A method and system for quantitatively scoring an electronic sand table by fusing sand tool features and topological associations

Abstract

The present application relates to the technical field of psychological assessment, and particularly relates to an electronic sand table quantitative scoring method and system fusing sand tool features and topological association, comprising the following steps: S1, obtaining operation log data of an electronic psychological sand table session; S2, constructing a sparse sand table relationship graph based on the operation log data; S3, inputting the sparse sand table relationship graph into a graph constraint Transformer regression model, and configuring an encoder of the graph constraint Transformer regression model to constrain the information propagation path between nodes based on the topological structure of the sparse sand table relationship graph, and to fuse node features and edge features for joint representation learning; and S4, obtaining a quantitative scoring prediction value of the electronic psychological sand table session based on the output of the graph constraint Transformer regression model. The present application structures the sand tool and its association by constructing a sparse sand table relationship graph, and designs a graph constraint Transformer regression model, and jointly introduces topological constraints and edge feature modulation in the attention mechanism, so as to realize more accurate scoring prediction.

Description

Technical Field

[0001] This invention relates to the field of psychological assessment technology, specifically to an electronic sandplay quantitative scoring method and system that integrates sandplay features and topological correlations. Background Technology

[0002] Sandplay therapy is a widely used projective intervention technique in clinical and counseling psychology. Clients construct scenarios by placing objects in a sandbox, while therapists interpret these scenarios and their formation process to gain insight into the client's inner world. With the development of digital technology, electronic sandplay systems have become available, capable of recording complete operation logs. This makes it possible to quantitatively analyze and intelligently assist in the assessment of the sandplay process. In particular, predictive expert ratings of sandplay creations are of significant value for therapist training, supervision quality control, and decision support.

[0003] However, transforming this task into a computable model faces a series of key challenges, and existing technical solutions have failed to adequately and effectively address the following issues: 1. Insufficient modeling of the relationship structure between sandplay objects makes it difficult to correspond to the evaluation dimensions of clinical "relationship organization".

[0004] Current computational analyses of electronic sandplay primarily focus on extracting features from the overall image or statistically summarizing individual sandplay objects. These methods fail to explicitly and structurally model the complex interactive networks formed between sandplay objects in terms of spatial arrangement and operational sequence during a session. However, in clinical interpretation, "how sandplay objects are organized and associated" is just as important as "which sandplay objects were used." Due to the lack of a specialized representation of this relational structure, existing methods struggle to support the quantitative assessment of this core evaluation dimension of "relationship organization," resulting in a gap between the information captured by the predictive model and the cues used by experts for scoring.

[0005] 2. Insufficient use of semantic knowledge clues such as names and symbols weakens the understanding of symbolic information.

[0006] The significance of sandplay objects extends far beyond their physical form, heavily relying on their names and symbolic meanings. Existing neural network methods, when applied to sandplay data, typically lack effective mechanisms to deeply integrate this prior semantic knowledge into the model's learning process. They may primarily rely on visual or statistical features, failing to fully grasp the symbolic narratives or themes behind the sandplay combinations. This limits the model's ability to understand the higher-level semantics of sandplay works, making it difficult to reproduce the symbolic and narrative reasoning processes relied upon by experts when rating them.

[0007] 3. The general model architecture fails to adapt to the characteristics of small sample size, high noise, and structural sparsity of sand table data.

[0008] First, the annotation of sand table data is highly dependent on experts, with inherent constraints such as limited sample size and high annotation costs, requiring models to have good robustness to small-sample learning. Second, existing general model architectures themselves have problems that are incompatible with the characteristics of sand table data: although graph neural networks can model relationships, they are usually not designed to effectively integrate continuous multi-dimensional features on the edges (such as distance, time interval, semantic similarity) into the message passing mechanism, resulting in a relatively coarse characterization of relationship strength; while if the powerful Transformer architecture is directly applied, its fully connected attention mechanism ignores the inherent sparsity of interactions in the sand table scene, allowing strong coupling between arbitrarily distant and unrelated sand table objects, which not only does not conform to physical and operational logic, but is also very likely to lead to severe overfitting with limited data.

[0009] In summary, existing technical solutions have significant limitations in three aspects: structured relationship modeling, semantic knowledge fusion, and adaptation to data characteristics. Therefore, there is an urgent need for a novel quantitative scoring method specifically designed for the inherent characteristics of electronic psychological sandplay data, capable of deeply integrating individual characteristics of sandplay figures with topological association information, in order to achieve more accurate and robust expert score predictions. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a quantitative scoring method and system for electronic sand table that integrates sand table features and topological relationships. This method constructs a sparse sand table relationship graph to structurally represent sand table features and their relationships, and designs a graph-constrained Transformer regression model. It also introduces topological constraints and edge feature modulation into the attention mechanism to achieve more accurate score prediction.

[0011] This invention is achieved through the following technical solution:

[0012] A quantitative scoring method for electronic sand tables that integrates sand table features and topological correlations is provided, including the following steps: S1. Obtain the operation log data of the electronic psychological sandplay session; S2. Based on the operation log data, construct a sparse sandbox relationship graph G=(V,E); where each node in the node set V represents a sandbox tool and is accompanied by a multi-dimensional node feature vector containing information such as the sandbox tool's location, operation sequence, and semantics; each edge in the edge set E represents the topological relationship between two sandbox tools and is accompanied by a multi-dimensional edge feature vector containing information such as spatial distance, operation time interval, name similarity, and symbolic similarity; the edge construction adopts a composite strategy including distance truncation, degree restriction, and semantic filtering to ensure the sparsity and rationality of the graph.

[0013] S3. Input the sparse sand table relationship graph into the pre-trained graph-constrained Transformer regression model. The core of this model is the graph-constrained Transformer encoder, which performs feature learning through the following mechanisms: Topological constraint attention mask: Generate an attention mask based on the adjacency matrix and valid node mask of the relationship graph, forcing information to propagate only on the actual associated edges and the nodes themselves on the graph.

[0014] Edge feature additive bias: Map the multi-dimensional features of the edges into additive bias terms in each attention head through a learnable linear layer, directly modulating the attention scores to encode continuous information about the relationship strength.

[0015] S4. Pool all the valid node representations output by the encoder to obtain a graph-level global representation, and then output the predicted value of the quantization score of this sand table session through the regression head.

[0016] Furthermore, the construction method of the sparse sand table relationship graph includes: Based on the three-dimensional Euclidean distance d between all pairs of sand tools ij , calculate the distance truncation threshold τ = Quantile(D, q) at a given quantile level q, where D = {d ij} is the set of distances between pairs of sand tools in the session, and Quantile(D, q) is the q-th quantile of D (0 < q < 1); and initially retain all candidate edges that satisfy d ij ≤τ.

[0017] For each node i, among its candidate edges truncated by distance, retain at most K adjacent edges in ascending order of distance.

[0018] Perform semantic screening on the above-retained candidate edges. Only when the edge (i, j) satisfies the condition s ij name ≥θ or s ij sym ≥θ will it be finally retained, where s ij name and s ij sym are the name and symbol similarity respectively, and θ is a preset threshold.

[0019] To ensure the connectivity of the graph, for nodes whose degrees are still lower than the preset minimum D min after the above steps, supplement edges in ascending order of distance from their candidate edges that satisfy the distance truncation and Top-K conditions but do not pass the semantic screening until reaching D min .

[0020] Furthermore, the core of the graph-constrained Transformer regression model is the graph-constrained Transformer encoder. The computation process of graph-constrained multi-head self-attention in each layer of the encoder is as follows: For the attention head α, its additive bias term The calculation is as follows: , Where w α and b α These are learnable parameters; E attr [p, q, :] represents the feature vector of the edge between sequence positions (p, q) (corresponding to the graph edge (i, j)|i=p,j=q), which at least contains the spatial distance d. pq Average operating time interval Δt pq Name similarity s pq name Symbolic similarity s pq sym .

[0021] An attention mask M is generated based on the graph's topology. For a query position p and a key position q, the attention mask M... pq The value can be: ; Where A pq m is the element at position (p, q) in the adjacency matrix A; p m q These are the components of mask m at positions p and q, respectively. k =1 indicates that the k-th position is a real sand figure, m k =0 is used for padding); N large A sufficiently large positive number (such as 10) 9 This makes the corresponding weights approach zero after softmax.

[0022] Furthermore, before inputting the node features into the encoder, learnable feature scaling is performed. Let the original node feature matrix be... The scaling process involves two steps: Apply the dimension-wise scaling parameter s: , s∈R77, For Hadamard product; Apply a global scaling scalar β to the semantic feature blocks: Where σ(·) is the sigmoid function, and Indicating implementation of The feature X is divided into two segments, before and after. inThe initial node representation is obtained after linear projection and position encoding. β can be initially set to 0. This applies a reduction scaling to semantic blocks during the early stages of training; thereafter, β is updated along with other parameters.

[0023] Furthermore, training methods for graph-constrained Transformer regression models include: Using Huber loss as the training objective function, it is defined as follows:

[0024] Where, r= -y represents the model's predicted value. The residual between the expert's actual rating y and the Huber loss threshold is a preset hyperparameter that is greater than 0.

[0025] The beneficial effects of this invention are: I. Strong structured representation capability: By constructing a sparse sand table relationship diagram, it is the first to unify the structured representation of "sand table objects" and "relationships" in the sand table scene, laying a solid foundation for the model to understand complex placement and organization patterns.

[0026] II. Highly targeted model design: The proposed graph-constrained Transformer encoder innovatively integrates topological constraints (through masks) and relation strength information (through edge feature bias) into the attention mechanism simultaneously, enabling the model to both follow the sparse priors of sandbox interaction and finely characterize the importance of different associations, significantly improving the effectiveness of feature learning.

[0027] Third, it is highly robust to noise and samples: Through a series of designs such as sparse graphing, feature scaling and training with Huber loss, it effectively alleviates the problems of overfitting and training instability caused by the strong subjectivity of expert scoring and the limited data samples.

[0028] IV. High prediction accuracy: Experiments on real electronic sand table datasets show that the method of this invention significantly outperforms traditional summary statistical methods, multilayer perceptrons, graph convolutional networks, graph attention networks, and standard sequence Transformers in multiple regression indicators such as the coefficient of determination (R²), Pearson correlation, mean absolute error (MAE), and root mean square error (RMSE), demonstrating its superiority. Attached Figure Description

[0029] Figure 1 This is a diagram illustrating the overall framework of the graph-constrained Transformer regression model based on a sandbox relationship diagram, as described in this invention.

[0030] Figure 2 This is a schematic diagram illustrating the construction of the electronic sand table scene to the sand table relationship diagram in this invention.

[0031] Figure 3 This is a scatter plot of the main method's predicted scores and the expert's actual scores on the test set in this invention.

[0032] Figure 4 This is a graph showing the variation of training and validation loss of the model with epochs in an embodiment of the present invention.

[0033] Figure 5 This is a graph showing the variation of the model validation set R² with epochs in an embodiment of the present invention.

[0034] Figure 6 This is a distribution diagram of the test set residuals relative to the true scores in this invention. Detailed Implementation

[0035] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to describe the solution.

[0036] This invention aims to address the shortcomings of existing methods in predicting expert ratings for electronic mental sandplay sessions. These methods suffer from insufficient modeling of the relationships between sandplay objects, inadequate utilization of semantic knowledge, and difficulty in adapting general models to small, noisy data. The core process of this invention is as follows: First, a sparse relational graph integrating sandplay object features and topological relationships is parsed and constructed from the operation log. Then, this graph is input into a specially designed graph-constrained Transformer regression model for representation learning. Finally, the model outputs continuous rating predictions for the current sandplay session.

[0037] Symbol explanation: L is the upper bound of the fixed-length node sequence length. i, j∈{1,…,L} are only used in the graph construction stage, representing the position of the sand tool node sequence. p, q∈{1,…,L} are only used in self-attention calculation, representing the sequence positions on the query side and the key side, respectively. m=(m1,…,m L )∈{0,1} L For the effective node mask, m k =1 indicates that the k-th position is a real sandplay figure, m k =0 indicates the fill position.

[0038] like Figure 2 As shown, the unstructured sandbox operation logs are transformed into structured graph data, which forms the basis for subsequent model processing.

[0039] First, parse the operation logs of a single sandplay session. Each log records the timestamp, three-dimensional spatial coordinates, sand tool ID, etc. when the sand tool is picked up, moved, and placed. Set the maximum number of nodes L (L = 128), and generate node features for all sand tools used in a session. Each sand tool node is represented by a 77-dimensional feature vector, which combines its normalized coordinates, size, operation timestamp features, as well as the name semantic vector and symbolic semantic vector obtained from the external knowledge base. The features of all nodes are stacked to form a node feature matrix. If the actual number of sand tools is less than L, it is filled with zero vectors. At the same time, generate a valid node mask vector m=(m1,…,m L ), where m i =1 indicates that the i-th position is a real sand tool, and m i =0 indicates a filled position.

[0040] Calculate the association relationship between any two sand tool nodes i and j to form a four-dimensional edge feature vector ; where: d ij is the three-dimensional Euclidean distance between the two in the sand tray; Δt ij is the average operation time interval; s ij name and s ij sym are the cosine similarities based on the name semantic vector and the symbolic semantic vector respectively.

[0041] Construct a sparse undirected graph, and the rules ensure that the graph reflects both spatial proximity and semantic relevance: Distance truncation: Calculate the set of distances between all pairs of sand tools in this session, set the quantile level q (e.g., q = 0.40), and calculate the distance truncation threshold according to τ = Quantile(D,q), where Quantile(D,q) is the q-th quantile of D (0 < q < 1); retain all candidate edges that satisfy d ij ≤τ. D = {d ij |1 ≤ i, j ≤ L, i ≠ j} is the set of spatial distances between pairs of sand tools in the session, and Quantile(D,q) represents the value obtained by taking the q-th quantile after arranging the elements in D in ascending order; Adjacent edge limit: For each node i, among its candidate edges passing the distance truncation, retain at most K adjacent edges (K = 6) in ascending order of distance, and make the undirected graph symmetric.

[0042] Semantic screening: For the candidate edges (i, j) passing the above two steps, impose semantic constraints, and only retain them finally when s ij name ≥θ or s ij sym ≥θ, where θ is the preset similarity threshold (e.g., θ = 0.3).

[0043] Minimum Degree Guarantee: To prevent isolated nodes after graph construction, any node whose degree is still below the preset minimum value D after processing according to the above rules will be guaranteed. min (e.g., D) min For nodes with degree = 1, supplement edges from the candidate edges that "satisfy distance truncation and adjacent edge constraints but fail semantic filtering" in order of distance, until the degree reaches D. min .

[0044] Through the above steps, the sparse undirected graph G=(V,E) of this session is obtained, and its topology is recorded using a symmetric adjacency matrix A (if edge (i,j) is preserved, then A is...). ij =1 otherwise 0). Finally, a single session is fully represented as a tuple G = {X, E}. attr ,A,m}.

[0045] The number of nodes is uniformly padded or truncated to L=128, and any insufficient number is padded with zero vectors, and an effective node mask m∈{0,1} is generated. L Its k-th component is m k (m) k =1 indicates a real sandplay figure, m k =0 indicates padding, with the same meaning as before). Generate a symmetric adjacency matrix A based on the final retained edges.

[0046] like Figure 1 As shown, the core innovation of the model structure in this invention lies in the graph-constrained attention mechanism in the encoder.

[0047] To ensure that components of different properties (such as numerical coordinates and semantic vectors) in node features are scaled in harmony before entering the model, learnable feature scaling is introduced.

[0048] First, apply the dimension-wise scaling parameter. : ,in For Hadamard products.

[0049] Then, a global scaling scalar β is applied to the last 60 dimensions of the feature matrix (semantic feature blocks): ; Where σ(·) is the sigmoid function. and Indicating implementation of The division is into two parts, before and after. β can be initially set to 0. This applies a reduction scaling to semantic blocks during the early stages of training; thereafter, β is updated along with other parameters.

[0050] Scaled feature X inThrough a linear projection layer W in The initial node representation is obtained by mapping to the hidden layer dimension d (d=192), superimposing sinusoidal position encoding, and then undergoing layer normalization.

[0051] In this invention, the encoder is composed of L layer It consists of three identical layers stacked together. Each layer contains a graph-constrained multi-head self-attention sublayer and a feedforward network sublayer. The sublayer outputs all employ residual connections and layer normalization.

[0052] Taking the l-th layer as an example, the calculation of its graph-constrained multi-head self-attention is key to this invention. For attention heads α (total number of heads nhead, for example 4): Calculate the query, key, and value vector: ; Among them, h (l-1) W represents the output of the (l-1)th layer encoder; q (α) W k (α) W v (α) These are the learnable projective weight matrices corresponding to the attention head α.

[0053] Calculate the basic scaling point integral: ; Where, d k =d / nhead is the dimension of each attention head; d is the hidden layer dimension of the node representation; nhead is the number of attention heads; S (α) Based on the scaled dot product attention score matrix; (K q (α) ) T This represents the transpose of the key matrix. At the query position p and the key position q, S... (α) The corresponding element is denoted as S. pq (α) ,Right now , where p and q∈{1,…,L} represent the query position and the key position, respectively.

[0054] Additive bias on edge features: Mapping the multidimensional continuous features of edges to additive biases on attention scores to finely characterize relation strength. ; Among them, E attr [p,q,:] is the multidimensional relational feature vector of the edges between sequence positions (p,q) (corresponding to the graph edge (i,j)|i=p,j=q); w α b α B is a learnable parameter of the attention head α; (α)For each element B pq (α) The additive bias matrix formed; This is the corrected attention score matrix after introducing edge feature bias.

[0055] Topologically constrained attention mask: Generates a mask M based on the graph's topology, forcing information to propagate only along actual edges in the graph. ; Among them, A pq m is the element at position (p, q) in the adjacency matrix A; p m q These are the components of the effective node mask m at positions p and q, respectively. k =1 indicates that the k-th position is a real sand figure, m k =0 indicates the fill position); N large It is a sufficiently large positive number so that the weight of the masked position approaches zero after the subsequent softmax operation.

[0056] The final attention weights and the head output are calculated as follows: ; Where, α pq (α) Normalized attention weights for key position q (calculated along the q-dimensional softmax); O p (α) V is the output vector of the attention head at position p; q (α) Let q be the value vector at the key position.

[0057] The nhead heads are concatenated and projected to the output to obtain O, which is then processed into the first residual and layer normalization: ; Where O represents the result of splicing and linear projection of the outputs of each attention head; LN1 represents the first layer normalization; and h̃ represents the intermediate representation after residual connection and normalization.

[0058] The feedforward subnetwork employs a two-layer linear transformation, with a hidden layer width of 2d and an activation function of ReLU, denoted as: ; Where h(l) is the output of the l-th layer encoder; LN2 is the normalization of the second layer.

[0059] The output of this coding layer is: ; Repeat the above calculations until... ,get .

[0060] Among them, L layer This represents the total number of layers in the encoder. This is the node representation matrix output by the final layer.

[0061] Expert scoring is at the session level, therefore the representations of all nodes need to be aggregated. Masked mean pooling is used to obtain the graph-level global representation vector h. （G） : ; in, Let m be the node representation of the i-th sequence position after encoding; i Let be the i-th component of the mask m.

[0062] h （G） Input a shallow multilayer perceptron regression head (For example, two linear transformations, using SiLU activation function and Dropout in between), outputting the final continuous rating prediction value. : .

[0063] Where fθ(·) is the regression head with parameter θ; This refers to the continuous score prediction values ​​output by the model.

[0064] Huber loss is used as the training objective function to enhance the model's robustness to labeled noise and outliers. Huber loss L δ (r) is defined as follows: ; Where, r= -y represents the model's predicted value. The residual between the model score and the expert's true score y is δ, which is a threshold parameter (δ=1.0 in this embodiment). The AdamW optimizer is used to optimize all parameters of the model end-to-end. During training, the coefficient of determination R² on the validation set is used as an early termination monitoring indicator. If it does not improve for several consecutive cycles (e.g., 40 rounds), training is terminated early to prevent overfitting.

[0065] The dynamics of the training process can be visually observed through monitoring curves. For example... Figure 4 As shown, the Huber loss of both the training and validation sets steadily decreases with increasing training epochs, and the difference between the two remains moderate, indicating that the model learns effectively and does not suffer from severe overfitting. Figure 5 The curve shows the trend of the validation set determination coefficient R2 as the number of training rounds increases and eventually stabilizes. This curve is used as the basis for model selection and early stopping judgment to ensure that the model with the best generalization performance is obtained.

[0066] To verify the effectiveness of the method of this invention, experiments were conducted on a dataset containing N=2939 valid electronic sandbox session samples. The dataset was randomly divided into training, validation, and test sets in a ratio of 70%:15%:15%.

[0067] Comparative experiment The method of this invention (Sand-RGTR) was compared with several baseline models on the same data partitioning and feature basis, and the results are shown in Table 1.

[0068] Table 1 Comparative Experimental Results

[0069] As shown in Table 1, this invention achieves the best results across all four metrics: R² is 0.183±0.035, Pearson correlation is 0.448±0.021, MAE and RMSE are 0.479±0.003 and 0.661±0.014, respectively. Ridge and linear mean pooling remain relatively weak (especially linear mean pooling RMSE of approximately 1.26), indicating that relying solely on summary statistics or single-layer readouts is insufficient to align with expert scores. MLP introduces nonlinearity after masked mean pooling, increasing R² to approximately 0.13, but still significantly lower than the main method. GCN and GAT outperform the non-graph sequence baseline under the same topology and node characteristics, but message passing does not explicitly utilize the four-dimensional continuous properties of the edge table, resulting in an overall performance lower than Sand-RGTR. The sequence Transformer has an R² of 0.146 ± 0.035 and an RMSE of 0.791 ± 0.011: self-attention, which is not subject to adjacency constraints and has no edge additive bias, still lags behind the master method that combines topology and edge continuity cues on the graph.

[0070] Visualization and analysis of prediction results: Figure 3 This provides a case-by-case, intuitive demonstration of the predictive performance of the method of this invention on the test set. For example... Figure 3 As shown, the scatter cloud is mainly distributed along the diagonal, indicating that the model's predicted values ​​show a consistent upward trend as the expert's actual score increases. This is consistent with the high Pearson correlation coefficient (0.448) in Table 1. Some discrete points deviating from the diagonal correspond to cases with higher prediction difficulty, which may be due to annotation noise or rare sand table placement patterns. This helps to understand the main sources of model error.

[0071] Figure 6This demonstrates the distribution of model prediction residuals relative to the true scores on the same test set. The residual points are evenly distributed on both sides of the zero line on the vertical axis, without exhibiting a significant systematic skew or "funnel" shape as the true scores change. This indicates that the model does not consistently overestimate or underestimate scores across different scoring intervals, and the overall calibration effect is good. Figure 3 and Figure 6 Combining these two aspects—trend consistency and error balance—the model's good predictive performance is confirmed, demonstrating its effectiveness in trend rationality and error balance.

[0072] To verify the contribution of each key design in the model, a systematic ablation experiment was conducted, and the results are shown in Table 2 below.

[0073] Table 2 Ablation Experiment

[0074] As shown in Table 2, the complete model of this application achieves R²=0.183±0.035, Pearson=0.448±0.021, MAE=0.479±0.003, and RMSE=0.661±0.014, which are the best among the ablation settings in all four indicators.

[0075] When semantic edge filtering is turned off, R² decreases to 0.165±0.042 and RMSE increases to 0.763±0.002. The decrease is relatively mild compared to the full-blooded model, indicating that in addition to geometric rules such as distance and Top-K, using name / symbol consistency to filter edges helps to make sparse topology fit the sandbox context better, but it is not the only decisive factor.

[0076] When the semantic features of the sand figurines (name and symbol similarity in node names / symbol vectors and edge attributes are set to zero before being fed into the model) are removed, R² decreases to 0.141±0.022 and RMSE increases to 0.774±0.007, which is more significant than the degradation of semantic-only screening, indicating that semantic cues on the node side and edge side still make a substantial contribution to score alignment.

[0077] When the average operation time interval on the edges is removed, R² decreases to 0.118±0.012, and RMSE increases to 0.785±0.012, indicating that procedural information such as interaction rhythm, after entering the attention bias through edge features, is still relatively sensitive to regression.

[0078] When the edge bias is removed (the linear mapping from edge attributes to the logits of each attention head is turned off), R² approaches 0 (0.019±0.010), and RMSE increases significantly to 1.648±0.006, which is the largest drop among all ablation methods. This indicates that injecting multidimensional edge continuous features into the attention scoring is particularly critical to the final performance.

[0079] This invention constructs a sparse sandplay relationship graph to structurally represent sandplay objects and their complex relationships, and innovatively designs a graph-constrained Transformer encoder, jointly introducing topological constraints and edge feature modulation into the attention mechanism. This method fully integrates the individual characteristics and topological relationships of the sandplay objects. Experiments on publicly available datasets demonstrate that it can achieve more accurate and robust automated quantitative prediction of expert ratings for electronic psychological sandplay, providing an effective technical solution for the intelligentization of psychological assessment.

[0080] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A quantitative scoring method for electronic sand table that integrates sand table features and topological correlations, characterized in that: Includes the following steps: S1. Obtain the operation log data of the electronic psychological sandplay session; S2. Based on the operation log data, construct a sparse sand table relationship graph. The nodes of the sparse sand table relationship graph represent sand objects and contain multi-dimensional features of the sand objects. The edges of the sparse sand table relationship graph represent the topological relationships between sand objects and contain multi-dimensional features of the relationships. S3. Input the sparse sand table relationship graph into the graph-constrained Transformer regression model. The encoder of the graph-constrained Transformer regression model is configured to: constrain the information propagation path between nodes based on the topological structure of the sparse sand table relationship graph, and fuse node features and edge features for joint representation learning. S4. Based on the output of the graph-constrained Transformer regression model, obtain the quantitative score prediction value of the electronic psychological sandplay session.

2. The electronic sand table quantitative scoring method integrating sand table features and topological correlation as described in claim 1, characterized in that: In S2, a sparse sandbox relationship graph is constructed, including: Based on the spatial distance between sand table pairs, distance truncation is performed, retaining candidate edges whose distance is less than or equal to the truncation threshold τ; where τ = Quantile(D,q), D = {d ij |1≤i,j≤L,i≠j} is the set of spatial distances between sand table pairs within the conversation, and Quantile(D,q) represents the value of the qth quantile after sorting the elements in D in ascending order; For each node, among the candidate edges truncated by distance, retain at most K neighboring edges in ascending order of distance; Semantic filtering is applied to the retained candidate edges. Only when the name similarity or symbolic similarity between sand table pairs is not lower than a preset threshold θ, the corresponding edge is retained in the sparse sand table relationship graph.

3. The electronic sand table quantitative scoring method integrating sand table features and topological correlation as described in claim 2, characterized in that: Constructing a sparse sandbox relationship diagram also includes: If, after distance truncation, neighbor preservation, and semantic filtering, there exists a node whose degree is lower than the preset minimum degree D... min Then, for this node, from the candidate edges that meet the distance truncation and adjacent edge retention conditions but fail the semantic filtering, edges are added in ascending order of distance until its degree reaches D. min .

4. The electronic sand table quantitative scoring method integrating sand table features and topological correlation as described in claim 1, characterized in that: In S3, the graph-constrained Transformer regression model includes a graph-constrained Transformer encoder. The encoder's attention calculation process includes: The adjacency matrix A and the effective node mask vector m∈{0,1} are based on the sparse sandbox relationship graph. L Its k-th component is m k Generate an attention mask M; for query position p and key position q, the attention mask M... pq The value can be: ; ； The multidimensional relational features of edges E attr [p, q, :] is mapped to the additive bias term B of the attention head α. pq (α) : ; Where w α and b α These are learnable parameters; Based on attention mask M and additive bias term B pq (α), calculate the attention weights between nodes and update the node representation.

5. The electronic sand table quantitative scoring method integrating sand table features and topological correlation as described in claim 4, characterized in that: Multidimensional relational features of edges E attr [i, j, :] includes at least spatial distance, average operation time interval, name similarity, and symbolic similarity.

6. The electronic sand table quantitative scoring method integrating sand table features and topological correlation as described in claim 1, characterized in that: Also includes: Training a Transformer regression model with graphical constraints includes: Obtain a training dataset containing multiple electronic mental sandplay sessions and their corresponding expert ratings; Using Huber loss as the objective function, the parameters of the graph-constrained Transformer regression model are optimized; where Huber loss... Defined as: ; Where: r= -y is the difference between the predicted value and the actual value, δ is the Huber loss threshold, and δ>

0.

7. The electronic sand table quantitative scoring method integrating sand table features and topological correlation as described in claim 1, characterized in that: Before inputting node features into the graph to constrain the Transformer regression model, learnable feature scaling is performed on the node features. Feature scaling includes applying a global scaling scalar β to the semantic feature blocks.

8. A scoring system using the electronic sand table quantitative scoring method based on the integration of sand table features and topological correlation as described in any one of claims 1 to 7, characterized in that: include: The data acquisition module is used to acquire operation log data of electronic psychological sandplay sessions; The relationship graph construction module is used to build a sparse sandbox relationship graph based on operation log data; The rating prediction module includes a graph-constrained Transformer regression model, which receives a sparse sandbox relationship graph and outputs a quantitative rating prediction value.

Images