A method and system for automatically identifying customer intentions based on NLP

Abstract

The present invention relates to the field of natural language processing technology, and in particular to a method and system for automatically identifying customer intentions based on NLP. The present invention collects interaction data of an interactive interface and analyzes state characteristics based on the interaction data; determines an intention representation coefficient for the sender based on the state characteristics during the interaction duration to classify the intention degree category for the sender; calls the interaction data corresponding to the sender, and adaptively analyzes the called interaction data based on the intention degree category. The present invention can promptly focus on potential customers with strong intentions, improve the accuracy of customer intention identification, and at the same time, can more quickly match the real needs of customers with fuzzy intentions, improve the overall operational efficiency of the responding end, and provide customers with a good communication experience.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a method and system for automatically identifying customer intentions based on NLP. Background Art

[0002] Against the backdrop of increasingly fierce market competition, product providers are increasingly prioritizing insights into their customers in order to better meet their needs and enhance their satisfaction and loyalty. Therefore, identifying customer intentions has become an important step for product providers to increase their product sales rates. Natural language processing (NLP) technology is utilized in the process of identifying customer intentions. By training on a large amount of customer communication texts, NLP can deeply learn customer intentions in different expressions, improve its ability to process complex language structures and meanings, and provide more powerful technical support for identifying customer purchase intentions. This helps product providers quickly and accurately understand customer purchase intentions, and provides decision-making support for product providers' marketing, product development, and customer service.

[0003] Chinese patent application publication number: CN114118080A, discloses a method and system for automatically identifying customer intentions from sales conversations, belonging to the field of natural language processing technology, and solves the problems of low accuracy and recall rate in existing customer intention identification methods. A method for automatically identifying customer intentions from sales conversations, including: dividing the sales conversation to obtain multiple conversation segments with identification role tags; performing answer matching on all conversation segments to obtain context conversation segments with answer relationships; treating each conversation segment except the context segment as an independent conversation segment, and judging whether each context conversation segment and independent conversation segment has an intention point respectively, and if so, also identifying the customer intention emotion of the current context conversation segment or independent conversation segment; summarizing the intention points and customer intention emotions of all context conversation segments and independent conversation segments as the customer intention identification result of the sales conversation.

[0004] However, the prior art still has the following problems:

[0005] When identifying customer intentions during online interactions, we mainly rely on the surface meaning of the message text sent by the customer to determine whether the customer has intentions and the degree of intentions. This simple reliance on subjective judgment cannot fully determine the customer's potential intentions, thereby reducing the accuracy of customer intention identification. Summary of the Invention

[0006] To this end, the present invention provides a method and system for automatically identifying customer intentions based on NLP, so as to overcome the problem in the prior art that, when identifying customer intentions during network interactions, the method mainly relies on the surface meaning of the message text sent by the customer during the interaction to distinguish whether the customer has intentions and the degree of intentions. This simple reliance on subjective judgment cannot comprehensively determine the customer's potential intentions, thereby reducing the accuracy of customer intention recognition.

[0007] To achieve the above objectives, the present invention provides a method and system for automatically identifying customer intentions based on NLP, which includes:

[0008] Collecting interaction data of the interactive interface, and analyzing state characteristics based on the interaction data, including the number of times the input state of the sending end switches during the interaction and the duration of the first input state;

[0009] Determining an intention representation coefficient for the sender based on the state characteristics during the interaction duration to classify the intention degree category for the sender;

[0010] Calling the interaction data corresponding to the sending end, and analyzing the called interaction data according to the intention level category of the sending end, including:

[0011] Performing logical coherence analysis on each text segment sent by the sender in the interaction data, identifying intentionally coherent text segment groups, determining logical closeness based on semantic similarity of the intentionally coherent text segment groups, and analyzing a representation value of the sender's intention intensity in combination with the number of occurrences of product detail words to determine whether to mark the sender;

[0012] Alternatively, the text segment is segmented to obtain a number of keywords, and the matching degree between the product detail words corresponding to each product in the product library of the responding end and each keyword is determined to determine whether to mark the sending end.

[0013] Furthermore, the input state includes a first input state and a second input state, and the input state is determined based on the instruction input device of the sending end, including:

[0014] If the instruction input device continues to input instructions, it is determined that the input state is the first input state;

[0015] If the instruction input device does not continuously input instructions, determining that the input state is the second input state;

[0016] The continuously input command means that the time interval between the input adjacent command signals is not less than a predetermined time interval threshold.

[0017] Furthermore, the process of determining the intention representation coefficient for the sending end based on the state characteristics includes:

[0018] The ratio of the number of switching times of the input state of the sending end during the interaction duration to the switching number threshold is used as the first intention feature;

[0019] taking a ratio of a duration of the first input state within the interaction duration to a duration threshold as a second intention feature;

[0020] The first intention feature and the second intention feature are weighted and summed to determine an intention representation coefficient.

[0021] Furthermore, the intention degree categories for the sender are divided into:

[0022] If the intention representation coefficient is greater than or equal to the intention representation coefficient threshold, classifying the intention degree of the sending end into a clear intention degree category;

[0023] If the intention representation coefficient is less than the intention representation coefficient threshold, the intention degree of the sending end is classified into a fuzzy intention degree category.

[0024] Further, according to the intention level category of the sending end, the called interaction data is analyzed, including:

[0025] If the sender's intention level is in the clear intention level category, performing a logical coherence analysis on each text segment sent by the sender in the interaction data, identifying groups of intentionally coherent text segments, determining logical closeness based on semantic similarity of the groups of intentionally coherent text segments, and analyzing the sender's intention intensity representation value in combination with the number of occurrences of product detail words to determine whether to mark the sender;

[0026] If the intention degree of the sender is a fuzzy intention degree category, the text segment is segmented to obtain several keywords, and the matching degree between the product detail words corresponding to each product in the responding end product library and each keyword is determined to determine whether to mark the sender.

[0027] Furthermore, the process of performing logical coherence analysis on each text segment sent by the sending end in the interactive data and identifying the intention coherent text segment group includes:

[0028] Identify the key words of each text segment through natural language models;

[0029] Determine the matching degree between adjacent text segments corresponding to each subject word;

[0030] identifying logically coherent phrases in adjacent text segments, and determining the number of occurrences of the logically coherent phrases in the adjacent text segments;

[0031] If the adjacent text segments meet the coherent text condition, the adjacent text segments are determined to be an intended coherent text segment group;

[0032] The coherent text condition is that the matching degree between the corresponding subject words of adjacent text segments is greater than or equal to the matching degree threshold and the number of occurrences of logically coherent words in adjacent text segments is greater than or equal to the number of occurrences threshold.

[0033] Furthermore, the process of analyzing the intensity representation value of the intention of the sender includes:

[0034] Extracting intentionally coherent text segments to identify logical closeness;

[0035] taking a ratio of the logic closeness to a logic closeness threshold as a first degree feature;

[0036] The ratio of the number of occurrences of product detail words to the number of occurrence threshold is used as the second degree feature;

[0037] Determining the sum of the first degree feature and the second degree feature as the intention intensity representation value;

[0038] The logical closeness is the proportion of identical or synonymous words in adjacent text segments, and the product detail words are the words included in the product details page.

[0039] Further, determining whether to mark the sending end includes:

[0040] If the intention intensity representation value is greater than or equal to the intention intensity representation threshold, the sending end is marked.

[0041] Further, determining the matching degree between the product details words corresponding to each product in the product library of the responding end and each keyword determines whether to mark the sending end, including

[0042] If the matching degree is greater than or equal to the matching degree threshold, the sending end is marked.

[0043] Furthermore, a system for applying a method for automatically identifying customer intentions based on NLP is provided, which includes:

[0044] An interaction collection module, configured to collect interaction data of the interaction interface and analyze state characteristics based on the interaction data, including the number of times the input state of the sending end switches during the interaction and the duration of the first input state;

[0045] an intention recognition module connected to the interaction acquisition module, configured to determine an intention representation coefficient for the sender based on the state characteristics during the interaction duration, so as to classify the intention degree category for the sender;

[0046] An intention analysis module is connected to the interaction acquisition module and the intention identification module respectively, and is used to call the interaction data corresponding to the sending end, and analyze the called interaction data according to the intention level category of the sending end, including:

[0047] Performing logical coherence analysis on each text segment sent by the sender in the interaction data, identifying intentionally coherent text segment groups, determining logical closeness based on semantic similarity of the intentionally coherent text segment groups, and analyzing a representation value of the sender's intention intensity in combination with the number of occurrences of product detail words to determine whether to mark the sender;

[0048] Alternatively, the text segment is segmented to obtain a number of keywords, and the matching degree between the product detail words corresponding to each product in the product library of the responding end and each keyword is determined to determine whether to mark the sending end.

[0049] Compared with the existing technology, the present invention collects interaction data of the interactive interface and analyzes state characteristics based on the interaction data; determines the intention representation coefficient for the sender based on the state characteristics during the interaction duration to divide the intention degree category for the sender; calls the interaction data corresponding to the sender, and adaptively analyzes the called interaction data according to the intention degree category, and provides a system that uses a method based on NLP to automatically identify customer intentions. The present invention can promptly focus on potential customers with strong intentions, improve the accuracy of customer intention identification, and at the same time, can more quickly match the real needs of customers with vague intentions, improve the overall operational efficiency of the responding end, and provide customers with a good communication experience.

[0050] In particular, the present invention determines the intention representation coefficient for the sender based on the state characteristics within the duration of the interaction. In the actual language interaction to reach an agreement on the purchase of a product, the number of times the sender's input state switches can, to a certain extent, reflect the sender's thinking or language organization during the interaction process, and the first input state can indicate that the sender is continuously inputting text. If the interactive interface is in this state for a long time, it means that the sender may be organizing language to communicate about the relevant product and has a high interest and willingness in the relevant content of the product. Therefore, the present application represents the sender's purchase intention for the product mentioned in the current communication with the responding end based on the intention representation coefficient for the sender. In addition, the present invention's method of determining the sender's intention based on state characteristics is more objective, and does not simply rely on subjective feelings to guess whether the sender has intention and the degree of intention. It provides more objective data support for the subsequent classification of the sender's intention degree category, and then adaptively analyzes the called interaction data to more accurately measure the sender's overall intention.

[0051] In particular, the present invention conducts targeted analysis based on senders of different intention degree categories. For senders of clear intention degree category, it identifies intention coherent text groups, and determines whether the sender's intention is targeted through the logical correlation of the text segment content of the sender received by the response end. Therefore, whether the sender has targeted intention is determined based on the logical correlation of the text segment content, and the text segments with the above characteristics in the adjacent text segments sent by the sender are determined as intention coherent text segment groups. The logical closeness is determined based on the semantic similarity of the content of the intention coherent text segment group and combined with the product details words. The number of occurrences is used to analyze the intensity of the sender's intention. On the premise of determining a logically coherent text segment group, logical closeness can reflect the degree of semantic connection between text segments. The semantic similarity between the text segments can further determine the intensity of the sender's intention. At the same time, the number of occurrences of product detail words can more intuitively reflect the sender's attention to the specific features of the product. Combining these features to determine the intensity of the intention represents the positivity and intensity of the sender's intention, providing data support for subsequent decisions on whether to tag the sender.

[0052] For senders with fuzzy intention levels, the text segment content received from the sender is segmented to obtain keywords to quickly extract key information from the text segment. When the sender's intention is relatively vague, this method can be used to capture key product-related words from the fragment's expression, avoiding omissions due to the sender's obscure expression or unclear needs, while the sender actually has a certain interest in the product. This saves computing power resources and helps the responding end seize the opportunity to quickly carry out subsequent recommendation follow-up work. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of the steps of a method for automatically identifying customer intentions based on NLP according to an embodiment of the invention;

[0054] Figure 2 A logical decision diagram for classifying the degree of intention of a sender according to an embodiment of the invention;

[0055] Figure 3 A logical decision graph for identifying groups of intentionally coherent text segments for an embodiment of the invention;

[0056] Figure 4 This is a functional module diagram of a system for automatically identifying customer intentions based on NLP in an embodiment of the invention. DETAILED DESCRIPTION

[0057] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0058] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0059] It should be noted that, in the description of the present invention, terms such as "upper" and "lower" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0060] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the term "connection" should be understood in a broad sense. For example, it can refer to a fixed connection, a detachable connection, or an integral connection; it can refer to a mechanical connection or an electrical connection. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0061] See also Figure 1 , which is a schematic diagram of the steps of a method for automatically identifying customer intentions based on NLP according to an embodiment of the present invention. The method for automatically identifying customer intentions based on NLP according to an embodiment of the present invention includes:

[0062] Collecting interaction data of the interactive interface, and analyzing state characteristics based on the interaction data, including the number of times the input state of the sending end switches during the interaction and the duration of the first input state;

[0063] Determining an intention representation coefficient for the sender based on the state characteristics during the interaction duration to classify the intention degree category for the sender;

[0064] Calling the interaction data corresponding to the sending end, and analyzing the called interaction data according to the intention level category of the sending end, including:

[0065] Performing logical coherence analysis on each text segment sent by the sender in the interaction data, identifying intentionally coherent text segment groups, determining logical closeness based on semantic similarity of the intentionally coherent text segment groups, and analyzing a representation value of the sender's intention intensity in combination with the number of occurrences of product detail words to determine whether to mark the sender;

[0066] Alternatively, the text segment is segmented to obtain a number of keywords, and the matching degree between the product detail words corresponding to each product in the product library of the responding end and each keyword is determined to determine whether to mark the sending end.

[0067] Specifically, the input state includes a first input state and a second input state, and the input state is determined based on the instruction input device of the sending end, including:

[0068] If the instruction input device continues to input instructions, it is determined that the input state is the first input state;

[0069] If the instruction input device does not continuously input instructions, determining that the input state is the second input state;

[0070] The continuously input command means that the time interval between the input adjacent command signals is not less than a predetermined time interval threshold.

[0071] The predetermined time interval threshold is selected in the interval [5s, 10s].

[0072] Specifically, the instruction input device can be any virtual device or hardware device that can input control instructions, for example, a virtual keyboard or physical keyboard at the sending end, which will not be described in detail here.

[0073] Specifically, the purpose of setting the time interval is to characterize the continuity of the instruction input, so as to represent the situation of the user end thinking and organizing language in actual conditions. Preferably, the predetermined time interval threshold is selected within the interval [4s, 7s].

[0074] It can be understood that the responding end refers to the merchant terminal that sells products on several platforms, and the sending end refers to the user terminal that initiates communication with the responding end regarding product-related aspects or receives product recommendation information sent by the responding end.

[0075] It is understandable that in order to protect and respect the privacy of the sender, the collection of interactive data on the interactive interface must be subject to the sender's consent and authorization before it can be collected. This will not be repeated here.

[0076] In this embodiment, a collection prompt pop-up window is sent to the sending end, and the interaction data of the interactive interface is collected according to the authorization status of the sending end. For example, if the sending end clicks the "Agree" button after following the prompt of the collection prompt pop-up window, the interaction data is collected; if the sending end clicks the "Reject" button after following the prompt of the collection prompt pop-up window, the interaction data cannot be collected. Of course, other methods can also be used, which will not be repeated here.

[0077] In this embodiment, the duration of the session is the time from when either the responding end or the sending end sends a message to the other party and delivers it to the other party's interactive interface to when either end closes the medium providing the interactive interface, where the medium can be an application software or a mobile web page, which will not be repeated here.

[0078] Specifically, the process of determining the intention representation coefficient for the sending end based on the state characteristics includes:

[0079] The ratio of the number of switching times of the input state of the sending end during the interaction duration to the switching number threshold is used as the first intention feature;

[0080] taking a ratio of a duration of the first input state within the interaction duration to a duration threshold as a second intention feature;

[0081] The first intention feature and the second intention feature are weighted and summed to determine an intention representation coefficient.

[0082] In this implementation, when performing weighted summation, the weight of the first intention feature is set to 0.4, and the weight of the second intention feature is set to 0.6;

[0083] Among them, the switching number threshold of the input state and the duration threshold of the first input state are pre-set. Since the purpose of setting the switching number threshold of the input state and the duration threshold of the first input state in this embodiment is to characterize the clarity of the intention of the sender, therefore, by obtaining the interaction data between the responding end and several sending ends during the continuous conversation, calling the switching number data of the input state and the duration data of the first input state, solving the average number of switching times and the average duration of the first input state, the switching number threshold is set to 1.01 times to 1.05 times the average number of switching times, and the duration threshold of the first input state is set to 1.05 times to 1.16 times the average duration of the first input state.

[0084] The present invention determines the sender's intention representation coefficient based on state characteristics over the duration of the interaction. In actual verbal interactions aimed at reaching a purchase agreement, the number of times the sender's input state switches can, to a certain extent, reflect the sender's thinking or language organization during the interaction. For example, frequent state switching may indicate that the sender is making multiple trade-offs and describing and communicating about the product. The first input state can indicate that the sender is continuously inputting text. If the interactive interface remains in this state for a long time, it indicates that the sender may be organizing language to communicate about the relevant product and has a high interest and willingness in the relevant content of the product. Therefore, the present application uses the sender's intention representation coefficient to represent the sender's purchase intention for the product currently being communicated with the responder. Furthermore, the present invention's method of determining the sender's intention based on state characteristics is more objective, rather than relying solely on subjective perception to guess whether the sender has intention and the degree of intention. This provides more objective data support for the subsequent classification of the sender's intention level. Furthermore, it adaptively analyzes the invoked interaction data to more accurately measure the sender's overall intention.

[0085] Specifically, see Figure 2, which is a logical decision diagram for classifying the intention degree categories for the sender according to an embodiment of the present invention, and the classification of the intention degree categories for the sender includes:

[0086] If the intention representation coefficient is greater than or equal to the intention representation coefficient threshold, classifying the intention degree of the sending end into a clear intention degree category;

[0087] If the intention representation coefficient is less than the intention representation coefficient threshold, the intention degree of the sending end is classified into a fuzzy intention degree category.

[0088] The intention representation coefficient threshold is selected in the interval [1.23, 1.35].

[0089] Specifically, according to the intention level category of the sender, the called interaction data is analyzed, including:

[0090] If the sender's intention level is in the clear intention level category, performing a logical coherence analysis on each text segment sent by the sender in the interaction data, identifying groups of intentionally coherent text segments, determining logical closeness based on semantic similarity of the groups of intentionally coherent text segments, and analyzing the sender's intention intensity representation value in combination with the number of occurrences of product detail words to determine whether to mark the sender;

[0091] If the intention degree of the sender is a fuzzy intention degree category, the text segment is segmented to obtain several keywords, and the matching degree between the product detail words corresponding to each product in the responding end product library and each keyword is determined to determine whether to mark the sender.

[0092] See also Figure 3 , which is a logical decision diagram for identifying intentionally coherent text segment groups according to an embodiment of the present invention. A logical coherence analysis is performed on each text segment sent by the sender in the interactive data. The process of identifying intentionally coherent text segment groups includes:

[0093] Identify the key words of each text segment through natural language models;

[0094] Determine the matching degree between adjacent text segments corresponding to each subject word;

[0095] identifying logically coherent phrases in adjacent text segments, and determining the number of occurrences of the logically coherent phrases in the adjacent text segments;

[0096] If the adjacent text segments meet the coherent text condition, the adjacent text segments are determined to be an intended coherent text segment group;

[0097] The coherent text condition is that the matching degree between the corresponding subject words of adjacent text segments is greater than or equal to the matching degree threshold and the number of occurrences of logically coherent words in adjacent text segments is greater than or equal to the number of occurrences threshold.

[0098] It is understandable that the purpose of using logically coherent phrases in the process of dialogue interaction is to connect sentences or paragraphs and make the logic between the text contents clearer and more coherent. Some commonly used logically coherent phrases include words that express causal relationships, such as because and so, words that express transitional relationships, such as but and on the contrary, words that express parallel relationships, such as and and at the same time, words that express progressive relationships, such as especially and even, words that express contrasting relationships, such as in comparison and rather than, and other logically coherent words, which will not be repeated here.

[0099] In this embodiment, an open source natural language processing toolkit is used to extract keywords from each text segment. The keywords are pre-set keywords of several categories, which include keywords that characterize product performance, keywords that characterize product names, keywords that characterize product uses, and keywords that characterize product effects. Specific keywords can be set by technical personnel in this field according to needs, and will not be repeated here.

[0100] In this embodiment, the matching degree threshold of the subject words corresponding to the adjacent text segments and the threshold of the number of occurrences of the logically coherent words in the adjacent text segments are pre-set. Interaction data corresponding to several transactions between the responding end and the sending end are extracted, and the matching degree data of the subject words corresponding to the adjacent text segments and the data of the number of occurrences of the logically coherent words in the adjacent text segments are called. The average matching degree of the subject words corresponding to the adjacent text segments and the average number of occurrences of the logically coherent words in the adjacent text segments are solved. The matching degree threshold of the subject words corresponding to the adjacent text segments is set to 1.02 to 1.08 times the matching degree average, and the threshold of the number of occurrences of the logically coherent words in the adjacent text segments is set to 1.07 to 1.18 times the average number of occurrences.

[0101] The selection of adjacent text segments is based on the text segments sent by the sending end, and two text segments that are adjacent to each other in the space position of the interactive interface are determined as adjacent text segments.

[0102] In this embodiment, the matching degree of the subject words can be calculated using cosine similarity. For example, a BERT model can be used to vectorize the subject words and then calculate the cosine similarity of the vectors between the subject words as the matching degree.

[0103] Specifically, the process of analyzing the intensity representation value of the intention of the sender includes:

[0104] Extracting intentionally coherent text segments to identify logical closeness;

[0105] taking a ratio of the logic closeness to a logic closeness threshold as a first degree feature;

[0106] The ratio of the number of occurrences of product detail words to the number of occurrence threshold is used as the second degree feature;

[0107] Determining the sum of the first degree feature and the second degree feature as the intention intensity representation value;

[0108] The logical closeness is the proportion of identical or synonymous words in adjacent text segments, and the product detail words are the words included in the product details page.

[0109] It can be understood that product detail words refer to words that describe the specific characteristics, attributes, dimensions, parameter indicators, etc. of various aspects of the product. Usually, words in this category are displayed on the product details interface to help the sender understand the characteristics of the product. The product detail words of the corresponding product details interface can be collected in advance for statistics, which will not be repeated here.

[0110] In this embodiment, the logical closeness threshold and the product detail word occurrence threshold are pre-set. The logical closeness data and product detail word data of the interaction data of several transactions between the responding end and the sending end are obtained, and the average logical closeness and the average occurrence of the product detail words are solved. The logical closeness threshold is set to 1.07 times to 1.15 times the average logical closeness, and the product detail word occurrence threshold is set to 1.03 times to 1.08 times the average occurrence.

[0111] Specifically, there is no specific limitation on the method of identifying synonyms. Synonymous words can be determined by training a word vector model. Of course, other methods can also be used, which will not be described in detail here.

[0112] Specifically, determining whether to mark the sending end includes:

[0113] If the intention intensity representation value is greater than or equal to the intention intensity representation threshold, the sending end is marked;

[0114] If the intention intensity representation value is less than the intention intensity representation threshold, there is no need to mark the sending end.

[0115] The threshold for strong intention characterization is selected in the interval [1.74, 1.85].

[0116] In implementation, the marked sender is a high-intent sender, which in turn provides effective support for the platform's subsequent related processing, such as timely access to manual replies, pushing advertising information to the senders of relevant marks, etc.

[0117] In some possible implementations, several senders may be analyzed in advance, and marked high-intent senders may be manually processed, while the remaining senders may be responded to by customer service robots.

[0118] The present invention conducts targeted analysis based on senders of different intention degree categories. For senders of clear intention degree category, it identifies intention coherent text groups and determines whether the sender's intention is targeted based on the logical correlation of the content of the sender's text segments received by the responding end. For example, whether the products that the sender has purchase intention are frequently mentioned in the text segments. If the logical correlation between the text groups is not strong, it may indicate that the sender's intention for the product mentioned in the previous text segment has decreased during the interaction process. However, since the sender is still interacting with the responding end, there is a possibility that the sender will develop an interest in other products. However, it is impossible to accurately determine the sender's intention.

[0119] Therefore, whether the sender has targeted intentions is determined based on the logical correlation of the text segment content, and the text segments with the above-mentioned characteristics in the preceding and following adjacent text segments sent by the sender are determined as intention-coherent text segment groups. The logical closeness is determined based on the semantic similarity of the contents of the intention-coherent text segment group, and the sender's intention intensity representation value is analyzed in combination with the number of occurrences of product detail words. On the premise of determining a text segment group with logical coherence, the logical closeness can reflect the degree of semantic connection between text segments, and the sender's intention intensity can be further determined based on the semantic similarity between each text segment. At the same time, the number of occurrences of product detail words can more intuitively reflect the sender's attention to the specific characteristics of the product. The intention intensity representation value determined in combination with the above-mentioned characteristics represents the positivity and intensity of the sender's intention, providing data support for the subsequent determination of whether to mark the sender.

[0120] Specifically, determining the matching degree between the product details words corresponding to each product in the product library of the responding end and each keyword determines whether to mark the sending end, including

[0121] If the matching degree is greater than or equal to the matching degree threshold, marking the sending end;

[0122] If the matching degree is less than the matching degree threshold, the sending end does not need to mark.

[0123] In this embodiment, the matching degree threshold between the product detail words corresponding to each product and each keyword is pre-set, and the relevant information of several products listed by the responding end and the interactive data with the sending end on the purchase agreement of the used products are obtained to determine the matching degree data between the product detail words corresponding to the product and each keyword, and solve the average matching degree between the product detail words corresponding to the product and each keyword, and set the matching degree threshold between the product detail words corresponding to the product and each keyword to be 1.03 times to 1.15 times the average matching degree between the product detail words corresponding to the product and each keyword.

[0124] For senders of fuzzy intention degree categories, the present invention performs word segmentation on the text segment content received from the sender to obtain keywords, so as to quickly extract key information from the text segment. In the case that the sender's intention is relatively fuzzy, it can capture key words related to the product from the expression of the fragment in this way, avoiding the situation where the sender has a certain interest in the product due to the sender's vague expression or unclear needs. For example, the sender extracts words related to the functional features of the product, analyzes the sender's potential intention through word segmentation and matching with the product details words in the product library of the responding end, quickly screens words related to the product from the overall interactive data between the sender and the responding end, accurately determines the sender with a certain degree of intention and marks it, saves computing power resources, and helps the responding end seize the opportunity to quickly carry out subsequent recommendation follow-up work.

[0125] Specifically, see Figure 4 , which is a functional module diagram of a system for automatically identifying customer intentions based on NLP according to an embodiment of the present invention, and provides a system for applying a method for automatically identifying customer intentions based on NLP, which includes:

[0126] An interaction collection module, configured to collect interaction data of the interaction interface and analyze state characteristics based on the interaction data, including the number of times the input state of the sending end switches during the interaction and the duration of the first input state;

[0127] an intention recognition module connected to the interaction acquisition module, configured to determine an intention representation coefficient for the sender based on the state characteristics during the interaction duration, so as to classify the intention degree category for the sender;

[0128] An intention analysis module is connected to the interaction acquisition module and the intention identification module respectively, and is used to call the interaction data corresponding to the sending end, and analyze the called interaction data according to the intention level category of the sending end, including:

[0129] Performing logical coherence analysis on each text segment sent by the sender in the interaction data, identifying intentionally coherent text segment groups, determining logical closeness based on semantic similarity of the intentionally coherent text segment groups, and analyzing a representation value of the sender's intention intensity in combination with the number of occurrences of product detail words to determine whether to mark the sender;

[0130] Alternatively, the text segment is segmented to obtain a number of keywords, and the matching degree between the product detail words corresponding to each product in the product library of the responding end and each keyword is determined to determine whether to mark the sending end.

[0131] The system further includes a tag processing module connected to the intention analysis module, configured to push product-related content to each of the sending terminals based on the determination result of the intention analysis module;

[0132] In this embodiment, for the senders marked with different degrees of intention, the responding end will subsequently perform targeted recommendation follow-up work: if the sender is marked with a strong intention label, the detailed information of the product will be sent to the sender. In this case, the sender is targeted at the intended product determined in the interaction process and has a strong purchase intention. By sending the specific detailed information of the intended product to the sender, the intensity of the sender's intention can be strengthened and the execution of the product purchase agreement can be promoted; if the sender is marked with a weak intention label, the responding end will perform targeted recommendation follow-up work based on several products in the product library of the responding end and the products that the sender has a weak intention to. By comparison, the derivative products of the product with slight intention are determined, and the detailed information of the product is extracted and sent to the sending end. In this case, the sending end does not have a clear and specific intended product. It may be that there is hesitation in choosing different models of the same product, or it may be that it cannot clearly express its own needs. Therefore, the characteristic information of the slightly intended product and its derivative products is sent to the sending end, so that the sending end can have a comprehensive understanding of the various products that the responding end can provide. It can not only help the sending end to clarify the purchase intention, but also more accurately identify the sender's intention level, and increase the probability of the product purchase agreement being executed.

[0133] Specifically, there is no limitation on the specific structure of the interaction acquisition module, intention recognition module, intention analysis module and marking processing module. They themselves or each unit therein can be composed of logical components or a combination of logical components, and the logical components include field programmable processors, computers or microprocessors in computers.

[0134] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for automatically identifying customer intentions based on NLP, characterized in that: include: Collecting interaction data of the interactive interface, and analyzing state characteristics based on the interaction data, including the number of times the input state of the sending end switches during the interaction and the duration of the first input state; Determining an intention representation coefficient for the sender based on the state characteristics during the interaction duration to classify the intention degree category for the sender; Calling the interaction data corresponding to the sending end, and analyzing the called interaction data according to the intention level category of the sending end, including: Performing logical coherence analysis on each text segment sent by the sender in the interaction data, identifying intentionally coherent text segment groups, determining logical closeness based on semantic similarity of the intentionally coherent text segment groups, and analyzing a representation value of the sender's intention intensity in combination with the number of occurrences of product detail words to determine whether to mark the sender; Alternatively, the text segment is segmented to obtain a number of keywords, and the matching degree between the product detail words corresponding to each product in the product library of the responding end and each keyword is determined to determine whether to mark the sending end; The input state includes a first input state and a second input state, and the input state is determined based on the instruction input device of the sending end, including: If the instruction input device continues to input instructions, it is determined that the input state is the first input state; If the instruction input device does not continuously input instructions, determining that the input state is the second input state; The continuously input command means that the time interval between the input adjacent command signals is not less than a predetermined time interval threshold.

2. The method for automatically identifying customer intentions based on NLP according to claim 1, characterized in that: The process of determining the intention representation coefficient for the sending end based on the state characteristics includes: The ratio of the number of switching times of the input state of the sending end during the interaction duration to the switching number threshold is used as the first intention feature; taking a ratio of a duration of the first input state within the interaction duration to a duration threshold as a second intention feature; The first intention feature and the second intention feature are weighted and summed to determine an intention representation coefficient.

3. The method for automatically identifying customer intentions based on NLP according to claim 1, characterized in that: Classify the intention level of the sender into categories, including: If the intention representation coefficient is greater than or equal to the intention representation coefficient threshold, classifying the intention degree of the sending end into a clear intention degree category; If the intention representation coefficient is less than the intention representation coefficient threshold, the intention degree of the sending end is classified into a fuzzy intention degree category.

4. The method for automatically identifying customer intentions based on NLP according to claim 1 is characterized in that ,Analyze the called interaction data according to the intention level category of the sender, including, If the sender's intention level is in the clear intention level category, performing a logical coherence analysis on each text segment sent by the sender in the interaction data, identifying groups of intentionally coherent text segments, determining logical closeness based on semantic similarity of the groups of intentionally coherent text segments, and analyzing the sender's intention intensity representation value in combination with the number of occurrences of product detail words to determine whether to mark the sender; If the intention degree of the sender is a fuzzy intention degree category, the text segment is segmented to obtain several keywords, and the matching degree between the product detail words corresponding to each product in the responding end product library and each keyword is determined to determine whether to mark the sender.

5. The method for automatically identifying customer intentions based on NLP according to claim 1 is characterized in that ,The process of performing logical coherence analysis on each text segment sent by the sender in the interactive data and identifying the intention coherent text segment group includes, Identify the key words of each text segment through natural language models; Determine the matching degree between adjacent text segments corresponding to each subject word; identifying logically coherent phrases in adjacent text segments, and determining the number of occurrences of the logically coherent phrases in the adjacent text segments; If the adjacent text segments meet the coherent text condition, the adjacent text segments are determined to be an intended coherent text segment group; The coherent text condition is that the matching degree between the corresponding subject words of adjacent text segments is greater than or equal to the matching degree threshold and the number of occurrences of logically coherent words in adjacent text segments is greater than or equal to the number of occurrences threshold.

6. The method for automatically identifying customer intentions based on NLP according to claim 1 is characterized in that ,The process of analyzing the intention intensity representation value of the ,sender includes, Extracting intentionally coherent text segments to identify logical closeness; taking a ratio of the logic closeness to a logic closeness threshold as a first degree feature; The ratio of the number of occurrences of product detail words to the number of occurrence threshold is used as the second degree feature; Determining the sum of the first degree feature and the second degree feature as the intention intensity representation value; The logical closeness is the proportion of identical or synonymous words in adjacent text segments, and the product detail words are the words included in the product details page.

7. The method for automatically identifying customer intentions based on NLP according to claim 1 is characterized in that , determine whether to mark the sending end, including, If the intention intensity representation value is greater than or equal to the intention intensity representation threshold, the sending end is marked.

8. The method for automatically identifying customer intentions based on NLP according to claim 7 is characterized in that , determine the matching degree between the product details words corresponding to each product in the product library of the responding end and each keyword to determine whether to mark the sending end, including If the matching degree is greater than or equal to the matching degree threshold, the sending end is marked.

9. A system using the method for automatically identifying customer intentions based on NLP according to any one of claims 1 to 8, characterized in that: include: An interaction collection module, configured to collect interaction data of the interaction interface and analyze state characteristics based on the interaction data, including the number of times the input state of the sending end switches during the interaction and the duration of the first input state; an intention recognition module connected to the interaction acquisition module, configured to determine an intention representation coefficient for the sender based on the state characteristics during the interaction duration, so as to classify the intention degree category for the sender; An intention analysis module is connected to the interaction acquisition module and the intention identification module respectively, and is used to call the interaction data corresponding to the sending end, and analyze the called interaction data according to the intention level category of the sending end, including: Performing logical coherence analysis on each text segment sent by the sender in the interaction data, identifying intentionally coherent text segment groups, determining logical closeness based on semantic similarity of the intentionally coherent text segment groups, and analyzing a representation value of the sender's intention intensity in combination with the number of occurrences of product detail words to determine whether to mark the sender; Alternatively, the text segment is segmented to obtain a number of keywords, and the matching degree between the product detail words corresponding to each product in the product library of the responding end and each keyword is determined to determine whether to mark the sending end.

Images