A digital employee method based on generative artificial intelligence technology
By analyzing the semantic bias values and key semantic fragments of user input data and adjusting word weights, the problem of discrepancies between the digital employee generation results and user needs was solved, achieving higher accuracy in understanding needs and interaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG DECIMAL POINT INFORMATION TECH CO LTD
- Filing Date
- 2026-03-12
- Publication Date
- 2026-05-15
AI Technical Summary
The results generated by digital employees deviate from the actual needs of users, especially when using LSTM technology to extract the needs of users at different stages. This can lead to a mismatch between the user's output and the characteristics of the real-time analysis environment.
By acquiring users' current and historical input data, analyzing the semantic deviation values of word sequences, extracting key semantic segments, adjusting word weights based on semantic deviation values, and using a generative model to generate the desired results.
This reduces the discrepancy between the results generated by digital employees and the actual needs of users, and improves the accuracy of understanding needs and the precision of interaction.
Smart Images
Figure CN121835694B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, specifically a digital employee method based on generative artificial intelligence technology. Background Technology
[0002] In the current business processing domain, digital employees, with their planning and orchestration capabilities, can connect complex and fragmented business processes, thereby enabling them to solve comprehensive problems. Among these capabilities, planning and orchestration, as the scheduling and collaboration hub of digital employees, are key to achieving their autonomy and goal-oriented behavior. Currently, the industry mainly employs two types of techniques in this field: dynamic task planning and static orchestration. Specifically, dynamic task planning technology uses a large language model as the decision-making core of the intelligent agent, using prompt words to standardize the large model's understanding of tasks. Regarding dynamic model memory planning, existing technologies update the purpose of the current task as the digital employee continuously receives new customer requests, thus achieving real-time understanding of complex and changing customer requests. However, when directly calling LSTM technology to extract user needs at different stages, a mismatch occurs between the user's output and the characteristics of the real-time analysis environment. This directly leads to a matching discrepancy between the intelligent model and the output, resulting in a deviation between the content generated by the intelligent employee and the user's needs. Summary of the Invention
[0003] The purpose of this application is to provide a digital employee method based on generative artificial intelligence technology to solve the technical problem that the results of digital employee generation deviate from the actual needs of users.
[0004] To achieve the above objectives, this application provides the following technical solution: a digital employee method based on generative artificial intelligence technology, comprising:
[0005] Obtain the user's current input request data;
[0006] Based on the current input demand data, a first vocabulary sequence and a second vocabulary sequence are obtained; the first vocabulary sequence is the vocabulary sequence corresponding to the current input demand data; the second vocabulary sequence is the vocabulary sequence corresponding to the previous input demand data of the current input demand data.
[0007] Based on the first word sequence and the second word sequence, obtain the semantic bias value that corresponds one-to-one with each word in the first word sequence;
[0008] Based on each semantic shift value, key semantic segments are extracted from the first word sequence; the key semantic segments are segments in the first word sequence whose semantic shift values change significantly.
[0009] Based on the key semantic segment, a semantic deviation value is obtained; the semantic deviation value is at least used to characterize the magnitude of the difference in semantic deviation values between each word in the key semantic segment.
[0010] Based on the semantic deviation value, the input weight of each word in the first word sequence is obtained; the input weight is used to characterize at least the importance of the corresponding word in the first word sequence.
[0011] Based on the weights of each input, obtain the required results.
[0012] As a specific technical solution in this application, the step of obtaining the first word sequence based on the current input demand data includes:
[0013] Stop words are removed from the current input data based on stop word cleaning rules to obtain cleaned data;
[0014] The cleaned data is split using a word segmentation tool to obtain the first word sequence.
[0015] As a specific technical solution in this application, the step of obtaining semantic shift values corresponding one-to-one with each word in the first word sequence based on the first word sequence and the second word sequence includes:
[0016] Based on the first word sequence, a first word is obtained; the first word is any word in the first word sequence for which a semantic bias value has not been obtained.
[0017] Based on the first word, a first value and a second value are obtained; the first value is the maximum similarity value between the first word and the remaining words in the first word sequence, and the second value is the maximum similarity value between the first word and each word in the second word sequence; or, the first value is the number of second words, and the second value is the number of third words; the second word is any word in the first word sequence whose similarity to the first word is greater than a first preset value; the third word is any word in the second word sequence whose similarity to the first word is greater than a first preset value.
[0018] Based on the first value and the second value, obtain the semantic bias value of the first word.
[0019] As a specific technical solution in this application, after obtaining the semantic shift value of the first word based on the first value and the second value, the method further includes:
[0020] Based on the first word, a first quantity is obtained; the first quantity is the number of sub-networks in the digital employee that the first word can activate.
[0021] Based on the first word sequence, a second quantity is obtained; the second quantity is the total number of sub-networks in the digital employee that each word in the first word sequence can activate.
[0022] Based on the first quantity and the second quantity, obtain the correction coefficient;
[0023] The semantic bias value is corrected based on the correction coefficient.
[0024] As a specific technical solution in this application, the step of extracting key semantic segments from the first word sequence based on various semantic shift values includes:
[0025] Based on the first word sequence, a first deflection value sequence is obtained; the position of the speech deflection value in the first deflection value sequence is the same as the position of the corresponding word in the first word sequence;
[0026] Based on the first deflection value sequence, a first difference sequence is obtained; the sequence value in the first difference sequence is equal to the difference between two adjacent sequence values in the first deflection value sequence.
[0027] Based on the first difference sequence, an absolute value sequence is obtained; the absolute value sequence is a sequence formed by arranging the absolute values of each sequence value in the first difference sequence in descending order;
[0028] Based on the absolute value sequence, a second difference sequence is obtained; the sequence value in the second difference sequence is equal to the difference between two adjacent sequence values in the absolute value sequence.
[0029] Based on the second difference sequence, obtain the maximum value;
[0030] Based on the maximum value, the maximum difference value is obtained from the absolute value sequence; the maximum difference value is the minuend of the absolute value sequence from which the maximum value is obtained.
[0031] Based on the maximum difference value, a semantic jump word group is obtained from the first word sequence;
[0032] Based on the semantic jump vocabulary group, key semantic fragments are extracted from the first vocabulary sequence.
[0033] As a specific technical solution in this application, based on the maximum difference value, a semantic jump word group is obtained from the first word sequence, including:
[0034] Based on the maximum difference value, a first significant deflection value is obtained from the first difference sequence; the first significant deflection value is the sequence value in the first difference sequence corresponding to the maximum difference value;
[0035] Based on the first obvious deflection value, multiple second obvious deflection values are obtained; the second obvious deflection value is any sequence value in the first difference sequence that is less than or equal to the first obvious deflection value;
[0036] Based on the first obvious deflection value and the second obvious deflection value, a semantic jump word group is obtained from the first word sequence; the semantic jump word group is a combination of two words in the first word sequence corresponding to the first obvious deflection value or the second obvious deflection value.
[0037] As a specific technical solution in this application, obtaining the semantic deviation value based on the key semantic fragment includes:
[0038] Traverse the key semantic segments to obtain the fourth word; the fourth word is any word in the key semantic segments for which the semantic deviation has not been obtained;
[0039] Based on the fourth word, a first semantic deviation value is obtained; the first semantic deviation value is the semantic deviation value corresponding to the fourth word;
[0040] Based on the key semantic segments, a first average value and a first standard deviation are obtained; the first average value is the average of the semantic shift values corresponding to each word in the key semantic segments; the first standard deviation is the standard deviation of the semantic shift values corresponding to each word in the key semantic segments.
[0041] Based on the first semantic deviation value, the first average value, and the first standard deviation, the semantic deviation of the fourth word is obtained;
[0042] Based on the semantic deviation degree, obtain the semantic deviation value.
[0043] As a specific technical solution in this application, obtaining the input weights of each word in the first word sequence based on the semantic deviation value includes:
[0044] Traverse the first word sequence to obtain the fifth word; the fifth word is any word in the first word sequence for which no input weight has been obtained; and the semantic bias value corresponding to the fifth word is negative.
[0045] Based on the fifth word, a second semantic deviation value and a total deviation value are obtained; the second semantic deviation value is the semantic deviation value corresponding to the fifth word; the total deviation value is the sum of the semantic deviation values of all sixth words in the key semantic segment corresponding to the fifth word; the sixth word is any word in the corresponding key semantic segment with a negative semantic deviation value;
[0046] The input weight of the fifth word is obtained based on the semantic deviation value, the second semantic deflection value, and the total value of the deflection values.
[0047] As a specific technical solution in this application, the step of obtaining the required result based on each input weight includes:
[0048] A generative model is used to obtain a sequence of word vectors based on the first word sequence; the generative model is pre-built.
[0049] Based on each input weight, obtain the weight vector sequence;
[0050] After replacing the attention weights in the generative model with the weights in the weight vector sequence, a weighted sentence representation is generated;
[0051] Based on the weighted sentence representation, the required result is obtained.
[0052] As a specific technical solution in this application, the step of obtaining the weight vector sequence based on each input weight includes:
[0053] Traverse the first word sequence to obtain the seventh word; the seventh word is any word in the first word sequence for which no weight value has been obtained.
[0054] Based on the seventh word, a corrected weight is obtained; if the input weight of the seventh word is empty, the corrected weight is equal to the second preset value; if the input weight of the seventh word is not empty, the corrected weight is equal to the sum of the second preset value and the input weight corresponding to the seventh word.
[0055] The corrected weights are normalized to obtain the weight values of the seventh word;
[0056] Based on the weight values, the weight vector sequence is obtained.
[0057] Compared with the prior art, the beneficial effects of this application are:
[0058] This application obtains the demand changes (i.e., semantic deviation values) of each word in the current input demand data by comparing current and historical input demand data. Then, it identifies and extracts key semantic segments that represent user intent using these semantic deviation values, eliminating interfering information. Finally, based on the degree of semantic deviation of each key semantic segment (i.e., semantic deviation value), it adjusts the weight vector of each word, giving high-contribution words a higher proportion in the output, thereby reducing the discrepancy between the results generated by digital employees and the user's actual needs. Attached Figure Description
[0059] Figure 1This is a flowchart illustrating a digital employee method based on generative artificial intelligence technology proposed in an embodiment of this application. Detailed Implementation
[0060] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0061] The terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. For example, the first and second word sequences mentioned below belong to different word sequences. It should be understood that such names can be used interchangeably where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. The division of modules in the embodiments of this application is merely a logical division. In actual applications, there may be other division methods. For example, multiple modules may be combined into or integrated into another system, or some features may be ignored or not performed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interface, and the indirect coupling or communication connection between modules may be electrical or other similar forms. None of these are limited in the embodiments of this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed among multiple circuit modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the embodiments of this application.
[0062] Before understanding the embodiments presented in this application, it is necessary to understand the following technical terms:
[0063] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.
[0064] AI technology is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.
[0065] Computer vision (CV) is the science that studies how to enable machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes for target recognition, tracking, and measurement, and then performs image processing to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include adversarial perturbation generation, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), and common biometric recognition technologies such as facial recognition and fingerprint recognition.
[0066] Generative AI is an important branch of artificial intelligence that autonomously generates logically coherent new content (e.g., text, images, audio, video, code) through algorithms and models, distinguishing it from the analytical and classification functions of traditional AI. It relies on technologies such as multimodal models, generative adversarial networks (GANs), diffusion models, and the Transformer architecture to learn data distribution patterns, capture inherent patterns, and achieve creative output. The specific process includes data learning (inductively summarizing probability distributions from massive amounts of data), pattern capture (abstracting patterns through adversarial training or denoising algorithms), and content generation (generating new content based on input prompts or noise). Unlike traditional discriminative AI (e.g., spam detection), generative AI focuses on "creation," outputting probabilistic original content rather than deterministic results. It is widely used in content creation (e.g., text generation, image design), industry empowerment (e.g., healthcare, education), and social services (e.g., government affairs, finance), but it also faces challenges such as data dependence, insufficient controllability, high computational costs, and ethical controversies (e.g., misinformation, copyright issues). As a key force driving digital transformation, generative AI is reshaping content production and industry landscape, and its future development requires a balance between technological innovation and regulation.
[0067] A digital employee is an intelligent virtual work role built upon technologies such as artificial intelligence, robotic process automation, and big data. Essentially, it simulates the operational logic and decision-making abilities of a human employee through software programs, assisting or replacing humans in completing repetitive, rule-based, or intelligent tasks. It serves as both a tool for enterprise digital transformation and an important vehicle for the "human-machine collaboration" model in the future workplace.
[0068] To address the technical problem mentioned in the background art where the generated digital employees deviate from the actual needs of users, this application proposes an embodiment of a digital employee method based on generative artificial intelligence technology. For example... Figure 1 As shown, the digital employee method based on generative artificial intelligence technology includes steps S100 to S700.
[0069] Step S100: Obtain the user's current input requirement data.
[0070] It is important to understand that digital employees' understanding of the evolution of user needs and intentions is based on the decomposition and analysis of the semantic state reflected by the user's input content in the current round (i.e., the current input demand data). As users continue to provide subsequent information in multiple rounds of interaction, new semantic inputs will cause dynamic changes in the contextual understanding state. By evaluating the differences between states and the migration trajectory between states, the evolution process of the user's true intentions and demand tendencies can be assessed.
[0071] In this embodiment, the current input demand data is the specific information input actively provided by the user in a certain round of multi-round interaction (such as text, voice, or image, etc.), and its core is to reflect the clear demand intention or semantic state of the user at present. As the "real-time anchor point" for the digital employee to understand the user's needs, it is both the direct basis for intention parsing in single-round interaction and the key link for context association in multi-round dialogue. By combining with historical input demand data, the new semantic input will trigger dynamic adjustment of the context understanding state. Its type is not restricted and can cover various forms such as question consultation, task instructions, demand description, etc., and supports multi-modal expressions such as text and voice. The digital employee gradually approaches the evolution context of the user's real needs through real-time analysis of this data and state migration evaluation, and finally achieves accurate capture and response of the needs.
[0072] Step S200: Based on the current input demand data, obtain a first vocabulary sequence and a second vocabulary sequence.
[0073] In this embodiment, the first vocabulary sequence is the vocabulary sequence corresponding to the current input demand data. The second vocabulary sequence is the vocabulary sequence corresponding to the previous input demand data of the current input demand data.
[0074] It should be clear that the first vocabulary sequence and the second vocabulary sequence can be obtained in any reasonable way. For example, in step S200, obtaining the first vocabulary sequence based on the current input demand data includes step S210 and step S220.
[0075] Step S210: Remove stop words from the current input demand data based on stop word cleaning rules to obtain cleaned data.
[0076] It should be clear that for the current input demand data (such as the text of the user's question), based on preset stop word cleaning rules (such as filtering high-frequency function words without actual semantics such as "de", "shi", "zai", etc., or redundant words irrelevant to the field), redundant stop words in the text are removed to generate "cleaned data". The core purpose of this step is to reduce the interference of invalid information, reduce the complexity of subsequent processing, and improve the accuracy of semantic analysis. For example, when the user inputs "Help me check the weather condition in G City today", keywords such as "check", "today", "G City", "weather", "condition" will be retained after cleaning.
[0077] Step S220: Split the cleaned data based on a word segmentation tool to obtain the first vocabulary sequence.
[0078] It's important to understand that after cleaning, the data undergoes structured segmentation using word segmentation tools (such as jieba or NLTK) to divide continuous text into sequences of words with independent semantic meaning, known as the "first word sequence." For example, "check today's weather in G city" would be segmented into "check / today / G city / weather / situation." Word segmentation is fundamental to semantic understanding. By converting text into discrete word units, digital engineers can further analyze the logical relationships between words (e.g., subject-verb-object structures) and thus infer the user's true intent.
[0079] It should be noted that the process of obtaining the second word sequence can refer to steps S210 and S220, which will not be elaborated here.
[0080] Step S300: Based on the first word sequence and the second word sequence, obtain the semantic bias value that corresponds one-to-one with each word in the first word sequence.
[0081] In this embodiment, the core purpose of obtaining semantic deviation values is to quantify the degree and direction of dynamic changes in user needs across multiple rounds of interaction, helping digital employees accurately capture the evolution logic of user intent and providing a scientific basis for adjusting response strategies. Specifically, its role is reflected in three key dimensions: First, tracking the "dynamic trajectory" of needs. By integrating the domain concentration of current vocabulary (reflecting changes in the domain to which the need belongs), semantic deviation from previous rounds (measuring the jump or continuity of needs), semantic effectiveness (assessing the vocabulary's ability to express the need), and historical usage (determining whether the need has been effectively expressed), the "direction of change" and "intensity" of user needs from the past to the present are comprehensively characterized. For example, when a user shifts from "checking order status" to "applying for after-sales return," dimensions such as domain span and semantic jump will significantly increase the deviation, intuitively reflecting the new task attributes of the need. Second, assisting in judging the "independence" and "continuity" of needs. High deviation (e.g., large domain differences or high semantic deviation) usually indicates the initiation of a new task or the refinement of existing needs (e.g., "change delivery address"); low deviation indicates the continuation of needs (e.g., from "check the weather in G city" to "check the weather in H city"), helping the system quickly locate the core intent and avoid response deviations. Third, optimize the "accuracy" of need understanding. By weighting the domain-centric evaluation value, invalid need expressions (e.g., low-value words lacking semantic proximity or not used historically) are filtered out, ensuring the system focuses on the user's true needs. In summary, the semantic deviation value is a "quantitative scale" of the dynamic changes in needs during multi-turn interactions. Through multi-dimensional information integration, it enables digital employees to more intelligently capture intent evolution, improving the targeting and accuracy of interactions. Based on this, in this embodiment, step S300, based on the first word sequence and the second word sequence, obtains the semantic deviation value corresponding one-to-one with each word in the first word sequence, including steps S310 to S330.
[0082] Step S310: Obtain the first word based on the first word sequence.
[0083] In this embodiment, the first word is any word in the first word sequence for which no semantic bias value has been obtained. That is, in this embodiment, the method for obtaining the semantic bias value corresponding to any word in the first word sequence is the same as the method for obtaining the semantic bias value corresponding to the first word sequence.
[0084] Step S320: Based on the first word, obtain the first value and the second value.
[0085] In the embodiments of this application, there are no restrictions on the first and second values, as long as they reflect the similarity between the first word and each word in the first and second word sequences. For example, in the embodiments of this application, the first value is the maximum similarity value between the first word and the remaining words in the first word sequence, and the second value is the maximum similarity value between the first word and each word in the second word sequence. Alternatively, the first value can be the number of second words, and the second value can be the number of third words. The second word is any word in the first word sequence whose similarity to the first word is greater than a first preset value; the third word is any word in the second word sequence whose similarity to the first word is greater than a first preset value.
[0086] In the embodiments of this application, a first preset value can be set according to requirements, for example, the first preset value can be 0.8 or 0.9, etc.
[0087] Step S330: Based on the first value and the second value, obtain the semantic shift value of the first word.
[0088] It's important to understand that the greater the difference between the first and second values, the more significant the semantic state generated by the current input request data differs from that generated by historical input request data in the current interaction round. This indicates a stronger novelty or independence of the current user intent. This means that the current input request data is less likely to maintain a high semantic match with recent historical input request data, and is more likely to represent a new user task, topic switch, or a significant refinement of the original need (for example, in a business consultation process, a user suddenly switches from "checking order status" to "applying for after-sales return," and this semantic jump reflects the independence of the intent).
[0089] In embodiments of this application, the semantic shift value of the first word can be obtained based on the first value and the second value in any reasonable manner. For example, the semantic shift value of the first word can be the ratio or difference between the first value and the second value.
[0090] It is important to note that for digital employees, the large model pre-assigns and trains sub-networks for different functions or themes by domain (e.g., "logistics inquiry," "after-sales return," "weather information," etc.), with each sub-network focusing on processing the needs of its corresponding domain. When a user inputs a word sequence into the model, only the sub-networks related to its domain are activated (e.g., "check express tracking number" activates the "logistics inquiry" sub-network). If the word deviates from a domain (e.g., "return" does not activate the "weather information" sub-network), the corresponding sub-network will not respond. The more sub-networks a word can activate, the stronger its relevance to the specific business domain, and the better it helps digital employees quickly pinpoint the core direction of user needs, avoiding misjudgments of intent due to domain confusion, and improving the accuracy and efficiency of understanding needs in multi-turn interactions. Based on this, after obtaining the semantic bias value of the first word based on the first and second values in step S330, the method further includes steps S340 to S370.
[0091] Step S340: Based on the first vocabulary, obtain the first quantity.
[0092] In this embodiment, the first quantity is the number of sub-networks in the digital employee that the first word can activate.
[0093] Step S350: Based on the first word sequence, obtain the second quantity.
[0094] In this embodiment, the second quantity is the total number of sub-networks in the digital employee whose words in the first word sequence can activate.
[0095] Step S360: Obtain the correction coefficient based on the first quantity and the second quantity.
[0096] In this embodiment, the correction coefficient can be obtained based on the first quantity and the second quantity in any reasonable manner. For example, the correction coefficient can be the difference between the first quantity and the second quantity.
[0097] In this embodiment, the directionality of a word to a certain domain can be measured by calculating the difference (i.e., the correction coefficient) between the number of sub-networks activated by a certain word (i.e., the first word) and the total number of activated sub-networks in the first word sequence. The smaller the difference, the more focused the first word is on the needs of that domain; the larger the difference, the more likely the first word involves multiple domain intersections or domain deviations.
[0098] Step S370: Correct the semantic deviation value based on the correction coefficient.
[0099] In a specific embodiment of this application, step S370, the calculation formula for correcting the semantic deviation value based on the correction coefficient, is as follows:
[0100]
[0101] in, This represents the corrected semantic shift value; Indicates the first quantity; Indicates the second quantity; This indicates the semantic bias value before correction.
[0102] Step S400: Extract key semantic segments from the first word sequence based on each semantic bias value.
[0103] It is important to understand that the core purpose of extracting key semantic segments is to accurately locate the core semantic boundaries of user needs, filter high-value content, and assess the stability of intent, thereby helping digital employees efficiently capture users' true needs, reduce processing errors, and improve interaction accuracy. In other words, in this embodiment, the key semantic segments are segments in the first word sequence where the semantic shift value changes significantly. Based on this, step S400 involves extracting key semantic segments from the first word sequence based on each semantic shift value, including steps S410 to S480.
[0104] Step S410: Based on the first word sequence, obtain the first deflection value sequence.
[0105] In this embodiment, the position of the semantic deflection value in the first deflection value sequence is the same as the position of the corresponding word in the first word sequence.
[0106] To facilitate understanding by those not skilled in the art, this application provides a specific embodiment (hereinafter referred to as the first embodiment). In this embodiment, the current input demand data is the text data "I want to know about travel insurance, such as whether it covers sudden illnesses, and what the price difference is between overseas travel and domestic travel?" After processing in step S200, the first word sequence obtained is ["want to know", "travel insurance", "sudden illness", "overseas travel", "domestic travel", "price difference"]. In this embodiment, the first deflection value sequence is [2.5, 2.2, 0.2, 2.0, 1.8, 2.1].
[0107] Step S420: Obtain the first difference sequence based on the first deflection value sequence.
[0108] In this embodiment, the sequence value in the first difference sequence is equal to the difference between two adjacent sequence values in the first deflection value sequence. It should be noted that obtaining the corresponding difference sequence from a given sequence is a mature technology and will not be elaborated upon here.
[0109] In the first embodiment, the first difference sequence is [0.3, 0.2, -1.8, 0.2, -0.3].
[0110] Step S430: Based on the first difference sequence, obtain the absolute value sequence.
[0111] In this embodiment, the absolute value sequence is a sequence formed by arranging the absolute values of each sequence value in the first difference sequence in descending order. It should be noted that arranging the absolute values of various values in descending order to form a sequence is a mature technology in the computer field, and will not be elaborated upon here.
[0112] In the first embodiment, the absolute value sequence is [1.8, 0.3, 0.3, 0.2, 0.2].
[0113] Step S440: Based on the absolute value sequence, obtain the second difference sequence.
[0114] In this embodiment, the sequence value in the second difference sequence is equal to the difference between two adjacent sequence values in the absolute value sequence.
[0115] In the first embodiment, the second difference sequence is [1.8-0.3=1.5,0.3-0.3=0,0.3-0.2=0.1,0.2-0.2=0].
[0116] Step S450: Obtain the maximum value based on the second difference sequence.
[0117] It should be clear that obtaining the maximum value from multiple values is a mature technique, which will not be elaborated here.
[0118] Step S460: Based on the maximum value, obtain the maximum difference value from the absolute value sequence.
[0119] In this embodiment, the maximum difference value is the minuend of the absolute value sequence from which the maximum value is obtained.
[0120] In the first embodiment, the maximum value is 1.5, and the minuend for obtaining the maximum value (i.e., 1.5) is 1.8 (i.e., the maximum difference value) in the absolute value sequence.
[0121] Step S470: Based on the maximum difference value, obtain the semantic jump word group from the first word sequence.
[0122] It is important to understand that when a word becomes "semantically ineffective" (i.e., the word no longer represents the user's core needs; for example, when a user shifts from "consulting on coverage" to "consulting on price," related terms gradually become ineffective), the semantic shift value of the words before and after it will increase significantly. The location where this semantic shift value increases significantly (i.e., the location corresponding to the maximum difference value) means that the meanings of two adjacent words are clearly different (e.g., "sudden illness" and "overseas travel"—the former belongs to the "coverage" topic, while the latter belongs to the "price difference" topic, and their meanings are clearly different). Based on this, step S470, based on the maximum difference value, obtains a semantic jump word group from the first word sequence, including steps S471 to S473.
[0123] Step S471: Based on the maximum difference value, obtain the first significant deflection value from the first difference sequence.
[0124] In this embodiment, the first obvious deflection value is the sequence value in the first difference sequence that corresponds to the maximum difference value.
[0125] In the first embodiment, the maximum difference value is 1.8, and the sequence value corresponding to the maximum difference value in the first difference sequence is -1.8. That is, in the first embodiment, the first significant deflection value is -1.8.
[0126] Step S472: Based on the first obvious deflection value, obtain a plurality of second obvious deflection values.
[0127] In this embodiment, the second obvious deflection value is any sequence value in the first difference sequence that is less than or equal to the first obvious deflection value.
[0128] In the first embodiment, the first obvious deflection value is -1.8, and there is no sequence value in the first difference sequence that is less than or equal to the first obvious deflection value. That is, there is no second obvious deflection value in the first embodiment.
[0129] Step S473: Based on the first obvious deflection value and the second obvious deflection value, obtain the semantic jump word group from the first word sequence.
[0130] In this embodiment, the semantic jump vocabulary group is a combination of two words in the first vocabulary sequence corresponding to the first obvious deflection value or the second obvious deflection value.
[0131] In the first embodiment, since there is only a first obvious deflection value and no second obvious deflection value, there is only one semantic jump vocabulary group. That is, the semantic jump vocabulary group consists of two words corresponding to the first vocabulary sequence and the first obvious deflection value (i.e., -1.8), which are "sudden illness" and "overseas travel".
[0132] Step S480: Extract key semantic segments from the first word sequence based on the semantic jump vocabulary group.
[0133] As mentioned above, since the meanings of the two words in the semantic jump vocabulary group have undergone a significant change, in this embodiment, key semantic segments can be extracted from the first word sequence through the semantic jump vocabulary group. For example, in the first embodiment, since it only has one semantic jump vocabulary group, two key semantic segments can be extracted from the first word sequence. The two extracted key semantic segments are "want to know, travel insurance, sudden illness" and "overseas travel, domestic travel, price difference".
[0134] Step S500: Based on the key semantic fragment, obtain the semantic deviation value.
[0135] In this embodiment, the semantic deviation value is used at least to characterize the magnitude of the difference in semantic deviation values between the words in the key semantic segment.
[0136] It should be clear that, in the embodiments of this application, any reasonable method can be used to obtain the semantic deviation value based on the key semantic segment. For example, the semantic deviation value can be the variance or standard deviation of the semantic deviation values corresponding to each word in the key semantic segment.
[0137] Step S500: Based on the key semantic fragment, obtain the semantic deviation value, including steps S510 to S550.
[0138] Step S510: Traverse the key semantic segments to obtain the fourth word.
[0139] In this embodiment, the fourth word is any word in the key semantic segment for which no semantic deviation has been obtained. That is, in this embodiment, the method for obtaining the semantic deviation of any word in the key semantic segment is the same as the method for obtaining the semantic deviation of the fourth word.
[0140] Step S520: Based on the fourth vocabulary, obtain the first semantic bias value.
[0141] In this embodiment, the first semantic deflection value is the semantic deflection value corresponding to the fourth word.
[0142] Step S530: Based on the key semantic fragment, obtain the first mean and the first standard deviation.
[0143] In this embodiment, the first average value is the average of the semantic shift values corresponding to each word in the key semantic segment. The first standard deviation is the standard deviation of the semantic shift values corresponding to each word in the key semantic segment.
[0144] Step S540: Based on the first semantic deviation value, the first average value, and the first standard deviation, obtain the semantic deviation of the fourth word.
[0145]
[0146] in, Indicates the semantic deviation of the fourth word; Indicates the semantic shift value of the fourth word; This represents the first average value; denoted as the first standard deviation; 'a' represents the zero-prevention coefficient, used to avoid a denominator of 0; the zero-prevention coefficient 'a' can be any value close to 0, for example, the zero-prevention coefficient 'a' can be 0.01 or 0.001, etc.
[0147] Step S550: Obtain the semantic deviation value based on the semantic deviation degree.
[0148] In embodiments of this application, the semantic deviation value corresponding to a key semantic segment can be the sum or average of the semantic deviations of each word in the key semantic segment. In a specific embodiment of this application, step S550, based on the semantic deviation, obtains the following formula for calculating the semantic deviation value:
[0149]
[0150] in, This represents the semantic deviation value corresponding to the key semantic segment; n represents the number of words in the key semantic segment. This represents the semantic deviation of the i-th word in the key semantic segment.
[0151] It's important to note that evaluating the semantic deviation value of key semantic segments reveals their actual deviation. When the semantic content of a key semantic segment largely aligns with user needs, the vocabulary within that segment is less likely to be modified, allowing for a more complete transmission of vocabulary to the next round of requirements. Therefore, a larger semantic deviation value indicates more words in the segment that generate a semantic deviation greater than the average (i.e., the first average), reflecting a semantic deviation in the vocabulary information. Conversely, a smaller semantic deviation value indicates a more consistent semantic deviation among the words in the segment, meaning the segment demonstrates stronger demand targeting in the current round.
[0152] Step S600: Based on the semantic deviation value, obtain the input weight of each word in the first word sequence.
[0153] In this embodiment, the input weights are used at least to characterize the importance of corresponding words in the first word sequence. It is easy to understand that if a word has a greater impact on the user's needs, then the weight of that word in the first word sequence should be heavier. Based on this, step S600, based on the semantic deviation value, obtains the input weights of each word in the first word sequence, including steps S610 to S630.
[0154] Step S610: Traverse the first word sequence to obtain the fifth word.
[0155] In this embodiment, the fifth word is any word in the first word sequence that has not received an input weight, and the semantic bias value corresponding to the fifth word is negative. As mentioned above, if the semantic bias value corresponding to a word is negative, it indicates that the customer demand represented by that word has undergone a "reverse change" in the current input demand data compared to the historical input demand data. In other words, in this embodiment, the smaller the semantic bias value corresponding to the fifth word, the greater its impact on the perceived user demand of the corresponding digital employee.
[0156] Step S620: Based on the fifth word, obtain the second semantic deviation value and the total deviation value.
[0157] In this embodiment, the second semantic deviation value is the semantic deviation value corresponding to the fifth word. The total deviation value is the sum of the semantic deviation values of all sixth words in the key semantic segment corresponding to the fifth word. The sixth word is any word in the corresponding key semantic segment with a negative semantic deviation value.
[0158] Step S630: Based on the semantic deviation value, the second semantic deflection value, and the total value of the deflection values, obtain the input weight of the fifth word.
[0159] In this embodiment, the input weight of the fifth word can be obtained in any reasonable way based on the semantic deviation value, the second semantic deviation value, and the total value of the deviation values. For example, in step S630, the formula for calculating the input weight of the fifth word based on the semantic deviation value, the second semantic deviation value, and the total value of the deviation values can be as follows:
[0160]
[0161] in, This indicates the input weight of the fifth word; Indicates the second semantic bias value; This represents the total deflection value; This indicates the semantic deviation value of the key semantic segment corresponding to the fifth word.
[0162] Step S700: Obtain the required results based on each input weight.
[0163] In the embodiments of this application, the input weights of the corresponding words can be used as the word weights in the large model corresponding to the digital employee, thereby obtaining the required results. Based on this, step S700, obtaining the required results based on each input weight, includes steps S710 to S740.
[0164] Step S710: Obtain a word vector sequence based on the first word sequence using a generative model.
[0165] In this embodiment, the generative model is pre-established. In the field of computer science, converting a word sequence (i.e., the first word sequence) into a vector sequence (i.e., a word vector sequence) is a mature technique, and will not be elaborated upon here. In the embodiments of this application, no restrictions are placed on the generative model; for example, the generative model can be a GPT model or a BERT model, etc.
[0166] Step S720: Obtain the weight vector sequence based on each input weight.
[0167] In the field of computer science, converting a weight sequence (i.e., the sequence formed by the individual input weights) into a vector sequence (i.e., a weight vector sequence) is a mature technique, which will not be elaborated here.
[0168] In a specific embodiment of this application, step S720, obtaining a weight vector sequence based on each input weight, includes steps S721 to S724.
[0169] Step S721: Traverse the first word sequence to obtain the seventh word.
[0170] In this embodiment, the seventh word is any word in the first word sequence for which no weight value has been obtained. That is, in this embodiment, the weight value of any word in the first word sequence is obtained in the same way as the weight value of the seventh word.
[0171] Step S722: Obtain the corrected weight based on the seventh vocabulary.
[0172] In this embodiment, if the input weight of the seventh word is empty, the corrected weight is equal to the second preset value. If the input weight of the seventh word is not empty, the corrected weight is equal to the sum of the second preset value and the input weight corresponding to the seventh word.
[0173] In the embodiments of this application, a second preset value can be set as needed. For example, the second preset value can be 1 or 2, etc.
[0174] Step S723: Normalize the corrected weights to obtain the weight values of the seventh word.
[0175] It is important to understand that normalization of numerical values is a well-established technique in the field of computer science, and will not be elaborated upon here.
[0176] Step S724: Obtain the weight vector sequence based on the weight values.
[0177] In the field of computer science, converting weight values into weight vectors is a mature technique, which will not be elaborated here.
[0178] Step S730: After replacing the attention weights in the generative model with the weights in the weight vector sequence, a weighted sentence representation is generated.
[0179] It's important to note that weighted sentence representation, in natural language processing, is a composite vector obtained by weighting the importance of different words in a sentence when converting it into a vector representation. Specifically, different words in a sentence often contribute differently to expressing the overall meaning. For example, in the sentence "I like to eat sweet apples," "like," "sweet," and "apple" might convey the core meaning of the sentence more effectively than "I" and "eat."
[0180] Step S740: Obtain the required result based on the weighted sentence representation.
[0181] It is important to understand that the technology of digital employees generating corresponding requirement results based on weighted sentence representations is a mature technology, which will not be elaborated upon here.
[0182] The proposed digital employee method based on generative artificial intelligence technology involves obtaining the demand changes (i.e., semantic deviation values) of each word in the current input demand data through current and historical input demand data. Then, the semantic deviation values are used to identify and segment key semantic segments that represent user intent, eliminating interfering information. Finally, based on the degree of semantic deviation of each key semantic segment (i.e., semantic deviation value), the weight vectors of each word are adjusted, allowing high-contribution words to account for a higher proportion in the output, thereby reducing the deviation between the generated digital employee and the user's actual needs.
[0183] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0184] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the methods, apparatuses, and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0185] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between devices or modules may be electrical, mechanical, or other forms.
[0186] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0187] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.
[0188] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.
[0189] The computer program product includes one or more computer instructions. When the computer program is loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video optical disc), or a semiconductor medium (e.g., solid-state drive (SSD)).
[0190] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles of this application.
Claims
1. A digital employee method based on generative artificial intelligence technology, characterized in that, include: Obtain the user's current input request data; Based on the current input demand data, a first vocabulary sequence and a second vocabulary sequence are obtained; the first vocabulary sequence is the vocabulary sequence corresponding to the current input demand data; the second vocabulary sequence is the vocabulary sequence corresponding to the previous input demand data of the current input demand data. Based on the first word sequence and the second word sequence, obtain the semantic bias value that corresponds one-to-one with each word in the first word sequence; Based on each semantic shift value, key semantic segments are extracted from the first word sequence; the key semantic segments are segments in the first word sequence whose semantic shift values change significantly. Based on the key semantic segments, obtain the semantic deviation value; The semantic deviation value is used at least to characterize the magnitude of the semantic deviation value difference between each word in the key semantic segment; Based on the semantic deviation value, the input weight of each word in the first word sequence is obtained; The input weights are used at least to characterize the importance of corresponding words in the first word sequence; Based on the weights of each input, obtain the required results.
2. The digital employee method based on generative artificial intelligence technology according to claim 1, characterized in that, The step of obtaining the first vocabulary sequence based on the current input demand data includes: Stop words are removed from the current input data based on stop word cleaning rules to obtain cleaned data; The cleaned data is split using a word segmentation tool to obtain the first word sequence.
3. The digital employee method based on generative artificial intelligence technology according to claim 1, characterized in that, The step of obtaining semantic shift values corresponding one-to-one with each word in the first word sequence based on the first word sequence and the second word sequence includes: Based on the first word sequence, a first word is obtained; the first word is any word in the first word sequence for which a semantic bias value has not been obtained. Based on the first word, a first value and a second value are obtained; the first value is the maximum similarity value between the first word and the remaining words in the first word sequence, and the second value is the maximum similarity value between the first word and each word in the second word sequence; or, the first value is the number of second words, and the second value is the number of third words; the second word is any word in the first word sequence whose similarity to the first word is greater than a first preset value; the third word is any word in the second word sequence whose similarity to the first word is greater than a first preset value. Based on the first value and the second value, obtain the semantic bias value of the first word.
4. The digital employee method based on generative artificial intelligence technology according to claim 3, characterized in that, After obtaining the semantic shift value of the first word based on the first value and the second value, the method further includes: Based on the first word, a first quantity is obtained; the first quantity is the number of sub-networks in the digital employee that the first word can activate. Based on the first word sequence, a second quantity is obtained; the second quantity is the total number of sub-networks in the digital employee that each word in the first word sequence can activate. Based on the first quantity and the second quantity, obtain the correction coefficient; The semantic bias value is corrected based on the correction coefficient.
5. The digital employee method based on generative artificial intelligence technology according to claim 1, characterized in that, The step of extracting key semantic segments from the first word sequence based on each semantic bias value includes: Based on the first word sequence, a first deflection value sequence is obtained; the position of the semantic deflection value in the first deflection value sequence is the same as the position of the corresponding word in the first word sequence; Based on the first deflection value sequence, a first difference sequence is obtained; the sequence value in the first difference sequence is equal to the difference between two adjacent sequence values in the first deflection value sequence. Based on the first difference sequence, an absolute value sequence is obtained; the absolute value sequence is a sequence formed by arranging the absolute values of each sequence value in the first difference sequence in descending order; Based on the absolute value sequence, a second difference sequence is obtained; the sequence value in the second difference sequence is equal to the difference between two adjacent sequence values in the absolute value sequence. Based on the second difference sequence, obtain the maximum value; Based on the maximum value, the maximum difference value is obtained from the absolute value sequence; the maximum difference value is the minuend of the absolute value sequence from which the maximum value is obtained. Based on the maximum difference value, a semantic jump word group is obtained from the first word sequence; Based on the semantic jump vocabulary group, key semantic fragments are extracted from the first vocabulary sequence.
6. The digital employee method based on generative artificial intelligence technology according to claim 5, characterized in that, Based on the maximum difference value, a semantic jump vocabulary group is obtained from the first vocabulary sequence, including: Based on the maximum difference value, a first significant deflection value is obtained from the first difference sequence; the first significant deflection value is the sequence value in the first difference sequence corresponding to the maximum difference value; Based on the first obvious deflection value, multiple second obvious deflection values are obtained; the second obvious deflection value is any sequence value in the first difference sequence that is less than or equal to the first obvious deflection value; Based on the first obvious deflection value and the second obvious deflection value, a semantic jump word group is obtained from the first word sequence; the semantic jump word group is a combination of two words in the first word sequence corresponding to the first obvious deflection value or the second obvious deflection value.
7. The digital employee method based on generative artificial intelligence technology according to claim 6, characterized in that, The step of obtaining semantic deviation values based on the key semantic fragments includes: Traverse the key semantic segments to obtain the fourth word; the fourth word is any word in the key semantic segments for which the semantic deviation has not been obtained; Based on the fourth word, a first semantic deviation value is obtained; the first semantic deviation value is the semantic deviation value corresponding to the fourth word; Based on the key semantic segments, a first average value and a first standard deviation are obtained; the first average value is the average of the semantic shift values corresponding to each word in the key semantic segments; the first standard deviation is the standard deviation of the semantic shift values corresponding to each word in the key semantic segments. Based on the first semantic deviation value, the first average value, and the first standard deviation, the semantic deviation of the fourth word is obtained; Based on the semantic deviation degree, obtain the semantic deviation value.
8. The digital employee method based on generative artificial intelligence technology according to claim 7, characterized in that, The step of obtaining the input weights of each word in the first word sequence based on the semantic deviation value includes: Traverse the first word sequence to obtain the fifth word; the fifth word is any word in the first word sequence for which no input weight has been obtained; and the semantic bias value corresponding to the fifth word is negative. Based on the fifth word, a second semantic deviation value and a total deviation value are obtained; the second semantic deviation value is the semantic deviation value corresponding to the fifth word; the total deviation value is the sum of the semantic deviation values of all sixth words in the key semantic segment corresponding to the fifth word; the sixth word is any word in the corresponding key semantic segment with a negative semantic deviation value; The input weight of the fifth word is obtained based on the semantic deviation value, the second semantic deflection value, and the total value of the deflection values.
9. The digital employee method based on generative artificial intelligence technology according to claim 8, characterized in that, The process of obtaining the required result based on each input weight includes: A generative model is used to obtain a word vector sequence based on the first word sequence; the generative model is pre-built. Based on each input weight, obtain the weight vector sequence; After replacing the attention weights in the generative model with the weights in the weight vector sequence, a weighted sentence representation is generated; Based on the weighted sentence representation, the required result is obtained.
10. The digital employee method based on generative artificial intelligence technology according to claim 9, characterized in that, The process of obtaining a weight vector sequence based on each input weight includes: Traverse the first word sequence to obtain the seventh word; the seventh word is any word in the first word sequence for which no weight value has been obtained. Based on the seventh word, a corrected weight is obtained; if the input weight of the seventh word is empty, the corrected weight is equal to the second preset value; if the input weight of the seventh word is not empty, the corrected weight is equal to the sum of the second preset value and the input weight corresponding to the seventh word. The corrected weights are normalized to obtain the weight values of the seventh word; Based on the weight values, the weight vector sequence is obtained.