Scientific research scheme evaluation method and system based on capability characterization

Abstract

The application discloses a scientific research scheme evaluation method and system based on capability representation, and through a three-path Transform architecture with a flexible fusion mechanism, author information, capability and research ideas are integrated; a double-layer architecture is also introduced to learn capability representation according to author information and ideas; a dataset covering personal information, capability and research ideas / lines of thought is systematically constructed, high-dimensional feature vectors are extracted through deep feature coding technology, and advanced fusion mechanisms are adopted to realize effective interaction of multi-source features, thereby improving the accuracy, objectivity and automation level of scientific research idea evaluation; the application is used for predicting the quality of final works or evaluation results in the preliminary conception stage, thereby reducing unnecessary resource investment.

Description

Technical Field

[0001] This invention relates to the field of capability prediction technology, and more specifically to a method for evaluating research programs based on capability characterization.

[0002] The present invention also relates to a research program evaluation system based on capability characterization. Background Technology

[0003] With the development of large language models, many researchers have attempted to use them to generate research ideas and have found their potential to produce highly innovative and applicable ideas. However, even with research ideas, the time and financial costs of research work can still be high, especially in areas involving data collection or large-scale computation. Currently, most idea evaluation and acceptance prediction work is based on completed papers that have entered the pre-review stage or on reviewers' comments. Therefore, it is necessary to develop methods to assess the quality of research ideas and whether they are compatible with the capabilities of research teams.

[0004] Existing research on paper acceptance prediction primarily relies on analyzing completed manuscripts with known experimental results. Existing approaches incorporate the PeerReadv1 dataset, which includes draft papers and peer reviews, and is widely used for acceptance prediction and reviewer scoring. They also propose a multi-instance learning network, MILAM, with a summary-based memory mechanism to predict overall recommendations such as acceptance / rejection and identify sentiment polarity in peer reviews. Furthermore, they explore using reviewer sentiment embedded in peer review texts to predict manuscript acceptance or rejection, proposing a deep neural architecture that integrates paper content, review comments, and sentiment polarity. Other approaches utilize machine learning methods to predict whether scientific papers will be accepted by top AI conferences; a modular hierarchical attention network, MHAN, is introduced for automatic academic paper scoring (AAPR), achieving a 65.33% accuracy in predicting acceptance decisions. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for evaluating research proposals based on ability representation. By extracting personal information and research ideas / thoughts from existing public databases, ability data is represented. Then, corresponding features are extracted through feature encoding and fused through a self-attention mechanism to obtain the predicted score of the final evaluated person.

[0006] The technical solution adopted in this invention is as follows: a research program evaluation method based on capability characterization, the method comprising: Construct a dataset; the dataset includes the personal information and research ideas / concepts of the evaluated individuals; Ability representation involves characterizing the ability data of the assessed individuals using their personal information and research ideas / reasons, and adding the assessed individuals' ability data to the dataset. Feature encoding involves encoding the dataset to obtain personal information feature vectors, ability feature vectors, and research idea / thought feature vectors. Fusion assessment involves fusing extracted personal information feature vectors, ability feature vectors, and research idea / idea feature vectors. The fused feature vectors are then used for assessment and prediction to obtain the ability representation of the person being assessed.

[0007] Furthermore, the capability representation specifically includes: Let the paper's metrics be... The author metrics in the author list are ,author A specific capability vector is In the thesis j The group capabilities demonstrated in the process are denoted as Information from the person being evaluated The ability vector of all those being evaluated and the representation of the research ideas / line of the paper The decision is made jointly, as follows: (1) In the formula, , All are Transformer encoder models.

[0008] Furthermore, the construction of the Transformer encoder model specifically includes: Model pre-training involves extracting templated information about the evaluated individuals and research ideas / thoughts from the papers, and then using this information as input to train the model. Model construction involves training a two-stage Transformer encoder model, which uses information from the evaluator. and research ideas / approaches As input, it is used to predict capability representation.

[0009] Furthermore, the training of the two-stage Transformer encoder model specifically includes: In the first stage, the textual descriptions of all authors in the current paper's author list are used as input, with descriptions between different authors separated by [SEP][CLS]. After vectorization, these descriptions are input into the first-stage Transformer encoder. The feature vectors at the [CLS] positions corresponding to each author in the output feature sequence of the first-stage encoder are taken as the author's capability vectors. Each capability vector is concatenated with the embedding of research ideas / thoughts in author order, and after adding positional encoding, it is input into the second-stage Transformer encoder. The research idea / thought embedding is obtained by inputting idea / thought tags into the embedding layer of the second-stage Transformer encoder. The concatenated representation sequence is fed into the second-stage Transformer encoder, and its first output is used to predict capability representations, thus completing the prediction of capability representations.

[0010] Furthermore, the capability representation prediction model is trained through a multi-objective learning mechanism, with the loss function as follows: (2) In the formula, For cosine similarity, For the aforementioned ability representation and prediction ability representation, 'a' represents the top-level vector of the author information model pre-trained on the same instance. The estimated and original capability representations are respectively predicted by a shared linear layer, representing the labels or outputs. These are the over-parameters to be adjusted.

[0011] Furthermore, the feature encoding specifically includes: The personal information, abilities, and research ideas / approaches of the evaluated individuals in the dataset are encoded using three independent Transformer encoders to obtain corresponding feature vectors for personal information, abilities, and research ideas / approaches.

[0012] Furthermore, the fusion evaluation specifically includes: using a self-attention mechanism to perform feature fusion calculation to obtain the final prediction score.

[0013] Furthermore, the aforementioned personal information feature vector, ability feature vector, and research idea / thought feature vector are stacked into a sequence matrix, as shown in the following formula: (3) In the formula, , and These are respectively the personal information feature vector, the ability feature vector, and the research idea / thought feature vector. Given the length of each eigenvector, concatenate the three eigenvectors side by side to form... 3D matrix; According to the sequence matrix Perform self-attention mechanism calculations: (4) In the formula, , , For a trainable parameter matrix, The calculation results are aggregated into the output representation sequence corresponding to the three input representations as follows: (5) In the formula, The output of equation (2) The three output representation vectors correspond to the components of each input representation. The aggregation results Project the results to obtain the final evaluation score. As shown in the following formula: (6) In the formula, These are trainable parameters.

[0014] The second technical solution adopted in this invention is a research proposal evaluation system based on capability representation, employing the aforementioned research proposal evaluation method based on capability representation. The system includes: The competency representation module collects personal information and research ideas / approaches from the assessed individuals using a large language model, and then represents their competency data based on this information. Dataset module; constructs a dataset from the acquired personal information of the evaluated individuals, research ideas / reasoning, and ability representations; The feature encoding module encodes the dataset to obtain personal information feature vectors, ability feature vectors, and research idea / thought feature vectors. The fusion assessment module integrates extracted personal information feature vectors, ability feature vectors, and research idea / idea feature vectors, and then performs assessment and prediction to obtain the ability representation of the person being assessed.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of this application are: This invention presents a research proposal evaluation method based on capability representation, utilizing only author information and research ideas, without requiring full text or experimental results to predict paper acceptance and scoring. Through a three-way Transformer architecture with a flexible fusion mechanism, it integrates author information, capability representation, and research ideas. This invention also introduces a two-layer architecture for learning capability representations based on author information and ideas. Experiments show that the method of this invention, by fine-tuning BERT-base and BERT-large models, significantly outperforms single-way models, and capability prediction significantly improves the prediction accuracy of the final model. The proposed method can be applied to predicting early research outcomes and allocating scientific resources, and can achieve optimal matching between research ideas and author teams to maximize the probability of success. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the review result prediction model of the method of the present invention; Figure 3 This is a diagram illustrating the AI-generated suggestions used in the method of this invention. Figure 4 This is a diagram illustrating the data collected in the verification example of the method of the present invention; Figure 5 This is a schematic diagram of the prompt words used to extract the abilities of the person being assessed in the method of the present invention; Figure 6 This is a summary diagram illustrating the performance of the fine-tuned BERT model under different input settings in the verification examples of the method of this invention; Figure 7 This is a summary of the performance of fine-tuning the BERT model for acceptance rate prediction under different settings in the verification examples of the method of this invention; Figure 8 This is a schematic diagram of the network structure for capability characterization prediction in the method of the present invention; Figure 9 This is a summary graph of regression coefficients on the test set when different models predict the average score in the verification example of this invention. Figure 10 This is a comparison chart of the performance of different models on the test set in the verification examples of this invention; Figure 11 This is a performance diagram illustrating the different fusion methods in admission prediction in the verification examples of this invention; Figure 12 This is a comparison chart of the performance of capability model replacement in the verification examples of this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings.

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] The purpose of this invention is to predict the acceptance and rating of research papers using only author information and research ideas, without the need for full text or experimental results, thereby predicting the outcome before investing major effort and solving the problem that existing methods rely heavily on completed manuscripts or peer reviews.

[0020] Example This embodiment provides a research program evaluation method based on capability characterization, such as... Figure 1 As shown, a joint architecture is proposed to evaluate research ideas and author-idea matching degree for paper quality prediction. The model in this embodiment encodes author information, abilities, and research ideas into independent representations, and then fuses them to predict review results. Unlike the single-encoder Transformer, the method in this embodiment can better capture the interactions between these different text components, similar to user-item embedding interactions in recommendation systems. To predict review results based on author information, ability representations, and research ideas, this embodiment employs a self-attention mechanism to model the interactions between feature vectors, as detailed below: First, the ability data is characterized based on personal information, skills, and research ideas / approaches, specifically including: Let the paper's metrics be... The author metrics in the author list are ,author A specific capability vector is In a specific paper j The group capabilities demonstrated in the process are denoted as Information from the person being evaluated The ability vector of all those being evaluated and the representation of the research ideas / line of the paper The decision is made jointly, as follows: (1) In the formula, , All are Transformer encoder models.

[0021] The model architecture specifically includes two phases: Model training such as Figure 8 As shown, this includes capability prediction, training a two-stage Transformer encoder model, the model using information from the evaluator. and research ideas / approaches As input to predict capability representation, in the first stage, the textual description information of all authors in the current paper's author list is used as input, with descriptions between different authors separated by [SEP][CLS]. After vectorization, this information is input into the first-stage Transformer encoder. The feature vector corresponding to the [CLS] position before each author in the feature sequence output by the first-stage encoder is taken as the author's capability vector. Each capability vector is concatenated with the embedding of research ideas / thoughts in author order, and after adding positional encoding, it is input into the second-stage Transformer encoder. The research idea / thought embedding is obtained by inputting idea / thought tags into the embedding layer of the second-stage Transformer encoder. The concatenated representation sequence is fed into the second-stage Transformer encoder, and its first output is used to predict capability representation, thus completing the prediction of capability representation.

[0022] The capability representation prediction model is trained through a multi-objective learning mechanism, and the loss function is as follows: (2) In the formula, For cosine similarity, For the aforementioned ability representation and prediction ability representation, 'a' represents the top-level vector of the author information model pre-trained on the same instance. The estimated and original capability representations are respectively predicted by a shared linear layer, representing the labels or outputs. These are the over-parameters to be adjusted.

[0023] The first term aims to maximize the similarity between the predicted representation and the ability representation, while the second term aims to minimize the similarity between the predicted representation and the author information representation to avoid multicollinearity. The second term pushes the similarity between the predicted ability representation and the author representation towards 0.

[0024] After constructing the dataset using AI big data and constructing prompt words, such as... Figure 2 As shown, feature encoding is performed using three independent Transformer encoders to obtain corresponding personal information feature vectors, ability feature vectors, and research idea / thought feature vectors. Given three feature representations , and First, stack them into a sequence matrix, as shown below: (3) In the formula, , and These are respectively the personal information feature vector, the ability feature vector, and the research idea / thought feature vector. Given the length of each eigenvector, concatenate the three eigenvectors side by side to form... 3D matrix; Then, the self-attention mechanism is calculated: (4) In the formula, , , For a trainable parameter matrix, The calculation results are aggregated into the output representation sequence corresponding to the three input representations as follows: (5) In the formula, The output of equation (2) The three output representation vectors correspond to the components of each input representation. The aggregation results Project the results to obtain the final evaluation score. As shown in the following formula: (6) In the formula, For trainable parameters, in regression problems, To predict evaluation scores, training is performed using mean squared loss (MSE) given the true rating labels. In binary classification problems, The sigmoid function provides the probabilities of different classifications. Given the true evaluation decision labels, binary cross-entropy loss is used for training, and the classification result can be given by the sign or a threshold. Furthermore, training parameters can be selected. At this time, you can The relative magnitudes of different components are used as the classification criterion, and the probabilities of different categories are obtained using softmax. Cross-entropy loss is used for model training.

[0025] Verification Example In constructing the dataset for this invention, the design of AI prompts specifically includes: This invention uses prompts in a standardized format to collect author capabilities and research ideas, excluding information from completed research. Prompts for extracting research ideas include... Figure 3 As shown, this involves extracting the research question, the method applied to solving the problem in the paper, and the innovative aspects of that method. This invention does not include experimental results in its research ideas, ensuring that it can be evaluated at an early stage or during research design. Cue words for capability extraction include... Figure 5As shown, this invention uses a template to extract the authors' skills and proficiency, considering six general skills, including mathematical derivation, theoretical analysis / proof, model / architecture design, data collection, and experimental design, as well as 5-10 areas of expertise specific to the author group. Simultaneously, computational resources, financial and time budgets in GPU type / hour were extracted, as computational power and financial budgets can also be considered as the author group's capabilities, which can also be planned in the research design.

[0026] 25,889 PDF submissions were downloaded from Openreview, including submissions to ICLR 2024 / 2025 and NeurIPS 2023 / 2024. Statistics are as follows: Figure 4 As shown; although most of the rejected papers from NeurIPS were not open-source, their papers were merged because the acceptance criteria for these two conferences are similar; since the method in this application involves training author capabilities, this application screened approximately 16,712 papers whose first authors appeared twice in all papers. This application used Grobid to parse the downloaded PDF files and used the DeepseekR1 large language model to extract research ideas and author capabilities according to the aforementioned predefined cue word format. In addition, this application directly collected author information, titles, abstracts, scores, and acceptance decisions from the Openreview API; for average score prediction, this application also used conference information such as year and conference name; for acceptance prediction, conference information was not used because most of the available open-source NeurIPS papers are already accepted papers, such as... Figure 4 As shown, the merged dataset was shuffled using a random seed of 42, and then divided into training, validation, and test sets in an 8:1:1 ratio. That is, the first 80% of the shuffled dataset was used as the training set, the next 10% as the validation set, and the last 10% as the test set. For each experiment, five trials were conducted using different random seeds (42, 0, 1, 2, 3) to initialize model parameters and the data loader, select the model checkpoint with the minimum validation loss, and report the median or mean and standard deviation of the corresponding model performance on the test set. The experiments were conducted in a PyTorch 2.5.1 + Transformers 4.54.1 environment using two GeForce RTX 4090 GPUs.

[0027] The Transformer encoder's encoding process involves fine-tuning the BERT-base and BERT-large models on different input combinations to predict paper acceptance and average scores: (a) title, (b) title + abstract, (c) author information, (d) ability, (e) ideas, (f) author + ideas, (g) ability + ideas, and (h) author + ability + ideas (ACI); to account for differences in score distribution across different conferences and years, conference information is included in the average score prediction across all settings; this application employs a discriminative fine-tuning strategy: the backbone network parameters use... The learning rate is updated, while the prediction head uses... The learning rate is optimized using AdamW (weight decay of 0.01) and a linear warm-up scheduler; such as Figure 6 The report presents the rating prediction performance (MSE / MAE). Models using capability-aware inputs or combinations thereof, such as ACI, achieved slightly lower errors than models using authors or ideas alone. The BERT-large model using ACI showed modest gains. Although all models performed similarly, this suggests that capability-aware inputs offer a subtle benefit. Figure 7 The results summarize the recruitment predictions. Capability-based settings, such as ACI, yielded higher accuracy and F1 scores, with BERT-large / ACI achieving 64.55% accuracy on non-text inputs. Text-based inputs with summaries performed best overall, with BERT-base achieving 68.48% accuracy, highlighting the complementary role of full-text information and capability features.

[0028] Statistical Analysis: This application collected the predicted values ​​of all Transformer encoder models trained on (a), (b), (c), (d), and (e), evaluated their correlation matrices, and established a linear regression to fit the average score. To avoid overfitting, the regression model was fitted on the test set. For the fitted models, their predicted values ​​on the test set were collected. The outputs of different models trained on different text information were used as independent variables, and the average score was selected as the dependent variable. The OLS summary is as follows: Figure 9 As shown, after controlling for other independent variables, ability remains a significant factor; author identity, ideas, and ability all show significant effects; this ensures that they should all be included in the joint three-way model; compared to author identity and ideas, ability is shown to have a higher slope and is therefore more significant. Furthermore, the ability output predicted based on author information is also a significant factor, regardless of whether the ability model's output exists.

[0029] The choice of top-level fusion mechanism is crucial to the model's predictive performance. In this experiment, the compared prediction models included: simple averaging / summing, a single-layer Transformer encoder with 1 / 2 / 4 / 8 heads (TF-1l-1 / 2 / 4 / 8h), self-attention mechanisms SA1 / SA2 with / without layer normalization and residual connections, linear fusion mechanism R1, and weighted averaging with gating. Meanwhile, BERT-base and BERT-large models, which handle overall textual information such as author identity, ability, and ideas, were used as baselines. The results of the average rating prediction and the acceptance prediction are as follows: Figure 10 and Figure 11 As shown. Figure 10 The performance of various top-level fusion mechanisms in average score prediction was compared. The proposed three-way fusion model (s3, SA-1, SA-2, R1) consistently outperforms the BERT and BERT-large baselines that handle concatenated inputs, achieving lower MSE and MAE. The Transformer-based fusion method (8 attention heads, TF-11-8h-abs) achieved the best results when using the summary as idea information (MSE: 0.9990, MAE: 0.7761). For extracted ideas, SA-1 performed best, demonstrating the effectiveness of explicit interaction modeling between author, ability, and idea representation. Residual linear fusion R1 also showed competitiveness, providing a simpler yet effective alternative. Figure 11 This paper demonstrates the performance of different fusion strategies in admission prediction. Overall, when using summaries as ideas, the 1-layer 8-head Transformer encoder performs best as the fusion mechanism, achieving an accuracy of 70.45%. When using extracted ideas, the self-attention mechanism SA1 with residual connections achieves the highest accuracy (65.29%) and precision (64.26%), outperforming the BERT baseline and other fusion methods. Simultaneously, the 1-layer 8-head Transformer encoder provides the best recall (46.67%) and F1 score (52.57%). The residual linear fusion R1 also demonstrates strong and balanced performance in precision and recall. These results confirm that structured fusion of multi-source inputs such as authors, abilities, and ideas improves prediction accuracy compared to single-encoder models, with the proposed SA-1 providing the most robust fusion scheme for admission decision prediction.

[0030] Results of the capability model predictions: This invention fine-tunes a two-layer architecture that uses author information and ideas to predict ability representations, applying the AdamW optimizer with a learning rate of 0.00002 and no weight decay. It is compared with existing methods: (a) a three-way BERT model using a 2-layer 8-head Transformer as the fusion mechanism (author + idea + ability); (b) a two-way BERT model using a 2-layer 8-head Transformer as the fusion mechanism (author + ability); (c) in the three-way BERT model, the ability model is replaced with an untrained two-layer ability prediction architecture; (d) in the three-way BERT model, the ability model is replaced with a pre-trained two-layer architecture. In all settings, the model is trained for a total of 8 epochs. The AdamW optimizer with the same settings is used, with an effective batch size of 64.

[0031] like Figure 12 As shown, the impact of replacing explicit competence descriptions with predicted competence representations was evaluated. The three-way model using predicted competence (pre-trained) achieved competitive score predictions (MSE: 1.0392, MAE: 0.7980) and admission predictions (accuracy: 62.32%, F1: 42.02%), approaching the performance of the model using directly extracted competences (three-way-based). This demonstrates the feasibility of inferring author competences from author information and research ideas, enabling early predictions even when explicit competence descriptions are unavailable, with minimal performance degradation.

[0032] This article uses specific embodiments to illustrate the principles and implementation methods of the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A research proposal evaluation method based on capability representation, characterized in that, The method includes: Construct a dataset; the dataset includes the personal information and research ideas / concepts of the evaluated individuals; Ability representation involves characterizing the ability data of the assessed individuals using their personal information and research ideas / reasons, and adding the assessed individuals' ability data to the dataset. Feature encoding involves encoding the dataset to obtain personal information feature vectors, ability feature vectors, and research idea / thought feature vectors. Fusion assessment involves fusing extracted personal information feature vectors, ability feature vectors, and research idea / idea feature vectors. The fused feature vectors are then used for assessment and prediction to obtain the ability representation of the person being assessed.

2. The research program evaluation method based on capability representation according to claim 1, characterized in that, The capability representation specifically includes: Let the paper's metrics be... The author metrics in the author list are ,author A specific capability vector is In the thesis j The group capabilities demonstrated in the process are denoted as Information from the person being evaluated The ability vector of all those being evaluated and the representation of research ideas / reasoning in the paper The decision is made jointly, as follows: （1） In the formula, , All are Transformer encoder models.

3. The research program evaluation method based on capability representation according to claim 2, characterized in that, The construction of the Transformer encoder model specifically includes: Model pre-training involves extracting templated information about the evaluated individuals and research ideas / thoughts from the papers, and then using this information as input to train the model. Model construction involves training a two-stage Transformer encoder model, which uses information from the evaluator. and research ideas / approaches As input, it is used to predict capability representation.

4. The research program evaluation method based on capability representation according to claim 3, characterized in that, The training of the two-stage Transformer encoder model specifically includes: In the first stage, the textual descriptions of all authors in the current paper's author list are used as input, with descriptions between different authors separated by [SEP][CLS]. After vectorization, these descriptions are input into the first-stage Transformer encoder. The feature vectors at the [CLS] positions corresponding to each author in the output feature sequence of the first-stage encoder are taken as the author's capability vectors. Each capability vector is concatenated with the embedding of research ideas / thoughts in author order, and after adding positional encoding, it is input into the second-stage Transformer encoder. The research idea / thought embedding is obtained by inputting idea / thought tags into the embedding layer of the second-stage Transformer encoder. The concatenated representation sequence is fed into the second-stage Transformer encoder, and its first output is used to predict capability representations, thus completing the prediction of capability representations.

5. The research program evaluation method based on capability representation according to claim 4, characterized in that, The capability representation prediction model is trained through a multi-objective learning mechanism, and the loss function is as follows: （2） In the formula, For cosine similarity, For the aforementioned ability representation and prediction ability representation, 'a' represents the top-level vector of the author information model pre-trained on the same instance. The estimated and original capability representations are respectively predicted by a shared linear layer, representing the labels or outputs. These are the hyperparameters to be adjusted.

6. The research program evaluation method based on capability representation according to claim 1, characterized in that, The feature encoding specifically includes: The personal information, abilities, and research ideas / approaches of the evaluated individuals in the dataset are encoded using three independent Transformer encoders to obtain corresponding feature vectors for personal information, abilities, and research ideas / approaches.

7. The research program evaluation method based on capability representation according to claim 1, characterized in that, The fusion evaluation specifically includes: using a self-attention mechanism to perform feature fusion calculation to obtain the final prediction score.

8. The research program evaluation method based on capability representation according to claim 7, characterized in that, The personal information feature vector, ability feature vector, and research idea / brain feature vector are stacked into a sequence matrix, as shown in the following formula: （3） In the formula, , and These are respectively the personal information feature vector, the ability feature vector, and the research idea / thought feature vector. Given the length of each eigenvector, concatenate the three eigenvectors side by side to form... 3D matrix; According to the sequence matrix Perform self-attention mechanism calculations: （4） In the formula, , , For a trainable parameter matrix, The calculation results are aggregated into the output representation sequence corresponding to the three input representations as follows: （5） In the formula, The output of equation (2) The three output representation vectors correspond to the components of each input representation. The aggregation results Project the results to obtain the final evaluation score. As shown in the following formula: （6） In the formula, These are trainable parameters.

9. A research proposal evaluation system based on competency representation, employing the research proposal evaluation method based on competency representation as described in any one of claims 1 to 8, characterized in that, The system includes: The competency representation module collects personal information and research ideas / approaches from the assessed individuals using a large language model, and then represents their competency data based on this information. Dataset module; constructs a dataset from the acquired personal information of the evaluated individuals, research ideas / reasoning, and ability representations; The feature encoding module encodes the dataset to obtain personal information feature vectors, ability feature vectors, and research idea / thought feature vectors. The fusion assessment module integrates extracted personal information feature vectors, ability feature vectors, and research idea / idea feature vectors, and then performs assessment and prediction to obtain the ability representation of the person being assessed.

Images