KR102418260B1 - Method for analyzing customer consultation record - Google Patents

Abstract

A computer program stored in a computer-readable storage medium is disclosed according to an embodiment of the present disclosure. When the computer program is executed in one or more processors, the following operations are performed for analyzing customer consultation records using a neural network, and the operations are: In a deep learning-based transformation model trained to convert voice information into text information obtaining text data by inputting the received customer call voice data; generating a training data set based on the text data and product contract information; learning the product recommendation model to generate product recommendation information by inputting the training data set into a deep learning-based product recommendation model; and inputting new customer call voice data into the learned transformation model to obtain new text data, and inputting new data based thereon into the product recommendation model to generate product recommendation information. It includes, and the product contract information may include information on at least one product and a result of whether each product is contracted.

Description

METHOD FOR ANALYZING CUSTOMER CONSULTATION RECORD

The present disclosure relates to data analysis technology, and more particularly, to technology for analyzing customer consultation records.

As the sales of financial products and insurance products through the Internet and mobile increase, there is a demand in the industry to collect customer consultations and use them for marketing. For example, if it is known that the probability of a contract will increase when a customer recommends a product to a customer based on customer consultation data in a call center, the effect can be maximized while reducing marketing costs and human resources. There is a need in the art for a method and a marketing method that can efficiently process consultation data for maximizing marketing targeting and marketing efficiency as described above.

Korean Patent Registration No. KR2110495 discloses “a system for providing discount marketing services using big data-based influencer advertisements”.

The present disclosure has been devised in response to the above-described background technology, and is intended to provide a method for analyzing customer consultation records.

A computer program stored in a computer-readable storage medium is disclosed according to an embodiment of the present disclosure for realizing the above-described problems. When the computer program is executed in one or more processors, the following operations are performed for analyzing customer consultation records using a neural network, and the operations are: In a deep learning-based transformation model trained to convert voice information into text information obtaining text data by inputting the received customer call voice data; generating a training data set based on the text data and product contract information; learning the product recommendation model to generate product recommendation information by inputting the training data set into a deep learning-based product recommendation model; and inputting new customer call voice data into the learned transformation model to obtain new text data, and inputting new data based thereon into the product recommendation model to generate product recommendation information. It includes, and the product contract information may include information on at least one product and a result of whether each product is contracted.

In an alternative embodiment of computer program operations for performing the following operations for analyzing a customer consultation record, a preprocessing operation comprising tokenizing the text data; may further include.

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, at least one of morphological analysis of the text data, parsing tree-based syntax analysis, and stopword processing on the text data A preprocessing operation comprising; may further include.

In an alternative embodiment of computer program operations for performing the following operations for analyzing a customer consultation record, generating vector data based on at least one token obtained as a result of a tokenization operation on the text data It further includes, and the vector data may be generated to reflect the semantic similarity between tokens.

In an alternative embodiment of the computer program operations for performing the following operations for analyzing the customer consultation record, the operation may further include generating an array by extracting a predetermined number of factors for a sentence included in the text data. .

In an alternative embodiment of computer program operations for performing the following operations for analyzing a customer consultation record, the predetermined number of factors may be extracted based on a computational contribution.

In an alternative embodiment of computer program operations for performing the following operations for analyzing a customer consultation record, the arrangement comprises: inputting a predetermined magnitude into a convolutional neural network; may further include.

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, the product recommendation model is trained to extract features of utterances included in the learning data, and the characteristics of utterances include: It may include the importance of individual data included in the training data.

In an alternative embodiment of the computer program operations for performing the following operations for analyzing a customer consultation record, a pre-processing operation including stopword processing based on a characteristic of the utterance content; and calculating a calculation contribution based on the characteristics of the utterance content. It may further include at least one of.

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, the product recommendation model may be trained to generate product recommendation information based on vector data included in the learning data.

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, the product recommendation model may be formed of a combination of at least one or more convolutional neural networks and at least one or more recurrent neural networks. can

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, the recurrent neural network may be applied with a Long Short Term Memory (LSTM) model.

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, the product recommendation model extracts a keyword from a parsing tree-based syntax analysis information of the learning data, and the keyword Product recommendation information may be generated based on the .

In an alternative embodiment of computer program operations for performing the following operations for analyzing customer consultation records, when a customer contracts a product based on the product recommendation information, the product recommendation model is created using the text data of the customer can relearn.

According to an embodiment of the present disclosure for realizing the above-described problem, a method for analyzing customer consultation records using a neural network is disclosed. The customer consultation record analysis method includes: acquiring text data by inputting the received customer call voice data into a deep learning-based transformation model trained to convert voice information into text information; generating a training data set based on the text data and product contract information; learning the product recommendation model to generate product recommendation information by inputting the training data set into a deep learning-based product recommendation model; and inputting new customer call voice data into the learned transformation model to obtain new text data, and inputting new data based thereon into the product recommendation model to generate product recommendation information, and the product contract information includes: It may include information on at least one product and a result of whether each product is contracted.

According to an embodiment of the present disclosure for realizing the above-described problem, a server for analyzing customer consultation records is disclosed. The server includes a processor including one or more cores; and a memory, wherein the processor acquires text data by inputting the received customer call voice data into a deep learning-based transformation model trained to convert voice information into text information, and based on the text data and product contract information to generate a training data set, and input the training data set to a deep learning-based product recommendation model to learn the product recommendation model to generate product recommendation information, and input new customer call voice data to the learned transformation model. Obtaining new text data and inputting new data based thereon into the product recommendation model to generate product recommendation information, and the product contract information may include information on at least one product and a result of whether each product is contracted.

The present disclosure may provide a method for analyzing customer consultation records.

1 is a diagram illustrating a block diagram of a computing device that performs an operation for calculating a target for marketing according to an embodiment of the present disclosure.
2 is a flowchart of a customer consultation record analysis system according to an embodiment of the present disclosure.
3 is an exemplary diagram of a parsing tree according to an embodiment of the present disclosure.
4 is an exemplary diagram of an arrangement related to text data according to an embodiment of the present disclosure.
5 is an exemplary diagram illustrating some data of a final result of a product recommendation model according to an embodiment of the present disclosure.

Various embodiments are now described with reference to the drawings. In this specification, various descriptions are presented to provide an understanding of the present disclosure. However, it is apparent that these embodiments may be practiced without these specific descriptions.

The terms “component,” “module,” “system,” and the like, as used herein, refer to a computer-related entity, hardware, firmware, software, a combination of software and hardware, or execution of software. For example, a component can be, but is not limited to being, a process running on a processor, a processor, an object, a thread of execution, a program, and/or a computer. For example, both an application running on a computing device and the computing device may be a component. One or more components may reside within a processor and/or thread of execution. A component may be localized within one computer. A component may be distributed between two or more computers. In addition, these components can execute from various computer readable media having various data structures stored therein. Components may communicate via a network such as the Internet with another system, for example, via a signal having one or more data packets (eg, data and/or signals from one component interacting with another component in a local system, distributed system, etc.) may communicate via local and/or remote processes depending on the data being transmitted).

In addition, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless otherwise specified or clear from context, "X employs A or B" is intended to mean one of the natural implicit substitutions. That is, X employs A; X employs B; or when X employs both A and B, "X employs A or B" may apply to either of these cases. It should also be understood that the term “and/or” as used herein refers to and includes all possible combinations of one or more of the listed related items.

Also, the terms "comprises" and/or "comprising" should be understood to mean that the feature and/or element in question is present. However, it should be understood that the terms "comprises" and/or "comprising" do not exclude the presence or addition of one or more other features, elements and/or groups thereof. Also, unless otherwise specified or unless it is clear from context to refer to a singular form, the singular in the specification and claims should generally be construed to mean “one or more”.

And, the term "at least one of A or B" should be interpreted as meaning "when including only A", "when including only B", and "when combined with the configuration of A and B".

Those skilled in the art will further appreciate that the various illustrative logical blocks, configurations, modules, circuits, means, logics, and algorithm steps described in connection with the embodiments disclosed herein may be implemented in electronic hardware, computer software, or combinations of both. It should be recognized that they can be implemented with To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, configurations, means, logics, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application. However, such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

Descriptions of the presented embodiments are provided to enable those skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art of the present disclosure. The generic principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present invention is not limited to the embodiments presented herein. This invention is to be interpreted in its widest scope consistent with the principles and novel features presented herein.

In the present disclosure, a network function, an artificial neural network, and a neural network may be used interchangeably.

1 is a diagram illustrating a block diagram of a computing device that performs an operation for calculating a target for marketing according to an embodiment of the present disclosure.

The configuration of the computing device 100 shown in FIG. 1 is only a simplified example. In an embodiment of the present disclosure, the computing device 100 may include other components for performing the computing environment of the computing device 100 , and only some of the disclosed components may configure the computing device 100 .

The computing device 100 may include a processor 110 , a memory 130 , and a network unit 150 .

Customer call voice data throughout this specification includes voice files in which conversations are recorded. The conversation includes all kinds of business conversations between the agent and the customer. The voice data may be stored in any file format for storing digital audio data. The arbitrary file format includes formats such as WAV, FLAC, MP3, TTA, and VOX. The above-described file format type is merely an example and does not limit the present invention.

2 is a flowchart of a method for analyzing customer consultation records according to an embodiment of the present disclosure. The processor 110 may receive 210 the customer's call voice data through the network unit 150 or a separate receiving unit (not shown). Thereafter, the processor 110 may acquire text data by inputting the voice data into the learned deep learning-based voice-to-text conversion model ( 230 ). The processor 110 may generate 250 a training data set based on the text data and product contract information. The processor 110 may input the training data set into a deep learning-based product recommendation model to train the product recommendation model to generate product recommendation information (270). When the learning is completed, finally, the processor 110 may sequentially input new customer call voice data into the conversion model and the product recommendation model to generate product recommendation information (290), and recommend products to the customer based thereon.

According to an embodiment of the present disclosure, the processor 110 may acquire text data by inputting the received customer call voice data into a deep learning-based transformation model trained to convert voice information into text information. The learned deep learning-based transformation model may include at least one or more network functions. The conversion model may divide a continuously input voice into a predetermined time unit, receive the input, and convert it into a text. The transformation model may include a network function having a many-to-many input/output structure in order to output an input voice sequence as a character sequence. The transformation model may be formed of a recurrent neural network including at least one hidden layer to convert voice information into text information. The processor 110 may store the text data obtained through the transformation model in the memory 130 . The conversion model for converting the voice information into text information may be implemented by STT (Speech To Text) technology widely known in the art.

According to an embodiment of the present disclosure, the processor 110 may generate a training data set based on the text data and product contract information. At this time, a training data set can be created by labeling each text data with product contract information. The processor 110 may further include an operation of preprocessing the text data to generate a training data set. The preprocessing operation may include tokenization of text data, morphological analysis of text data, parsing tree-based syntax analysis of text data, and stopword processing of text data.

The processor 110 may perform a pre-processing operation including tokenization of text data. The tokenization operation may mean decomposing continuous text data into units called tokens. The token may be a unit for separating text data including a word, a phrase, a sentence, or a character string having a predetermined length included in the text data. The token may be a basic unit of learning of a deep learning-based product recommendation model, which will be described later. The tokenization operation may be performed based on a space for spacing in a character string included in the given text data. For example, if tokenization is performed based on a space for spaces for the sentence “It is a pure insurance type with no cancellation refund, and a large insurance is possible with a low premium”, ['Cancellation', 'Refund is ', 'no', 'pure', 'insured,', 'low', 'as premium', 'large', 'guaranteed', 'can be tokenized'].

The processor 110 may perform a preprocessing operation including at least one of morpheme analysis on text data, parsing tree-based syntax analysis, and stopword processing on text data.

A morpheme is a unit of speech with a meaning, and can be divided into independent morpheme and dependent morpheme depending on whether it can be used independently or not. Alternatively, depending on the presence or absence of a lexical meaning, it can be divided into a real morpheme and a formal morpheme. As is commonly known in the field of natural language processing technology, in the case of Korean, it can be classified into 5 words and 9 parts of speech. The five words and nine parts of speech may mean an adjective (noun, pronoun, rhetoric), a modifier (adjective, adverb), a related word (exposition), an independent word (exclamation), and a verb (verb, adjective). Also, verbs and adjectives corresponding to verbs may be composed of a combination of a stem and an ending. In the present disclosure, morphological analysis of text data may be performed to become the basis of the tokenization operation, and after tokenizing text data based on the classification of the morpheme, labeling may be performed for each token. . The tokenization and labeling may be performed by a separate network function including at least one artificial neural network. That is, the separate network function decomposes text data and the text data to receive a training data set in which classification results are labeled for tokens of each morpheme unit, and can be learned to enable tokenization and labeling operations in units of morphemes. have. As an example of tokenization and labeling based on morpheme analysis, when morpheme analysis of the sentence "It is a pure guarantee type with no cancellation refund, and a large guarantee is possible with a low insurance premium", ['cancellation' (noun), 'refund' ( Noun), 'silver' (auxiliary), 'without' (adjective), 'is' (tubular ending), 'pure' (noun), 'guaranteed' (noun), 'form' (noun derivation suffix), as ' '(adverbial investigation), ',' (comma), 'day' (adjective), 'silver' (auxiliary), 'insurance' (noun), 'lo' (adverbial investigation), 'k' (adjective), ' ㄴ' (tubular ending), 'guarantee' (noun), 'i' (nominative participle), 'possible' (root), 'ha' (adjective derivation suffix), 'b' (final ending), '.' (period) ] can be tokenized and labeled. The above-described standards and methods of morphological analysis are merely exemplary and do not limit the present disclosure.

The processor 110 may perform parsing tree-based syntax analysis in data preprocessing. In the parsing tree, the inner node includes information about the structure of the sentence (eg, noun phrase, verb phrase, adjective phrase, etc.), and the outer node is a token of actual text data. can be The internal node may mean a node other than a leaf node, and the external node may mean a leaf node. The processor 110 may recursively visit a node of a tree representing each sentence for parsing tree-based syntax analysis. The data of each node may include information of the node itself and information of child nodes. Alternatively, information on the grandchild node may be additionally included.

Hereinafter, a parsing tree will be exemplarily described with reference to FIG. 3 . 3 is an exemplary diagram of a parsing tree according to an embodiment of the present disclosure. For example, when the processor 110 performs the parsing tree-based parsing analysis on a sentence such as "good agent's product", the root node 311 includes the entire sentence data, and the left child node ( 313) includes a noun phrase having text data “of a good agent”, and the right child node 315 of the root node may include a noun phrase having text data “product”. At this time, the processor 110 may determine that parsing can no longer be performed on the "product" in the right child node 315 of the root node, and determine this node as a leaf node. You can then continue making recursive calls to parse the parsing tree for other nodes. The left child node 313 of the root node is continuously subjected to parsing tree-based parsing by the processor 110, so that the adjective phrase “good” is the left child node 317 and the noun phrase “agent’s” to the right It may have as a child node 319 . The processor 110 may perform additional syntactic analysis on the adjective phrase "good" or the noun phrase "of the agent", but if the number of syllables is less than a predetermined criterion or is no longer split into the minimum unit of parsing, You can stop parsing and determine the node as a leaf node. The processor 110 may use a phrase expressed as a set of morphemes and words for learning through the parsing tree-based syntax analysis. In other words, not only can the contract probability of products possessed by individual words, morphemes or tokens be considered for learning, but also the contract probability of products having the structure of the phrase itself can be considered for learning or the importance of text data based on syntax analysis By analyzing the input text data from various aspects, the performance of the product recommendation model is finally improved. How the deep learning-based product recommendation model uses text data preprocessed by parsing tree-based syntax analysis for training will be described later in detail. The above-described parsing tree-based parsing example is merely an example and does not limit the present disclosure.

The processor 110 may perform stopword processing in data preprocessing. Stop words include words that do not significantly affect the processing of text data. The processor 110 may delete stopwords from the text data in order to minimize the learning time by reducing the size of the training data and to save required memory resources. The stopword processing may be performed by deleting a token related to the stopword from the tokenized text data. A list related to stopwords to be processed in text data may be pre-stored in the memory 130 as an arbitrary text file. The stopword list may be a set of words treated as stopwords in an insurance business, a finance business, or a marketing-related field. As another embodiment, the stopword may be calculated and managed by an artificial neural network. That is, through the artificial neural network, it is possible to manage a word or a set of words that have low correlation with the contract probability of all products in full-text data as stopwords. As a specific example of stop-word processing, when the processor 110 performs stop-word processing for a sentence such as "a pure guarantee type without cancellation refund, a large guarantee is possible with a low insurance premium." Probes such as 'to' or verb phrases such as 'do' included in 'it is possible' do not significantly affect data analysis, so they can be recognized as stopwords and deleted from the corresponding text data. The deletion may include a method of deleting a token of the corresponding data from the tokenized text data. In another embodiment of the stop word processing by the processor 110 , the processor 110 performs morpheme analysis and then morphemes other than a noun, a pronoun, a verb including a verb, and a verb including an adjective (for example, an interjection) It is judged that all independent words including , or relational words including proposition) do not have a great meaning in text data analysis, so that stopword processing can be performed. The above-described stopword processing method is only an example and does not limit the present disclosure. Data that appears at a very high rate in the entire text data through the stopword processing of the processor 110, but has no great value in extracting the features of the actual text content (for example, 'silver', 'is', 'this', By deleting the subject proposition such as 'a'), the speed of the entire product recommendation process can be improved or the waste of memory resources can be prevented.

The pre-processing operation by the processor 110 may include a deletion operation of data unnecessary for learning. The deletion operation may include a noise removal operation of text data. The noise of the text data includes both a case in which the noise of the voice information itself is converted into text as it is and noise generated in the conversion process. The above-described pre-processing operations are merely examples of pre-processing operations for data purification and do not limit the present disclosure, and further, the processor 110 performs natural language in addition to the above-mentioned pre-processing operations to effectively learn a deep learning-based product recommendation model and improve model performance. Any pretreatment technique commonly used in the art of treatment may be included.

According to an embodiment of the present disclosure, the processor 110 may generate a learning data set by labeling product contract information in text data, wherein the product contract information includes information on at least one product and a result of each product contract. may include The number of products included in the product contract information may be one. Information on the product may include insurance products, financial products, and the like. The result of whether the product contract has been made may include whether the final contract has been made, the product contract period, product contract conditions, the time of contract completion, whether or not to cancel, personal information of the contractor, and the like.

According to an embodiment of the present disclosure, the processor 110 may generate vector data representing the text data. That is, in generating a training data set for input to a deep learning-based artificial neural network based on text data and product contract information, vector data representing text data is generated to label product contract information, and then the training data set can be used as The processor 110 may generate vector data based on at least one token obtained as a result of a tokenization operation on the text data. In this case, the vector data may be generated to reflect the semantic similarity between tokens. After the processor 110 tokenizes the text data for each word of each word, each token may be vectorized. Alternatively, the text data may be vectorized for each token after the text data is separated by a predetermined criterion according to the above-described tokenization operation example. Furthermore, the text data may be tokenized according to morphological analysis, and vector data corresponding to each token may be generated. The processor 110 may vectorize each word or token to reflect semantic similarity between words or tokens to generate vector data representing text data. The dimension of the vector may be predetermined as an arbitrary size such as 100 or 1000 according to the size of text data and the type of word used. Hereinafter, a specific example of generating vector data reflecting semantic similarity will be described. When the dimension of the vector is 2, the processor 110 may generate vector data of (10,10) for the word “join”. Thereafter, with respect to the input word "cancellation", the processor 110 may generate vector data based on a semantic similarity to the word "subscription". That is, since the meaning and usage of "subscription" and "cancellation" are opposite to each other, the processor 110 may assign (-10, -10) to the "cancellation". Furthermore, since the word "cancellation" has a high degree of similarity to the word "cancellation", the processor 110 may generate vector data having values of (-9.5, -10) in response to the word "cancellation". The above-described dimension of vector data and a method of generating vector data are merely examples and do not limit the present disclosure. As the vectorization method, a one-hot encoding method or a word-to-vector (Word2Vec) method generally known in the art may be used. In general, the concept of similarity in meaning between words is a property inherent in text data that cannot be calculated mathematically. However, according to the present disclosure, as described above, by converting text data into vector data by reflecting the similarity of meaning between tokens, it is possible to provide for learning of artificial neural networks by quantifying, quantifying, and expressing the similarity of meanings between words. In addition, it has the effect of quickly comparing and judging the degree of similarity in meaning between words due to the difference in the relative distances of the vector coordinates according to the numerical comparison.

According to an embodiment of the present disclosure, the processor 110 may generate an array by extracting a predetermined number of factors from a sentence included in the text data. The factor may be a syllable, a word, a word, or a token. The processor 110 may generate an array by extracting factors from the preprocessed text data based on a predetermined number. Here, the predetermined number may be the maximum number of factors that can be extracted. The processor 110 may recognize a sentence by recognizing a period, a dot, etc. in the given text data, and may generate an array based thereon. This will be described with reference to FIG. 4 . 4 is an exemplary diagram of an arrangement related to text data according to an embodiment of the present disclosure. “It is a non-dividend A integrated insurance. It is a product with a premium of 100,000 won and a 25-year payment at the age of 100. In addition, driver's insurance is also included." This is a specific example of creating an array by extracting a predetermined number of factors for a sentence with exemplary text data. In an example, the array may be a two-dimensional array 400 having a row 411 and a column 413 , and the position of each data 415 included in the array may be expressed as an ordered pair (i, j). In an example, the predetermined number may be set to five. In this case, since the predetermined number is 5, the number (j) of the columns of the example sentence is set to 5. The processor 110 first decomposes the text data into sentences based on a period, a dot, and the like, and then extracts five factors from each sentence. When a sentence is composed of the number of factors from which 5 factors cannot be extracted, the value of the remaining column 413 is padded with NULL or <PAD>. The value filled in the array for the padding may be an arbitrary value selected by the user. In the example two-dimensional array 400, the i-th row 411 generated for the sentence "Non-dividend A is integrated insurance." has only three extracted factors: “Non-dividend”, “A”, and “Comprehensive insurance”. Therefore, the remaining column 413 is filled with <PAD>. In the sentence "non-dividend A integrated insurance.", the factor "is." may be deleted in the pre-processing process by processing a stopword by the processor 110 . The i+1-th row 411 and the i+2-th row 411 may also be results of the above-described preprocessing such as stopword processing and morpheme analysis for each sentence by the processor 110 in the same manner. The size and dimension of the above arrangement are merely exemplary and do not limit the present disclosure. The array may be a one-dimensional linear array.

The processor 110 may extract the predetermined number of factors based on the calculation contribution. When the processor 110 decomposes an arbitrary sentence to generate each row 411 and column 413 of the two-dimensional array 400, when deriving the product contract probability according to each sentence, the computational contribution to the model is large You can create an array by extracting the top 5 arguments. The calculation contribution may be calculated by a separate network function. The calculation contribution may be fed from a result of a product recommendation model to be described later. In addition, it is possible to calculate the calculation contribution based on the characteristics of the utterances extracted from the product recommendation model. Specifically, the order of importance of individual data in the text data included in the feature of the speech content may be utilized as the calculation contribution. The method of extracting features of speech content will be described later in detail. At this time, when the contribution of each factor is fed from the result of the product recommendation model, when the factor is extracted from the text data to generate an array, only the factors that the actual product recommendation model is used to draw conclusions in the inference process are the input data. can be provided, which has the effect of improving the accuracy of inference. In addition, since an array can be configured by extracting only a predetermined number of factors, the size of the training data can be adjusted according to a user's selection. After creating the array, the processor 110 may replace data existing in the array with vector data. The array may have vector data as values of each matrix. In other words, since an array having vector data as a value is a matrix that can be operated mathematically, inter-matrix addition and matrix multiplication are possible, making it possible to operate between matrices. It helps to extract meaning from the learning data of

The processor 110 may input the array into a convolutional neural network that receives a predetermined size as an input. In general, a convolutional artificial neural network receives training data of a fixed size. Therefore, data cannot be input to the convolutional neural network unless there is a separate operation for sequential data including sentences, texts, etc., that is, meaningful in order, and it is difficult to expect a normal output even if it is input. However, as described above in the present disclosure, when the processor 110 generates an array by extracting a predetermined number of factors from a sentence included in the text data, the processor 110 circulates the continuous text data (Recurrent) Not only can it be input to the neural network, but it can also be input to the convolutional neural network by arranging text data into an array of a predetermined size as described above. In other words, the processor 110 adjusts the number of rows 411 and the number of columns 413 of the two-dimensional array 400 to generate arrays of various sizes, and thus supports input data of different sizes. Several artificial neural networks can be used as learning models. Also, filters having different window sizes can be applied to the arranged text data. As a result, by being able to generate various feature maps, the product recommendation model can extract more features from text data, and finally, the performance of the model is improved.

According to an embodiment of the present disclosure, the processor 110 may input the training data set into a deep learning-based product recommendation model to train the product recommendation model to generate product recommendation information. The learning data consists of text data and product contract information labels for each text data, and using this, the processor 110 learns the deep learning-based product recommendation model to derive product recommendation information for input of new text data. have. The product recommendation information generated by the product recommendation model may include product contract probabilities calculated for one or more products. The contract probability may be expressed as a real value existing within an arbitrary section. The arbitrary interval may be [0,1]. The product recommendation information may include information on whether to recommend a product by comparing a contract probability for one or more products calculated by a deep learning-based product recommendation model based on text data with a predetermined probability value. The predetermined probability value may be different for each product, and may be determined as an average value, a median value, or a quartile value of contract probabilities calculated by a model for each product. For example, product A is an unpopular product that does not lead to an actual contract, and the product A contract probability distribution obtained as a result of the calculation of the product recommendation model may be concentrated in an overall low section. In this case, if the first quartile value for the entire probability distribution is used as the predetermined probability value for product A, rather than a fixed value, the product is provided to at least 1/4 of the customers who have entered the data, who are judged to have a high contract probability. It is possible to recommend the product, so that it is not recommended at all or the product is abolished even if the probability of contracting the product is low. Furthermore, when the processor 110 compares the contract probability, which is the output of the recommendation model for each product, with a predetermined probability value to derive whether to recommend the product, the final output may be displayed as a binary value indicating recommendation or non-recommendation.

According to an embodiment of the present disclosure, the processor 110 may train the product recommendation model to extract features of utterances included in the learning data. The characteristics of the utterance content may include the importance of individual data included in the learning data. The importance of the individual data may include an importance for each word, an importance for each token, or an importance for each phrase included in the text data. The processor 110 may extract the importance of individual data based on an error back-propagation method that is calculated when learning a deep learning-based product recommendation model. Specifically, in the course of learning the product recommendation model, adjustments may be made for each epoch with respect to the internal weight and bias of the artificial neural network. The adjustment of the weight and bias values can be propagated back to the chain rule. In this case, the adjustment of the weight and bias values is performed in the reverse order from the output end to the input end for each layer and filter inside the artificial neural network, and finally The first input is reached. In this case, the processor 110 may determine the importance of the individual data based on the weight of the input terminal corresponding to each individual data included in the initial input data. The processor 110 may determine the importance of individual data as the sum of weights corresponding to each individual data of the input data. For example, in the question "What are the conditions for joining the new product A?", texts such as "new" and "condition" repeatedly appear for multiple products and have a weak correlation with product A recommendation information. may be a low word. On the other hand, if the customer directly mentions product A as “product A” or uses a word such as “subscription”, it may be a highly important word having a high correlation with product A recommendation information, so the processor 110 recommends product When learning the model, it can be recognized and extracted as a feature of the speech content. In addition, in the above example, in addition to words that have a strong positive correlation with the contract probability, such as "join", words that have a strong negative correlation with the contract probability, such as "withdrawal", "cancellation", "recommend other products", or The sentence is also recognized as a high-importance word as a meaning not to subscribe to the product and can be extracted as a feature of the utterance. The above-described words and sentences are merely examples and do not limit the present disclosure.

The processor 110 may perform a pre-processing operation including stopword processing based on the characteristics of the utterance content or calculate the contribution to calculation based on the characteristics of the utterance content. Here, the calculation contribution may be an extraction criterion when the processor 110 generates an array by extracting a predetermined number of factors from text data as described above. In the present disclosure, extracting the features of the utterances included in the learning data while the processor 110 learns the product recommendation model has the effect of increasing the user's understanding of the final output of the model. In addition, by acquiring the characteristics of speech content that have a strong correlation with the actual product contract, it can be provided to the above-mentioned pre-processing and used for data purification including stopword processing. For example, if words such as "new" and "condition" were not treated as stopwords in the text data preprocessing process, but were classified as low-importance words as a result of extracting utterances from the product recommendation model, they are added to the stopword list in the preprocessing process. It can also be added to effectively compress data and reduce pre-processing time and inference time. Furthermore, the learning data actually used by the product recommendation model and the pre-processing process for the learning data are organically used by the product recommendation model to calculate the calculation contribution when generating an array for the input text data based on the importance of the individual data included in the characteristics of the speech content. Combined, the performance of the model can be improved.

The processor 110 may generate product recommendation information based on the vector data included in the learning data. The product recommendation model may be a deep learning-based model including at least one or more network functions. Hereinafter, a specific embodiment in which a deep learning-based product recommendation model generates product recommendation information based on vector data according to the present disclosure will be described.

According to an embodiment of the present disclosure, the product recommendation model may be formed of a combination of at least one or more convolutional neural networks and at least one or more recurrent neural networks. In this case, the recurrent neural network may be a recurrent neural network to which a Long Short Term Memory (LSTM) model is applied. For example, the product recommendation model may have a structure in which an artificial neural network of a CNN structure and an artificial neural network of an RNN structure are connected. Specifically, the processor 110 may input an array related to vector data representing text data to the product recommendation model. The array may be a one-dimensional linear array, a simple arrangement of each factor of text data, or a two-dimensional array as a matrix as described above. The processor 110 may input an arrangement related to the text data to at least one or more convolutional neural networks included in the product recommendation model, and then input the output to at least one or more recurrent neural networks.

The final result value of the product recommendation model will be described with reference to FIG. 5 . 5 is an exemplary diagram illustrating some data among final results of a product recommendation model according to an embodiment of the present disclosure. The processor 110 compares the contract probability 510 for each product included in the product recommendation information, which is the output of the product recommendation model, with whether or not the actual contract is 530 for each product included in the product contract information of the learning data. The recommendation model can be backpropagated to learn the difference as an error. Accordingly, by including the convolutional neural network, the product recommendation model can not only understand and analyze the characteristics of input text data in a variety and depth, but also sequentially input it into the recurrent neural network and derive the results, thereby the relationship between continuous text data. It is possible to output the product contract probability by considering up to . In general, when an artificial neural network is configured only with a convolutional neural network, the analysis ability for input data having a specific time including an image increases, but there is a limit in interpreting the mutual influence of different input data at different time points. On the other hand, when an artificial neural network is composed of only a recurrent neural network, it shows high performance in analyzing the mutual influence and meaning of data in analyzing sequential data according to time order including text data, but deep analysis of individual data is limited. Therefore, in the present disclosure, as described above, the processor 110 uses a combination of at least one or more convolutional neural networks and at least one or more recurrent neural networks as a product recommendation model to configure a neural network alone or with a recurrent neural network alone. The analysis efficiency for text data is increased. In other words, since deep analysis and semantic extraction are performed on individual units of text data, and then cross-semantic analysis is also performed with before and after text data, the performance of the product recommendation model can be increased.

According to an embodiment of the present disclosure, the processor 110 may train the product recommendation model to extract a keyword from parsing tree-based syntax analysis information of the learning data and generate product recommendation information based on the keyword. have. As described above, the parsing tree may hold elements (noun phrases, verb phrases, adjective phrases, etc.) of each sentence constituting text data as internal node values. Accordingly, the processor 110 may learn the deep learning-based product recommendation model based on the sentence structure included in the analysis result of the parsing tree. For example, "Hello", "Thank you." A sentence in which a verb phrase is combined with a noun phrase such as “Is it possible” or “I want a contract” can have more meaning in relation to a product contract than a sentence composed of simple verb phrases such as “Is it possible?” Therefore, the processor 110 inputs the structure of each sentence together with the text data as a result of the parsing tree-based syntax analysis into the deep learning-based product recommendation model, and the product recommendation model places a higher weight on the sentence structure and word to be recognized. You can provide a location and extract keywords from it. The keywords include sentences, phrases, words, morphemes, and the like. In addition, the processor 110 may extract the keyword and generate final product recommendation information based on the number of keywords included in the text data, the frequency of appearance, and the like. As described above, by additionally inputting or feeding the sentence structure according to the parsing tree-based syntax analysis to the product recommendation model, the product recommendation model can extract important sentences and word candidates faster than other parts when a large amount of text data is input. This has the effect of improving the performance and output speed of the recommendation model.

According to an embodiment of the present disclosure, the processor 110 inputs new customer call voice data into the learned transformation model to obtain new text data, and inputs new data based on this into the product recommendation model to generate product recommendation information. can do. By obtaining text data from voice data in the learned transformation model and obtaining product recommendation information from text data by the product recommendation model, the present disclosure provides a method for directly creating text data or analyzing the text data one by one. There is an effect that product recommendation becomes possible only by receiving voice data without any need. In another embodiment, when a customer contracts for a product based on the product recommendation information, the processor 110 may re-learn the product recommendation model using text data of the customer. The model re-learning may be performed by repeating the learning of the model at least once by inputting the text data of the customer who finally made the contract as training data. As described above, model re-learning allows the model to learn the utterances of customers who have succeeded in contracting as a result of marketing from a large amount of text data several more times, so that the product recommendation model can better understand the characteristics of the utterances of customers with high contract probability. It has the effect of improving the performance of the model by understanding it and giving it a high weight.

One of ordinary skill in the art of this disclosure will understand that information and signals may be represented using any of a variety of different technologies and techniques. For example, data, instructions, instructions, information, signals, bits, symbols, and chips that may be referenced in the above description are voltages, currents, electromagnetic waves, magnetic fields or particles, optical field particles or particles, or any combination thereof.

Those of ordinary skill in the art of the present disclosure will recognize that the various illustrative logical blocks, modules, processors, means, circuits, and algorithm steps described in connection with the embodiments disclosed herein include electronic hardware, (convenience For this purpose, it will be understood that it may be implemented by various forms of program or design code (referred to herein as software) or a combination of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. A person skilled in the art of the present disclosure may implement the described functionality in various ways for each specific application, but such implementation decisions should not be interpreted as a departure from the scope of the present disclosure.

The various embodiments presented herein may be implemented as methods, apparatus, or articles of manufacture using standard programming and/or engineering techniques. The term article of manufacture includes a computer program, carrier, or media accessible from any computer-readable storage device. For example, computer-readable storage media include magnetic storage devices (eg, hard disks, floppy disks, magnetic strips, etc.), optical disks (eg, CDs, DVDs, etc.), smart cards, and flash drives. memory devices (eg, EEPROMs, cards, sticks, key drives, etc.). Also, various storage media presented herein include one or more devices and/or other machine-readable media for storing information.

It is to be understood that the specific order or hierarchy of steps in the presented processes is an example of exemplary approaches. Based on design priorities, it is to be understood that the specific order or hierarchy of steps in the processes may be rearranged within the scope of the present disclosure. The appended method claims present elements of the various steps in a sample order, but are not meant to be limited to the specific order or hierarchy presented.

The description of the presented embodiments is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present disclosure is not intended to be limited to the embodiments presented herein, but is to be construed in the widest scope consistent with the principles and novel features presented herein.

Claims (16)

A computer program stored in a computer-readable storage medium, wherein the computer program, when executed on one or more processors, performs the following operations for analyzing customer consultation records using a neural network, the operations comprising:
obtaining text data by inputting the received customer call voice data into a deep learning-based transformation model trained to convert voice information into text information;
generating a training data set in which product contract information is labeled in the text data obtained through the deep learning-based transformation model;
learning the product recommendation model to generate product recommendation information by inputting the training data set into a deep learning-based product recommendation model; and
inputting new customer call voice data into the learned transformation model to obtain new text data, and inputting new data based thereon into the product recommendation model to generate product recommendation information;
includes, and
The product contract information is
Including information on at least one product and the result of each product contract,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
a preprocessing operation including tokenization of the text data after the operation of acquiring the text data;
further comprising,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
a preprocessing operation including at least one of morpheme analysis on the text data, parsing tree-based syntax analysis, and stopword processing on the text data after the operation of acquiring the text data;
further comprising,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
after obtaining the text data, generating vector data based on at least one token obtained as a result of a tokenization operation on the text data;
further comprising,
The vector data is generated to reflect the semantic similarity between tokens,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
generating an array by extracting a predetermined number of factors from a sentence included in the text data after the operation of acquiring the text data;
further comprising,
A computer program stored on a computer-readable storage medium.
6. The method of claim 5,
The predetermined number of factors are extracted based on the calculation contribution,
A computer program stored on a computer-readable storage medium.
6. The method of claim 5,
after generating the array, inputting the array into a convolutional neural network receiving a predetermined size as an input;
further comprising,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
The product recommendation model mentioned above,
It is learned to extract the characteristics of the speech content included in the learning data,
The characteristics of the utterance include the importance of individual data included in the learning data,
A computer program stored on a computer-readable storage medium.
9. The method of claim 8,
After the operation of generating the learning data set, a pre-processing operation including stopword processing based on the characteristics of the utterance content; and
calculating a calculation contribution based on the characteristics of the utterance content;
further comprising at least one of
A computer program stored on a computer-readable storage medium.
The method of claim 1,
The product recommendation model mentioned above,
Learned to generate product recommendation information based on the vector data included in the learning data,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
The product recommendation model mentioned above,
Consisting of a combination of at least one or more convolutional neural networks and at least one or more recurrent neural networks,
A computer program stored on a computer-readable storage medium.
12. The method of claim 11,
The recurrent neural network is a Long Short Term Memory (LSTM) model applied,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
The product recommendation model mentioned above,
Extracting a keyword from the parsing tree-based syntax analysis information of the learning data, and generating product recommendation information based on the keyword,
A computer program stored on a computer-readable storage medium.
The method of claim 1,
re-learning the product recommendation model using text data of the customer when a customer contracts for a product based on the product recommendation information;
further comprising,
A computer program stored on a computer-readable storage medium.
As a method for analyzing customer consultation records using a neural network performed on a server,
acquiring text data by inputting the received customer call voice data into a deep learning-based transformation model trained to convert voice information into text information;
generating a training data set in which product contract information is labeled in the text data obtained through the deep learning-based transformation model;
learning the product recommendation model to generate product recommendation information by inputting the training data set into a deep learning-based product recommendation model; and
generating product recommendation information by inputting new customer call voice data into the learned transformation model to obtain new text data and inputting new data based thereon into the product recommendation model;
includes, and
The product contract information is
Including information on at least one product and the result of each product contract,
A method of analyzing customer consultation records using a neural network performed on the server.
As a server for analyzing customer consultation records using a neural network,
a processor including one or more cores; and
Memory;
including,
The processor is
Input the received customer call voice data to a deep learning-based transformation model trained to convert voice information into text information to acquire text data,
generating a training data set in which product contract information is labeled in the text data obtained through the deep learning-based transformation model;
By inputting the training data set into a deep learning-based product recommendation model, the product recommendation model learns to generate product recommendation information,
Input new customer call voice data to the learned transformation model to obtain new text data, and input new data based on this into the product recommendation model to generate product recommendation information, and
The product contract information is
Including information on at least one product and the result of whether each product is contracted,
Server for customer consultation record analysis using neural network.

Images