIG1 is a structural diagram of a knowledge recommendation system 1 according to an embodiment of the present invention. Referring to FIG1 , the knowledge recommendation system 1 includes a memory 11, a processor 12, and a display 13, wherein the processor 12 is coupled to the memory 11 and the display 13. In practice, the knowledge recommendation system 1 can be implemented by a computer device, such as a desktop computer, a laptop computer, a tablet computer, a workstation, or other computer device with computing functions, display functions, and networking functions, and the present invention is not limited thereto.

The memory 11 stores multiple modules, including a conversation storage module 111, a knowledge data module 112, a voice similarity calculation module 113, a keyword recognition module 114, a keyword deadline module 115, a keyword comparison module 116, a keyword weight module 117, and a knowledge recommendation sorting module 118. In practice, the memory 11 is, for example, a static random access memory (SRAM), a dynamic random access memory (DRAM), or other memory, and the present invention is not limited thereto.

The processor 12 is used to read multiple modules stored in the memory 11 and execute the multiple modules to generate knowledge recommendation sorting results. In practice, the processor 12 can be a central processing unit (CPU), a microprocessor, or an embedded controller built into the computer device that implements the knowledge recommendation system 1, and the present invention is not limited to this.

FIG2 is a schematic diagram of a knowledge recommendation system 1 executing a knowledge recommendation method according to an embodiment of the present invention, that is, a schematic diagram of a processor 12 of the knowledge recommendation system 1 reading multiple modules in a memory 11 and then executing the knowledge recommendation method through multiple modules. Please refer to FIG1 and FIG2 simultaneously.

The conversation storage module 111 is used to store voice data. The knowledge recommendation system 1 has an external sound receiving device (not shown in the figure), which receives sound signals and stores them in the conversation storage module 111 in the form of voice data. The knowledge recommendation system 1 can be connected to two external sound receiving devices (such as microphones), one of which receives the sound signal of the first speaker (such as a customer who conducts business at the counter), and the other receives the sound signal of the second speaker (such as a counter business staff). The voice data is composed of multiple sound signals received by the two sound receiving devices. When the two parties are talking, the conversation record of the two parties will be stored in the conversation storage module 111 in the form of voice data. The conversation storage module 111 can also attach labels according to the sound signals received by different sound receiving devices. For example, the content of the conversation between the two parties is as follows:

Customers who come to the counter to do business: "I would like to inquire about the after-sales service of the product."

Counter clerk: "What product is this?"

Customers who come to the counter to conduct business: "I would like to inquire about the price of East Asia roaming."

Counter clerk: "Which country do you want to inquire about?"

The dialogue storage module 111 will attach a first label to "I want to inquire about the product after-sales service." and "I want to inquire about the price of East Asia roaming.", and attach a second label to "Which product is it?" and "Which country do you want to inquire about?" to distinguish the sound signals received by different radio devices.

The knowledge data module 112 is used to store multiple knowledge numbers, knowledge names corresponding to each knowledge number, knowledge descriptions, and multiple knowledge keyword groups, wherein each knowledge keyword group has a corresponding weight and includes at least one knowledge keyword. Multiple knowledge keyword groups include category keyword groups, subcategory keyword groups, and general keyword groups. Table 1 is an example of knowledge numbers, knowledge names corresponding to each knowledge number, knowledge descriptions, and multiple knowledge keyword groups stored in the knowledge data module 112.

In one embodiment, the category keyword group corresponds to the first weight, the subcategory keyword group corresponds to the second weight, and the general keyword group corresponds to the third weight, wherein the first weight is greater than the second weight, and the second weight is greater than the third weight.

The knowledge data module 112 is further used to receive at least one newly added knowledge keyword corresponding to at least one of the knowledge keyword groups. For example, if the user wants to add the keyword "Hanoi" to the subcategory keyword group of knowledge number 4, the user can input "Hanoi" through an input device (such as a keyboard) connected to the processor 12, and the processor 12 stores the keyword "Hanoi" in the knowledge data module 112 in the memory 11.

The voice similarity calculation module 113 is used to calculate the similarity between the voice data and the knowledge name corresponding to each knowledge number, and select multiple candidate knowledge numbers from the knowledge numbers according to the similarity. The voice similarity calculation module 113 is further used to calculate the similarity between the voice data and the knowledge name corresponding to each of the knowledge numbers through a string matching algorithm.

In one embodiment, the voice similarity calculation module 113 can calculate the similarity between the voice data and the knowledge name corresponding to each knowledge number through a string comparison algorithm. The string comparison algorithm is, for example, Levenshtein distance, cosine similarity, Pearson Similiarity, etc. As long as it is a string comparison algorithm, it is within the protection scope of the present invention. For example, Levenshtein distance is calculated based on the number of times a string is added, deleted, or modified (replaced) into another string and divided by the sum of the two compared strings as the similarity, and the formula is ( A - B )/ A , where A is SUM ( String ₁ , String ₂ ), and B is MinimumEditDistance ( String ₁ , String ₂ ).

The keyword recognition module 114 is used to recognize the voice data in real time based on all the knowledge keywords in the keyword group corresponding to each knowledge number stored in the knowledge data module 112 using named entity recognition, and to mark multiple voice keywords in the voice data.

The keyword deadline module 115 is used to add a time tag to the voice keyword to form multiple deadline keywords. In other words, each deadline keyword corresponds to each voice keyword and the time when the voice keyword appears in the voice data.

The keyword matching module 116 is used to compare the knowledge keywords in the knowledge data module 112 with the term keywords generated by the keyword term module 115, and to filter out multiple knowledge numbers to be recommended from multiple candidate knowledge numbers based on the term keywords that match the knowledge keywords.

The keyword weight module 117 is used to calculate the keyword weight score of each knowledge number to be recommended selected by the keyword matching module 116 according to the weight corresponding to each knowledge keyword group. For example, the category keyword group corresponds to the first weight (e.g., 2), the subcategory keyword group corresponds to the second weight (e.g., 1), and the general keyword group corresponds to the third weight (e.g., 0.5). The keyword weight score W(k _N ) of the knowledge number N to be recommended is calculated as follows: W ( k _N )= X _N ×2+ Y _N ×1+ Z _N ×0.5; wherein X _N is the number of times each term keyword generated by the keyword term module 115 matches the category keyword corresponding to the knowledge number N to be recommended, Y _N is the number of times each term keyword generated by the keyword term module 115 matches the subcategory keyword corresponding to the knowledge number N to be recommended, and Z _N is the number of times each deadline keyword generated by the keyword deadline module 115 matches the sub-category keyword corresponding to the knowledge number N to be recommended.

In one embodiment, the knowledge recommendation system 1 may provide the user with the ability to filter voice data to capture voice data within a set time period (e.g., 0:08:00~0:11:00, three minutes) in the voice data. The time limit keywords within the set time period will be used for subsequent processing, and the time limit keywords outside the set time period will be ignored, so that the knowledge recommendation system 1 is more efficient in processing. Therefore, before the keyword weight module 117 calculates the keyword weight score corresponding to each knowledge number to be recommended according to the weight corresponding to each knowledge keyword group, the keyword deadline module 115 is further used to capture part of the knowledge numbers to be recommended corresponding to part of the deadline keywords with time tags within a set time period.

The knowledge recommendation ranking module 118 is used to sort the knowledge numbers to be recommended from high to low according to the keyword weight scores and similarities of the knowledge numbers to be recommended to generate knowledge recommendation ranking results. In one embodiment, the knowledge recommendation ranking results may only include some of the recommended knowledge numbers ranked in front of the knowledge numbers to be recommended.

After the knowledge recommendation ranking module 118 generates the knowledge recommendation ranking result, the display 13 is used to display the recommended knowledge description corresponding to each recommended knowledge number in the knowledge recommendation ranking result.

FIG. 3 is a flow chart of a knowledge recommendation method 3 according to an embodiment of the present invention, i.e., a flow chart of the processor 12 of the knowledge recommendation system 1 shown in FIG. 2 executing the knowledge recommendation method 3 through multiple modules. The knowledge recommendation method 3 includes steps S310 to S380. Please refer to FIG. 2 to FIG. 3 at the same time.

In step S310, a plurality of knowledge numbers, a knowledge name corresponding to each knowledge number, and a plurality of knowledge keyword groups are read from the knowledge data module 112, wherein each knowledge keyword group has a corresponding weight and includes at least one knowledge keyword.

In step S320, the similarity between the voice data and the knowledge name corresponding to each knowledge number is calculated, and multiple candidate knowledge numbers are selected from the knowledge numbers according to the similarity.

In step S330, multiple voice keywords in the voice data are identified and labeled based on the keyword group corresponding to each knowledge number.

In step S340, a time tag is added to the voice keyword to form a plurality of time-limited keywords.

In step S350, at least one knowledge keyword is compared with a deadline keyword, and a plurality of knowledge numbers to be recommended are screened from the candidate knowledge numbers.

In step S360, the keyword weight score corresponding to each knowledge number to be recommended is calculated according to the weight corresponding to each knowledge keyword group.

In step S370, the knowledge numbers to be recommended are sorted from high to low according to the keyword weight score and similarity of each knowledge number to be recommended to generate a knowledge recommendation sorting result.

In step S380, the recommended knowledge description corresponding to each recommended knowledge number is displayed.

Next, an example will be given to illustrate the implementation of the knowledge recommendation method 3 by the processor 12 of the knowledge recommendation system 1 through multiple modules.

Embodiment 1 is an example of knowledge recommendation in which the context of voice data is only current data and the keyword weights are equal. Assume that consumer A goes to a counter of a certain telecommunication store to inquire about telecommunication services. After drawing a number card, he goes to a counter to inquire. The content of the conversation is shown in Table 2:

The counter staff can manually trigger the knowledge recommendation system 1 to make knowledge recommendations. This conversation will first undergo a phonetic-to-text conversion process and be stored in the conversation storage module 111. Afterwards, the speech similarity calculation module 113 will be used to compare the knowledge name corresponding to each knowledge number stored in the knowledge data module 112, see Table 1.

Taking Levenshtein proportional distance as an example, the speech similarity calculation module 113 calculates six similarities based on the knowledge names corresponding to the conversation and knowledge numbers 1 to 6, which are 0.25, 0.375, 0.375, 0.25, 0.2353 and 0.2353, respectively. In one embodiment, the user sets in advance to retain the knowledge numbers with similarities greater than or equal to the threshold value of 0.25. The speech similarity calculation module 113 will filter out the knowledge numbers with similarities greater than or equal to the threshold value of 0.25 from the six similarities to generate multiple candidate knowledge numbers.

Since the similarity between the knowledge name "Korea Roaming" corresponding to knowledge number 2 and the knowledge name "Thailand Roaming" corresponding to knowledge number 3 is 0.375, both knowledge numbers 2 and 3 are included in the candidate knowledge numbers. In addition, the similarity between the knowledge name "Japan Roaming" corresponding to knowledge number 1 and the knowledge name "Vietnam Roaming" corresponding to knowledge number 4 is 0.25, so both knowledge numbers 1 and 4 are also included in the candidate knowledge numbers. When the number of candidate knowledge numbers exceeds 3, the knowledge numbers with similarity lower than 0.25 will no longer be used. Therefore, the speech similarity calculation module 113 selects candidate knowledge numbers 1, 2, 3 and 4 from knowledge numbers 1 to 6.

Next, the keyword recognition module 114 uses the named entity to recognize the voice data stored in the real-time recognition dialogue storage module 111 based on all the knowledge keywords in the keyword group corresponding to the knowledge numbers 1 to 6 stored in the knowledge data module 112, and labels multiple voice keywords in the voice data, namely "roaming" and "price".

The keyword deadline module 115 adds time tags to the voice keywords "roaming" and "price" to form multiple deadline keywords. In addition, the keyword deadline module 115 will also capture the voice data with time tags within the set time period set by the counter staff (for example: 0:08:00~0:11:00, five minutes) for filtering. Since the conversation in the voice data was captured at 0:08:05, the related keywords "roaming" and "price" of this conversation will be retained.

The keyword matching module 116 compares the knowledge keywords in the knowledge data module 112 with the term keywords "roaming" and "price" generated by the keyword term module 115, and selects the knowledge numbers 1, 2, 3 and 4 to be recommended from the candidate knowledge numbers 1, 2, 3 and 4 based on the term keywords "roaming" and "price" that match the knowledge keywords.

The keyword weight module 117 calculates the keyword weight scores W(k1), W(k2), W(k3) and W(k4 ) of the knowledge numbers 1 , ₂ , ₃ and 4 selected by the keyword matching module 116 according to the weights 2, ₁ and 0.5 corresponding to the category keyword _group , subcategory keyword group and general keyword group respectively.

Since the number of times X _N that the term keywords "roaming" and "price" generated by the keyword term module 115 match the category keywords corresponding to the knowledge number 1 to be recommended is 0, the number of times Y _N that the term keywords "roaming" and "price" generated by the keyword term module 115 match the subcategory keywords corresponding to the knowledge number 1 to be recommended is 1, and Z _N is the number of times Z _N that the term keywords "roaming" and "price" generated by the keyword term module 115 match the subcategory keywords corresponding to the knowledge number 1 to be recommended. Therefore, the keyword weight score W(k ₁ ) is calculated as follows: W ( k ₁ )=0×2+1×1+1×0.5=1.5

By analogy, the keyword weight scores W(k ₂ ), W(k ₃ ) and W(k ₄ ) are calculated as follows: W ( k ₂ )=0×2+1×1+1×0.5=1.5

W ( k ₃ )=0×2+1×1+1×0.5=1.5

W ( k ₄ )=0×2+1×1+1×0.5=1.5

Next, the knowledge recommendation ranking module 118 ranks the knowledge numbers 1, 2, 3 and 4 to be recommended from high to low according to their keyword weight scores W(k ₁ ), W(k ₂ ), W(k ₃ ) and W(k ₄ ) to generate a knowledge recommendation ranking result. Since the keyword weight scores W(k ₁ ), W(k 2 ), W(k ₃ ) and W(k ₄ ) calculated by the keyword weight module 117 for the knowledge numbers 1 , 2, ₃ and 4 to be recommended are the same, the counter staff can set to recommend only the top three knowledge numbers to be recommended.

The knowledge recommendation ranking module 118 ranks the knowledge numbers 1, 2, 3 and 4 according to their keyword weight scores W( _k1 ), W(k2 ₎ , W(k3 ₎ and W(k4 ₎ and similarities 0.25, 0.375, 0.375 and 0.25 from high to low to generate a knowledge recommendation ranking result including the recommended knowledge numbers.

In detail, since the keyword weight scores of the recommended knowledge numbers 1, 2, 3 and 4 are the same, but after sorting the recommended knowledge numbers 1, 2, 3 and 4 from high to low according to their similarity, the knowledge recommendation ranking results can be obtained as recommended knowledge number 2, recommended knowledge number 3, recommended knowledge number 1, and recommended knowledge number 4. However, since the number sequence of recommended knowledge number 1 is before recommended knowledge number 4, the final knowledge recommendation ranking result only includes recommended knowledge number 2, recommended knowledge number 3, and recommended knowledge number 1.

After the knowledge recommendation ranking module 118 generates the knowledge recommendation ranking result, the display 13 displays the recommended knowledge descriptions corresponding to the recommended knowledge number 2, the recommended knowledge number 3, and the recommended knowledge number 1 in the knowledge recommendation ranking result. The counter staff can convey the recommended knowledge descriptions corresponding to the recommended knowledge number 2, the recommended knowledge number 3, and the recommended knowledge number 1 to the consumer A from the display 13.

Embodiment 2 is an example of correct knowledge recommendation based on contextual keywords through the knowledge recommendation system 1 and knowledge recommendation method 3 of the present invention. Assume that consumer B goes to a counter of a certain telecommunications store to inquire about telecommunications services. After drawing a number card, he goes to a certain counter to inquire. The content of the conversation is as shown in Table 3:

The counter staff can manually trigger the knowledge recommendation system 1 to make knowledge recommendations. This conversation will first undergo a phonetic-to-text conversion process and be stored in the conversation storage module 111. Afterwards, the speech similarity calculation module 113 will be used to compare the knowledge name corresponding to each knowledge number stored in the knowledge data module 112, see Table 1.

Taking the Levenshtein proportional distance as an example, the speech similarity calculation module 113 calculates six similarities of each dialogue based on the knowledge names corresponding to these dialogues and knowledge numbers 1 to 6, as shown in Table 4.

In one embodiment, the user sets in advance to retain knowledge numbers with similarities greater than or equal to the threshold value of 0.25, and the voice similarity calculation module 113 will filter out the knowledge numbers with similarities greater than or equal to the threshold value of 0.25 from the six similarities of each dialogue in order according to the sequence of dialogue time to generate multiple candidate knowledge numbers.

Specifically, the six similarities calculated by the voice similarity calculation module 113 based on the dialogue at 0:08:05 and the knowledge names corresponding to the knowledge numbers 1 to 6 are all greater than or equal to the threshold value, so all of them are taken out. The six similarities calculated by the voice similarity calculation module 113 based on the dialogue at 0:08:06 and 0:08:08 and the knowledge names corresponding to the knowledge numbers 1 to 6 do not exceed the threshold value of 0.25, so they are discarded. The voice similarity calculation module 113 calculates the similarity based on the dialogue at 0:08:07 and the knowledge names corresponding to knowledge numbers 1 to 6. Only the similarity with knowledge number 1 is higher than the threshold value of 0.25, so knowledge number 1 is taken out. Finally, the voice similarity calculation module 113 takes the union of the similarities of all dialogues to filter out candidate knowledge numbers 1 to 6.

Next, the keyword recognition module 114 uses the named entity to recognize the voice data stored in the real-time recognition dialogue storage module 111 based on all the knowledge keywords in the keyword group corresponding to the knowledge numbers 1 to 6 stored in the knowledge data module 112, and labels multiple voice keywords in the voice data, namely "roaming", "price", and "Japan".

The keyword deadline module 115 adds time tags to the voice keywords "roaming", "price", and "Japan" to form multiple deadline keywords. In addition, the keyword deadline module 115 will also capture the voice data with time tags within the set time period set by the counter staff (for example: 0:08:00~0:11:00, five minutes) for filtering. Since the conversation in the voice data was captured at 0:08:05, the related keywords "roaming", "price", and "Japan" of this conversation will be retained.

The keyword matching module 116 compares the knowledge keywords in the knowledge data module 112 with the term keywords "roaming", "price", and "Japan" generated by the keyword term module 115, and selects the knowledge numbers 1 to 6 to be recommended from the candidate knowledge numbers 1 to 6 based on the term keywords "roaming", "price", and "Japan" that match the knowledge keywords.

The keyword weight module 117 calculates the keyword weight scores W(k 1 ), W(k 2 ), W(k 3 ), W(k ₄ ) , W (k ₅ ), W(k ₆ ) of the knowledge numbers 1 to 6 selected by the keyword matching module 116 according to the weights 2, 1, and 0.5 corresponding to the category keyword group, the subcategory keyword group, and _{the general keyword group} , respectively .

Since the number of times X _N that the term keywords "roaming", "price", and "Japan" generated by the keyword term module 115 match the category keywords corresponding to the knowledge number 1 to be recommended is 1, the number of times Y _N that the term keywords "roaming", "price", and "Japan" generated by the keyword term module 115 match the subcategory keywords corresponding to the knowledge number 1 to be recommended is 1, and Z _N is the number of times Z N that the term keywords "roaming", "price", and "Japan" generated by the keyword term module 115 match the subcategory _keywords corresponding to the knowledge number 1 to be recommended is 1. Therefore, the keyword weight score W(k ₁ ) is calculated as follows: W ( k ₁ )=1×2+1×1+1×0.5=3.5

By analogy, the keyword weight scores W(k ₂ ), W(k ₃ ), W(k ₄ ), W(k ₅ ), and W(k ₆ ) are calculated as follows: W ( k ₂ )=0×2+1×1+1×0.5=1.5

W ( k ₃ )=0×2+1×1+1×0.5=1.5

W ( k ₄ )=0×2+1×1+1×0.5=1.5

W ( k ₅ )=0×2+1×1+1×0.5=1.5

W ( k ₆ )=0×2+1×1+1×0.5=1.5

Next, the knowledge recommendation ranking module 118 ranks the knowledge numbers 1 to 6 to be recommended from high to low according to the keyword weight scores W(k ₁ ) to W(k ₆ ) of the knowledge numbers 1 to 6 to be recommended to generate a knowledge recommendation ranking result.

The counter staff sets the recommendation to be made only for the top three recommended knowledge numbers. The keyword weight score W(k ₁ ) of recommended knowledge number 1 is the highest. Therefore, in the knowledge recommendation ranking result generated by the knowledge recommendation ranking module 118, recommended knowledge number 1 is the top priority. Since the keyword weight scores of recommended knowledge numbers 2 to 6 are the same, the knowledge recommendation ranking module 118 ranks the recommended knowledge numbers from high to low according to the keyword weight scores W(k ₂ ) to W(k ₆ ) of recommended knowledge numbers 2 to 6 and the similarity, so recommended knowledge numbers 2 and 3 are left.

After the knowledge recommendation ranking module 118 generates the knowledge recommendation ranking result, the display 13 displays the recommended knowledge descriptions corresponding to the recommended knowledge number 1, the recommended knowledge number 2, and the recommended knowledge number 3 in the knowledge recommendation ranking result. The counter staff can convey the recommended knowledge descriptions corresponding to the recommended knowledge number 1, the recommended knowledge number 2, and the recommended knowledge number 3 from the display 13 to the consumer B.

In summary, the knowledge recommendation system and knowledge recommendation method provided by the present invention can provide real-time service knowledge recommendations for counter personnel to select based on the content of customer counter conversations, assist counter personnel in recalling service processes, and improve customer satisfaction. In addition to reducing training time, it can also reduce corporate training costs.

3: Knowledge recommendation method

S310~S380: Steps

Claims (14)

A knowledge recommendation system includes: a memory storing a plurality of modules; a processor coupled to the memory for reading and executing the modules to generate a knowledge recommendation ranking result, wherein the modules include: a knowledge data module for storing a plurality of knowledge numbers, a knowledge name corresponding to each of the knowledge numbers, and a plurality of knowledge keyword groups, wherein each of the knowledge keyword groups has a corresponding weight and includes at least one knowledge keyword; a voice similarity calculation module for calculating a voice data The method comprises the steps of: determining a similarity between the knowledge name corresponding to each of the knowledge numbers and filtering out a plurality of candidate knowledge numbers from the knowledge numbers according to the similarity, wherein the voice data is composed of a plurality of sound signals received by two microphones; a keyword recognition module for recognizing and labeling a plurality of voice keywords in the voice data based on the keyword groups corresponding to each of the knowledge numbers; a keyword expiration module for adding time tags to the voice keywords to form a plurality of expiration keywords; and a keyword matching module for matching the at least one knowledge keyword with the time limit keywords, and screening out a plurality of knowledge numbers to be recommended from the candidate knowledge numbers based on the time limit keywords matching the at least one knowledge keyword; a keyword weight module for calculating a keyword weight score corresponding to each of the knowledge numbers to be recommended according to the weight corresponding to each of the knowledge keyword groups; and a knowledge recommendation ranking module for ranking the knowledge numbers to be recommended according to the keyword weight score of each of the knowledge numbers to be recommended. and the similarity is sorted from high to low to generate the knowledge recommendation sorting result; wherein the knowledge keyword groups include category keyword groups, subcategory keyword groups and general keyword groups; wherein the category keyword group corresponds to a first weight, the subcategory keyword group corresponds to a second weight, and the general keyword group corresponds to a third weight, wherein the first weight is greater than the second weight, and the second weight is greater than the third weight; wherein each of the knowledge numbers to be recommended ( N ) is calculated as follows: W ( k _N ) = _{X N} ×2+ Y _N ×1+ Z _N ×0.5 , where X _N is the number of times each of the term keywords generated by the keyword term module matches the category keyword corresponding to the knowledge number (N) to be recommended, Y _N is the number of times each of the term keywords generated by the keyword term module matches the sub-category keyword corresponding to the knowledge number (N) to be recommended, and Z _N is the number of times each of the term keywords generated by the keyword term module matches the sub-category keyword corresponding to the knowledge number (N) to be recommended. The knowledge recommendation system as described in claim 1, wherein the modules further include a dialogue storage module for storing the voice data. The knowledge recommendation system as described in claim 1, wherein the voice similarity calculation module is further used to calculate the similarity between the voice data and the knowledge name corresponding to each of the knowledge numbers through a string matching algorithm. The knowledge recommendation system as described in claim 1, wherein before the keyword weight module calculates the keyword weight score corresponding to each of the knowledge numbers to be recommended according to the weight corresponding to each of the knowledge keyword groups, the keyword deadline module is further used to capture part of the knowledge numbers to be recommended corresponding to part of the deadline keywords with the time tag within a set time period. A knowledge recommendation system as described in claim 1, wherein the knowledge data module is further used to receive at least one newly added knowledge keyword corresponding to at least one of the knowledge keyword groups. A knowledge recommendation system as described in claim 1, wherein the knowledge recommendation ranking result includes some of the recommended knowledge numbers that are ranked first among the knowledge numbers to be recommended. The knowledge recommendation system as described in claim 6 further includes: a display coupled to the processor for displaying the recommended knowledge description corresponding to each of the recommended knowledge numbers. A knowledge recommendation method includes: executing, by a processor: reading a plurality of knowledge numbers, knowledge names corresponding to each of the knowledge numbers, and a plurality of knowledge keyword groups from a knowledge data module, wherein each of the knowledge keyword groups has a corresponding weight and includes at least one knowledge keyword; calculating the similarity between a voice data and the knowledge name corresponding to each of the knowledge numbers, and filtering out a plurality of candidate knowledge numbers from the knowledge numbers according to the similarity, wherein the voice data is composed of a plurality of sound signals received by two microphones; identifying and labeling the candidate knowledge numbers based on the keyword groups corresponding to each of the knowledge numbers; a plurality of voice keywords in the voice data; adding time tags to the voice keywords to form a plurality of time limit keywords; comparing the at least one knowledge keyword with the time limit keywords, and selecting a plurality of knowledge numbers to be recommended from the candidate knowledge numbers based on the time limit keywords that match the at least one knowledge keyword; and selecting a plurality of knowledge numbers to be recommended based on each of the knowledge numbers. The weight corresponding to the keyword group is used to calculate the keyword weight score corresponding to each of the knowledge numbers to be recommended; and the knowledge numbers to be recommended are sorted from high to low according to the keyword weight score of each of the knowledge numbers to be recommended and the similarity to generate a knowledge recommendation sorting result; wherein the knowledge keyword group includes a category keyword group, a subcategory keyword group and a general keyword group; wherein the category keyword group corresponds to a first weight, the subcategory keyword group corresponds to a second weight, and the general keyword group corresponds to a third weight, wherein the first weight is greater than the second weight, and the second weight is greater than the third weight; wherein each of the knowledge numbers to be recommended ( N ) is calculated as follows: W ( k _N ) = X _N ×2+ Y _N ×1+ Z _N ×0.5 where X _N is the number of _times each of the term keywords generated by the keyword term module matches the category keyword corresponding to the knowledge number (N) to be recommended, Y _N is the number of times each of the term keywords generated by the keyword term module matches the sub-category keyword corresponding to the knowledge number (N) to be recommended, and Z _N is the number of times each of the term keywords generated by the keyword term module matches the sub-category keyword corresponding to the knowledge number (N) to be recommended. The knowledge recommendation method as described in claim 8 further includes: reading the voice data from the dialogue storage module by the processor. The knowledge recommendation method as described in claim 8 further includes: the processor calculates the similarity between the voice data and the knowledge name corresponding to each of the knowledge numbers through a string matching algorithm. The knowledge recommendation method as described in claim 8 further includes: before the processor calculates the keyword weight score corresponding to each of the knowledge numbers to be recommended according to the weight corresponding to each of the knowledge keyword groups, extracting part of the knowledge numbers to be recommended corresponding to part of the time limit keywords within a set time period of the time tag. The knowledge recommendation method as described in claim 8 further includes: receiving at least one newly added knowledge keyword corresponding to at least one of the knowledge keyword groups by the processor. The knowledge recommendation method as described in claim 8, wherein the knowledge recommendation ranking result includes some of the recommended knowledge numbers that are ranked first among the knowledge numbers to be recommended. The knowledge recommendation method as described in claim 13 further includes: displaying the recommended knowledge description corresponding to each of the recommended knowledge numbers by a display.

Images