A dynamic assessment system and method for elementary school writing level

Abstract

The application belongs to the technical field of data analysis, and discloses a primary school student writing level dynamic evaluation system and method, which comprises a data analysis unit, receives composition text data to be evaluated, and performs semantic unit decomposition to form a semantic unit sequence; performs context dependency analysis on the semantic unit sequence to generate a semantic representation vector; a semantic anchoring unit takes the semantic representation vector as input, performs semantic node matching in a pre-constructed cultural background knowledge graph to obtain a first anchoring result; takes the first anchoring result as a constraint condition, performs ability node reasoning in a preset student ability target graph to generate a second anchoring result; a causal structure modeling unit takes the first anchoring result and the second anchoring result as constraints, performs causal relationship extraction on the semantic unit sequence, and constructs a directed acyclic causal graph; dynamic, interpretable and quantifiable evaluation of the writing level of students is realized.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and more specifically, to a dynamic assessment system and method for primary school students' writing skills. Background Technology

[0002] In primary school Chinese language teaching, composition assessment is a core component for evaluating students' language expression, logical thinking, and knowledge application abilities. Its scientific rigor and accuracy directly impact the efficiency of teaching feedback and the effectiveness of improving students' writing skills. With the development of educational informatization, traditional manual assessment, due to its low efficiency and strong subjectivity, is gradually being replaced by automated composition assessment systems. However, existing technologies still face many key technical bottlenecks in practical applications, making it difficult to meet the needs of refined teaching.

[0003] Existing dynamic assessment systems and methods for writing proficiency mainly suffer from the following problems:

[0004] The inability to effectively distinguish between "causal information" and "decorative information" in the text, with assessments often remaining at the level of keyword matching or semantic similarity, makes it impossible to determine which semantic units truly drive the evaluation results and which are merely irrelevant or weakly related components. This results in the inability to locate the core semantic units that drive the assessment of writing ability, leading to a lack of causal explanatory power in the evaluation results, and making it difficult for teachers and students to identify the root causes of their ability shortcomings.

[0005] Existing essay assessment technologies typically only analyze the surface features of the text or score its overall quality. They lack the ability to break down the causal hierarchy of the mechanisms that lead to writing quality, cannot pinpoint "which content factors" influence "which abilities" through the causal propagation chain, and cannot quantitatively explain the true contribution of different factors to the final ability assessment results. This results in assessments that are unexplainable, uninterventional, and unusable for learning guidance.

[0006] It is difficult to systematically characterize the causal relationship between semantic factors and writing ability dimensions, and it is impossible to identify the complete causal link from semantic nodes to evaluation target nodes. It is also impossible to quantify the specific impact of each semantic factor. It is impossible to trace "which semantic chain caused the decline in ability" or quantify the intensity of the impact of each semantic factor on ability. As a result, the output results are only "conclusions" rather than "conclusions with paths and reasons". This leads to insufficient targeting of the assessment results and makes it difficult to support personalized teaching guidance and precise improvement of students' writing ability.

[0007] In view of this, the present invention proposes a dynamic assessment system and method for primary school students' writing level to solve the above problems. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the existing technology and to achieve the above objectives, the present invention provides the following technical solution: a dynamic assessment system for primary school students' writing skills, comprising:

[0009] The data parsing unit receives the essay text data to be evaluated and decomposes it into semantic units to form a semantic unit sequence; it then performs context dependency analysis on the semantic unit sequence to generate a semantic representation vector.

[0010] The semantic anchoring unit takes the semantic representation vector as input, performs semantic node matching in a pre-constructed cultural background knowledge graph to obtain the first anchoring result, and performs ability node reasoning in a pre-set student ability target graph to generate the second anchoring result.

[0011] The causal structure modeling unit uses the first and second anchoring results as constraints to extract causal relationships from the semantic unit sequence and construct a directed acyclic causal graph.

[0012] The causal path identification unit identifies all causal paths in the directed acyclic causal graph, calculates the comprehensive causal strength of each path, filters out key causal paths, and generates counterfactual text based on the key causal paths.

[0013] The reasoning and analysis unit performs reasoning and comparative analysis on the essay text data and counterfactual text in a directed acyclic causal graph. Based on the results of the comparative analysis, a writing level assessment report is generated.

[0014] Preferably, the method for obtaining the semantic representation vector includes:

[0015] Upload images of students' handwritten essays to be evaluated and perform optical character recognition to obtain essay text data; then, by performing lexical segmentation, syntactic structure recognition, and semantic role labeling on the essay text data, the essay text data is decomposed into a sequence of semantic units containing word-level, phrase-level, and sentence-level semantic units.

[0016] Context dependency modeling is performed on the semantic unit sequence to capture the referential relationships, logical connections, and text connection clues between semantic units, thereby obtaining context dependency representations. The obtained context dependency representations are then input into a preset semantic projection network for feature fusion and compression mapping to generate semantic representation vectors.

[0017] Preferably, the method for obtaining the first anchoring result includes:

[0018] The semantic representation vector is input into a pre-constructed cultural background knowledge graph, and node retrieval and matching are performed. The pre-constructed cultural background knowledge graph consists of semantic nodes and relation edges.

[0019] By calculating the cosine similarity between the semantic representation vector and each semantic node in the cultural background knowledge graph, the semantic node with the highest cosine similarity is identified as the semantic node that matches the semantic representation vector; all matched semantic nodes are collected to form a candidate node set; the semantic neighborhood nodes associated with the candidate node set in the cultural background knowledge graph are extracted and integrated to form a semantic anchoring subgraph, thereby obtaining the first anchoring result.

[0020] Preferably, the method for obtaining the second anchoring result includes:

[0021] The pre-defined student competency target graph consists of competency nodes and competency-related edges. The semantic anchoring subgraph corresponding to the first anchoring result is used as a semantic constraint. For each competency node in the student competency target graph, the importance of the competency node is weighted and aggregated based on the consistency between the weight of the related edges of the corresponding neighboring competency nodes and the semantic constraint, thereby obtaining the importance value of the competency node. The importance values ​​of all competency nodes are sorted, and competency nodes with importance values ​​higher than the preset importance matching threshold are selected to form a candidate competency node set. The candidate competency node set and the corresponding neighboring competency nodes are extracted to form a competency anchoring subgraph, forming the second anchoring result.

[0022] Preferably, the method for constructing a directed acyclic causal graph includes:

[0023] Using the first anchoring result as a semantic constraint and the second anchoring result as a capability constraint, causal relationship extraction and directed acyclic causal graph construction are performed on the semantic unit sequence corresponding to the essay text data. The first anchoring result is used to limit the scope of semantic units that can participate in causal relationship extraction, and only semantic units within the coverage of the first anchoring result are processed. The second anchoring result is used to limit the capability association scope, and causal relationship identification is performed only on semantic units that are associated with the capability nodes corresponding to the capability anchoring subgraph.

[0024] Under the dual constraints of semantics and capability, explicit or implicit causal relationships between semantic units are detected. When a causal relationship is detected, the two semantic units involved in the causal relationship are used as causal nodes, and directed dependency edges are established between the causal nodes, thereby gradually forming a semantic causal graph structure. An acyclicity constraint is applied to the formed semantic causal graph structure to obtain a directed acyclic causal graph.

[0025] Preferably, the method for screening key causal paths includes:

[0026] In a directed acyclic causal graph, each causal node is used as the starting node to initialize the path search, and the path set is initialized to empty. The starting causal node is added to the current path, and it is checked whether it is the target capability node. If it is, the current path is recorded in the path set. If not, a recursive search is performed along all the directed dependent edges of the starting causal node. Each recursion adds the corresponding visited adjacent causal nodes to the current path, and continues to advance towards the target capability node. After all the directed dependent edges of the starting causal node have been recursively searched, all directed paths from the starting causal node to the target capability node are obtained, forming a path set.

[0027] For each path in the path set, based on the intervention effect of each causal node in the path, the structural causal model is used to perform intervention calculations, calculate the average causal effect of each directed dependency edge in the path, and weight it according to the preset weights to obtain the comprehensive causal strength of each path; the path set is sorted according to the comprehensive causal strength, and the paths with a comprehensive causal strength greater than or equal to the preset comprehensive causal strength threshold are selected as key causal paths.

[0028] Preferably, the method for obtaining the counterfactual text includes:

[0029] The initial causal node in the key causal path is identified as the counterfactual intervention object; the corresponding semantic unit in the essay text data is explicitly locatable; for the counterfactual intervention object, an alternative expression that is opposite to or zeroed out to the original semantic expression is constructed; the semantic unit corresponding to the counterfactual intervention object is located in the original essay text data and rewritten through the alternative expression, thereby generating a counterfactual text containing counterfactual premises.

[0030] Preferably, the method for reasoning and performing comparative analysis on the essay text data and counterfactual text separately includes:

[0031] Using essay text data and generated counterfactual text as input, reasoning is performed node by node along key causal paths in a directed acyclic causal graph. The reasoning results are mapped to the ability dimensions in the student ability target map to perform comparative analysis, generating quantitative scores for essay text data and counterfactual text on each ability dimension, and forming an ability deviation matrix.

[0032] Preferably, the method for obtaining the writing level assessment report includes:

[0033] By analyzing the differences in different ability dimensions caused by text variations using an ability deviation matrix, the core semantic units affecting writing level are identified, and causal attribution explanations are generated. The ability deviation matrix and causal attribution explanations are then integrated to generate a writing level assessment report.

[0034] A dynamic assessment method for primary school students' writing skills includes:

[0035] S1. Receive the essay text data to be evaluated and decompose it into semantic units to form a semantic unit sequence; perform context dependency analysis on the semantic unit sequence to generate a semantic representation vector;

[0036] S2. Using the semantic representation vector as input, perform semantic node matching in the pre-constructed cultural background knowledge graph to obtain the first anchoring result; using the first anchoring result as a constraint, perform ability node reasoning in the preset student ability target graph to generate the second anchoring result.

[0037] S3. Using the first and second anchoring results as constraints, perform causal relationship extraction on the semantic unit sequence and construct a directed acyclic causal graph;

[0038] S4. Identify all causal paths in the directed acyclic causal graph, calculate the comprehensive causal strength of each path, filter out key causal paths, and generate counterfactual text based on the key causal paths.

[0039] S5. In the directed acyclic causal graph, reasoning is performed on the essay text data and counterfactual text data respectively, and a comparative analysis is performed. Based on the results of the comparative analysis, a writing level assessment report is generated.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] By enumerating all directed paths from semantic nodes to capability nodes in a directed acyclic causal graph and selecting paths with sufficient transmission effects, we can accurately pinpoint "which semantic component, along which causal chain," affects writing ability, solving the problem of existing methods that only provide conclusions without tracing the causes. By calculating and weighting the intervention effects of causal relationships within the paths, we obtain quantifiable and comparable path-level comprehensive strength, enabling the ranking and selection of influencing factors and improving the accuracy and robustness of causal explanations. Key causal paths also serve as a constraint source for counterfactual interventions. By performing simulated interventions at the path's starting point and constructing counterfactual texts, we can verify the direction and magnitude of changes in capability output, transforming assessment from "static diagnosis" to "operable verification," providing a basis for personalized training.

[0042] Compared to arbitrary text replacement, counterfactual intervention targets only the starting nodes of critical paths, ensuring that the rewriting has genuine contribution to ability. Alternative expressions are constructed under conditions of clear semantic positioning, syntactic feasibility, and consistency with the context, ensuring the assessment results possess explanatory credibility and reproducibility. The final output is not an abstract score, but a structured explanation with causal chains and contribution strength criteria, presenting a closed-loop logic of "causal path—semantic components—ability differences." This upgrades the assessment system from "static scoring" to "intervention-based dynamic assessment," enhancing its transparency, pedagogical usability, and learning guidance value. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the structure of a dynamic assessment system for primary school students' writing skills according to the present invention;

[0044] Figure 2 This is a schematic diagram of a dynamic assessment method for primary school students' writing skills according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1

[0047] Please see Figure 1 As shown in the figure, this embodiment provides a dynamic assessment system for primary school students' writing skills, which specifically includes the following steps:

[0048] The data parsing unit receives the essay text data to be evaluated and decomposes it into semantic units to form a semantic unit sequence; it then performs context dependency analysis on the semantic unit sequence to generate a semantic representation vector.

[0049] The semantic anchoring unit takes the semantic representation vector as input, performs semantic node matching in a pre-constructed cultural background knowledge graph to obtain the first anchoring result, and performs ability node reasoning in a pre-set student ability target graph to generate the second anchoring result.

[0050] The causal structure modeling unit uses the first and second anchoring results as constraints to extract causal relationships from the semantic unit sequence and construct a directed acyclic causal graph.

[0051] The causal path identification unit identifies all causal paths in the directed acyclic causal graph, calculates the comprehensive causal strength of each path, filters out key causal paths, and generates counterfactual text based on the key causal paths.

[0052] The reasoning and analysis unit performs reasoning and comparative analysis on the essay text data and counterfactual text in a directed acyclic causal graph. Based on the results of the comparative analysis, a writing level assessment report is generated.

[0053] Methods for obtaining semantic representation vectors include:

[0054] Upload images of students' handwritten essays to be evaluated and perform optical character recognition (OCR), including noise filtering (removing garbled characters, redundant spaces, or invalid symbols caused by OCR errors) and format standardization (such as standardizing punctuation and correcting typos), to obtain essay text data; then, by performing lexical segmentation, syntactic structure recognition, and semantic role labeling on the essay text data, the essay text data is decomposed into a sequence of semantic units containing word-level, phrase-level, and sentence-level semantic units;

[0055] For example, the sentence "Because it was cold, I didn't go out" can be parsed to generate semantic units: "Because √ (causal marker) / Weather (environmental conditions) / Cold (property description) / So √ (causal sequence) / I (agent) / Did not go out (action and result)", thus making semantic roles and discourse logic explicit. Among them, "agent" is a term in semantic role annotation, corresponding to the English "Agent", which refers to the participant who actually initiates the action and actively performs the behavior in the event or action described in the sentence. It has two key characteristics: first, it is the initiator of the behavior; second, the action is actively performed by it, rather than being passively generated by it.

[0056] Context dependency modeling is performed on semantic unit sequences to capture referential relationships, logical connections, and discourse connection clues between semantic units, thereby obtaining context dependency representations. For example, when pronouns or appositive structures such as "he," "this matter," and "the above reasons" appear in the semantic unit sequence, the system will automatically align them with the corresponding entities in the preceding text and identify conjunctions that indicate progression (such as "in addition" and "and"), contrast (such as "but" and "however"), and causality (such as "therefore" and "so") in discourse logic, thereby forming context dependency representations that reflect inter-sentence dependencies and logical connections.

[0057] The acquired contextual dependency representations are input into a pre-defined semantic projection network for feature fusion and compression mapping, generating semantic representation vectors that characterize the semantic structure and textual relationships of the essay. The pre-defined semantic projection network is a pre-trained semantic representation model, trained on a training dataset constructed from textbook texts, reading comprehension data, and standard essay corpora. This model learns how semantic features at different levels are fused in the vector space, thereby obtaining semantic representation vectors suitable for subsequent semantic anchoring and causal structure modeling.

[0058] The pre-defined semantic projection network includes an input encoding layer, a semantic feature mapping layer, a semantic fusion layer, and a vector compression mapping layer. First, a training dataset is constructed based on textbook texts, reading comprehension data, and standard essay corpora. Second, the training dataset is input into the pre-defined semantic projection network, and forward propagation is performed sequentially through the input encoding layer, semantic feature mapping layer, semantic fusion layer, and vector compression mapping layer to obtain the corresponding semantic representation vectors. Simultaneously, contrastive loss is used as the training loss function to measure the deviation between the generated semantic representation vectors and the target semantic representation, thus obtaining the training loss value.

[0059] Based on the training loss value, the parameters of each layer in the preset semantic projection network are updated through backpropagation. Specifically: the Adam optimizer is used to update the parameters, with the initial learning rate set to 0.001, which is decayed to 0.8 times the original rate every 10 iterations; the gradient of each layer parameter is calculated through backpropagation, and all trainable parameters are updated based on gradient descent to minimize the training loss value; the above steps are repeated until the training loss value meets the preset convergence condition (minimizing the training loss value), and the trained preset semantic projection network is obtained.

[0060] The methods for obtaining the first anchoring result include:

[0061] The semantic representation vector is input into a pre-constructed cultural background knowledge graph, and node retrieval and matching are performed. The pre-constructed cultural background knowledge graph consists of semantic nodes and relation edges. Each semantic node consists of a corresponding node semantic vector, which is used to represent the semantic features of cultural elements, writing themes, semantic images and rhetorical patterns in the semantic vector space. Relation edges are used to represent cultural associations between semantic nodes (such as the cultural co-occurrence relationship between "Spring Festival" and "reunion"), logical deduction (such as the logical implication between "ambition" and "success") or semantic inheritance relationships (such as "parallelism" belonging to the expression strategy of "emphasis").

[0062] By calculating the cosine similarity between the semantic representation vector and each semantic node in the cultural background knowledge graph, the semantic node with the highest cosine similarity is identified as the semantic node that matches the semantic representation vector; all matched semantic nodes are collected to form a candidate node set; the semantic neighborhood nodes associated with the candidate node set in the cultural background knowledge graph are extracted and integrated to form a semantic anchoring subgraph, thereby obtaining the first anchoring result.

[0063] The semantic anchoring subgraph consists of a candidate node set and its corresponding semantic neighborhood nodes. The semantic neighborhood nodes are semantic nodes that are directly connected to the candidate node set through relation edges. The two types of nodes form a locally closed semantic association structure through relation edges. The semantic anchoring subgraph is used to limit the landing point and semantic scope of the semantic representation vector in the cultural background knowledge graph. By preserving the semantic association structure between the candidate node set and its corresponding semantic neighborhood nodes, subsequent association reasoning and causal node extraction are completed within the local effective semantic domain, thereby avoiding semantic dilution and mismatch caused by full graph reasoning and improving the stability and interpretability of the anchoring results.

[0064] The methods for obtaining the second anchoring result include:

[0065] The pre-designed student competency target map consists of competency nodes and competency-related edges. The competency nodes are used to represent writing competency elements, such as language expression, text structure, logical causality, emotional expression, cultural fit, and creative expression. The competency-related edges are used to represent the transformation dependence, hierarchical evolution, or mutual constraint relationships between different competency nodes, forming a competency representation space with structural hierarchy and logical orientation.

[0066] Using the semantic anchoring subgraph corresponding to the first anchoring result as a semantic constraint, for each ability node in the student ability target graph, the importance of the ability node is weighted and aggregated based on the consistency between the weight of the associated edges of the corresponding neighboring ability nodes and the semantic constraint, thereby obtaining the importance value of the ability node. ; This represents the ability nodes in the student ability target map; Represents capability nodes Importance value; This represents the damping coefficient, which can be set manually based on expert experience. The value is between 0 and 1 and is used to suppress unstable oscillations or excessively rapid accumulation and amplification of importance values ​​during iterative updates, making the convergence of importance values ​​smoother. Representation and capability nodes Associated neighborhood capability nodes; Represents capability nodes The set of nodes with adjacency capability; Represents neighborhood capability nodes Importance value; Indicates from point to The weight of the associated edge (reflecting the strength of the association between the two). Indicates the measurement of associated edges With semantic anchoring subgraph The consistency function can be obtained by analyzing the semantic content similarity between the associated edges and the corresponding paths in the semantic anchoring subgraph; Represents a semantic anchoring subgraph; Represents capability nodes With neighboring capability nodes The associated edges between them;

[0067] Sort all capability nodes by importance value, select capability nodes with importance value higher than the preset importance matching threshold to form a candidate capability node set; extract the candidate capability node set and the corresponding neighboring capability nodes to form a capability anchoring subgraph, forming the second anchoring result.

[0068] Methods for constructing directed acyclic causal graphs include:

[0069] Using the first anchoring result as a semantic constraint and the second anchoring result as a capability constraint, causal relationship extraction and directed acyclic causal graph construction are performed on the semantic unit sequence corresponding to the essay text data. The first anchoring result is used to limit the scope of semantic units that can participate in causal relationship extraction, and only semantic units within the coverage of the first anchoring result are processed. The second anchoring result is used to limit the capability association scope, and only semantic units that are associated with the capability nodes corresponding to the capability anchoring subgraph are used to perform causal relationship identification, thereby avoiding causal modeling of text fragments that are unrelated to writing ability.

[0070] Under the dual constraints of semantics and ability, the explicit or implicit causal relationship between semantic units is detected. When a causal relationship is detected, the two semantic units involved in the causal relationship are used as causal nodes, and directed dependency edges are established between the causal nodes, thereby gradually forming a semantic causal graph structure. The acyclicity constraint is applied to the formed semantic causal graph structure, and then a directed acyclic causal graph reflecting the logical chain of the composition content and the logic of the ability association is obtained.

[0071] Explicit causality includes structures with connectors such as "because...therefore...", "due to...leading to...", and "therefore..."; implicit causality includes logical recursive relationships such as action-result (causal connection), premise-conclusion (logical implication), and purpose-measure (sequential structure).

[0072] The specific implementation method for applying acyclicity constraints to the formed semantic causal graph structure is as follows: To ensure that the semantic causal graph structure is a directed acyclic structure, loop detection is performed every time a new directed dependency edge is added: By performing a depth-first search (DFS) on the current semantic causal graph structure, it is determined whether the addition of a new directed dependency edge leads to a closed loop in the semantic causal graph structure (it can be traced back to the origin from the target causal node through the existing directed dependency edges); if a closed loop is detected, the addition of the corresponding new directed dependency edge is rejected; after all directed dependency edges are established, a directed acyclic causal graph is obtained, while retaining the original logical chain of the essay and the semantic association of ability defined by the second anchoring result, thus providing a structured foundation for subsequent writing ability assessment.

[0073] Methods for identifying key causal pathways include:

[0074] In a directed acyclic causal graph, each causal node is used as the starting node to initialize the path search, and the path set is initialized to empty. The starting causal node is added to the current path, and it is checked whether it is the target capability node. If it is, the current path is recorded in the path set. If not, a recursive search is performed along all the directed dependent edges of the starting causal node. Each recursion adds the corresponding visited adjacent causal nodes to the current path, and continues to advance towards the target capability node. After all the directed dependent edges of the starting causal node have been recursively searched, all directed paths from the starting causal node to the target capability node are obtained, forming a path set.

[0075] For each path in the path set, based on the intervention effect of each causal node in the path, the structural causal model is used to perform intervention calculations, calculate the average causal effect of each directed dependency edge in the path, and weight it according to the preset weights to obtain the comprehensive causal strength of each path; the path set is sorted according to the comprehensive causal strength, and the paths with a comprehensive causal strength greater than or equal to the preset comprehensive causal strength threshold are selected as key causal paths.

[0076] It should be noted that the specific intervention calculation process is as follows: by applying human intervention to the initial causal node, different intervention situations are constructed respectively, and the expected output difference of the target capability node under different intervention situations is calculated, and then the average causal effect of each directed dependency edge in the path is obtained.

[0077] The average causal effect is: ;in, Indicates the causal node from the beginning To the target capability node The average causal effect, representing the intervention In the case of The expected change; This represents the initial causal node in a directed acyclic causal graph. This represents the target capability node in a directed acyclic causal graph, and represents the writing capability dimension (e.g., "logical coherence"). This represents the mathematical expectation operation, which calculates the average value of a random variable under given conditions. Indicates intervention settings, will The value is manually set to 1 to indicate the state of the causal node; Indicates intervention settings, will The value was set to 0 to control the calculation of causal effects;

[0078] For example, a certain essay contains the following causal chain expression:

[0079] "Because I kept a diary for a long time, my language expression ability gradually improved, and I eventually won the first prize in the essay competition."

[0080] One of the following directed acyclic causal image segments has been constructed through the preceding steps:

[0081] C1: "Keep writing a diary" → C2: "Improve language expression" → C3: "Win an award for writing"; where, the ability node corresponding to C3 = "Language expression ability" + "Cultural fit / expression effectiveness";

[0082] Initialize the search using C1 as the starting node of the path:

[0083] The current path [C1] -- C1 is not the target capability node, continue the recursive search along the directed dependency edges;

[0084] Recursive search enters C2:

[0085] The current path [C1→C2] -- C2 is not the target capability node, continue the recursive search;

[0086] Recursive search enters C3:

[0087] Current path: [C1→C2→C3] -- C3 corresponds to the ability node (related to writing ability). The recursive search ends, and the current path is added to the path set to form the final complete path set {P1=[C1→C2→C3]}.

[0088] Causal effect calculations were performed on the path. For the two directed edges in path P1: C1→C2 (the effect of consistent diary writing on improved language expression ability) and C2→C3 (the effect of improved language expression ability on winning essay competitions), intervention calculations were performed:

[0089] ACE(C1→C2)=E[C2|do(C1=1)]−E[C2|do(C1=0)];

[0090] ACE(C2→C3)=E[C3|do(C2=1)]−E[C3|do(C2=0)];

[0091] Assume that C1 = whether or not one persists in writing a diary (1 = persist, 0 = not persist), C2 = language expression ability score (normalized from 0 to 1); C3 = probability of winning an essay competition (values ​​from 0 to 1). Using a structural causal model based on historical experience, we obtain the expected value of the language expression ability score, E[C2|do(C1=1)] = 0.72 under the intervention setting C1 = 1 (mandatory diary writing); and the expected value of the language expression ability score, E[C2|do(C1=0)] = 0.55 under the intervention setting C1 = 0 (prohibition of diary writing). Therefore, the average causal effect ACE(C1→C2) = 0.72 - 0.55 = 0.17. Explanation: Persisting in writing a diary improves language expression ability by an average of 0.17 (on a normalized scale).

[0092] When the intervention setting C2=1 (forced improvement in language expression ability), the expected value of the probability of winning the essay competition is E[C3|do(C2=1)]=0.41; when the intervention setting C2=0 (forced decrease in language expression ability), the expected value of the probability of winning the essay competition is E[C3|do(C2=0)]=0.19; therefore, the average causal effect ACE(C2→C3)=0.41-0.19=0.22; Explanation: The average increase in the probability of winning the essay competition due to improved language expression ability is 0.22.

[0093] In summary, the path set {P1=[C1→C2→C3]} shows that "keeping a diary" will improve "language expression ability," which in turn will affect "the result of winning the essay competition." The impact of keeping a diary can be explained by a chain causal effect, where the improvement in language expression ability (0.17) is then passed on to the improvement in the probability of winning the essay competition (0.22).

[0094] This approach addresses the following technical problems of existing technologies: Current methods for evaluating elementary school students' compositions typically rely solely on feature extraction or statistical indicators, making it difficult to systematically characterize the causal relationship between semantic factors and writing ability dimensions within the composition text, and failing to identify the complete causal path from semantic nodes to the target node of the writing evaluation. Traditional methods, when assessing abilities, often depend on simple weighting or scoring rules, failing to quantify the specific impact of each semantic factor on the ability dimension, and also failing to identify key factors. Existing technologies cannot distinguish which semantic nodes or paths have a significant impact on the writing ability dimension based on causal effects, resulting in evaluation results lacking specificity and interpretability.

[0095] Compared to existing technologies, the advantages are as follows: By enumerating all directed paths from semantic nodes to capability nodes in a directed acyclic causal graph and filtering out paths with sufficient transitive effects, it is possible to accurately identify "which semantic factor, along which logical chain," ultimately leads to an increase or decrease in capability performance, thus solving the problem that existing assessments can only see the results but cannot trace the causes; by calculating and weighting the intervention effects of causal relationships within the paths, comparable path-level comprehensive strengths can be obtained, making the causes affecting capability performance not only "visible" but also "quantifiable and ranked," thereby supporting a critical path screening mechanism based on strength thresholds and improving the accuracy and robustness of causal explanations;

[0096] Key causal paths are the foundation for constructing counterfactual texts and implementing intervention assessments. By conducting simulated interventions at the starting points of these paths, the direction and magnitude of changes in ability levels after modifying essay content can be verified. This transforms assessment from a "diagnostic" approach to an "operable trial-and-error guidance" approach, providing an actionable basis for personalized writing training. It also enhances the interpretability and pedagogical usability of the assessment. The final output is not an abstract model result, but a set of key paths with clear causal chains and contribution strength criteria. This can intuitively present the logical closed loop of "causal chain - semantic unit - ability difference," improving the transparency, credibility, and practical value of the assessment system.

[0097] Methods for obtaining counterfactual texts include:

[0098] The initial causal node in the key causal path is identified as the counterfactual intervention object. The counterfactual intervention object corresponds to the semantic unit that can be explicitly located in the essay text data, such as the action event "persist in writing a diary", the psychological state "worried about failure", and the environmental condition "bad weather". For the counterfactual intervention object, an alternative expression that is opposite to or zeroed out to the original semantic expression is constructed. The semantic unit corresponding to the counterfactual intervention object is located in the original essay text data and rewritten through the alternative expression, thereby generating a counterfactual text containing counterfactual premises.

[0099] For example, replacing "it's very cold" with "it's warm" corresponds to a reverse environmental intervention; replacing "emotionally tense" with "no tense feelings" corresponds to an emotion suppression intervention.

[0100] This solution addresses the following technical problems of existing technologies: Existing essay assessment systems cannot distinguish between "causal information" and "decorative information" in text content, often relying on keyword matching or semantic similarity, and cannot determine which semantic units truly drive the assessment of writing ability, resulting in a lack of causal explanatory power in the assessment; existing methods evaluate only from static text, and cannot test whether the level would change "without a certain semantic factor," nor can they construct a comparative baseline relative to the original text, thus failing to determine whether a certain semantic component is a decisive factor in the assessment of writing ability; they cannot achieve explicit localization and interventionizable generation of "ability-related key semantics," cannot locate operable sources of ability contribution (such as psychological state, behavior, and background conditions) from the written sentences, and cannot construct scientifically reasonable counterfactual experimental conditions.

[0101] Compared to existing technologies, the advantages are as follows: By locating causal paths, the "interpretable" assessment of writing ability is achieved. The counterfactual intervention target is limited to the starting node of the key causal path, ensuring that the rewritten object has real significance in terms of ability impact, rather than arbitrary sentence modification, so that the assessment results have a structured interpretive basis; by constructing counterfactual text, the "verifiable" assessment of ability contribution is achieved. Semantic units are replaced with opposite or zero-valued expressions to construct a control text, so that the system can compare the difference in ability output "before and after intervention", thereby verifying whether the semantic component is a core factor in ability contribution.

[0102] This approach upgrades from "static text scoring" to "intervention-based dynamic assessment," moving beyond simply grading original essays to generating counterfactual data to identify ability biases and attributions, thus providing a feasible path for instructional guidance and ability diagnosis. The counterfactual generation retains semantic consistency constraints, enhancing the credibility and interpretability of the results. The generated counterfactual texts are not randomly rewritten but are completed under conditions of semantic location, syntactic plausibility, and no distortion of the context, ensuring that the assessment conclusions are acceptable and reproducible for both teachers and students.

[0103] Methods for reasoning and performing comparative analysis on essay text data and counterfactual texts separately include:

[0104] Using essay text data and generated counterfactual text as input, reasoning is performed node by node along key causal paths in a directed acyclic causal graph. The reasoning results are mapped to the ability dimensions in the student ability target map to perform comparative analysis, generating quantitative scores for essay text data and counterfactual text on each ability dimension, and forming an ability deviation matrix.

[0105] Specific implementation steps: Receive the original essay text, for example, a student's sentence: "Xiaoming gets up on time every morning to go to school." Simultaneously, generate a corresponding counterfactual text, such as: "Xiaoming does not get up on time every morning to go to school." Input the original essay text and the counterfactual text into a directed acyclic causal graph, and reason node by node along the key causal path to determine the contribution strength of each semantic unit to the ability node; for example: in the original text, the node "gets up on time" contributes 0.9 to the "time management ability node" and 0.8 to the "sense of responsibility node"; in the counterfactual text, the node "does not get up on time" reduces its contribution to the "time management ability node" to 0.3 and its contribution to the "sense of responsibility node" to 0.4.

[0106] The inference results are mapped to ability dimensions in the student ability target map, such as "time management ability," "sense of responsibility," and "planning execution ability." A quantitative score for each ability dimension is calculated based on the node activation intensity.

[0107] Original text: Time management ability = 0.85, sense of responsibility = 0.80, plan execution ability = 0.75;

[0108] Counterfactual text: Time management ability = 0.30, sense of responsibility = 0.40, plan execution ability = 0.50;

[0109] A capability bias matrix is ​​generated by comparing the capability scores of the original text and the counterfactual text. For example:

[0110] Table 1 Ability Deviation Matrix

[0111] Capability Dimension Original text score Counterfactual text score Difference Time management skills 0.85 0.30 0.55 Sense of responsibility 0.80 0.40 0.40 Plan execution capability 0.75 0.50 0.25

[0112] The ability deviation matrix can be used to intuitively analyze the differences in different ability dimensions caused by text changes, providing a quantitative basis for assessing students' writing level.

[0113] Methods for obtaining a writing proficiency assessment report include:

[0114] By analyzing the differences in different ability dimensions caused by text variations using an ability deviation matrix, the core semantic units affecting writing level are identified, and causal attribution explanations are generated. For example, "Completing assignments on time in the original text demonstrates good planning and execution ability, while failure to complete assignments on time in the counterfactual text leads to a decline in planning and execution ability and sense of responsibility, indicating that planning and execution ability has a key impact on relevant ability dimensions." The ability deviation matrix and causal attribution explanations are integrated to generate a writing level assessment report, enabling dynamic evaluation of writing level.

[0115] For example, the score range of each ability dimension in the ability deviation matrix is ​​limited to 0 to 1, and the score range is divided into grade intervals to intuitively reflect the student's writing level. Based on the grade intervals that the score range falls into, combined with causal attribution explanations, a writing level assessment report is generated.

[0116] Table 2 Writing Level Mapping Table

[0117] Score range Level range illustrate 0.85～1.0 excellent The writing ability was outstanding, and all key semantic units were correctly represented. 0.70～0.84 good The writing ability was good, with a few minor shortcomings. 0.50～0.69 medium Writing ability was average, with significant room for improvement. 0.30～0.49 Poor The writing ability was insufficient, and key semantic units were not effectively represented. 0～0.29 Difference Writing skills are severely lacking and require intensive guidance.

[0118] The preset importance matching threshold is set by staff. It is obtained by collecting the importance values ​​of different ability nodes and taking the average of multiple importance values ​​as the preset importance matching threshold. Similarly, preset weights and preset comprehensive causal strength thresholds are set. The pre-constructed cultural background knowledge graph, preset student ability target graph, and preset semantic projection network can be set by experts or by staff according to specific usage needs.

[0119] This embodiment, by enumerating all directed paths from semantic nodes to capability nodes in a directed acyclic causal graph and selecting paths with sufficient transmission effects, can accurately pinpoint "which semantic component, along which causal chain," affects writing ability, solving the problem of existing methods that only provide conclusions without tracing the causes. By calculating and weighting the intervention effects of causal relationships within the paths, a quantifiable and comparable path-level comprehensive strength is obtained, enabling the ranking and selection of influencing factors and improving the accuracy and robustness of causal explanations. Key causal paths also serve as a constraint source for counterfactual interventions. By performing simulated interventions at the path's starting point and constructing counterfactual texts, the direction and magnitude of changes in capability output can be verified, transforming the assessment from "static diagnosis" to "operable verification," providing a basis for personalized training.

[0120] Compared to arbitrary text replacement, counterfactual intervention targets only the starting nodes of critical paths, ensuring that the rewriting has genuine contribution to ability. Alternative expressions are constructed under conditions of clear semantic positioning, syntactic feasibility, and consistency with the context, ensuring the assessment results possess explanatory credibility and reproducibility. The final output is not an abstract score, but a structured explanation with causal chains and contribution strength criteria, presenting a closed-loop logic of "causal path—semantic components—ability differences." This upgrades the assessment system from "static scoring" to "intervention-based dynamic assessment," enhancing its transparency, pedagogical usability, and learning guidance value.

[0121] Example 2

[0122] Please see Figure 2 As shown, the parts not described in detail in this embodiment are described in Embodiment 1. This embodiment provides a dynamic assessment method for primary school students' writing skills, including:

[0123] S1. Receive the essay text data to be evaluated and decompose it into semantic units to form a semantic unit sequence; perform context dependency analysis on the semantic unit sequence to generate a semantic representation vector;

[0124] S2. Using the semantic representation vector as input, perform semantic node matching in the pre-constructed cultural background knowledge graph to obtain the first anchoring result; using the first anchoring result as a constraint, perform ability node reasoning in the preset student ability target graph to generate the second anchoring result.

[0125] S3. Using the first and second anchoring results as constraints, perform causal relationship extraction on the semantic unit sequence and construct a directed acyclic causal graph;

[0126] S4. Identify all causal paths in the directed acyclic causal graph, calculate the comprehensive causal strength of each path, filter out key causal paths, and generate counterfactual text based on the key causal paths.

[0127] S5. In the directed acyclic causal graph, reasoning is performed on the essay text data and counterfactual text data respectively, and a comparative analysis is performed. Based on the results of the comparative analysis, a writing level assessment report is generated.

[0128] Since the electronic device described in this embodiment is the electronic device used to implement the dynamic assessment system and method for primary school students' writing level in the embodiments of this application, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the dynamic assessment system and method for primary school students' writing level described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the dynamic assessment system and method for primary school students' writing level in the embodiments of this application falls within the scope of protection of this application.

[0129] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0130] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic assessment system for primary school students' writing skills, characterized in that, include: The data parsing unit receives the essay text data to be evaluated and performs semantic unit decomposition to form a semantic unit sequence; Perform context dependency analysis on the sequence of semantic units to generate semantic representation vectors; The semantic anchoring unit takes the semantic representation vector as input, performs semantic node matching in a pre-constructed cultural background knowledge graph to obtain the first anchoring result, and performs ability node reasoning in a pre-set student ability target graph to generate the second anchoring result. The causal structure modeling unit uses the first and second anchoring results as constraints to extract causal relationships from the semantic unit sequence and construct a directed acyclic causal graph. The method for constructing a directed acyclic causal graph includes: Using the first anchoring result as a semantic constraint and the second anchoring result as a capability constraint, causal relationship extraction and directed acyclic causal graph construction are performed on the semantic unit sequence corresponding to the essay text data. The first anchoring result is used to limit the scope of semantic units that can participate in causal relationship extraction, and only semantic units within the coverage of the first anchoring result are processed. The second anchoring result is used to limit the capability association scope, and causal relationship identification is performed only on semantic units that are associated with the capability nodes corresponding to the capability anchoring subgraph. Under the dual constraints of semantics and capability, explicit or implicit causal relationships between semantic units are detected. When a causal relationship is detected, the two semantic units involved in the causal relationship are used as causal nodes, and directed dependency edges are established between the causal nodes, thereby gradually forming a semantic causal graph structure. An acyclicity constraint is applied to the formed semantic causal graph structure to obtain a directed acyclic causal graph. The causal path identification unit identifies all causal paths in the directed acyclic causal graph, calculates the comprehensive causal strength of each path, filters out key causal paths, and generates counterfactual text based on the key causal paths. The methods for identifying key causal pathways include: In a directed acyclic causal graph, each causal node is used as the starting node to initialize the path search, and the path set is initialized to empty. The starting causal node is added to the current path, and it is checked whether it is the target capability node. If it is, the current path is recorded in the path set. If not, a recursive search is performed along all the directed dependent edges of the starting causal node. Each recursion adds the corresponding visited adjacent causal nodes to the current path, and continues to advance towards the target capability node. After all the directed dependent edges of the starting causal node have been recursively searched, all directed paths from the starting causal node to the target capability node are obtained, forming a path set. For each path in the path set, based on the intervention effect of each causal node in the path, the structural causal model is used to perform intervention calculations, calculate the average causal effect of each directed dependency edge in the path, and weight it according to the preset weights to obtain the comprehensive causal strength of each path; the path set is sorted according to the comprehensive causal strength, and the paths with a comprehensive causal strength greater than or equal to the preset comprehensive causal strength threshold are selected as key causal paths. The reasoning and analysis unit performs reasoning and comparative analysis on the essay text data and counterfactual text in a directed acyclic causal graph. Based on the results of the comparative analysis, a writing level assessment report is generated.

2. The dynamic assessment system for primary school students' writing level according to claim 1, characterized in that, The method for obtaining the generated semantic representation vector includes: Upload images of students' handwritten essays to be evaluated and perform optical character recognition to obtain essay text data; then, by performing lexical segmentation, syntactic structure recognition, and semantic role labeling on the essay text data, the essay text data is decomposed into a sequence of semantic units containing word-level, phrase-level, and sentence-level semantic units. Context dependency modeling is performed on the semantic unit sequence to capture the referential relationships, logical connections, and text connection clues between semantic units, thereby obtaining context dependency representations. The obtained context dependency representations are then input into a preset semantic projection network for feature fusion and compression mapping to generate semantic representation vectors.

3. The dynamic assessment system for primary school students' writing level according to claim 2, characterized in that, The method for obtaining the first anchoring result includes: The semantic representation vector is input into a pre-constructed cultural background knowledge graph, and node retrieval and matching are performed. The pre-constructed cultural background knowledge graph consists of semantic nodes and relation edges. By calculating the cosine similarity between the semantic representation vector and each semantic node in the cultural background knowledge graph, the semantic node with the highest cosine similarity is identified as the semantic node that matches the semantic representation vector; all matched semantic nodes are collected to form a candidate node set; the semantic neighborhood nodes associated with the candidate node set in the cultural background knowledge graph are extracted and integrated to form a semantic anchoring subgraph, thereby obtaining the first anchoring result.

4. The dynamic assessment system for primary school students' writing level according to claim 3, characterized in that, The method for obtaining the second anchoring result includes: The pre-defined student competency target graph consists of competency nodes and competency-related edges. The semantic anchoring subgraph corresponding to the first anchoring result is used as a semantic constraint. For each competency node in the student competency target graph, the importance of the competency node is weighted and aggregated based on the consistency between the weight of the related edges of the corresponding neighboring competency nodes and the semantic constraint, thereby obtaining the importance value of the competency node. The importance values ​​of all competency nodes are sorted, and competency nodes with importance values ​​higher than the preset importance matching threshold are selected to form a candidate competency node set. The candidate competency node set and the corresponding neighboring competency nodes are extracted to form a competency anchoring subgraph, forming the second anchoring result.

5. The dynamic assessment system for primary school students' writing level according to claim 4, characterized in that, The method for obtaining the counterfactual text includes: The initial causal node in the key causal path is identified as the counterfactual intervention object; the corresponding semantic unit in the essay text data is explicitly locatable; for the counterfactual intervention object, an alternative expression that is opposite to or zeroed out to the original semantic expression is constructed; the semantic unit corresponding to the counterfactual intervention object is located in the original essay text data and rewritten through the alternative expression, thereby generating a counterfactual text containing counterfactual premises.

6. The dynamic assessment system for primary school students' writing level according to claim 5, characterized in that, The method for reasoning and performing comparative analysis on essay text data and counterfactual text data includes: Using essay text data and generated counterfactual text as input, reasoning is performed node by node along key causal paths in a directed acyclic causal graph. The reasoning results are mapped to the ability dimensions in the student ability target map to perform comparative analysis, generating quantitative scores for essay text data and counterfactual text on each ability dimension, and forming an ability deviation matrix.

7. The dynamic assessment system for primary school students' writing level according to claim 6, characterized in that, The methods for obtaining the writing proficiency assessment report include: By analyzing the differences in different ability dimensions caused by text variations using an ability deviation matrix, the core semantic units affecting writing level are identified, and causal attribution explanations are generated. The ability deviation matrix and causal attribution explanations are then integrated to generate a writing level assessment report.

8. A method for dynamically assessing the writing level of primary school students, implemented through a dynamic assessment system for primary school students' writing level as described in any one of claims 1 to 7, characterized in that, include: S1. Receive the essay text data to be evaluated and decompose it into semantic units to form a semantic unit sequence; perform context dependency analysis on the semantic unit sequence to generate a semantic representation vector; S2. Using the semantic representation vector as input, perform semantic node matching in the pre-constructed cultural background knowledge graph to obtain the first anchoring result; using the first anchoring result as a constraint, perform ability node reasoning in the preset student ability target graph to generate the second anchoring result. S3. Using the first and second anchoring results as constraints, perform causal relationship extraction on the semantic unit sequence and construct a directed acyclic causal graph; S4. Identify all causal paths in the directed acyclic causal graph, calculate the comprehensive causal strength of each path, filter out key causal paths, and generate counterfactual text based on the key causal paths. S5. In the directed acyclic causal graph, reasoning is performed on the essay text data and counterfactual text data respectively, and a comparative analysis is performed. Based on the results of the comparative analysis, a writing level assessment report is generated.

Images