CN-121998176-A - Creativity assessment situation controllable automatic generation method and system based on evolution optimization

CN121998176ACN 121998176 ACN121998176 ACN 121998176ACN-121998176-A

Abstract

The invention discloses a creativity assessment situation controllable automatic generation method and system based on evolution optimization, and belongs to the technical field of intelligent education. The method adopts a closed loop flow of layering constraint planning, monte Carlo sentence level generation, evolution variation screening and simulation evaluation feedback iteration, firstly plans a situation outline, regenerates a seed text, optimizes diversity and quality through Map-Elites, and combines virtual evaluation feedback iteration screening of a high-quality text. The system comprises a user request, a data loading module, a model generating module and an interaction module. The invention solves the problems of expert dependence, low efficiency and insufficient controllability of the existing situation generation, automatically generates the evaluation situation with complete structure and strong suitability, reduces the development cost and facilitates the large-scale standardized evaluation.

Inventors

QIAN HONG
WANG YIXUAN
GUO JIAJUN
DING YIFEI
WANG WENKAI
LIU ZHI
CHEN NAO
ZHOU AIMIN

Assignees

华东师范大学

Dates

Publication Date: 20260508
Application Date: 20260109

Claims (10)

1. The creativity assessment situation controllable automatic generation method based on evolution optimization is characterized by comprising the following steps of: Step 1, according to the appointed theme With predefined rule sets Hierarchical planning of test points and hyperlinking from a plurality of candidates Selecting the optimal one Refining is continued; Step 2, from Middle extraction decomposable node Expanding, wherein the nodes correspond to situation evaluation points to be refined and use a language model To select the node most suitable for expansion ; Step 3, repeating the step 1 and the step 2, performing multiple rounds of 'selection-expansion' operation on the hyperlinks in the situation supertree until all leaf nodes can not be further refined or reach a preset depth, constructing a complete situation outline structure, and passing through a language model Selecting an optimal hyperlink And taking the outline as a final planning outline; step 4, planning outline Construction of a State space for Monte Carlo Tree search for constraint And an action space Initializing search parameters; step 5, calling the text generation model in the expansion stage For the current state Generating candidate sets Performing quality evaluation on the candidate set by using a scoring device; step 6, in the backtracking stage, the quality scores of the candidate sentences are along the search track Returning, updating search parameters, and adjusting subsequent expansion depth according to the average value of the path; step 7, extracting behavior feature vectors for seed context text obtained by MCTS Calculating a fitness value, mapping the text and the characteristics thereof into a Map-Elites file structure, and storing or updating elite samples with optimal fitness in corresponding behavior positions; Step 8, based on mutation operator set The selected archive samples are mutated, new samples are generated, the fitness of the new samples is calculated, if the new samples are better than the existing samples in the corresponding behavior positions, the new samples are inserted or replaced according to rules, and the situation text diversity is improved; Step 9, constructing a student modeling and simulation answer module, inputting candidate situation texts in the files as stimulating materials, collecting answers and scores of simulated users on creative divergent thinking indexes, forming evaluation index vectors and feeding back to fitness functions and evolution processes; And 10, continuously optimizing the situation text generation and evolution flow according to the simulation evaluation feedback, gradually eliminating the situation text which is difficult to excite creative divergent thinking, and reserving high-quality situation text.
2. The method of claim 1, wherein in step 1, a given topic is With predefined rule sets In the pre-training language model Depth of reasoning Width and spread width Selecting an optimal context hyperlink under conditions including Initializing a supertree for a root node: , At each depth of layer The current supertree is mapped into a set of hyperlinks: , If the number of chains exceeds Scoring each super-chain by using a large model and keeping the super-chain before Strip: , selecting an optimal hyperlink therefrom And further planning and refining are carried out.
3. The method of claim 1, wherein in step 2, from an optimal hyperlinks Extracting decomposable nodes and expanding: Selecting a separable node set from candidate hyperlinks by a large model: , the relevant rules are then selected from the rule set and child nodes are generated: , 。 Finally, the child node is hung on the supertree: 。
4. The method of claim 1, wherein in step 3, the preset inference depth threshold is By at least one of Circularly executing the step 1 and the step 2 in the range, and supertree the situation Multiple rounds of selection-expansion operation are carried out on candidate hyperlinks in the tree, and when all leaf nodes can not be further refined or reach a preset depth threshold K, the cycle is terminated and a language model is utilized Comprehensively evaluating each candidate hyperlink in the supertree, and selecting the hyperlink with the highest score As a final context planning outline.
5. The method of claim 1, wherein in step 4, the search parameter comprises a node access count Status-action count Value estimation Search coefficient Sentence level generation logic for Monte Carlo tree search is to model a sequence of generated context text as a state Modeling a sentence or fragment to be selected as an action Node selection heuristic function in UCB form: , generating candidate sentences and calculating comprehensive scores by large language model in expansion stage Updated incrementally during the backtracking phase: , thereby giving consideration to both value evaluation and exploration efficiency.
6. The method of claim 1, wherein in step 6, the subsequent expansion depth is adjusted by adjusting the composite value of the candidate sentence along the search trajectory After layer-by-layer return, the average value of paths is compared And a preset threshold value When (1) Increasing the number of expansion steps at higher levels to enhance exploration of high value branches when Reducing the number of expansion steps at lower levels to reduce sampling overhead on low value branches, the expansion depth passing through a parameter Characterization.
7. The method of claim 1, wherein in step 7, the feature mapping function is passed through Extracting behavior feature vectors By a fitness function The storage rule of the Map-Elites archive structure is that only the elite sample with the highest fitness is stored or updated in the corresponding behavior lattice, so that the text quality is ensured, and meanwhile, the diversity coverage of the behavior space is kept.
8. The method of claim 1, wherein in step 8, the set of mutation operators Includes the operations of expanding writing, deleting and replacing, and the variation and replacement logic includes the steps of selecting file sample from Archive Performing mutation to generate new sample Calculate its fitness If the target behavior lattice is empty or For the current sample of the lattice, insert or use in Archive according to Archive replacement rules The existing sample is replaced, and the diversity of the situation text in the behavior space is improved on the premise of ensuring the fitness quality.
9. The method of claim 1, wherein in the step 9, the specific mode of simulating answering and scoring is that a large language model is called and a preset 'virtual student' prompting template is combined to perform batch simulation answering and scoring on candidate context texts in the Archive, the creative divergent thinking indexes comprise novelty, practicability and fluency, statistical mean values and distribution characteristics after scoring of the indexes are collected to form an evaluation index vector, and the evaluation index vector is used as an external feedback signal for correcting an adaptability function or updating evolution strategy parameters to promote the effectiveness of the candidate context texts to excite creative divergent thinking reactions; and/or the number of the groups of groups, In step 10, the specific logic of closed loop iterative optimization is to record the candidate situation set output by the Map-Elites module as an initial candidate pool In the first place In round iteration, candidate pool Each of the situations in (a) Calculating an average creativity score using a virtual assessment module In combination with a preset elimination threshold And maximum number of iterations Screening to obtain a retention set: , ; When (when) Or (b) Terminating the iteration, and For final high-quality situation set, otherwise for elimination set And generating new candidates by applying mutation operators and updating a candidate pool, and cycling the process.
10. An evolutionary optimization-based creativity assessment context-controllable automatic generation system, characterized in that the system implements the method of any of claims 1-9, the system comprising a user request module, a data loading module, a model generation module, and a user interaction module, wherein, The user request module is responsible for processing a request sent by a user and processing subject data contained in a request message body; the data loading module is used for representing according to standardized subjects Retrieving and loading rules and background knowledge related to the topic from a rule base, a knowledge base and a corpus, preprocessing the rules and background knowledge, and then taking the rules and background knowledge as topic constraint data generated subsequently; The model generation module is used for calling a large language model to execute hierarchical task planning, monte Carlo tree search generation, behavior space mapping and evolution screening and virtual evaluation feedback iterative optimization under the common constraint of the topic information and the topic constraint data, so as to generate a complex context text set matched with the current topic; The user interaction module presents the generated context to the user, where the user performs creative assessment, answers related content, and optionally receives user feedback for optimization of subsequent context generation effects.

Description

Creativity assessment situation controllable automatic generation method and system based on evolution optimization Technical Field The invention belongs to the technical field of intelligent education, and relates to a creativity assessment situation controllable automatic generation method and system based on evolution optimization. Background In the education field, creativity assessment is an important topic, and researchers in the related field propose numerous methods and indexes [1] for creativity assessment. In the contextualized evaluation process, an evaluation situation which is complex and is close to a real application scene is built, and the method has important significance [2] for efficient and effective creativity evaluation. In the traditional mode, the creativity assessment process adopts experts to evaluate most, takes a questionnaire form as a main, a user answers in a text form, data are manually collected, and each questionnaire is scored by the experts to serve as creativity potential scoring results [3] of the user. The method takes abstract characters as the main material, does not construct a specific scene related to real life, causes the disjoint of the evaluation scene and the actual application scene of creativity, lacks the 'substitution sense' and the 'participation sense' required by immersive experience, is easy to cause the problems of application of a derivative answer, false answer and the like, directly weakens the authenticity and the effectiveness of an evaluation result, and is difficult to accurately reflect the actual creativity level [4] of a tested person. The existing contextualized creativity assessment situation generation mainly has two forms, namely, the first is expert dominant situation design [5] represented by PISA. The core logic is based on the deep understanding of creativity by experts, and the targeted evaluation situation is designed in a fit way. The method can ensure the precise excitation of the situation on creativity, the scientificity and the strictness [6] of the evaluation process, and meanwhile, the research and development process strictly follows the standard flow of psychological evaluation, thereby ensuring the credibility and the effectiveness of the evaluation tool. However, the form has obvious limitation, the situation design is highly dependent on expert experience and expertise, the generation period is long, the efficiency is low, the batch requirement of large-scale evaluation is difficult to adapt quickly, and the suitability is insufficient when the method is used for coping with flexible scenes such as diversified evaluation subjects, differentiated evaluation objects and the like. The second is a method [7-9] of autonomously generating by large models, such as SS_GEN. The core of the method is based on the autonomous generation capability of a large language model, and the model is guided to directly output a complete evaluation situation by constructing a situation text example, defining a situation text structure specification, limiting an application scene boundary and other accurate prompt word designs. The method has the advantages that the situation generation flow is greatly simplified, the situation output efficiency is remarkably improved by means of the large-scale generation advantage of the model, wider situation topics can be covered by depending on the generalization capability of the model, and the application scene of contextualized evaluation is further expanded. However, the form still faces the problems that the controllability of situation generation is insufficient, and the generated text is difficult to completely meet the strict requirements of the situational creativity assessment in terms of structural integrity, logical self-consistency, assessment target adaptation and the like. Disclosure of Invention In order to solve the defects existing in the prior art, the invention aims to provide a creativity assessment situation controllable automatic generation method and system based on evolution optimization, and solves the problems of lack of creativity assessment situation and complex design in the education field by using a large language model and an evolution optimization algorithm. The method comprises the steps of firstly, conducting element and constraint structured hierarchical planning on a given subject by a subject hierarchical constraint planning module, outputting a situation outline comprising a field scale, key constraints and challenge clues, conducting sentence level progressive generation by adopting Monte Carlo tree search on the basis of the situation outline, conducting quality assessment by using multidimensional indexes to obtain a group of initial seed texts with complete structure and logical continuity, then conducting constraint variation and screening on the seed texts by introducing diversity-quality joint optimization based on elite files (Map-Elites), forming elite file