A real estate investment consultant dialogue quality inspection method, device and computer equipment

Abstract

The application discloses a real estate investment consultant dialogue quality inspection method and device and computer equipment, the method comprises the following steps: obtaining the related information and real-time communication text of the investment consultant and the customer; the communication text is cleaned, cut and constructed into a standard dialogue text data set; the dialogue text data set is determined according to the quality inspection composite dictionary, and the quality inspection label to which the standard word belongs is output; the key words in the dialogue text data set are extracted, and a sentence vector is generated; the similarity between the sentence vector and the corpus sentence vector under the quality inspection label is calculated; whether the similarity reaches a threshold value is judged, if the similarity reaches the threshold value, the quality inspection label is hit, and the dialogue process feedback is given according to the positive and negative of the label. The application improves the field adaptation ability of the dialogue quality inspection, reduces the post-maintenance cost of the dialogue quality inspection model, and realizes the dialogue quality inspection label determination of informal text under a small sample.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of dialogue quality inspection, in particular to a real estate consultant dialogue quality inspection method and device and computer equipment. BACKGROUND

[0002] With the gradual digitization of the real estate industry, the communication between real estate consultants and customers is gradually moving to WeChat for Work. With the deployment and expansion of WeChat for Work, the quality of communication by real estate consultants determines the income and customer satisfaction of the enterprise. Therefore, dialogue quality inspection and process monitoring during the communication between real estate consultants and customers are the key to enhancing the efficiency of private domain operation of enterprises. The mainstream "regular" mode dialogue quality inspection system has very serious missed detection and false detection, and has been difficult to meet the growing communication text quality inspection needs of real estate consultants.

[0003] Dialogue quality inspection can also be referred to as text quality inspection. The main task of text quality inspection is to find out the content that should not exist in the text (negative quality inspection) or the content that should exist but is missed (positive quality inspection), and the business can execute a punishment and reward mechanism according to needs. The current technical solutions for text quality inspection are divided into three categories: 1. Based on a dictionary, the method based on a dictionary is to use existing prior knowledge to construct a quality inspection label dictionary, and to determine whether the text hits the quality inspection label through word matching. 2. Based on a dictionary and regular expressions, the method based on a dictionary and regular expressions improves the generalization ability of model word recognition by designing regular expression templates, and adds rules between words to the dictionary of quality inspection labels. Whether the text hits the quality inspection label is determined by comparing the multiple words hit by the text with the preset rules. 3. Based on a deep learning algorithm model, the method based on a deep learning algorithm model treats text quality inspection as a semantic understanding task. A word-level self-encoding neural network is trained using a large amount of corpus to learn the semantics of each word in the corpus, and then used for text encoding. A deep learning algorithm model is used to perform secondary training on positive and negative example quality inspection texts that have been manually labeled with quality inspection labels, so that it can learn the semantic information of each quality inspection label and determine the quality inspection label of the text.

[0004] However, the above three methods have their own advantages and disadvantages. The explainability of the dictionary-based and dictionary-based and regular-based methods is the strongest, the deployment and the initial operation are relatively fast and the performance is very stable, and the accuracy is relatively stable. However, the two methods depend on the design of the dictionary or the size of the regular template. With the diversification of words, the informal structure of the text, and the gradual increase of the dictionary and the regular template, the recall rate and maintainability of the model will decrease rapidly, and the conditions of missed detection and false detection will continue to improve the maintenance cost. The deep learning algorithm model can construct text semantic coding from the text resources and statistical data obtained from a large amount of corpus, and realize the end-to-end model effect. To some extent, the effect is proportional to the richness of the corpus, so the problem is that a large amount of manual work is needed to perform field text quality inspection label annotation work in the industry field outside the general field. Without a certain amount of manpower and sample size, it is difficult to reach the standard of landing application.

[0005] In summary, the dictionary-based method has high accuracy and low recall, depends on the size of the dictionary or the sentence template, and lacks domain adaptation ability. The dictionary-based and regular-based method is not suitable for informal text with diversified text structure, and requires expert experience for regular template design and high maintenance cost. The deep learning algorithm model needs a large amount of manual work to perform field text quality inspection label annotation work, and it is difficult to reach the available level in the development stage of insufficient domain text level. SUMMARY

[0006] Therefore, the present application provides a real estate consultant dialogue quality inspection method, device and computer equipment to solve the problems of lack of domain adaptation ability, high maintenance cost and the need for a large amount of manual quality inspection label annotation in the prior art.

[0007] In order to achieve the above purpose, the present application provides the following technical solutions:

[0008] In a first aspect, a real estate consultant dialogue quality inspection method comprises:

[0009] Obtaining the relevant information and real-time communication text of the real estate consultant and the customer;

[0010] Cleaning, cutting and constructing the communication text into a standard dialogue text dataset;

[0011] Performing dialogue text judgment on the dialogue text dataset according to a quality inspection composite dictionary, and outputting the quality inspection label to which the standard word belongs;

[0012] Extracting the key words in the dialogue text dataset, and generating a sentence vector with the key words as an abstract;

[0013] Calculating the similarity between the sentence vector and the corpus sentence vector under the quality inspection label;

[0014] determining whether the similarity reaches a threshold value,

[0015] if the similarity reaches the threshold value, a quality inspection label is hit, and a flow feedback is given according to the positive and negative of the label.

[0016] Preferably, the quality inspection composite dictionary comprises a fixed word library and a corpus sentence library, and the fixed word library comprises constant words and vector words.

[0017] Preferably, the quality inspection composite dictionary is used to determine the flow text of the dialogue text data set, and the fixed word belongs to a quality inspection label, specifically as follows:

[0018] determining whether there is a constant word in the dialogue text data set according to the quality inspection composite dictionary;

[0019] if there is, the fixed word belongs to a quality inspection label is directly outputted;

[0020] if there is not, the dialogue text data set is segmented by a word vector model, and a word vector list is generated;

[0021] the similarity between the word vector list and the vector words of the fixed word library is calculated;

[0022] determining whether the similarity meets a set threshold value;

[0023] if it meets, the fixed word belongs to a quality inspection label is outputted;

[0024] if it does not meet, the dialogue text set does not have quality inspection label information.

[0025] Preferably, the word vector model is obtained by using a word2vec algorithm to train a real estate industry worker card and a telephone text as a basic corpus.

[0026] Preferably, the key words in the dialogue text data set are extracted by a TextRank algorithm and a high-low order of word scores.

[0027] Preferably, the sentence vector is generated by using the key words as an abstract, and is obtained by a first formula, wherein the first formula is:

[0028]

[0029] wherein, WS(V i ) represents the score of word i, W ij represents the similarity between word i and word j, WS(V j ) represents the score of iteration word j, In(V i ) represents all words associated with word i in the field text, and Out(V j) is the iteration word j in all associated words in the field text, d is the damping coefficient.

[0030] As a preference, the damping coefficient d is 0.85.

[0031] As a preference, it also includes:

[0032] The dialogue quality inspection score is calculated, and the proportion of the number of positive quality inspection labels in the range of the number of quality inspection process labels is calculated.

[0033] In a second aspect, a real estate property consultant dialogue quality inspection device comprises:

[0034] A communication text acquisition module is configured to acquire relevant information and real-time communication text of a property consultant and a customer.

[0035] A dialogue text dataset module is configured to clean and cut the communication text and build a standard dialogue text dataset.

[0036] A dialogue text determination module is configured to determine dialogue text of the dialogue text dataset according to a quality inspection composite dictionary and output a quality inspection label to which a standard word belongs.

[0037] A sentence vector generation module is configured to extract keywords in the dialogue text dataset and generate sentence vectors using the keywords as abstracts.

[0038] A similarity calculation module is configured to calculate the similarity between the sentence vectors and the sentence vectors of the corpus under the quality inspection label.

[0039] A judgment module is configured to determine whether the similarity reaches a threshold value. If the similarity reaches the threshold value, it is determined that the quality inspection label hits, and a dialogue process feedback is given according to the positive and negative nature of the label.

[0040] In a third aspect, a computer device comprises a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of the real estate property consultant dialogue quality inspection method are implemented.

[0041] Compared with the prior art, the present application has at least the following beneficial effects:

[0042] The application provides a real estate investment consultant dialogue quality inspection method and device and a computer device. The method comprises the following steps: obtaining relevant information and real-time communication text of an investment consultant and a customer; cleaning and cutting the communication text and constructing a standard dialogue text dataset; performing dialogue text determination on the dialogue text dataset according to a quality inspection composite dictionary, outputting a quality inspection label to which a standard word belongs; extracting keywords in the dialogue text dataset and generating a sentence vector; calculating the similarity between the sentence vector and a corpus sentence vector under the quality inspection label; judging whether the similarity reaches a threshold value, if the similarity reaches the threshold value, the quality inspection label is hit, and a dialogue process feedback is given according to the positive and negative nature of the label. The application improves the field adaptation capability of dialogue quality inspection, reduces the post-maintenance cost of the dialogue quality inspection model, and realizes the dialogue quality inspection label determination of informal text under a small sample. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more intuitively illustrate the prior art and the application, several exemplary drawings are given below. It should be understood that the specific shapes, structures shown in the drawings should not be regarded as limiting conditions in the implementation of the application; for example, based on the technical concepts and exemplary drawings disclosed in the application, those skilled in the art can easily make routine adjustments or further optimizations to the increase / decrease / ownership division, specific shape, positional relationship, connection mode, size ratio relationship, etc. of certain units (components).

[0044] Figure 1 A real estate investment consultant dialogue quality inspection method flowchart is provided for the first embodiment of the application.

[0045] Figure 2 A dialogue text dataset generation schematic diagram is provided for the first embodiment of the application.

[0046] Figure 3 A quality inspection composite dictionary structure schematic diagram is provided for the first embodiment of the application.

[0047] Figure 4 A dialogue text determination flowchart is provided for the first embodiment of the application.

[0048] Figure 5 A quality inspection label matching and feedback flowchart is provided for the first embodiment of the application. DETAILED DESCRIPTION

[0049] The application will be further described in detail below with reference to the accompanying drawings.

[0050] In the description of the present application: unless otherwise specified, the meaning of "a plurality" is two or more. The terms "first", "second", "third" and the like in the present application are intended to distinguish the objects referred to, and do not have a special meaning in the technical connotation aspect (for example, it should not be understood as emphasizing importance or order, etc.). The expressions such as "include", "contain", "have" and the like also mean "not limited to" (certain units, components, materials, steps, etc.).

[0051] The terms such as "upper", "lower", "left", "right", "middle" and the like cited in the present application are generally for the purpose of facilitating intuitive understanding by referring to the drawings, and are not an absolute limitation on the positional relationship in the actual product. Change of these relative positional relationships is also considered within the scope of the present application without departing from the technical concept disclosed in the present application.

[0052] Embodiment one

[0053] Please refer to Figure 1 The embodiment provides a real estate investment consultant dialogue quality inspection method, which comprises the following steps:

[0054] S1: obtaining the relevant information of the real estate consultant and the customer and the real-time communication text;

[0055] Specifically, the enterprise WeChat private deployment is customized and developed according to the api interface provided by the enterprise WeChat to obtain the relevant information of the real estate consultant and the customer and the real-time communication text.

[0056] S2: cleaning, cutting and constructing the communication text into a standard dialogue text data set;

[0057] Specifically, the dialogue text data set can refer to Figure 2 .

[0058] S3: determining the dialogue text according to the quality inspection composite dictionary to obtain the quality inspection label to which the standard word belongs;

[0059] Please refer to Figure 3 In order to improve the model recognition efficiency and reduce the cost of dictionary design, the present application designs a quality inspection composite dictionary, which is composed of a corpus sentence library and a standard word library.

[0060] Among them, the corpus sentence library is mainly composed of sentences, and the corpus sentence library is a small amount of artificially annotated dialogue text conforming to the quality inspection label.

[0061] The calibration word library is mainly composed of words. In order to further improve the generalization ability of word matching and cope with the diversification of text in informal text, the calibration word library is further divided into constant words and vector words to balance the accuracy and recall rate of the model. The vector word is a keyword similar to the quality inspection label and has a certain representative, which can be specified by the business personnel. The constant word is a keyword entity with a word frequency of top 10 extracted from the corresponding quality inspection label corpus sentence library by using the TP-IDF algorithm.

[0062] Please refer to Figure 4 Specifically, when the dialogue text data set is determined according to the quality inspection composite dictionary, it is divided into two levels of determination. The first level of determination uses the calibration word library to filter the constant words in the dialogue text data set, and the second level of determination uses jieba segmentation to segment the dialogue text data set. The word2vec algorithm is used to train the real estate industry's worker card and telephone text as the basic corpus as the word vector model of the application. The dialogue text is segmented and converted into a corresponding word vector by the word vector model. The cosine similarity of the word vector and the vector word of the calibration word library is calculated. The segmented text is filtered out after meeting the set threshold.

[0063] S301: determining whether there is a constant word in the dialogue text data set according to the quality inspection composite dictionary;

[0064] S302: if there is, directly outputting the quality inspection label to which the calibration word belongs;

[0065] S303: if there is not, segmenting the dialogue text set by the word vector model and generating a word vector list;

[0066] S304: calculating the similarity of the word vector list and the vector word of the calibration word library;

[0067] S305: determining whether the similarity meets the set threshold;

[0068] S306: if it meets, outputting the quality inspection label to which the calibration word belongs;

[0069] S307: if it does not meet, the dialogue text set does not have quality inspection label information.

[0070] S4: extracting keywords in the dialogue text data set and generating a sentence vector with the keywords as an abstract;

[0071] Specifically, the sentence vector used in the application is obtained by averaging the stacking of the segmented vector. Since there are redundant words in informal text that will affect the quality of the generation of the sentence vector, the application uses the abstract extraction method to preprocess the text. The keywords in the dialogue text are extracted by the TextRank algorithm and the high and low order of word scoring, and the sentence vector is generated with the keywords as an abstract. The algorithm formula is as follows:

[0072]

[0073] wherein, WS(V i ) represents the score of word i, W ij represents the similarity between word i and word j, WS(V j ) represents the score of iteration word j, In(V i ) represents all the words associated with word i in the field text, Out(V j ) is all the associated words of iteration word j in the field text, and d is the damping coefficient, generally 0.85.

[0074] Preferably, the sentence vector can also be generated by a deep learning algorithm.

[0075] S5: Calculate the similarity between the sentence vector and the corpus sentence vector under the corresponding label;

[0076] Specifically, the script text summary sentence vector generated in step S4 is calculated for similarity with the corpus sentence vector under the corresponding label.

[0077] S6: Determine whether the similarity reaches the threshold value;

[0078] S7: If the similarity reaches the threshold value, the quality inspection label is hit, and feedback is given to the script process according to the positive and negative nature of the label;

[0079] Please refer to Figure 5 , specifically, if the maximum similarity reaches the threshold value, it is determined that the quality inspection label is hit, and feedback is given to the script process according to the positive and negative nature of the label, triggering the enterprise micro data interface to be transmitted to the enterprise micro interface of the real estate consultant to prompt, if it is a positive quality inspection, it prompts that the specified process has been completed and gives a prompt information of the next process, if it is a negative quality inspection, it prompts the real estate consultant to trigger the prohibited label, and informs the consultant to execute the message withdrawal to correct the process error.

[0080] S7: Calculate the script quality inspection score and the process sequence completion degree;

[0081] Specifically, according to the business personnel demand, set the weight and score calculation formula of each quality inspection label, according to the number n of positive quality inspection labels touched by the real estate consultant and the customer in the communication process, calculate the proportion (n / m*100%) in the range of the specified quality inspection process label number m as the process sequence completion degree reached in the current customer communication, and take the process sequence completion degree as part of the performance reference.

[0082] The real estate investment consultant dialogue quality inspection method provided in the application needs to periodically sample model quality inspection results and manually label, calculate model accuracy and recall rate, if the accuracy rate appears to decline and is lower than the threshold set by the business, then the newly accumulated new work cards and telephone texts in recent period are added to the word2vec training corpus to continue training the word vector optimization model accuracy, if the recall rate appears to decline and is lower than the threshold set by the business, then the missed samples are expanded into the corpus sentence library to improve the constant word volume and optimize the model recall rate.

[0083] The application adopts text corpus generated by other business lines in the field to train word vectors, which improves the field adaptation ability of the dialogue quality inspection scheme; through the quality inspection composite dictionary optimization scheme and the unsupervised word vector model training method, the post-maintenance cost of the dialogue quality inspection model is reduced; the quality inspection composite dictionary achieves a balance between manpower investment, post-maintenance cost and model availability, realizes the dialogue quality inspection label determination of informal text under a small sample, and can be used not only for informal text, but also for formal text.

[0084] Embodiment Two

[0085] The embodiment provides a real estate investment consultant dialogue quality inspection device, which comprises:

[0086] A communication text acquisition module is configured to acquire relevant information and real-time communication texts of an investment consultant and a customer.

[0087] A dialogue text dataset module is configured to clean and cut the communication texts and construct standard dialogue text datasets.

[0088] A dialogue text determination module is configured to determine dialogue texts of the dialogue text datasets according to a quality inspection composite dictionary and output quality inspection labels to which the determined words belong.

[0089] A sentence vector generation module is configured to extract keywords in the dialogue text datasets and generate sentence vectors by taking the keywords as abstracts.

[0090] A similarity calculation module is configured to calculate the similarity between the sentence vectors and corpus sentence vectors under the quality inspection labels.

[0091] A judgment module is configured to judge whether the similarity reaches a threshold value, if the similarity reaches the threshold value, it is determined that the quality inspection label hits, and a dialogue process feedback is given according to the positive and negative nature of the label.

[0092] The specific limitations of the real estate investment consultant dialogue quality inspection device can be seen in the limitations of the real estate investment consultant dialogue quality inspection method in the foregoing, which will not be repeated here.

[0093] Embodiment Three

[0094] The embodiment provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor realizes the steps of the real estate investment consultant dialogue quality inspection method when executing the computer program.

[0095] Any combination of the technical features in the above embodiments can be made (as long as the combination of the technical features does not exist contradiction), in order to make the description simple, all possible combinations of each technical feature in the above embodiments are not described; these embodiments which are not explicitly written should also be considered as the scope of the present application.

[0096] The application is described in detail above through general description and specific embodiments. It should be understood that, based on the technical concept of the application, some conventional adjustments or further innovations can be made to these specific embodiments; as long as the technical concept of the application is not deviated, the technical solutions obtained by these conventional adjustments or further innovations also fall within the protection scope of the claims of the application.

Claims

1. A real estate agent script quality inspection method, characterized in that, The method comprises the following steps: Obtaining relevant information and real-time communication text of the real estate consultant and the customer; Cleaning, cutting and constructing the communication text into a standard dialogue text dataset; According to the quality inspection composite dictionary, the dialogue text dataset is determined for the text of the dialogue, and the quality inspection label to which the standard word belongs is output. The quality inspection composite dictionary is composed of a corpus sentence library and a standard word library. The corpus sentence library is mainly composed of sentences. The corpus sentence library is a small amount of artificially annotated text of dialogue with quality inspection labels. The standard word library is mainly composed of words, including constant words and vector words. The vector word is a key word similar to the quality inspection label and has a representative. The constant word is a key entity word with a word frequency Top10 extracted from the corpus sentence library corresponding to the quality inspection label by using a TP-IDF algorithm; Extracting key words in the dialogue text dataset and generating a sentence vector based on the key words as an abstract; Calculating the similarity between the sentence vector and the corpus sentence vector under the quality inspection label; Judging whether the similarity reaches a threshold value, If the similarity reaches the threshold value, the quality inspection label is hit, and a dialogue process feedback is given according to the positive and negative nature of the label; According to the quality inspection composite dictionary, the dialogue text dataset is determined for the text of the dialogue, and the quality inspection label to which the standard word belongs is output. Specifically, According to the quality inspection composite dictionary, it is judged whether there is a constant word in the dialogue text dataset; If there is, the quality inspection label to which the standard word belongs is directly output; If not, the dialogue text dataset is segmented by a word vector model, and a word vector list is generated; Calculating the similarity between the word vector list and the vector word of the standard word library; Judging whether the similarity meets a set threshold value; If it meets, the quality inspection label to which the standard word belongs is output; If it does not meet, the dialogue text dataset does not have quality inspection label information.

2. The real estate agent script quality checking method according to claim 1, wherein, The word vector model is obtained by using a word2vec algorithm to train the real estate industry's worker card and dialogue text as basic corpus.

3. The real estate agent script quality checking method according to claim 1, wherein, The key words in the dialogue text dataset are extracted by using a TextRank algorithm and word score sorting.

4. The real estate agent dialogue quality inspection method according to claim 1, wherein, The sentence vector is generated based on the key words as an abstract, which is obtained by a first formula. where WS(V i ) denotes the score of word i, W ij denotes the similarity between word i and word j, WS(V j ) represents the score of iteration word j, In(V i ) represents all the associated words of word i in the domain text, Out(V j ) is all the associated words of iteration word j in the domain text, and d is the damping coefficient.

5. The real estate agent dialogue quality inspection method according to claim 4, characterized in that, The damping coefficient d is 0.

85.

6. The real estate agent dialogue quality inspection method according to claim 1, wherein, Further comprising: Calculating the quality inspection score of the dialogue, and calculating the proportion of the positive quality inspection label in the range of the specified quality inspection process label number.

7. A real estate consultant dialogue quality inspection device, characterized in that, The method comprises the following steps: A communication text acquisition module is configured to obtain relevant information and real-time communication text of the real estate consultant and the customer; A dialogue text dataset module is configured to clean, cut and construct the communication text into a standard dialogue text dataset; The dialogue text judgment module is configured to perform dialogue text judgment on the dialogue text dataset according to a quality inspection composite dictionary, and output a quality inspection label to which a marking word belongs; the quality inspection composite dictionary is composed of a corpus sentence library and a marking word library, the corpus sentence library is mainly composed of sentences, the corpus sentence library is a small amount of artificially annotated dialogue text conforming to a quality inspection label, the marking word library is mainly composed of words, including constant words and vector words, the vector words are representative keywords similar to the quality inspection label, and the constant words are key entity words with a word frequency of Top 10 in the corpus sentence library corresponding to the quality inspection label and extracted by using a TP-IDF algorithm; The sentence vector generation module is configured to extract keywords in the dialogue text dataset, and generate a sentence vector by taking the keywords as an abstract; The similarity calculation module is configured to calculate the similarity between the sentence vector and a corpus sentence vector under the quality inspection label to which the sentence vector belongs; The judgment module is configured to judge whether the similarity reaches a threshold value, and if the similarity reaches the threshold value, it is determined that the quality inspection label is hit, and a dialogue process feedback is given according to the positive and negative nature of the label. The dialogue text judgment module is configured to perform dialogue text judgment on the dialogue text dataset according to a quality inspection composite dictionary, and output a quality inspection label to which a marking word belongs; the quality inspection composite dictionary is composed of a corpus sentence library and a marking word library, the corpus sentence library is mainly composed of sentences, the corpus sentence library is a small amount of artificially annotated dialogue text conforming to a quality inspection label, the marking word library is mainly composed of words, including constant words and vector words, the vector words are representative keywords similar to the quality inspection label, and the constant words are key entity words with a word frequency of Top 10 in the corpus sentence library corresponding to the quality inspection label and extracted by using a TP-IDF algorithm; 8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to realize the steps of the method in any one of claims 1 to 6. The processor executes the computer program to realize the steps of the method in any one of claims 1 to 6.

Images