Intelligent number-asking rapid construction optimization system for power production scene

Abstract

The invention, which relates to the technical field of artificial intelligence, discloses a power production scene-oriented intelligent question number rapid construction optimization system comprising a word vector management module, a question management module, a semantic extraction module, a semantic offset management module, a semantic offset management module, a feature word screening module and a feature word compensation module. By means of dynamically adjusting a semantic weight value, accurately judging semantic offset, strengthening key feature words, compensating word vector offset and the like, the accuracy of user intention tracking and the query result matching degree in multiple rounds of dialogues are remarkably improved. According to the method, query deviation caused by semantic offset or improper feature word weight distribution can be effectively avoided, and efficient and accurate data query requirements in an electric power production scene are met.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly to an intelligent question number rapid construction and optimization system for power production scenarios. Background Art

[0002] The intelligent question number system is an intelligent question and answer tool that combines technologies such as dynamic adjustment, context modeling, and domain knowledge enhancement. The system realizes the autonomous design of correlation words by obtaining the characteristics of the user's question and the dialogue scenario and matching the correlation word model, so as to guide the user's question, make up for the deficiency of manual natural language input, and can recognize and process free-form natural language input to improve the accuracy of data query tasks. Existing intelligent question and answer systems have certain limitations in the depth of semantic understanding and the accuracy of intention capture. Although the efficiency of keyword extraction can be improved through specific algorithms, in continuous dialogue scenarios, the system's ability to predict the multi-round intention changes of users still has room for improvement. In addition, when there is a semantic deviation in the user's question, the real-time detection and compensation mechanism of the system still needs to be further optimized, which may lead to a deviation between the query result and the user's needs. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent question number rapid construction and optimization system for power production scenarios to solve the problems raised in the prior art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: an intelligent question number rapid construction and optimization system for power production scenarios.

[0005] The system includes: a word vector management module, a question management module, a semantic extraction module, a semantic offset amount management module, a semantic offset management module, a feature word screening module, and a feature word compensation module.

[0006] The word vector management module is used to collect the historical records in the continuous dialogue scenario between the questioner and the intelligent agent, extract the feature words in the question information, and establish the corresponding relationship between the feature words and the word vectors through word vector marking; The question management module is used to collect the question records of the questioner to the intelligent agent in each round in several rounds of dialogue scenarios, and gather the questions raised by the questioner into a question sequence; The semantic extraction module is used to classify the feature words in the question, use the classification result as the semantic feature, and identify and label the semantic feature to obtain the feature words under each semantic feature; The semantic offset amount management module: is used to calculate the change value of the semantic feature in the question information, including the semantic weight value of the target semantic feature and the semantic change evaluation value of the same semantic feature in two adjacent rounds of questions; Semantic deviation management module: used to calculate the semantic deviation threshold and determine whether there is a semantic deviation in the questioner's question intention based on the semantic deviation threshold; The feature word screening module is used to screen the feature words that need to be strengthened in the question content, collect the feature words that need to be strengthened, and obtain a set of feature words; The feature word compensation module is used to calculate the vector offset and compensate the word vector to generate a new vector, and obtain the feature word corresponding to the new vector according to the corresponding relationship between the vector and the feature word, and output the feature word to the questioner for selection.

[0007] Furthermore, in the word vector management module, when collecting the historical records of the questioner in the continuous dialogue scenario with the intelligent agent, the feature words in the question information are parsed through natural language processing technology, and the feature words are converted into word vectors in a high-dimensional space by using the word embedding algorithm, and the mapping relationship between the feature words and the word vectors is established.

[0008] Furthermore, in the question management module, when collecting the question records of the questioner to the intelligent agent in each round, the question content is obtained through voice recognition or text input, and the question content is arranged in chronological order to form a question sequence; the question sequence is indexed by time stamps and supports retrieving and analyzing the question distribution rules in different time periods according to the time range.

[0009] Furthermore, in the semantic extraction module, when classifying the feature words in the question, based on the domain knowledge base and the semantic network model, the feature words are divided into categories such as device status class, operating parameter class, fault diagnosis class, and scheduling strategy class.

[0010] Furthermore, the semantic deviation amount management module includes: a semantic feature value calculation unit and a semantic change management unit; The semantic feature value calculation unit is used to calculate the i-th semantic feature in the t-th round of questions as the target semantic feature, and calculate the semantic weight value W of the target semantic feature t,i , , where q ti represents the number of feature words included in the i-th semantic feature in the t-th round of questions, and Q t represents the number of feature words included in the t-th round of questions; The semantic change management unit is used to calculate the semantic change evaluation value of the same semantic feature in two adjacent rounds of questions, obtain the semantic weight value W of the target semantic feature in the (t - 1)-th round of questions t-1,i , calculate the semantic change value R of the target semantics in the t-th round of questions t , .

[0011] The semantic deviation amount management module dynamically analyzes the semantic weight value to ensure that the importance of the target semantic feature in each round of questions can be accurately captured.

[0012] Furthermore, the semantic shift management module includes: a semantic change value prediction unit, a shift threshold management unit, and a shift judgment unit; The semantic change value prediction unit is used to calculate the semantic change prediction value, obtain the semantic change value of the target semantic feature in n rounds of questions, where n ≥ 3, and n rounds of questions end with the t-th round of questions. The semantic change prediction value R of the target semantics in the (t + 1)-th round is calculated through function fitting t+1,p ; The shift threshold management unit is used to calculate the semantic shift threshold, and calculate the semantic shift threshold S of the target semantics in the (t + 1)-th round of questions t+1 , , where α represents the weight value, satisfying the condition: 0 < α < 1, and ω t-1,t represents the similarity of the corresponding word vectors of all feature words in the (t - 1)-th round of questions compared with the t-th round of questions; The shift judgment unit is used to judge whether there is a shift in the questioner's question intention, and obtain the true semantic change value R of the (t + 1)-th round of questions t+1,a , when R t+1,a > S t+1 , it is judged that in the (t + 1)-th round of questions, a semantic shift of the target semantics occurs.

[0013] The semantic shift management module combines the semantic change prediction value and the context similarity to calculate the semantic shift threshold, effectively judging whether the semantics has shifted, so as to avoid query deviation caused by semantic shift.

[0014] Furthermore, the feature word screening module includes: a word frequency acquisition unit, a word evaluation value management unit, and a feature word set management unit; The word frequency acquisition unit is used to obtain the word frequency of each feature word in the target semantics in the (t + 1)-th round of questions, and the number of question rounds in which each feature word appears in the first (t + 1) rounds. Among them, in the (t + 1)-th round of questions, the word frequency of the k-th feature word of the target semantics in the (t + 1)-th round of questions is denoted as TF k , and the number of rounds in which the k-th feature word appears in the first (t + 1) rounds of questions is denoted as N k ; The word evaluation value management unit is used to calculate the word evaluation value B of the k-th feature word k , ; The feature word set management unit is used to arrange the feature words of the target semantics from high to low according to the word evaluation value, and include the first m feature words in the feature word set.

[0015] The feature word screening module screens key feature words by calculating the word evaluation value, ensuring that the feature word set can accurately reflect the core content of the target semantics.

[0016] Further, the feature word compensation module includes: a vector offset calculation unit, an offset coefficient calculation unit, a vector correction unit, and a feature word output unit; The vector offset calculation unit is used to calculate the vector offset. The vector offset is the average value of the word vectors of all the feature words included in the semantic features in the (t + 1)-th round of query, denoted as avg; The offset coefficient calculation unit is used to calculate the offset coefficient γ, ; The vector correction unit is used to calculate the correction vector of the word vector. Among them, the correction vector p of the k-th feature word is calculated k , , where v k represents the word vector corresponding to the k-th feature word; The feature word output unit is used to collect the feature words corresponding to the correction vectors and output the set of feature words corresponding to the correction vectors to the questioner.

[0017] The feature word compensation module generates new feature words by calculating the correction vectors, compensates for the word vector offset problem caused by semantic offset or improper feature word weight distribution, thereby improving the matching degree between the query result and the user's needs.

[0018] Compared with the prior art, the beneficial effects of the present invention are: by means of dynamically adjusting the semantic weight value, accurately judging the semantic offset, strengthening the key feature words, and compensating for the word vector offset, etc., the accuracy of user intention tracking and the matching degree of query results in multi-round conversations are significantly improved. This application can effectively avoid query deviation caused by semantic offset or improper feature word weight distribution, and meet the efficient and accurate data query requirements in the power production scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic structural diagram of an intelligent question number rapid construction and optimization system for a power production scenario according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] Embodiment: As Figure 1 shown, the present invention provides a technical solution, an intelligent question number rapid construction and optimization system for a power production scenario.

[0022] In the power production scenario, the equipment status monitoring task is one of the typical applications of the intelligent question-answering system. First, the historical records in the continuous dialogue scenario between the questioner and the intelligent agent are collected, and the feature words in the question information are parsed through natural language processing technology. For example, when the questioner asks about the operating status of a certain generator, the question content may contain keywords such as "generator", "temperature", "vibration", etc. These keywords are converted into word vectors in a high-dimensional space through the word embedding algorithm, and the mapping relationship between the feature words and the word vectors is established. The dimension of the word vector is determined by the pre-trained language model, and the update frequency is associated with the timeliness of the question information. This real-time update mechanism ensures that the word vector can dynamically adapt to the changes in the question content, thereby improving the response accuracy of the system.

[0023] The system includes: a word vector management module, a question management module, a semantic extraction module, a semantic offset management module, a semantic offset management module, a feature word screening module, and a feature word compensation module.

[0024] The word vector management module is used to collect the historical records in the continuous dialogue scenario between the questioner and the intelligent agent, extract the feature words in the question information, and establish the corresponding relationship between the feature words and the word vectors through word vector marking; Among them, in the word vector management module, when collecting the historical records of the questioner in the continuous dialogue scenario with the intelligent agent, the feature words in the question information are parsed through natural language processing technology, and the feature words are converted into word vectors in a high-dimensional space by using the word embedding algorithm, and the mapping relationship between the feature words and the word vectors is established.

[0025] The question management module is used to collect the question records of the questioner to the intelligent agent in each round in several rounds of dialogue scenarios, and pool the questions raised by the questioner into a question sequence; Among them, in the question management module, when collecting the question records of the questioner to the intelligent agent in each round, the question content is obtained through voice recognition or text input, and the question content is arranged in chronological order to form a question sequence; the question sequence is indexed by time stamps, and supports retrieving and analyzing the question distribution rules in different time periods according to the time range.

[0026] The semantic extraction module is used to classify the feature words in the question, use the classification result as the semantic feature, and identify and label the semantic feature to obtain the feature words under each semantic feature; Among them, in the semantic extraction module, when classifying the feature words in the question, based on the domain knowledge base and the semantic network model, the feature words are divided into categories such as equipment status class, operating parameter class, fault diagnosis class, and scheduling strategy class.

[0027] Semantic offset management module: used to calculate the change value of semantic features in the question information, including the semantic weight value of the target semantic feature and the semantic change evaluation value of the same semantic feature in two adjacent rounds of questions; Among them, the semantic offset management module includes: a semantic feature value calculation unit and a semantic change management unit.

[0028] Among them, the semantic feature value calculation unit is used to calculate the i-th semantic feature in the t-th round of questions as the target semantic feature, and calculate the semantic weight value W of the target semantic feature t,i , , where, q ti represents the number of feature words included in the i-th semantic feature in the t-th round of questions, and Q t represents the number of feature words included in the t-th round of questions; The semantic change management unit is used to calculate the semantic change evaluation value of the same semantic feature in two adjacent rounds of questions, obtain the semantic weight value W of the target semantic feature in the (t - 1)-th round of questions t-1,i , and calculate the semantic change value R of the target semantics in the t-th round of questions t , .

[0029] Semantic offset management module: used to calculate the semantic offset threshold, and judge whether there is a semantic offset in the questioner's question intention according to the semantic offset threshold; Among them, the semantic offset management module includes: a semantic change value prediction unit, an offset threshold management unit, and an offset judgment unit.

[0030] Among them, the semantic change value prediction unit is used to calculate the semantic change prediction value, obtain the semantic change value of the target semantic feature in n rounds of questions, n ≥ 3, where n rounds of questions end with the t-th round of questions, and calculate the semantic change prediction value R of the target semantics in the (t + 1)-th round through function fitting t+1,p ; The offset threshold management unit is used to calculate the semantic offset threshold, and calculate the semantic offset threshold S of the target semantics in the (t + 1)-th round of questions t+1 , , α represents the weight value, satisfying the condition: 0 < α < 1, where, ω t-1,t represents the similarity of the corresponding word vectors of all feature words in the (t - 1)-th round of questions compared with the t-th round of questions; The offset judgment unit is used to judge whether there is an offset in the questioner's question intention, obtain the true semantic change value R of the (t + 1)-th round of questions t+1,a , when R t+1,a > S t+1 , it is judged that in the (t + 1)-th round of questions, the target semantics has a semantic offset.

[0031] The feature word screening module is used to screen the feature words that need to be strengthened in the question content, gather the feature words that need to be strengthened, and obtain the feature word set; Among them, the feature word screening module includes: a word frequency acquisition unit, a word evaluation value management unit, and a feature word set management unit.

[0032] Among them, the word frequency acquisition unit is used to obtain the word frequency of each feature word in the target semantics in the (t + 1)-th round of questions, and the number of question rounds in which each feature word appears in the previous (t + 1) rounds. Among them, in the (t + 1)-th round of questions, the word frequency of the k-th feature word in the target semantics in the (t + 1)-th round of questions is denoted as TF k and the number of rounds in which the k-th feature word appears in the previous (t + 1) rounds of questions is denoted as N k ; The word evaluation value management unit is used to calculate the word evaluation value B of the k-th feature word k , ; The feature word set management unit is used to arrange the feature words of the target semantics from high to low according to the word evaluation value, and include the first m feature words in the feature word set.

[0033] The feature word compensation module is used to calculate the vector offset and compensate the word vector to generate a new vector, and obtain the feature word corresponding to the new vector according to the correspondence between the vector and the feature word, and output the feature word to the questioner for selection; Among them, the feature word compensation module includes: a vector offset calculation unit, an offset coefficient calculation unit, a vector correction unit, and a feature word output unit.

[0034] Among them, the vector offset calculation unit is used to calculate the vector offset. The vector offset is the average value of the word vectors of all feature words included in the semantic features in the (t + 1)-th round of questions, denoted as avg; The offset coefficient calculation unit is used to calculate the offset coefficient γ, ; The vector correction unit is used to calculate the correction vector of the word vector. Among them, the correction vector p of the k-th feature word is calculated k , , where v k represents the word vector corresponding to the k-th feature word; The feature word output unit is used to gather the feature words corresponding to the correction vectors and output the set of feature words corresponding to the correction vectors to the questioner.

[0035] In practical applications, the method of the present invention is applicable to multi-round dialogue systems in power production scenarios, such as scenarios of equipment operation status query, fault diagnosis information retrieval, and production data statistics. Taking a certain power company as an example, its operation and maintenance personnel need to obtain historical operation data or real-time monitoring information of specific equipment by asking questions in natural language. In this process, the system needs to be able to accurately capture the change of the user's intention and dynamically adjust the query strategy, so as to ensure a high degree of matching between the query result and the user's needs.

[0036] First, in the semantic extraction module, the user's question information is transmitted to the internal processing unit of the module through the interface. The processing unit splits the question content into multiple feature words according to the preset word segmentation algorithm and identifies the target semantic features therein. The determination of the target semantic features is based on the power production-related terms stored in the domain knowledge base and their semantic classification rules. For example, when the user asks "Query the change trend of the temperature of the A-phase winding of the main transformer in the past week", the processing unit will extract feature words such as "main transformer", "A-phase winding", "temperature", and "change trend", and classify them as target semantic features. Subsequently, the processing unit calculates the semantic weight value W of the target semantic features in the t-th round of questions t,i , and its formula is W t,i = q ti / Q t . Among them, q ti represents the number of feature words included in the target semantic feature, and Q t represents the total number of all feature words in the current question. For example, if the current question contains 10 feature words and the target semantic feature contains 3 feature words, the semantic weight value is 0.3. This weight value is stored in the semantic weight database for subsequent modules to call.

[0037] Next, the offset management module extracts the semantic weight values of the same semantic feature in two adjacent rounds of questions from the semantic weight database and calculates the semantic change evaluation value R t . The formula for the semantic change evaluation value is R t = (W t,i - W t-1,i ) / W t-1,i .

[0038] For example, if the weight values of a certain semantic feature in the 4th round of questions are 0.8, 0.75, 0.6, and 0.4, and the weight value in the t-th round is 0.4, the semantic change evaluation values are -0.0625, -0.2, and -0.33 respectively.

[0039] The offset management module generates a semantic change data set based on historical data and fits and predicts the semantic change prediction value R of the next round of questions through the linear regression algorithm t+1,p, it is -0.4525. At the same time, the module also calculates the similarity ω of the word vectors corresponding to all feature words in the previous two rounds of questions t-1,t , whose value range is from 0 to 1. The offset management module combines the semantic change prediction value R t+1,p and the word vector similarity ω t-1,t to calculate the semantic offset threshold S t+1 , The formula is S t+1 = α × R t+1,p + (1 - α) × ω t-1,t . Among them, α is the weight coefficient, which can be set to 0.6 according to actual needs. For example, if R t+1,p is -0.4525 and ω t-1,t is 0.6, then the semantic offset threshold S t+1 is 0.6 × (-0.4525) + 0.4 × 0.6 = -0.0315.

[0040] After completing the calculation of the semantic offset threshold, the offset management module obtains the true value R of the semantic change of the (t + 1)-th round of questions t+1,a , and compares it with the semantic offset threshold S t+1 . If R t+1,a is greater than S t+1 , it is determined that the target semantics has a semantic offset, and a semantic offset flag bit is generated. For example, if the weight value of a certain semantic feature in the (t + 1)-th round of questions suddenly rises from 0.4 to 0.5, then the true value of the semantic change is (0.5 - 0.4) / 0.4 = 0.25, exceeding the semantic offset threshold of 0.2815. At this time, the offset management module generates a semantic offset flag bit and transmits it to the feature word screening module.

[0041] After receiving the semantic offset flag bit, the feature word screening module counts the term frequency TF of each feature word in the target semantics in the (t + 1)-th round of questions k and the number of rounds N in which it appears in the previous (t + 1) rounds of questions k . The module sorts the feature words in descending order of the word evaluation value and selects the top m feature words to form a feature word set. For example, if m is set to 3, then the 3 feature words with the highest word evaluation value are selected as the feature word set and transmitted to the feature word compensation module.

[0042] After receiving the feature word set, the feature word compensation module calculates the average value avg of the word vectors corresponding to all feature words in the (t + 1)-th round of questions. The module combines the semantic offset threshold S t+1 and the true value R of the semantic change t+1,a to calculate the offset coefficient γ, and its formula is γ = S t+1 / R t+1,a .

[0043] For example, if the semantic deviation threshold is -0.0315 and the true value of semantic change is 0.25, then the deviation coefficient γ is (-0.0315) / 0.25 ≈ -0.126. The module calculates the corrected vector p of each feature word according to the deviation coefficient γ and the average value avg of word vectors k , and its formula is p k = 0.5×(v k + (1 - γ)×avg). Wherein, v k represents the original word vector corresponding to the k-th feature word.

[0044] For example, if the original word vector of a certain feature word is [0.1, 0.2, 0.3] and the average value of word vectors is [0.4, 0.5, 0.6], then its corrected vector is 0.5×([0.1, 0.2, 0.3] + 1.126×[0.17, 0.28, 0.36]) = [0.145, 0.257, 0.352], and the fine-tuning of the corresponding original word vector has been realized. The feature word compensation module generates new feature words according to the corrected vector and outputs the new feature words to the user interface for the user to select.

[0045] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed claims.

Claims

1. An intelligent question number rapid construction and optimization system for power production scenarios, characterized in that: The system includes: a word vector management module, a question management module, a semantic extraction module, a semantic offset management module, a semantic shift management module, a feature word screening module, and a feature word compensation module; The word vector management module is used to collect historical records in the continuous conversation scenario between the questioner and the intelligent agent, extract feature words in the question information, and establish the corresponding relationship between the feature words and the word vectors through word vector tagging; The question management module is used to collect the question records of the questioner to the intelligent agent in each round in several rounds of conversation scenarios, and assemble the questions raised by the questioner into a question sequence; The semantic extraction module is used to classify the feature words in the question, use the classification result as the semantic feature, identify and label the semantic feature, and obtain the feature words under each semantic feature; The semantic offset management module: is used to calculate the change value of the semantic feature in the question information, including the semantic weight value of the target semantic feature and the semantic change evaluation value of the same semantic feature in two adjacent rounds of questions; The semantic shift management module: is used to calculate the semantic shift threshold, and judge whether there is a semantic shift in the question intention of the questioner according to the semantic shift threshold; The feature word screening module is used to screen the feature words that need to be strengthened in the question content, assemble the feature words that need to be strengthened, and obtain a feature word set; The feature word compensation module is used to calculate the vector offset and compensate the word vector, generate a new vector, obtain the feature word corresponding to the new vector according to the corresponding relationship between the vector and the feature word, and output the feature word to the questioner for selection.

2. The intelligent question number rapid construction and optimization system for the power production scenario according to claim 1, characterized in that: In the word vector management module, when collecting the historical records of the questioner in the continuous conversation scenario with the intelligent agent, the feature words in the question information are parsed through natural language processing technology, and the feature words are converted into word vectors in the high-dimensional space by using the word embedding algorithm, and the mapping relationship between the feature words and the word vectors is established.

3. The intelligent question number rapid construction and optimization system for the power production scenario according to claim 1, characterized in that: In the question management module, when collecting the question records of the questioner to the intelligent agent in each round, the question content is obtained through voice recognition or text input method, and the question content is arranged in chronological order to form a question sequence; the question sequence is indexed by time stamps, and supports retrieving and analyzing the question distribution rules in different time periods according to the time range.

4. The intelligent question number rapid construction and optimization system for the power production scenario according to claim 1, characterized in that: In the semantic extraction module, when classifying the feature words in the question, based on the domain knowledge base and the semantic network model, the feature words are divided into categories such as equipment status class, operation parameter class, fault diagnosis class, and scheduling strategy class.

5. An intelligent question number rapid construction and optimization system for power production scenarios according to claim 1, characterized in that: The semantic offset management module includes: a semantic feature value calculation unit and a semantic change management unit; A semantic feature value calculation unit, which is used to calculate the i-th semantic feature in the t-th round of questions as the target semantic feature, and calculate the semantic weight value W of the target semantic feature t,i , , where q ti represents the number of feature words included in the i-th semantic feature in the t-th round of questions, and Q t represents the number of feature words included in the t-th round of questions; The semantic change management unit is used to calculate the semantic change evaluation value of the same semantic feature in two adjacent rounds of questions, and obtain the semantic weight value W of the target semantic feature in the (t-1)-th round of questions t-1,i , and calculate the semantic change value R of the target semantics in the t-th round of questions t , .

6. An intelligent question number rapid construction and optimization system for the power production scenario according to claim 5, characterized in that: The semantic shift management module includes: a semantic change value prediction unit, an offset threshold management unit, and an offset judgment unit; The semantic change value prediction unit is used to calculate the semantic change prediction value, obtain the semantic change value of the target semantic feature in n rounds of questions, where n≥3, and the n rounds of questions end with the t-th round of questions. The semantic change prediction value R of the target semantics in the (t + 1)-th round is calculated by function fitting t+1,p ; The offset threshold management unit is used to calculate the semantic offset threshold, and calculate the semantic offset threshold S of the target semantics in the (t + 1)-th round of questioning t+1 , , where α represents the weight value, satisfying the condition: 0 < α < 1, and ω t-1,t represents the similarity of the corresponding word vectors of all feature words in the (t - 1)-th round of questioning compared with the t-th round of questioning; The offset judgment unit is used to judge whether there is an offset in the questioning intention of the questioner and obtain the true value R of the semantic change of the (t + 1)-th round of questioning. t+1,a When R t+1,a > S t+1 it is judged that in the (t + 1)-th round of questioning, a semantic offset has occurred in the target semantics.

7. An intelligent question number rapid construction and optimization system for power production scenarios according to claim 6, characterized in that: The feature word screening module includes: a word frequency acquisition unit, a word evaluation value management unit, and a feature word set management unit; The word frequency acquisition unit is used to acquire the word frequency of each feature word in the target semantics in the (t + 1)-th round of questioning, and the number of rounds of questioning in which each feature word appears in the previous (t + 1) rounds. Among them, in the (t + 1)-th round of questioning, the word frequency of the k-th feature word of the target semantics in the (t + 1)-th round of questioning is denoted as TF k , and in the previous (t + 1) rounds of questioning, the number of rounds in which the k-th feature word appears is denoted as N k ; The word evaluation value management unit is used to calculate the word evaluation value B of the k-th feature word k , ; The feature word set management unit is used to arrange the feature words of the target semantics according to the word evaluation values from high to low, and include the top m feature words into the feature word set.

8. An intelligent question number rapid construction and optimization system for power production scenarios according to claim 7, characterized in that: The feature word compensation module includes: a vector offset calculation unit, an offset coefficient calculation unit, a vector correction unit, and a feature word output unit; The vector offset calculation unit is used to calculate the vector offset, and the vector offset is the average value of the word vectors of the feature words included in all semantic features in the (t + 1)-th round of questioning, denoted as avg; The offset coefficient calculation unit is used to calculate the offset coefficient γ, ; The vector correction unit is used to calculate the correction vector of the word vector, where the correction vector p of the k-th feature word is calculated k , , where v k represents the word vector corresponding to the k-th feature word; The feature word output unit is used to collect the feature words corresponding to the corrected vector and output the feature word set corresponding to the corrected vector to the questioner.

Images