A dream text similarity evaluation method and system based on multi-dimensional indexes

Abstract

This invention relates to the field of artificial intelligence technology, specifically to a method and system for evaluating the similarity of dream texts based on multidimensional indicators. The method includes: acquiring human dream texts and humanoid dream texts generated by generative language models; performing unified preprocessing on the texts to extract multidimensional text features such as narrative structure features, emotional valence features, lexical features, and semantic consistency features; calculating the multidimensional similarity assessment result between the humanoid dream texts and human dream texts based on the multidimensional text features; and then, based on the result, performing a difference analysis on the emotional and thematic dimensions of texts generated by different models and outputting the results. This invention, by constructing a multidimensional, hierarchical evaluation system and employing a weighting method combining subjective and objective approaches, achieves a systematic, refined, and quantifiable evaluation of dream text similarity, solving the problems of existing technologies having single evaluation dimensions and difficulty in quantifying model differences.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for evaluating the similarity of dream texts based on multi-dimensional indicators. Background Technology

[0002] With the rapid development of artificial intelligence technology, generative language models are increasingly being applied in text generation, content creation, and natural language processing. In psychology, neuroscience, and digital health, using generative language models to simulate or predict human dream content has become an emerging research direction. This type of research is of great significance for exploring human cognition, assisting in the analysis of psychological states, and even promoting personalized content generation. However, how to scientifically and objectively assess the similarity between these model-generated "human-like dream texts" and real human dream descriptions has become a key technical bottleneck.

[0003] Existing text similarity assessment methods, such as those based on perplexity, keyword matching, or relying solely on subjective human scoring, are significantly inadequate when dealing with dream texts. Dream texts are highly subjective, possess non-linear narrative structures, rich emotional shifts, and unique symbolic imagery. Perplexity metrics primarily measure the fluency of language models but fail to capture narrative logic and emotional consistency; keyword matching is too superficial, struggling to understand deeper semantic connections and plot coherence; while human scoring provides a comprehensive assessment, it is costly, inefficient, and heavily influenced by the evaluator's subjectivity, making it difficult to guarantee consistency and repeatability. More importantly, current technologies lack a systematic, multi-dimensional evaluation framework, failing to provide a refined and quantifiable analysis of the "anthropomorphism" of model-generated dream texts from multiple professional dimensions, including narrative structure, emotional valence, vocabulary usage, and semantic coherence. This makes it difficult to accurately measure and compare performance differences between different models, severely hindering the in-depth development and optimization of generative language models in dream-related research and applications. Therefore, there is an urgent need in this field for a technical solution that can comprehensively and deeply assess the similarity of dream texts.

[0004] Therefore, existing technologies still need further development. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a dream text similarity evaluation method and system based on multi-dimensional indicators to solve the problems existing in the prior art.

[0006] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a method for evaluating dream text similarity based on multi-dimensional indicators, comprising: S1. Obtain at least one human dream text and at least one human-like dream text generated by a generative language model; S2. Perform text preprocessing on the human dream text and the humanoid dream text; extract multi-dimensional text features from the preprocessed text, wherein the multi-dimensional text features include at least narrative structure features, emotional valence features, lexical features and semantic consistency features; S3. Based on the multidimensional text features, calculate the multidimensional similarity assessment result between the humanoid dream text and the human dream text; S4. Based on the multidimensional similarity evaluation results, perform a difference analysis on the emotion dimension and the theme dimension of the humanoid dream text generated by different generative language models, and output the evaluation results.

[0007] Specifically, the text preprocessing includes segmenting the human dream text and the humanoid dream text into sentences, removing stop words, performing part-of-speech tagging, and word vector representation to obtain a unified text representation format.

[0008] Specifically, the narrative structure features include event sequence features, plot coherence features, and syntactic structure distribution features; the emotional valence features are intensity vectors based on sentiment analysis of the text, containing multiple basic emotional dimensions; the lexical features include lexical richness, word frequency distribution, and lexical complexity; and the semantic consistency features include intra-sentence semantic similarity, cross-sentence semantic coherence, and thematic consistency.

[0009] Specifically, calculating the multidimensional similarity assessment result includes: For each extracted text feature, the similarity between the humanoid dream text and the human dream text in that feature dimension is calculated to form a similarity vector. Then, a weighting method based on information entropy is used, combined with a preset empirical adjustment factor, to assign weights to the similarity of each dimension in the similarity vector, so as to calculate a comprehensive multidimensional similarity evaluation score.

[0010] Specifically, the calculation of the similarity vector includes at least three levels of evaluation: Surface vocabulary layer evaluation is calculated based on the matching degree of surface vocabulary in the text; Deep semantic layer evaluation calculates semantic similarity based on the vector embedding representation of the text; The generative model heuristic evaluation layer uses a generative language model to score the consistency of two texts in the dream narrative logic.

[0011] Specifically, the human-like dream text generated by the generative language model is a predicted dream text generated based on multi-source information collected from the target individual during waking hours. The multi-source information during waking hours includes at least one of the following: individual experience information, emotional state information, daily life behavior information, social relationship and interpersonal interaction information, and information contact or media exposure information.

[0012] Specifically, the emotional valence features and narrative structure features in the multidimensional text features can be weighted or calibrated using biosignal features associated with the time of acquisition of the human dream text, which are extracted from electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, or electrodermal signals.

[0013] Specifically, the evaluation results generated by the method are further used to drive the multimodal content generation process of generating corresponding images or videos based on dream text content, or associated with the results of non-diagnostic state characteristic analysis of individuals based on waking information and dream characteristics, so as to form a component of a comprehensive technical framework for systematically analyzing, predicting, generating and quantitatively evaluating dream narratives.

[0014] Specifically, the method further includes: storing the multidimensional similarity evaluation results and difference analysis results, and associating the results with the corresponding generative language model identifiers, input conditions and text preprocessing parameters to construct an evaluation database for model performance comparison and optimization.

[0015] According to a second aspect of the present invention, a dream text similarity evaluation system based on multidimensional indicators is provided, comprising: The data acquisition module is configured to acquire at least one human dream text and at least one human-like dream text generated by a generative language model; The text preprocessing module is configured to perform unified text preprocessing on the human dream text and the humanoid dream text; The multidimensional feature extraction module is configured to extract multidimensional text features from the preprocessed text, including at least narrative structure features, emotional valence features, lexical features, and semantic consistency features. The similarity assessment module is configured to calculate the multidimensional similarity assessment result between the humanoid dream text and the human dream text based on the multidimensional text features; The results output module is configured to output the difference analysis results of human-like dream texts generated by different generative language models in terms of emotion and theme dimensions, based on the multidimensional similarity evaluation results.

[0016] Beneficial effects: Compared with existing technologies, the dream text similarity evaluation method and system based on multi-dimensional indicators provided by this invention have at least the following beneficial effects: First, this invention constructs a systematic, multi-dimensional evaluation framework, fundamentally overcoming the shortcomings of existing technologies that rely on a single evaluation dimension. By extracting features from four deeply related dimensions of dream expression characteristics—narrative structure, emotional valence, lexical features, and semantic consistency—a panoramic portrayal of dream texts can be achieved. Narrative structure features focus on the sequence of events and plot coherence; emotional valence features quantify complex emotional fluctuations; lexical features analyze the richness and specificity of language use; and semantic consistency features ensure the self-consistency of the theme and logic. This multi-dimensional evaluation not only calculates a comprehensive similarity score but also provides detailed reports for each dimension, thus revealing the advantages and disadvantages of generative models in simulating human dreams from all angles and perspectives, providing clear and precise directions for model optimization.

[0017] Secondly, this invention introduces an innovative weighting method that combines subjective and objective approaches, ensuring the scientific and professional nature of the evaluation results. Employing an objective weighting method based on information entropy, it automatically allocates weights according to the dispersion of the feature data in each dimension, avoiding the subjective arbitrariness of manually setting weights and making the evaluation results more objective and reliable. Simultaneously, it creatively introduces an experience-based adjustment factor preset by domain experts, integrating professional knowledge of dream research into the weighting system. Through a harmonic coefficient, objective entropy weights are organically combined with subjective experience, ensuring that the final comprehensive score is both faithful to the patterns revealed by the data itself and consistent with professional understanding in the field, greatly enhancing the authority and persuasiveness of the evaluation results.

[0018] Furthermore, this invention possesses high scalability and technological synergy, enabling it to be embedded into a broader technological ecosystem and form a value loop. Specifically, this evaluation method can be linked with "dream prediction based on lucid information" technology to provide a core quantitative benchmark for the accuracy of the prediction model; combined with "multimodal content generation based on dream text" technology, it can serve as a quality controller for selecting the optimal generated content; and its output multidimensional features can also serve as important inputs for "mental state-assisted analysis." In addition, by introducing biosignal features to calibrate the text analysis results, the physiological validity of the evaluation is further enhanced. By constructing a structured evaluation database, this solution can continuously accumulate data, driving iterative optimization of model performance and the establishment of standardized benchmark test sets. Therefore, this invention is not only an independent evaluation tool, but also a core quantitative engine connecting cutting-edge technologies such as prediction, generation, and analysis in the field of dream research, significantly enhancing the industrial application value and creative level of the overall technical solution. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the dream text similarity evaluation method based on multi-dimensional indicators provided in a specific embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0021] First, it should be noted that the technical solution involved in this invention is a component module of the DreamScore framework. DreamScore is a comprehensive artificial intelligence technology architecture proposed by the inventor for the systematic analysis, prediction, generation, and quantitative evaluation of dream narratives.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides a method for evaluating dream text similarity based on multi-dimensional indicators, including: S1. Obtain at least one human dream text and at least one human-like dream text generated by a generative language model.

[0024] It should be further explained that this method aims to provide a systematic and quantifiable framework (DreamScore) for evaluating the ability of generative language models to simulate human dream narratives. In step S1, the acquisition of human dream texts should be carried out through standardized collection protocols, such as requiring subjects to record their dream content in detail using natural language within a short period of time after waking up (e.g., within 5 minutes) to ensure the vividness of the memory and the integrity of the details. Human-like dream texts are generated by one or more generative language models to be evaluated. These models may include, but are not limited to, mainstream large language models such as GPT-4, Claude 3, Gemini, and Wenxin Yiyan. To ensure fair comparison, a uniform prompt template should be used when generating texts, such as: "Please describe a dream in detail from a first-person perspective. The dream content should include specific characters, scenes, events, and emotional changes." At the same time, to ensure the statistical significance of the evaluation, for the same prompt or the same set of waking-period input information, each evaluated model should generate multiple (e.g., 10) human-like dream texts.

[0025] S2. Perform text preprocessing on the human dream text and the humanoid dream text; extract multi-dimensional text features from the preprocessed text, wherein the multi-dimensional text features include at least narrative structure features, emotional valence features, lexical features and semantic consistency features.

[0026] It should be further noted that the text preprocessing in step S2 is the cornerstone of all subsequent analyses. Its primary goal is to transform unstructured, potentially noisy natural language text into clean, structured, machine-readable numerical representations. This process is pipeline-based, and the specific design includes: (1) Sentence splitting: Using regular expressions or dedicated natural language processing tools (such as NLTK's sent_tokenize or SpaCy's sentence splitter), split the long dream narrative into a list of independent sentences based on full stops, question marks, exclamation marks, semicolons, and line breaks. This lays the foundation for analyzing the semantics at the sentence level and the relationships between sentences.

[0027] (2) Stop word removal: Construct an enhanced stop word list suitable for the dream domain. This list not only includes common function words (such as "de", "le", "zai", "he"), but also specifically incorporates high-frequency but low-information guiding words and filler words in dream narratives, such as "I dreamed", "then", "next", "suddenly", "seemingly", etc. Removing these words helps to highlight the entity words, action words, and emotion words that carry the core semantics.

[0028] (3) Part-of-speech tagging and dependency parsing: Use tools such as Stanford CoreNLP, SpaCy, or Harbin Institute of Technology's LTP to perform part-of-speech tagging (POS Tagging) and dependency parsing on each sentence. Part-of-speech tagging is used to identify content words such as nouns, verbs, and adjectives. Dependency parsing can extract the core predicate (verb) in the sentence and its arguments such as the subject and object, which is crucial for constructing "event triples" to analyze the narrative structure later.

[0029] (4) Word vector and sentence vector representation: Convert the cleaned and tagged text into vectors in a deep semantic space. For words, use pre-trained language models (such as BERT, RoBERTa) to obtain the context-related vector representation of each word (or sub-word). For sentences and texts, the sentence vector of a fixed dimension can be obtained by performing mean pooling on the vectors of all words in the sentence, or directly using an optimized sentence encoding model (such as Sentence-BERT, SimCSE). Finally, a human dream text and a humanoid dream text are represented as a collection of a series of sentence vectors, or an overall document vector, preparing for the calculation of deep semantic features.

[0030] Furthermore, the multi-dimensional text features defined in step S2 are quantitative indicators specifically tailored to the characteristics of dream text, further refined from the aforementioned vectorized representation. These four dimensions of features together constitute an evaluation matrix: (1) Narrative structural features: Evaluate the organization and progression of the dream story. This includes: a) Event sequence features: The core "subject-verb-object" triples in each sentence are extracted as event units through dependency parsing. Then, the edit distance or dynamic time warping distance between the event sequences of human text and generated text is calculated to measure the similarity of narrative order. b) Plot coherence feature: Calculate the average cosine similarity between adjacent sentence vectors in the text, which reflects the smoothness of plot progression; c) Syntactic structure distribution characteristics: Statistically analyze the frequency distribution of various syntactic dependency relations (such as subject-verb, verb-object, and attributive-head) in the whole text, and evaluate the similarity of sentence style by comparing the Jensen-Shannon divergence of the two distributions.

[0031] (2) Emotional Valence Features: Quantifying the emotional content of dream texts. A six-dimensional basic emotion model is used, and a pre-trained language model fine-tuned on an emotion analysis dataset (e.g., RoBERTa fine-tuned on the ISEAR or GoEmotions dataset) is used to process the text, outputting a six-dimensional emotion intensity vector. Each dimension The value range is [0,1], representing the intensity of the emotion in the text. These six dimensions cover the basic emotions that are generally recognized in psychology, and can effectively depict the complex emotional atmosphere of dreams.

[0032] (3) Lexical features: Analyze the characteristics of word usage in dream texts. This includes: a) Lexical richness, measured by the Type-Token Ratio (TTR), which is the ratio of the number of unique terms to the total number of terms. The higher the value, the more diverse the vocabulary. b) Word frequency distribution: Statistically analyze the distribution of nouns, verbs, and adjectives, and compare it with the distribution of general corpora (such as news). Calculate the chi-square statistic to capture the special characteristics of dream vocabulary (such as more first-person pronouns and body part words). c) Lexical complexity: Calculate the average norm of all real word vectors in the text. In the pre-trained semantic space, abstract words usually have a larger norm, and this metric can indirectly reflect the degree of abstraction of the text.

[0033] (4) Semantic consistency features: assessing the degree of semantic unity and logical consistency within the text. This includes: a) Intra-sentence semantic similarity: For each sentence, calculate the cosine similarity between its subject component vector and predicate component vector, and then average them across the entire text to measure the semantic tightness within the sentence. b) Cross-sentence semantic coherence: The sentence vectors of the entire text are regarded as a graph, with nodes representing sentences and edge weights representing inter-sentence similarity. The global clustering coefficient of the graph is calculated to quantify the overall tightness of the semantic network of the whole text. c) Topic consistency: Use topic models (such as LDA) or directly cluster document vectors to analyze the core topic distribution of the text and calculate the Hellinger distance between the topic distribution of human and generated text. The smaller the distance, the more similar the topics are.

[0034] S3. Based on the multidimensional text features, calculate the multidimensional similarity assessment result between the humanoid dream text and the human dream text.

[0035] It should be further explained that the core of step S3 is to integrate the above multi-dimensional features into a holistic similarity evaluation result, namely, DreamScore. This process consists of two stages: The first stage is to construct a multi-dimensional similarity vector. For each feature dimension... Use a similarity metric function suitable for this feature type. To calculate human text With generated text Similarity score on this dimension For example, the sentiment vector uses cosine similarity, the distribution feature uses 1 minus the Jensen-Shannon divergence, and the scalar feature uses 1 minus the normalized absolute difference. This ultimately yields an m-dimensional similarity vector. , where m is the total number of feature dimensions.

[0036] The second stage involves a comprehensive weighting and scoring process based on entropy weighting combined with empirical adjustment factors. This is to objectively and scientifically assign weights to vectors. Aggregation into a single scalar (DreamScore). First, a certain number (n pairs) of evaluation samples are collected to form an initial similarity matrix. .right Normalization by column (feature dimension) yields the matrix. Next, calculate the first... Information entropy of each feature dimension The smaller the entropy value, the greater the difference between samples and the more information it contains; therefore, its weight should be greater. Objective weights based on information entropy are calculated accordingly. Then, an empirical moderating factor, pre-defined by domain experts based on prior knowledge of dream research, is introduced. This is to reflect the relative importance of different dimensions in professional evaluation (for example, "emotional valence" might be considered more important than "lexical richness"). Ultimately, the overall weight of each dimension is determined. The objective entropy weight and the subjective experience weight are harmonicized by a coefficient. (The preferred value is 0.7) Weighted harmonic summation yields: Ultimately, the DreamScore overall score was... The calculation is as follows: The higher the D value, the more similar the generated text is to human dream text as a whole.

[0037] S4. Based on the multidimensional similarity evaluation results, perform a difference analysis on the emotion dimension and the theme dimension of the humanoid dream text generated by different generative language models, and output the evaluation results.

[0038] Step S4 involves application-level analysis based on the calculation results of step S3. The system will organize and visualize the scores of different generative language models across various feature dimensions. ) and the final comprehensive DreamScore ( This generates detailed comparison reports. For example, the report might indicate that "Model A performs exceptionally well in emotional valence mimicry, but is significantly weaker than Model B in narrative coherence," or "All models struggle to simulate the specific distribution of dream vocabulary." This difference analysis not only provides users with a direct basis for model selection, but more importantly, it points to specific technical directions for further optimization of generative language models (e.g., adding more narratively coherent text to the training data, or specifically strengthening emotional consistency).

[0039] It is understandable that, through the highly specific and procedural implementation methods described above, the method protected by claim 1 of this invention has been transformed from a conceptual framework into a rigorously operable and reproducibly verifiable technical solution. From the standardized acquisition and cleaning of text, to the refined feature engineering of four deeply customized dimensions, to the weighted comprehensive algorithm based on information theory that combines subjective and objective methods, and finally to the difference analysis leading to practical applications, each step provides a clear, complete, and highly patentable technical path for evaluating the performance of generative language models in simulating the highly complex, subjective, and unstructured task of human dreams. This allows those skilled in the art to reproduce this invention based on this implementation method without excessive creative effort, and to reliably obtain quantitative evaluation results of dream text similarity, thereby solving the long-standing technical problem of how to scientifically measure the "human-likeness" of AI-generated dreams.

[0040] Specifically, the text preprocessing includes sentence splitting, stop word removal, part-of-speech tagging, and word vector representation for the human dream text and the humanoid dream text, so as to obtain a unified text representation form.

[0041] It should be further noted that text preprocessing is the basis for subsequent feature extraction and calculation, and its purpose is to convert unstructured natural language text into structured and standardized data forms. The specific steps are as follows: (1) Sentence splitting: Use punctuation marks such as full stops, question marks, exclamation marks, and line breaks to split long dream texts into independent sentence sequences. This helps with subsequent syntactic analysis and event order analysis.

[0042] (2) Stop word removal: Construct a stop word list applicable to the dream domain, and remove common function words such as "de", "le", "zai", etc., as well as high-frequency but low-information dream narrative guiding words such as "I dreamed", "then", etc., to highlight content words.

[0043] (3) Part-of-speech tagging: Use tools such as Stanford Core NLP, Jieba (Chinese), or SpaCy (English) to tag the part of speech (such as noun, verb, adjective, etc.) of each word in the text. This helps identify key entities (nouns) and actions (verbs) in the narrative.

[0044] (4) Word vector representation: Convert the words that have undergone the above processing into high-dimensional real number vectors through a pre-trained word embedding model (such as Word2Vec, GloVe, BERT's token embedding). Preferably, use the BERT model pre-trained on a large corpus to obtain context-related word vectors, which can better capture the semantics of words in a specific context. Through this step, a text is finally represented as a vector sequence or a document vector obtained through pooling (such as average pooling), providing input for subsequent deep semantic similarity calculation.

[0045] It can be understood that the standardized preprocessing process eliminates the noise in the text that is irrelevant to similarity evaluation, unifies the formats of texts from different sources, and converts them into a numerical form that can be directly processed by machine learning models, ensuring the accuracy and efficiency of subsequent feature extraction and calculation.

[0046] Specifically, the narrative structure features include event order features, plot coherence features, and syntactic structure distribution features; the emotional valence features are intensity vectors containing multiple basic emotion dimensions obtained by performing emotion analysis on the text; the lexical features include lexical richness, word frequency distribution, and lexical complexity; the semantic consistency features include intra-sentence semantic similarity, cross-sentence semantic coherence, and topic consistency.

[0047] It should be further noted that the multidimensional text features defined in this method are the evaluation pillars of the DreamScore framework. Each feature is carefully designed for the characteristics of dream text, and its specific calculation method is as follows: (1) Narrative structural features, specifically including: a) Event Sequence Features: Dependency parsing is used to extract the core verb and its subject and object from each sentence, forming a subject-verb-object triple as an event. The consistency of the event sequence is measured by comparing the edit distance or dynamic time warping (DTW) distance between the event triple sequences in human and generated texts. The smaller the distance, the more similar the narrative order.

[0048] b) Plot coherence feature: Calculate the average cosine similarity between adjacent sentences in the semantic vector space. Specifically, each sentence is converted into a vector using a sentence encoding model such as Sentence-BERT, and then the cosine similarity between adjacent sentence vectors is calculated and averaged. The higher the value, the more coherent the plot progression.

[0049] c) Syntactic structure distribution characteristics: The frequency of different syntactic tree structures in the entire text is statistically analyzed to form a distribution histogram. The Jensen-Shannon divergence between the syntactic distribution histograms of human text and generated text is compared to measure the similarity in sentence usage habits between the two.

[0050] (2) Emotional Valence Feature This feature uses a six-dimensional emotion model to quantify the text. Specifically, a pre-trained language model (such as RoBERTa) fine-tuned on a sentiment analysis dataset is used to analyze the text, outputting a six-dimensional emotion intensity vector: in, Represents the intensity of the emotion of "joy" in the text. Represents the intensity of the emotion of "sadness". Represents the intensity of the emotion "anger". Represents the intensity of the emotion "fear". Represents the intensity of the emotion of "surprise". This represents the intensity of the emotion of "disgust". The value of each dimension is in the range of [0,1], with larger values ​​indicating a stronger emotion. The six-dimensional model was chosen because it covers the basic emotions recognized in psychology and can comprehensively depict the complex and ever-changing emotional states in dreams.

[0051] (3) Lexical features, specifically including: a) Lexical richness: Measured by the Type-Token Ratio (TTR) of the text, which is the ratio of the number of unique words in the text to the total number of words. The higher the TTR value, the richer and more diverse the vocabulary.

[0052] b) Word frequency distribution: The word frequencies of nouns, verbs, and adjectives in the text are statistically analyzed and compared with the corresponding word frequencies in a standard everyday language corpus (such as news corpus), and the chi-square value is calculated. Dream texts often have unique word frequency distributions (such as more first-person pronouns and more emotional words), and this feature can capture this difference.

[0053] c) Lexical complexity: Calculate the average word vector norm of all words in the text. In the pre-trained word vector space, the vector norm of abstract words and technical terms is usually greater than that of common concrete words. This metric can reflect the abstractness and complexity of the words used in the text to a certain extent.

[0054] (4) Semantic consistency features, specifically including: a) Intra-sentence semantic similarity: For each sentence, calculate the cosine similarity between its subject vector and predicate vector, and then average this value across all sentences. This is used to assess the semantic coherence of components within a single sentence.

[0055] b) Cross-sentence semantic coherence: The entire text is treated as a graph, with each sentence as a node, and the edge weights between nodes representing the semantic similarity between sentences. The global clustering coefficient of this graph is calculated to quantify the tightness of the overall semantic network of the text; a higher value indicates a focus on specific topics and good coherence.

[0056] c) Topic Consistency: The LDA topic model is used to extract topics from the text, obtaining the topic distribution vector. Then, the Hellinger distance between the topic distribution vectors of the human text and the generated text is calculated. The smaller the distance, the closer the core topics discussed are.

[0057] Understandably, by extracting features from four dimensions deeply related to the expressive characteristics of dreams—narrative, emotion, vocabulary, and semantics—it is possible to comprehensively and from multiple perspectives depict the intrinsic attributes of a dream text. This provides a solid and rich quantitative basis for subsequent refined similarity comparisons, avoids the one-sidedness of a single text similarity index, and significantly improves the professionalism and depth of the assessment.

[0058] Specifically, calculating the multidimensional similarity assessment result includes: for each extracted text feature, calculating the similarity between the humanoid dream text and the human dream text on that feature dimension to form a similarity vector; and using an information entropy-based weighting method, combined with a preset empirical adjustment factor, assigning weights to the similarity of each dimension in the similarity vector to calculate a comprehensive multidimensional similarity assessment score.

[0059] It should be further explained that calculating the multidimensional similarity assessment result is the core step of the DreamScore framework, and its final output is a comprehensive score, namely the DreamScore. This process consists of two sub-steps: (1) Constructing a multidimensional similarity vector: It is assumed that m dimensions of features were extracted. For the i-th feature dimension (i=1,2,…,m), the human dream text is calculated respectively. and generating dream text Eigenvalues ​​in this dimension and Then, through a dimension-specific similarity function... Calculate the similarity between the two. : in, The function is chosen based on the feature type. For example, cosine similarity can be used for sentiment vectors; 1 minus Jensen-Shannon divergence can be used for distributional features (such as syntactic distribution); and 1 minus the standardized absolute difference can be used for scalar features (such as TTR). This ultimately yields an m-dimensional similarity vector. .

[0060] (2) Comprehensive weighting based on entropy weighting method and empirical adjustment factor in order to weight the multidimensional similarity vector To synthesize a single scalar score, DreamScore, appropriate weights need to be assigned to each dimension. This method uses an objective entropy weighting method combined with a subjective experience adjustment factor to determine the final weights. The specific design includes: a) First, collect n sets of evaluation samples (human text, generated text) to form a similarity matrix. ,in It is the similarity of the j-th sample in the i-th dimension.

[0061] b) For the matrix Standardization is performed to obtain the standardized matrix. For benefit-type indicators (the higher the similarity, the better), the standardized formula is: ,in and These are the minimum and maximum values ​​among all samples in the i-th dimension, respectively.

[0062] c) Calculate the contribution (weight) of the j-th sample under the i-th dimension. : d) Calculate the information entropy of the i-th dimension. : in, Used for standardization to ensure Between [0,1]. It is the information entropy of the i-th dimension, which measures the degree of dispersion of the data distribution in that dimension. It represents the contribution of the j-th sample in the i-th dimension.

[0063] e) Calculate the information entropy redundancy of the i-th dimension. : . The larger the value, the more information that dimension provides, and the greater its weight should be given in the evaluation.

[0064] f) Calculate the initial weights based on the entropy weight method : g) Introducing empirical adjustment factors . This is a pre-defined weighting system, assigned by domain experts based on the prior importance of each evaluation dimension in dream similarity assessment. For example, "emotional valence" and "semantic consistency" can be considered more important than "lexical features," thus assigning them greater weight. Value. All Must meet .

[0065] h) Calculate the final composite weight. It is a weighted harmonic sum of entropy weights and empirical adjustment factors: in, It is a harmonic coefficient between 0 and 1, with a preferred value of 0.7. (Select) The reason is that in similarity assessment, we tend to rely more on the objective importance revealed by the data itself (entropy weight method), while supplementing it with a small amount of domain knowledge for fine-tuning. The weight allocation of 0.7 has been shown in most experiments to achieve a good balance between objectivity and expert prior knowledge.

[0066] i) Ultimately, the DreamScore overall score Calculated by the following formula: The higher the value, the higher the overall similarity between the generated text and human text.

[0067] Understandably, the entropy weighting method objectively allocates basic weights based on the degree of variation in each dimension of data, avoiding the subjective arbitrariness of manually setting weights. At the same time, the introduction of an empirical adjustment factor ensures the guiding role of domain expertise in the final evaluation, making the DreamScore score both objective and reliable, and consistent with professional understanding in dream research. This combined subjective and objective weighting method significantly enhances the scientific rigor and persuasiveness of the evaluation results.

[0068] Specifically, the calculation of the similarity vector includes at least three levels of evaluation: surface vocabulary layer evaluation, which calculates the matching degree based on the surface vocabulary of the text; deep semantic layer evaluation, which calculates semantic similarity based on the vector embedding representation of the text; and generative model heuristic evaluation layer, which uses a generative language model to score the consistency of the two texts in the dream narrative logic.

[0069] It should be further explained that the similarity calculation function for each feature dimension mentioned in this invention... This method further implements it into three evaluation levels from shallow to deep, forming a multi-level, multi-granular similarity measurement system: (1) Surface lexical layer assessment This level focuses on lexical overlap and word order matching on the text surface. It is primarily used to calculate certain metrics within the "lexical features" framework, and its specific design includes: a) For comparing "vocabulary richness", the similarity of the TTR values ​​of the two can be directly calculated: . It is the ratio of human text types. It is the ratio of the types of text generated.

[0070] b) For simple word frequency distribution comparisons, the Jaccard similarity coefficient can be used to compare the overlap of the high-frequency word sets of the two.

[0071] This layer's evaluation calculation is simple and fast, and can capture the most intuitive language similarity, but it is not sensitive to synonym substitution, word order changes, etc.

[0072] (2) Deep semantic layer evaluation This layer, by mapping text to a deep semantic space for comparison, can capture deep features such as "semantic consistency" and "emotional valence." Specific design elements include: a) For the "emotional valence" similarity, calculate the six-dimensional emotion vector. and Cosine similarity: .

[0073] b) For “intra-sentence semantic similarity” and “cross-sentence semantic coherence”, which are already calculation results based on sentence vectors, they can be directly used as the output of this layer’s evaluation.

[0074] c) For deep semantic similarity of the entire text, the [CLS] token vectors from models such as BERT or the document vectors obtained by average pooling all token vectors can be used, and then cosine similarity can be calculated. This similarity can serve as a supplementary or alternative measure to "topic consistency". This layer of evaluation effectively overcomes lexical differences and understands the deeper meaning of the text, making it the core of similarity evaluation.

[0075] (3) Generative Model Heuristic Evaluation Layer: This layer utilizes the generative language model itself as a "judge" to evaluate the consistency between two texts in terms of more abstract narrative logic, rationality, and dream characteristics. This part is particularly suitable for evaluating "narrative structure features." In specific implementation, a specific prompt can be constructed, requiring the generative language model (such as GPT-4) to act as a rater and score the degree of conformity between the generated text and the reference human text from the perspectives of "the logic of event development," "the rationality of plot twists," and "whether it conforms to the absurdity of dreams." For example, the prompt could be: "Please rate the following 'generated dream' on a scale of 0 to 10, evaluating its similarity to the 'reference dream' in terms of narrative logic and dream style. Reference dream: [human text]. Generated dream: [generated text]. Please output only a score between 0 and 10." The obtained score is standardized to the [0,1] interval as the evaluation result of this layer. This layer of evaluation leverages the implicit knowledge of powerful generative models, enabling it to capture narrative similarities that are difficult to quantify in the first two layers and are more in line with human intuition.

[0076] Understandably, by integrating surface, deep, and heuristic evaluation layers, this method constructs a three-dimensional similarity measurement network. The surface layer ensures matching of basic linguistic forms, the deep layer guarantees semantic similarity, and the heuristic layer introduces higher-order evaluation based on the "intelligence" of a large model, thus improving the final similarity vector. It can comprehensively and robustly reflect the similarity between two dream texts from multiple cognitive levels, greatly enhancing the robustness of the assessment system and its relevance to human subjective judgment.

[0077] Specifically, the human-like dream text generated by the generative language model is a predicted dream text generated based on multi-source information collected from the target individual during waking hours. The multi-source information during waking hours includes at least one of the following: individual experience information, emotional state information, daily life behavior information, social relationship and interpersonal interaction information, and information contact or media exposure information.

[0078] It should be further clarified that the "human-like dream text" evaluated in this method is not limited to the model's free generation based on a simple thematic cue word. A more important application scenario is evaluating the model's ability to "predict" an individual's dream content based on their waking-time information. Specifically, this includes: (1) Individual experience information: refers to specific events that an individual experiences during the day, such as "attending an important meeting" or "having a dispute with a friend". Key elements of these events (people, places, activities) may enter the dream.

[0079] (2) Emotional state information: refers to an individual's persistent or primary emotional state during the day, such as "anxiety", "excitement" or "calm", which is obtained through heart rate variability (HRV) analysis using wearable devices or daily mood self-rating scales.

[0080] (3) Daily life behavior information: Data obtained through mobile phone sensors, calendar applications, etc., such as daily routine, geographical location movement trajectory, amount of exercise, etc.

[0081] (4) Social relations and interpersonal interaction information: Metadata from communication records and social media interactions are used to analyze daytime social activity, interaction targets, etc.

[0082] (5) Information exposure or media exposure: web page content browsed, video / movie types watched, books and articles read, etc. After being structured, this multi-source information is used as a rich contextual cue and input into a finely tuned generative language model. The model's task is to generate a predicted dream text that the individual might experience that night, based on information such as "the individual experienced event A on that day, the main emotion was B, there was interaction with C, and the individual watched a movie of type D in the evening...". Subsequently, the predicted text and the dream text (human text) that the individual actually recalled and recorded the next morning are input into the DreamScore evaluation system of this patent to calculate a similarity score, thereby scientifically quantifying the accuracy of the dream prediction model.

[0083] Understandably, this evaluation method achieves a closed loop from "prediction" to "evaluation." This makes the method more than just an isolated text similarity comparison tool; it becomes a key evaluation benchmark for measuring and driving progress in the cutting-edge research direction of "personalized dream prediction," greatly enhancing the application value and industrial relevance of this invention.

[0084] Specifically, the emotional valence features and narrative structure features in the multidimensional text features can be weighted or calibrated using biosignal features associated with the time of acquisition of the human dream text, which are extracted from electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, or electrodermal signals.

[0085] It should be further noted that, to further improve the physiological validity of emotion and narrative assessments, this method allows the introduction of biosignal data collected synchronously with dream reports to calibrate features obtained from text analysis. The specific implementation is as follows: (1) Biosignal acquisition: While the subject is reporting their dreams (orally or in writing), their multi-channel physiological signals are recorded, including but not limited to electroencephalogram (EEG), electrocardiogram (ECG / HRV), respiration, and skin conductance (GSR). It is preferred to record when the dream segment with "intense emotions" is reported.

[0086] (2) Feature extraction and association: a) Emotional Calibration: Frequency domain metrics of heart rate variability (HRV), such as the LF / HF ratio, are extracted from ECG signals, and skin conductance level (SCL) and skin conductance response (SCR) counts are extracted from GSR signals. Studies have shown that these metrics are correlated with emotional arousal. A simple linear or nonlinear model (such as support vector regression) can be constructed to correlate the overall strength of the six-dimensional emotion vector obtained from text analysis (e.g., the L2 norm of the vector) with biosignal features (such as the SCR count). In subsequent evaluation, if the "emotion vector strength" of a generated text does not match the "biosignal response" predicted based on its content, the emotional valence similarity of that text is determined. Apply appropriate penalties (such as multiplying by a decay factor less than 1, for example, 0.8).

[0087] b) Narrative Structure Calibration: Features related to cognitive load and memory retrieval are extracted from EEG signals, such as frontal lobe theta wave power and alpha wave desynchronization degree. In dream narratives, more complex and abrupt plots may correspond to specific EEG patterns. The correlation between the "plot coherence feature" value and synchronized EEG features in human dream text can be analyzed. If the "plot coherence" value of the generated text is abnormally high or low, but does not conform to typical dream EEG patterns, its narrative structure similarity is fine-tuned. For example, if EEG features show high cognitive load (possibly corresponding to bizarre plots), and the plot coherence calculation value of the generated text is also high (meaning very coherent), this may be contradictory, and its similarity weight on this feature needs to be reduced.

[0088] (3) Optimization of calibration coefficient: The intensity of biosignal calibration can be determined by a coefficient. (0≤) ≤1) is used for control. When At times, it relies entirely on text features; when At this time, biosignal calibration plays its most significant role. Experimental verification has shown that... A value between 0.3 and 0.6 can effectively improve the physiological relevance of the assessment without over-reliance on expensive biological equipment. The preferred value is 0.4.

[0089] Understandably, by introducing biosignals to cross-validate and calibrate text features, this method transcends the scope of pure text analysis, establishing a correlation between "subjective reports and objective physiological indicators." This allows similarity assessment to be based not only on "what was said" but also indirectly on "the mental and physical state at the time of speaking," thus more closely aligning with the essence of dreams as a product of the mind and body. This significantly enhances the scientific rigor and cross-modal validity of the assessment results, providing more reliable input for subsequent potential mental state-assisted analysis.

[0090] Specifically, the evaluation results generated by the method are further used to drive the multimodal content generation process of generating corresponding images or videos based on dream text content, or associated with the results of non-diagnostic state characteristic analysis of individuals based on waking information and dream characteristics, so as to form a component of a comprehensive technical framework for systematically analyzing, predicting, generating and quantitatively evaluating dream narratives.

[0091] It should be further explained that the dream text similarity evaluation method based on multi-dimensional indicators in this patent is not an isolated module, but a key link embedded in a larger, comprehensive technical framework that runs through the entire process of "dream narrative". Specifically: (1) The DreamScore generated by this method can be used as a quality indicator for filtering or ranking generated content. For example, when using a text-to-image / video model (such as Stable Diffusion, Sora) to generate visualizations of a dream description, multiple alternative visualizations can be generated simultaneously (corresponding to different random seeds or fine-tuning of prompts). Then, each visualization is reverse-engineered using an image description model (such as BLIP) to generate a text description, and this description is then input into this evaluation system to calculate the similarity with the original dream text. The visualization with the highest DreamScore is selected as the final output, thereby ensuring that the generated content is highly consistent with the original dream in semantics.

[0092] (2) The multidimensional text features extracted by this method (such as emotional valence, narrative coherence, and lexical complexity) and the final DreamScore (used to assess predicted dreams) can serve as important input features for the "mental state-assisted assessment" system. For example, a long-term dream report shows that the "fear" dimension in the emotional valence vector is consistently high and the narrative coherence is significantly reduced. Combined with its waking information, the system may provide a non-diagnostic state feature description such as "recent stress levels are high, and it is recommended to pay attention to anxiety." On the other hand, a consistently low DreamScore in a dream prediction model may reflect that the model has failed to capture subtle changes in an individual's recent cognitive or emotional state.

[0093] (3) Closed-loop framework: The above three stages (prediction, evaluation, and generation) together with state analysis constitute a complete technical closed loop of "collecting information during lucidity -> predicting dreams -> recording / evaluating actual dreams -> visualizing dreams -> analyzing state characteristics". The similarity evaluation method of this patent is the core bridge connecting prediction, generation and objective quantitative evaluation in this closed loop.

[0094] Understandably, placing the evaluation method of this invention within such a larger systematic framework greatly highlights its technological advancement and irreplaceability. It is no longer a simple text comparison tool, but has become a core quantitative engine connecting cutting-edge technologies in dream research (prediction, generation, and analysis), significantly expanding the scope of protection and application scenarios of this invention, and enhancing the overall solution's creativity and industrial value.

[0095] Specifically, the method further includes: storing the multidimensional similarity evaluation results and difference analysis results, and associating the results with the corresponding generative language model identifiers, input conditions and text preprocessing parameters to construct an evaluation database for model performance comparison and optimization.

[0096] It should be further noted that, in order to support large-scale model research, performance tracking, and iterative optimization, this method must include a structured evaluation database construction and management module. The specific steps are as follows: (1) Data storage structure: Each database record should contain the following fields: a) Unique record ID.

[0097] b) Generative language model identifier: Indicates the name and version number of the model used to generate the text (e.g., GPT-4-0125-preview).

[0098] c) Input conditions: including specific cue words used to generate the dream, associated waking-time multi-source information summaries (if any), and any special generation parameters (such as temperature, top_p).

[0099] d) Text preprocessing parameters: Record the specific tool version, stop word list version, word vector model name, etc. used in the preprocessing process.

[0100] e) Multidimensional feature values: Stored as structured data (such as JSON format), containing all feature values ​​extracted from human text and generated text.

[0101] f) Similarity vector : Storing the calculated similarity across various dimensions .

[0102] g) Final DreamScore : The overall score obtained from storage and calculation.

[0103] h) Difference analysis results: Store a summary of the comparative analysis results of different models in dimensions such as emotion and theme.

[0104] i) Timestamp: Records the time when the evaluation was completed.

[0105] (2) Database functions: a) Performance comparison: Users can quickly compare the average DreamScore and scores of each dimension of different models (such as GPT-4 vs Claude-3) under the same or different input conditions by querying the database, and generate a performance comparison report.

[0106] b) Optimization guidance: Model developers can analyze low-scoring cases of their models to identify weak dimensions (e.g., generally low scores on "narrative structure") and then make targeted adjustments to the model architecture, training data, or hint engineering.

[0107] c) Benchmark set construction: A batch of representative human dream texts and their corresponding multi-model generation results can be selected from the database to form a standard dream generation model benchmark set for unified use by academia and industry.

[0108] (3) Preferred implementation scheme: The evaluation database is preferably implemented using a relational database (such as MySQL, SQLite) or a document database (such as MongoDB), and provides a web-based query and visualization interface to facilitate interactive exploration of data by users.

[0109] Understandably, by systematically storing and managing all evaluation data, this method not only completes a one-time similarity assessment but also accumulates a valuable knowledge base of dream generation model performance. This provides a data foundation for continuously tracking the progress of generative language models in simulating human dreams, making model optimization based on evidence, and allowing the evaluation system itself to be iteratively updated with data accumulation (such as adjusting the empirical adjustment factor in the entropy weight method), forming a self-reinforcing virtuous cycle.

[0110] This invention provides another embodiment, which offers a dream text similarity evaluation system based on multidimensional indicators. The dream text similarity evaluation system based on multidimensional indicators includes: (1) Data acquisition module, configured to acquire at least one human dream text and at least one human-like dream text generated by a generative language model.

[0111] It should be further noted that this system is an implementation of the aforementioned methods in both software and hardware. Each module can be implemented through independent software functional units, hardware chips, or a combination thereof. The data acquisition module is a component with network interface and API call capabilities, capable of receiving or acquiring dream text data from local files, databases, or remote servers. For predicting dream scenarios, this module also needs to interface with the "Multi-Source Information Acquisition and Preprocessing System During Lucidity".

[0112] (2) Text preprocessing module, configured to perform unified text preprocessing on the human dream text and the humanoid dream text.

[0113] It should be further noted that the text preprocessing module integrates a sentence segmenter, a stop word filter, a part-of-speech tagger, and a word vector embedding model. During deployment, pre-trained model files (such as BERT model parameters) can be loaded to provide efficient text vectorization services.

[0114] (3) Multidimensional feature extraction module, configured to extract multidimensional text features from the preprocessed text, including at least narrative structure features, emotional valence features, lexical features and semantic consistency features.

[0115] It should be further explained that the multidimensional feature extraction module is a computationally intensive component, which can be further divided into narrative structure analysis, sentiment analysis, lexical analysis, and semantic analysis submodules. Each submodule runs a corresponding algorithm, such as dependency parsing, sentiment analysis model, lexical statistical function, sentence encoding model, etc., to process the input text in parallel or sequentially, and finally output a structured feature vector.

[0116] (4) Similarity assessment module, configured to calculate the multidimensional similarity assessment result between the humanoid dream text and the human dream text based on the multidimensional text features.

[0117] It should be further explained that the similarity assessment module consists of two core computational units. The first unit is responsible for calculating the multidimensional similarity vector, calling different similarity calculation functions (such as cosine similarity, edit distance, and Jensen-Shannon divergence). The second unit implements the entropy weight method for weight calculation and the DreamScore comprehensive score. It needs to preload or receive a historical evaluation sample set to calculate the initial entropy weight and store the preset empirical adjustment factor and harmonic coefficient β.

[0118] (5) The result output module is configured to output the difference analysis results of human-like dream texts generated by different generative language models in terms of emotion dimension and theme dimension based on the multidimensional similarity evaluation results.

[0119] It should be further explained that the results output module is responsible for outputting the calculated DreamScore, similarity scores for each dimension, and comparative analyses based on this data (such as generating bar charts, radar charts, and model ranking lists) in the form of visual reports (HTML pages, PDFs) or structured data (JSON). This module can also directly interface with the API of the "Dream Visualization Generation System" or the "Mental State Assisted Assessment System" to transmit the assessment results to downstream systems in real time.

[0120] Understandably, through modular design, this system decomposes the complex dream similarity evaluation process into highly cohesive, loosely coupled functional components, making the system easy to develop, maintain, and extend. Each module can be independently optimized and upgraded (e.g., upgrading to a better sentiment analysis model) without affecting other modules. The implementation of this system enables the aforementioned advanced multi-dimensional index-based dream text similarity evaluation method to be deployed in a stable, efficient, and serviceable manner, serving multiple application areas such as dream research, artificial intelligence model evaluation, and digital health.

[0121] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the multi-dimensional index-based dream text similarity evaluation method. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0122] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0123] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0124] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0125] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for evaluating dream text similarity based on multidimensional indicators, characterized in that, Includes the following steps: S1. Obtain at least one human dream text and at least one human-like dream text generated by a generative language model; S2. Perform text preprocessing on the human dream text and the humanoid dream text; extract multi-dimensional text features from the preprocessed text, wherein the multi-dimensional text features include at least narrative structure features, emotional valence features, lexical features and semantic consistency features; S3. Based on the multidimensional text features, calculate the multidimensional similarity assessment result between the humanoid dream text and the human dream text; S4. Based on the multidimensional similarity evaluation results, perform a difference analysis on the emotion dimension and the theme dimension of the humanoid dream text generated by different generative language models, and output the evaluation results.

2. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1, characterized in that, The text preprocessing includes segmenting the human dream text and the humanoid dream text into sentences, removing stop words, performing part-of-speech tagging, and word vector representation to obtain a unified text representation.

3. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1, characterized in that, The narrative structure features include event sequence features, plot coherence features, and syntactic structure distribution features; the emotional valence features are intensity vectors based on sentiment analysis of the text, containing multiple basic emotional dimensions; the lexical features include lexical richness, word frequency distribution, and lexical complexity; and the semantic consistency features include intra-sentence semantic similarity, cross-sentence semantic coherence, and thematic consistency.

4. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1 or 3, characterized in that, The calculation of the multidimensional similarity assessment results includes: For each extracted text feature, the similarity between the humanoid dream text and the human dream text in that feature dimension is calculated to form a similarity vector. Then, a weighting method based on information entropy is used, combined with a preset empirical adjustment factor, to assign weights to the similarity of each dimension in the similarity vector, so as to calculate a comprehensive multidimensional similarity evaluation score.

5. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 4, characterized in that, The calculation of the similarity vector includes at least three levels of evaluation: Surface vocabulary layer evaluation is calculated based on the matching degree of surface vocabulary in the text; Deep semantic layer evaluation calculates semantic similarity based on the vector embedding representation of the text; The generative model heuristic evaluation layer uses a generative language model to score the consistency of two texts in the dream narrative logic.

6. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1, characterized in that, The human-like dream text generated by the generative language model is a predicted dream text generated based on multi-source information collected from the target individual during waking hours. The multi-source information during waking hours includes at least one of the following: individual experience information, emotional state information, daily life behavior information, social relationship and interpersonal interaction information, and information contact or media exposure information.

7. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1, characterized in that, The emotional valence features and narrative structure features in the multidimensional text features can be weighted or calibrated using biosignal features associated with the time of acquisition of the human dream text, which are extracted from electroencephalogram (EEG), electrocardiogram (ECG), respiratory signals, or electrodermal signals.

8. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1, characterized in that, The evaluation results generated by the method are further used to drive the multimodal content generation process of generating corresponding images or videos based on dream text content, or to be associated with the results of non-diagnostic state characteristic analysis of individuals based on waking information and dream characteristics, so as to form a component of a comprehensive technical framework for systematically analyzing, predicting, generating and quantitatively evaluating dream narratives.

9. The dream text similarity evaluation method based on multi-dimensional indicators according to claim 1, characterized in that, The method further includes: The multidimensional similarity evaluation results and difference analysis results are stored, and the results are associated with the corresponding generative language model identifiers, input conditions, and text preprocessing parameters to construct an evaluation database for model performance comparison and optimization.

10. A dream text similarity evaluation system based on multi-dimensional indicators, characterized in that, include: The data acquisition module is configured to acquire at least one human dream text and at least one human-like dream text generated by a generative language model; The text preprocessing module is configured to perform unified text preprocessing on the human dream text and the humanoid dream text; The multidimensional feature extraction module is configured to extract multidimensional text features from the preprocessed text, including at least narrative structure features, emotional valence features, lexical features, and semantic consistency features. The similarity assessment module is configured to calculate the multidimensional similarity assessment result between the humanoid dream text and the human dream text based on the multidimensional text features; The results output module is configured to output the difference analysis results of human-like dream texts generated by different generative language models in terms of emotion and theme dimensions, based on the multidimensional similarity evaluation results.

Images