CN-121501970-B - Question and scheduling method and system

CN121501970BCN 121501970 BCN121501970 BCN 121501970BCN-121501970-B

Abstract

The invention relates to a questioning and scheduling method and a questioning and scheduling system, and belongs to the technical field of natural language processing and intelligent scheduling. The method comprises the steps of obtaining question interaction multi-source state data, setting quantifiable monitoring indexes of each dimension and three-level judging standards, constructing a hierarchical flow monitoring mechanism, preprocessing the multi-source data, constructing a multi-dimensional state identification model, judging state grades according to the standards and outputting a grinding and judging result, constructing a strategy set, continuously monitoring state changes, constructing state feature vectors and strategy parameter vectors, executing scheduling actions, obtaining feedback evaluation effects, storing related data in a correlated mode, calculating strategy matching degree, and iteratively optimizing the monitoring indexes and the judging standards. The intelligent questioning interaction method and device achieve accurate perception and scientific research and judgment of the full-flow multidimensional state of questioning interaction, solve the problems that scheduling is stiff, pain points without full-flow monitoring and iteration capability exist in the prior art, and adapt to various intelligent questioning interaction scenes.

Inventors

FANG XIAOLEI
DING QUN
LIU ZHIWEI

Assignees

上海近屿智能科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260114

Claims (10)

1. A questioning and scheduling method, comprising: s1, acquiring multisource state data of questioning interaction, setting quantifiable monitoring indexes for each dimension, setting three-level index judgment standards according to the multisource state data and historical interaction data, and constructing a hierarchical flow monitoring mechanism; the layered flow monitoring mechanism comprises an answer state layer, an environment state layer and an operation state layer; The method comprises the steps of presetting a core monitoring range and specific responsibilities of each layer, wherein the answer state layer is responsible for continuously tracking the integrity, answer attitude and emotion fluctuation of answer content of an evaluation object, the environment state layer is responsible for monitoring the type and influence range of an interference source in an evaluation scene in real time, and the running state layer is responsible for dynamically controlling the data transmission fluency and program running state in the interaction process; Setting a data synchronization triggering condition, and automatically synchronizing related data to other related levels when any layer monitors suspected abnormal data; s2, preprocessing the multi-source state data, constructing a multi-dimensional state recognition model based on the preprocessed data, judging the state grade according to the three-level index judgment standard, and outputting a state judging result and a corresponding state type; The specific process for constructing the multidimensional state identification model comprises the following steps: comprises a feature extraction layer, a feature fusion layer and a state classification layer; The feature extraction layer adopts a corresponding feature extraction mode aiming at different types of preprocessing data, extracts semantic features from the answer text, extracts Mel frequency related features from voice data, extracts interference features from environment data and extracts state features from operation data; the feature fusion layer integrates the features extracted from each dimension into a unified feature vector in a feature splicing mode, and redundant overlapping parts among the features are removed; The state classification layer adopts a classification algorithm to construct a classifier, and inputs the fused feature vectors into the classifier to carry out classification recognition on the answer state, the environment state and the running state; S3, constructing a strategy set containing parameterized strategies based on the state research and judgment result, continuously monitoring state changes in the questioning interaction process, constructing the state research and judgment result into state feature vectors containing a plurality of state feature dimensions, and constructing each scheduling strategy into corresponding strategy parameter vectors; And S4, executing a scheduling action according to the scheduling strategy, acquiring the scheduled interactive feedback data in real time and evaluating the scheduling effect, storing the multi-source state data, the scheduling strategy and the scheduling effect data in a correlated mode, calculating the matching degree of the scheduling strategy under different state scenes, and updating the scheduling strategy parameters under the corresponding state types when the matching degree is lower than a preset threshold after continuous scheduling.
2. The method of claim 1, wherein the specific process of setting the quantifiable monitoring indexes for each dimension is to divide a core monitoring dimension according to an answer state, an environment state and an operation state, set the quantifiable monitoring indexes for each dimension, set an answer integrity index by counting answer key information coverage rate of the answer state dimension, set a semantic fit index by comparing answer with a questioning core semantic, set an interference degree index by interference influence degree of the environment state dimension on interaction, and set a stability index by data transmission delay and fault occurrence frequency by the operation state dimension.
3. The method of claim 1, wherein the specific process of setting the three-level index judgment standard is to collect multi-source state data and historical interaction data of each dimension monitoring index, set normal, early warning and abnormal three-level index intervals for each dimension monitoring index respectively, preset index ranges corresponding to each level, and trigger recalculation or update of at least one other layer of corresponding state characteristics when the monitoring index of any layer meets the abnormal judgment condition in the hierarchical flow monitoring mechanism.
4. The method according to claim 1, wherein the specific process of preprocessing the multi-source state data is that the collected multi-source state data is cleaned to remove invalid data and redundant information, storage and expression formats of data with different sources and different formats are unified, core characteristic information is extracted from unstructured data and converted into standardized structured data, and consistency and usability of the data are checked.
5. The method of claim 1, wherein the specific process of constructing the policy set including the parameterized policies includes carding the processing requirements and the scheduling targets corresponding to the state research and judgment results, designing an adaptive scheduling action, disassembling the adaptive scheduling action into quantifiable policy parameters and defining a value constraint, establishing structural association according to the state types, and forming a standardized policy set with parameter expansion interfaces and quick retrieval indexes.
6. The method of claim 1, wherein the specific process of constructing each scheduling policy into a corresponding policy parameter vector includes screening quantifiable core parameters based on the influence weight of the scheduling policy on the state, unifying parameter quantification standards and dimension units, mapping parameters to a [0,1] interval to complete normalization, arranging the parameters according to a preset fixed dimension sequence and preserving indexes to form a policy parameter vector with unified structure.
7. The method of claim 1, wherein the specific process of calculating the state feature vector and the policy parameter vector includes taking a state improvement effect, a flow consistency and a question propulsion efficiency as core measurement dimensions, comparing state differences before and after scheduling, accounting for improvement amplitude of an abnormal state and an early warning state by a scheduling policy, counting a duration ratio of no interruption of a question flow in the scheduling process, measuring and calculating completion efficiency of the question propulsion according to a plan in a corresponding state scene, and calculating matching degree of the scheduling policy in different state scenes by integrating various accounting and statistics results.
8. The method of claim 1, wherein the specific process of executing the adjustment type scheduling policy by triggering the early warning is that the specific type of the early warning is matched, the adjustment action is correspondingly executed according to the specific early warning type, the response type early warning is used for slowing down the question speed and prolonging the response interval, the environment type early warning is used for suspending the question and is recovered after the interference is weakened, the operation type early warning is used for following the question sequence, and the response information missing type early warning is used for generating and pushing the additional question content according to the missing information.
9. The method of claim 1, wherein the specific process of updating the scheduling policy parameters under the corresponding state type is to extract the relevant history data to locate the core policy parameters when the matching degree under the single state type does not reach the preset threshold, iteratively adjust and verify the core policy parameters, update the corresponding scheduling policy parameters, analyze the monitoring index or three-level judgment standard suitability of the corresponding dimension when the matching degree after the single state type is updated by the policy parameters does not reach the standard, synchronously adjust the index setting or judgment standard range, link and optimize the corresponding policy parameters, and detect the mapping logic matching defect of the state feature vector and the policy parameter vector when the matching degree under the various different state scenes does not reach the preset threshold, and synchronously calibrate the scheduling policy parameters of each relevant state scene.
10. A questioning and scheduling system, comprising: the monitoring index construction module is used for acquiring multisource state data of questioning interaction, setting quantifiable monitoring indexes for each dimension, setting three-level index judgment standards according to the multisource state data and historical interaction data and constructing a layered flow monitoring mechanism; the layered flow monitoring mechanism comprises an answer state layer, an environment state layer and an operation state layer; The method comprises the steps of presetting a core monitoring range and specific responsibilities of each layer, wherein the answer state layer is responsible for continuously tracking the integrity, answer attitude and emotion fluctuation of answer content of an evaluation object, the environment state layer is responsible for monitoring the type and influence range of an interference source in an evaluation scene in real time, and the running state layer is responsible for dynamically controlling the data transmission fluency and program running state in the interaction process; Setting a data synchronization triggering condition, and automatically synchronizing related data to other related levels when any layer monitors suspected abnormal data; the state recognition and judgment module is used for preprocessing the multi-source state data, constructing a multi-dimensional state recognition model based on the preprocessed data, judging the state grade according to the three-level index judgment standard, and outputting a state judgment result and a corresponding state type; The specific process for constructing the multidimensional state identification model comprises the following steps: comprises a feature extraction layer, a feature fusion layer and a state classification layer; The feature extraction layer adopts a corresponding feature extraction mode aiming at different types of preprocessing data, extracts semantic features from the answer text, extracts Mel frequency related features from voice data, extracts interference features from environment data and extracts state features from operation data; the feature fusion layer integrates the features extracted from each dimension into a unified feature vector in a feature splicing mode, and redundant overlapping parts among the features are removed; The state classification layer adopts a classification algorithm to construct a classifier, and inputs the fused feature vectors into the classifier to carry out classification recognition on the answer state, the environment state and the running state; The scheduling rule matching module is used for constructing a strategy set containing parameterized strategies based on the state research and judgment result, continuously monitoring state change in the questioning interaction process, constructing the state research and judgment result into state feature vectors containing a plurality of state feature dimensions, and constructing each scheduling strategy into corresponding strategy parameter vectors; The scheduling iteration optimization module is used for executing scheduling actions according to the scheduling strategies, acquiring interaction feedback data after scheduling in real time and evaluating scheduling effects, storing the multi-source state data, the scheduling strategies and the scheduling effect data in an associated mode, calculating the matching degree of the scheduling strategies in different state scenes, and updating scheduling strategy parameters in corresponding state types when the matching degree is lower than a preset threshold after continuous scheduling.

Description

Question and scheduling method and system Technical Field The invention belongs to the technical field of natural language processing and intelligent scheduling, and particularly relates to a questioning and scheduling method and system. Background In the scenes of intelligent evaluation, talent selection, skill assessment and the like, which rely on questioning interaction to complete evaluation, the consistency, rhythm suitability and state controllability of the questioning process directly determine the evaluation efficiency and the result accuracy. Along with the development of intelligent interaction technology, the prior art has realized basic intelligent questioning function, namely, questioning contents are automatically output according to a preset list, but a remarkable technical short board exists on a dynamic control layer of the whole questioning process, and the intelligent assessment requirements of high efficiency and accuracy cannot be met, and the specific defects are as follows: The existing intelligent questioning system only focuses on a questioning output link, lacks a multidimensional state monitoring mechanism for the whole questioning interaction flow, and cannot capture answer states (such as answer completeness, semantic fit degree and emotion fluctuation) of evaluation objects, environmental states (such as noise interference and light abnormality) of evaluation scenes and running states (such as data transmission delay and program running faults) of interaction processes in real time. The lack of state perception leads to the fact that the system can only push questions according to fixed logic, and the problem that the questions are not matched with the state of an evaluation object cannot be perceived, for example, continuous high-frequency questions of the evaluation object with tension emotion can aggravate contradiction emotion, responses to information deficiency cannot be perceived in time, and finally evaluation experience and data quality are affected. In the prior art, a 'one-cut' fixed scheduling mode is generally adopted, namely, all evaluation objects are applicable to the same question interval, speech speed and sequence, the response speed and cognitive level difference of different evaluation objects are not considered, and real-time state dynamic adjustment is not combined. The stiff scheduling mode has obvious defects that for an evaluation object with a slower response, the fixed rhythm can cause the evaluation object to fail to fully think about the response to generate information omission, for an evaluation object with a faster response, the too slow rhythm can reduce the evaluation efficiency, and meanwhile, in the face of sudden situations such as environmental interference, equipment jamming and the like, the influence cannot be avoided through rhythm adjustment, so that the process confusion is further aggravated. The prior system does not establish a linkage mechanism of 'state research judgment-inquiry triggering', and when the answer of an evaluation object has the problems of information deficiency, deviation theme, semantic ambiguity and the like, the inquiry cannot be accurately triggered based on the real-time state, and the follow-up inquiry can only be continuously advanced according to a preset list. Meanwhile, the joint between the inquiry and the original inquiry flow lacks planning, and even if the inquiry is triggered manually, problems such as mismatching of inquiry contents and missing information, mismatching of inquiry time and the like easily occur, and the comprehensiveness of an evaluation result is further influenced. In the face of abnormal scenes such as evaluation interruption, equipment failure, serious environmental interference and the like, the existing intelligent questioning system lacks a mature emergency processing mechanism, can not quickly locate abnormal nodes and backup evaluation data, and can not generate a targeted flow recovery scheme. In case of abnormality, the whole evaluation flow is often required to be restarted, so that the early evaluation data is lost, the evaluation period is prolonged, the evaluation efficiency is reduced, and the continuity and reliability of the evaluation are affected due to the interference emotion of the evaluation object caused by the restarting of the flow. In summary, the existing intelligent questioning technology has core defects of state monitoring deficiency, scheduling stiffness, questioning and inquiring connection fault, abnormal emergency deficiency and the like, essentially neglects the dynamic control requirement of the questioning execution whole flow, and cannot realize closed-loop management. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a questioning and scheduling method and system, The aim of the invention can be achieved by the following technical scheme: Comprising the following steps: s1, acquiring multisourc