Instrument poem analysis method based on historical data

Abstract

The invention relates to the technical field of data analysis and artificial intelligence, in particular to a historical data-based instrumental poem analysis method, which comprises the following steps of: acquiring multi-modal original data containing audio signals, instrumental work symbols and text evaluation data from a historical instrumental work database; performing standardized preprocessing to form a structured analysis data set; based on the data set, extracting multi-dimensional features such as timbre variance, force gradient, duration offset, pitch distribution, rhythm entropy, chord distance, emotion score and vocabulary distribution statistics, and constructing a comprehensive poem concept feature vector; and training the training set marked with the poem concept style label by using a deep neural network, constructing a poem concept analysis model, finally realizing poem concept quantitative scoring and style similarity calculation of the instrumental works, and visually outputting a result. According to the method, the defects of strong subjectivity and single dimension of attribute analysis of abstract instrumental works in the prior art are overcome, and objective, multi-dimensional and interpretable automatic analysis of poems and concepts of the instrumental works is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data analysis and artificial intelligence, in particular to a method for analyzing poetic conception of instrumental music based on historical data. BACKGROUND

[0002] As an art form, the expression of "poetic conception" of instrumental music is one of the important dimensions to measure the artistic value of the work. Traditionally, the understanding and evaluation of the poetic conception of instrumental music highly depends on the subjective judgment of professional critics or scholars, and lacks standardized and quantifiable analysis methods. With the in-depth application of artificial intelligence technology in the field of instrumental music, especially the development of multi-modal data processing and machine learning methods, it has become possible to build a computational model to automatically identify and quantify the aesthetic features of instrumental music.

[0003] Currently, there have been some related technologies in the field of computational analysis of instrumental music. For example, feature extraction methods based on audio signals, such as MFCC, spectral centroid, and rhythm features, are widely used in instrumental music emotion recognition or style classification. Analysis methods based on symbolic data focus on statistical modeling of pitch, rhythm, and harmonic structure. In addition, sentiment analysis technology combined with text reviews is also used to assist in understanding the content of instrumental music. Some systems such as MusicBrainz and AcousticBrainz provide instrumental music feature extraction and retrieval services, and some research attempts to use multi-modal data for instrumental music recommendation or generation tasks.

[0004] However, the existing technology still has the following obvious limitations: most methods only rely on single modal data and fail to fully integrate multi-source information, resulting in limited ability to capture complex aesthetic attributes such as "poetic conception"; even with multi-modal data, there is a lack of specialized feature design for poetic conception expression, such as dynamic changes in timbre, rhythm freedom, and non-traditional harmonic proportion, which are not systematically integrated; existing models often focus on classification or retrieval tasks rather than providing interpretable quantitative scores and style similarity analysis, making it difficult to support in-depth instrumental music research and applications; these problems limit the practicality and accuracy of existing technology in the analysis of poetic conception of instrumental music. The present application is a systematic solution to these deficiencies. SUMMARY

[0005] In view of the above shortcomings of the prior art, the present application provides a method for analyzing the poetic conception of instrumental music based on historical data, which can effectively solve the problems mentioned in the prior art.

[0006] To achieve the above purpose, the present application is realized by the following technical scheme:

[0007] The application provides a method for instrumental poetic analysis based on historical data, comprising the following steps:

[0008] S1, obtaining original data of multiple instrumental works from a historical instrumental work database; the original data includes audio signal data, instrumental work symbol data, and text evaluation data related to the works;

[0009] S2, standardizing and preprocessing the original data to form an analysis data set, specifically comprising:

[0010] Cleaning the obtained original data to remove invalid or damaged data segments; for multiple data forms of the same instrumental work, performing time axis alignment processing to ensure the consistency of the audio signal, the instrumental work symbol, and the text evaluation data in time sequence;

[0011] Audio signal standardization: performing noise reduction, normalized volume amplitude processing on the audio signal data, and uniformly converting to the same sampling rate and bit depth to generate standard format audio waveform data;

[0012] Instrumental work symbol standardization: analyzing and converting the instrumental work symbol data into a unified internal data representation structure, converting all tonalities to C major or a minor to eliminate tonal differences, and quantizing all rhythm lengths to the smallest integer multiple of standard note values to generate standardized symbol sequence data;

[0013] Text data standardization: cleaning the text evaluation data to remove stop words and irrelevant symbols; using natural language processing technology for stem extraction and lemmatization, and converting all texts to unified encoding format lowercase texts to generate standardized text corpus data;

[0014] Associating and storing the preprocessed standardized audio waveform data, standardized symbol sequence data, and standardized text corpus data according to the unique identification of the works to form a structured analysis data set;

[0015] S3, based on the analysis data set, extracting a multi-dimensional feature set for quantifying instrumental poetic concepts; the multi-dimensional feature set includes:

[0016] Audio level features extracted from the audio signal data, including timbre variance features, dynamic gradient features, and duration offset features; structural level features extracted from the instrumental work symbol data, including interval distribution features, rhythm entropy value features, and chord distance features; semantic level features extracted from the text evaluation data, including emotion score features and vocabulary distribution statistical features;

[0017] S4, calculating a comprehensive poetic concept feature vector for each instrumental piece based on the multi-dimensional feature set; training the poetic concept feature vector based on a training set labeled with historical prior knowledge by using a machine learning algorithm, and constructing a poetic concept analysis model;

[0018] S5, inputting the instrumental piece to be analyzed into the trained poetic concept analysis model, outputting a quantitative score of the poetic concept of the piece, and performing poetic concept style similarity calculation based on the poetic concept feature vectors of multiple pieces;

[0019] S6, outputting the quantitative score of the poetic concept and the results of the style similarity analysis in a visual form.

[0020] Further, the multi-dimensional feature set for quantifying the poetic concept of the instrumental piece extracted in step S3 specifically includes:

[0021] The timbre variance feature is obtained by calculating the variance of the Mel Frequency Cepstral Coefficient (MFCC), and the specific calculation formula is:

[0022] ;

[0023] Wherein, is the timbre variance feature, is the MFCC coefficient vector of the t-th frame of audio, is the mean vector of all frame MFCC coefficients, is the total number of frames;

[0024] The dynamic gradient feature is obtained by calculating the gradient standard deviation of the root mean square (RMS) energy of the audio signal, and the specific calculation formula is:

[0025] ;

[0026] Wherein, is the dynamic gradient feature, is the energy change of adjacent frames, is the mean energy change, is the total number of frames;

[0027] The duration offset feature is obtained by calculating the standard deviation of the local time offset between the actual performance audio and the standard rhythm of the score, and the specific calculation formula is:

[0028] ;

[0029] Wherein, is the duration offset feature, is the difference between the actual performance duration and the standard duration of the i-th note, is the total number of notes, the mean value of the time offset;

[0030] The interval distribution feature is quantified by calculating the melody interval smoothness index, and the specific formula is:

[0031] ;

[0032] Wherein, is the interval distribution feature, is the pitch difference between adjacent notes, is the total number of melody notes;

[0033] The rhythm entropy value feature is quantified by calculating the information entropy of the rhythm pattern, and the specific formula is:

[0034] ;

[0035] Wherein, is the rhythm entropy value feature, is the frequency of the jth rhythm pattern in the music, and R is the total number of rhythm pattern categories;

[0036] The chord distance feature is quantified by calculating the proportion of non-traditional harmony, and the specific formula is:

[0037] ;

[0038] Wherein, is the chord distance feature, is the number of non-traditional harmony performances in the work, is the total number of chord performances in the work;

[0039] The emotion score feature is calculated by the following formula:

[0040] ;

[0041] Wherein, is the emotion score feature, is the lth text data, is the emotion score output by the BERT emotion analysis model, and L is the total number of texts;

[0042] The word distribution statistical feature is extracted by calculating the word distribution-inverse document frequency TF-IDF, and the calculation formula is:

[0043] ;

[0044] Wherein, is the word distribution statistical feature, is the frequency of word w in document d, and N is the total number of documents, The number of documents containing the word w.

[0045] Further, the calculation of the poetic concept feature vector specifically includes:

[0046] The extracted timbre variance feature, dynamic gradient feature, duration offset feature, interval distribution feature, rhythm entropy value feature, chord distance feature, emotion score feature and vocabulary distribution statistical feature are subjected to z-score standardization to eliminate dimensional differences, and the calculation formula is:

[0047] ;

[0048] wherein, is the original feature value, is the mean of the feature in the training set, is the standard deviation of the feature in the training set, is the standardized feature value;

[0049] All the standardized feature values are spliced in a predetermined order to form a high-dimensional and unified mathematical vector, which is defined as a poetic concept feature vector .

[0050] Further, the step of constructing the poetic concept analysis model specifically includes:

[0051] A training set is constructed, and a number of historical instrumental works are selected from a historical instrumental work database as training samples. Each instrumental work in the training set is labeled with a poetic concept style label. The poetic concept style label includes an era style label, a genre style label and a composer style label.

[0052] Feature matrix construction is performed. The multi-dimensional feature set is extracted from each instrumental work in the training set, and all feature values are subjected to z-score standardization to construct a feature matrix , wherein is the feature dimension, is the sample number;

[0053] A machine learning algorithm is used to train a poetic concept analysis model, with the poetic concept feature vector as the input and the poetic concept style label as the output target. The machine learning algorithm uses a deep neural network structure.

[0054] The cross-validation method is used to evaluate the model performance, and the model accuracy is optimized through hyperparameter tuning. The trained model parameters are saved to form a poetic concept analysis model that can be applied to the analysis of new instrumental works.

[0055] Further, the deep neural network structure adopts an input layer, three hidden layers and an output layer, wherein the number of nodes of the input layer corresponds to the feature dimension, the hidden layers successively adopt 256, 128 and 64 nodes and use ReLU activation functions, and the number of nodes of the output layer corresponds to the number of label categories and uses a Softmax activation function.

[0056] Further, the training poetic concept analysis model adopts a small batch gradient descent algorithm, the batch size is set to 64, the loss function is cross-entropy loss, the optimizer adopts an Adam optimization algorithm, the initial learning rate is set to 0.001, the number of training rounds is set to 100 and the early stopping method is adopted.

[0057] Further, the hyperparameters of the Adam optimizer are set to , , .

[0058] Further, step S5 specifically comprises the following steps:

[0059] After the feature vector of the instrumental work to be analyzed is extracted and normalized, the trained poetic concept analysis model is input, and the poetic concept style probability distribution is output , wherein represents the probability of belonging to the i-th poetic concept style label;

[0060] Based on the poetic concept style probability distribution, a comprehensive poetic concept quantitative score S is calculated, and the calculation formula is:

[0061] ;

[0062] , wherein is a preset weight coefficient of the i-th poetic concept style;

[0063] Based on the poetic concept feature vectors of multiple works, the poetic concept style similarity is calculated, the similarity calculation adopts a cosine similarity algorithm, and the calculation formula is:

[0064] ;

[0065] , wherein and are feature vectors of two instrumental works, is the poetic concept style similarity of the two instrumental works.

[0066] The technical solution provided by the application has the following beneficial effects compared with the known prior art:

[0067] The present application changes the past analysis mode which can only be described qualitatively but cannot be compared quantitatively by fusing multi-modal data such as audio, score and text comments, and designing a series of computable features with instrumental work learning significance, and converting instrumental works into specific mathematical vectors and scores, thereby providing objective data support for instrumental work research.

[0068] The present application can accurately capture subtle and complex style differences between different eras, genres and composers by training a poetic concept analysis model based on deep learning, and present them in the form of probability distribution, and combined with cosine similarity calculation, can efficiently perform style similarity retrieval and comparison between works in a large-scale music library, greatly surpassing traditional retrieval methods based on metadata or simple audio features, and providing a powerful and efficient automated tool for instrumental work recommendation, academic research and digital humanities application. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0070] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0072] The present application will be further described in combination with the embodiments.

[0073] Embodiments:

[0074] Reference Figure 1 A method for analyzing instrumental poetry based on historical data, comprising the following steps:

[0075] S1, obtaining original data of a plurality of instrumental works from a historical instrumental work database; the original data includes audio signal data, instrumental work symbol data and text evaluation data related to the works;

[0076] S2, standardizing preprocessing of the original data to form an analysis data set, specifically including:

[0077] Cleaning the obtained original data, eliminating invalid or damaged data segments; for multiple data forms of the same instrumental work, performing time axis alignment processing to ensure the consistency of audio signals, instrumental work symbols and text evaluation data in time sequence;

[0078] Audio signal standardization: denoising, normalizing volume amplitude of audio signal data, and uniformly converting to the same sampling rate and bit depth to generate standard format audio waveform data;

[0079] Instrumental work symbol standardization: analyzing and converting instrumental work symbol data into a unified internal data representation structure, converting all tonalities to C major or a minor to eliminate tonal differences, and quantizing all rhythm lengths to the smallest integer multiple of standard note values to generate standardized symbol sequence data;

[0080] Text data standardization: cleaning text evaluation data to remove stop words and irrelevant symbols; using natural language processing techniques for stem extraction and lemmatization, and converting all text to unified encoding format lowercase text to generate standardized text corpus data;

[0081] Associating and storing the preprocessed standardized audio waveform data, standardized symbol sequence data and standardized text corpus data according to the unique identification of the work to form a structured analysis data set;

[0082] S3, based on the analysis data set, extracting a multi-dimensional feature set for quantifying the poetic concept of instrumental music; the multi-dimensional feature set includes:

[0083] Audio level features extracted from audio signal data, including timbre variance features, dynamic gradient features and duration offset features; structural level features extracted from instrumental work symbol data, including interval distribution features, rhythm entropy value features and chord distance features; semantic level features extracted from text evaluation data, including emotion score features and word distribution statistical features;

[0084] S4, based on the multi-dimensional feature set, calculating a comprehensive poetic concept feature vector for each instrumental work; using machine learning algorithms, based on a training set annotated with historical prior knowledge, training the poetic concept feature vector to build a poetic concept analysis model;

[0085] S5, inputting the instrumental work to be analyzed into the trained poetic concept analysis model, outputting the poetic concept quantitative score of the work; and based on the poetic concept feature vectors of multiple works, calculating the poetic concept style similarity;

[0086] S6. Output the quantitative scoring of poetic concepts and the results of style similarity analysis in a visual form.

[0087] In one specific embodiment, during the data acquisition and preprocessing stage, multimodal data containing audio, sheet music, and text commentary can be obtained from public instrumental works databases or professional instrumental works institutions, such as MusicNet or MAESTRO databases. The data cleaning process includes high-pass filtering for noise reduction and amplitude normalization of the audio data, with a uniform sampling rate of 44.1kHz and a bit depth of 16bit; for the sheet music data, tools such as music21 are used to parse MIDI files, converting all keys to C major or A minor, and quantizing the rhythms to integer multiples based on quarter notes; for the text data, tools such as NLTK or Spacy are used for word segmentation, stop word removal, and lemmatization, uniformly encoding it to UTF-8 format and converting it to lowercase text. Furthermore, a dynamic time warping algorithm is used to align the audio and sheet music time points, ensuring the temporal consistency of the multimodal data.

[0088] The feature extraction stage encompasses features at three levels: audio, structure, and semantics. Audio features include extracting 13-dimensional MFCC coefficients using Librosa and calculating their variance as an indicator of timbre stability; calculating the standard deviation of the RMS energy gradient to capture dynamic variations; and calculating the standard deviation of note duration shifts after dynamic time warping alignment to reflect rhythmic freedom. Structural features include the interval smoothness index to measure melodic fluency; rhythmic entropy calculated using rhythmic pattern frequencies to assess rhythmic complexity; and chord distance quantifying the degree of harmonic deviation through the proportion of non-traditional harmonics. Semantic features employ a pre-trained BERT model to perform sentiment analysis on the text and take the average score as the sentiment score. Simultaneously, the TF-IDF algorithm is used to extract high-frequency word weights related to poetic concepts.

[0089] In the feature standardization and vector construction stage, all extracted features are z-score standardized, using the mean and standard deviation of the training set to eliminate dimensional differences. The standardized features are then concatenated in a fixed order into a high-dimensional mathematical vector, defined as the poetic concept feature vector. Its dimension depends on the specific design of each feature class; for example, if each feature class is single-dimensional, an 8-dimensional vector is formed.

[0090] During the model training phase, instrumental music experts were invited to label the training set works with multiple style tags, such as era style, genre style, or composer style. A deep neural network model was constructed, with the number of nodes in the input layer equal to the feature dimension. The hidden layers used 256, 128, and 64 nodes respectively, employing the ReLU activation function. The number of nodes in the output layer corresponded to the number of label categories, and the Softmax activation function was used to output the probability distribution. The training configuration used the Adam optimizer, and hyperparameters were set accordingly. , , , the initial learning rate is 0.001, the batch size is 64, the number of training rounds is 100, and the early stopping method is used to prevent overfitting, and the loss function is cross-entropy loss. The model performance is evaluated by 5-fold cross-validation to optimize the hyperparameters, and the trained model parameters are finally saved for subsequent analysis.

[0091] In the poetic concept analysis and visualization stage, the feature vectors of the works to be analyzed are extracted and standardized, and then input into the trained model to output the poetic concept style probability distribution. The comprehensive poetic concept quantitative score is obtained by weighting and summing according to the preset weight coefficient. The cosine similarity algorithm is used to calculate the similarity between the feature vectors of multiple works for style comparison and recommendation. Visualization output uses Matplotlib or D3.js to generate radar charts to display feature dimensions, heat maps to show style similarity, and score bar charts for easy integration into instrumental music education platforms or intelligent composition assistance systems.

[0092] In one specific embodiment, the poetic concepts of Western instrumental music works in the Romantic period and late Romanticism in the 19th century are determined and analyzed. First, from historical music databases such as PHENICX, The Romantic Era PDF Corpus, and IMSLP authorized recordings and scores, the instrumental works of representative composers in the 19th century, including Beethoven, Schubert, Chopin, Liszt, Brahms, Wagner, and Mahler, are obtained. The original data includes modern digital audio recordings, machine-readable score files, and related historical reviews, academic paper abstracts, concert program descriptions, and other text data of the above works.

[0093] In the data acquisition stage, a time filter is set to select only works created between 1800 and 1900 to ensure that the analysis objects are strictly within the target historical perspective. In the standardization preprocessing stage, all audio is denoised and amplitude normalized, and converted to standard waveform files with a sampling rate of 44.1 kHz and a depth of 16 bits. Additional repair processing is required for historical recordings.

[0094] The music21 tool library is used to parse the score data. To eliminate the interference of tonality on structural feature analysis, the tonality of all works is converted to C major or a minor. The rhythm time value is quantized to an integer multiple of the basic time value with 32nd note as the smallest unit to ensure the accuracy of rhythm calculation. For text evaluation data, the collected text data may contain multiple languages such as English, German, and French. The corresponding NLP tools are used for word segmentation, stop word removal, and lemmatization. All texts are converted to UTF-8 encoded lowercase format, and special attention is paid to the preservation of words highly related to poetic, expressive, emotional, narrative, title, sublime, melancholic, and other Romantic aesthetic concepts.

[0095] Based on the pre-processed dataset, features of multiple dimensions are extracted, which are closely related to the characteristics of 19th century music, such as personal emotional expression, color and form innovation. Audio-level features include timbre variance features, which quantify the instability and expressiveness of timbre in performance by calculating the variance of MFCC coefficients; dynamic gradient features, which capture the extreme dynamic contrast and delicate gradation characteristics of romantic music by calculating the standard deviation of RMS energy gradient; and duration offset features, which quantify the core element of 19th century performance practice, free tempo, by calculating the standard deviation of the performance and standard duration offset.

[0096] Structural-level features include interval distribution features, which measure the smoothness and jumpiness of the melody by calculating the smoothness index of the melody interval; rhythm entropy features, which evaluate the complexity and unpredictability of rhythm by calculating the information entropy of the rhythm pattern; and chord distance features, which quantify the degree of innovation of harmony by calculating the proportion of non-traditional harmony, which is the soul of romantic music.

[0097] Semantic-level features include sentiment score features, which use a pre-trained multilingual BERT sentiment analysis model to analyze text reviews and calculate the average sentiment score to capture the consensus of the audience and critics on the emotional intensity of the work; and vocabulary distribution statistics features, which use TF-IDF to extract key terms in the text corpus, especially focusing on the weight of words related to 19th century poetic concepts such as lyrical, dramatic, passionate, dreamy, tragic, nationalistic, etc. These high-weight words constitute the semantic tags of the poetic style of the work.

[0098] According to the history and aesthetics of 19th century music, three style labels are labeled for the works in the training set. Time style labels include early romanticism, high romanticism, late romanticism and national school; genre style labels include character pieces such as nocturne, intermezzo, symphonic poem, program symphony and absolute sonata; and composer style labels include Chopin style, Wagner style and Brahms style.

[0099] The deep neural network model is trained using the above annotated 19th century works dataset, and the model learns the complex mapping relationship between different styles and multi-dimensional features within 19th century music. In the analysis stage, input a 19th century instrumental work to be analyzed, such as the slow movement of Mahler's Fifth Symphony, and the model will output a poetic concept style probability distribution. The quantitative score of the poetic concept is set according to the research purpose, if you want to evaluate the late romanticism characteristics of the work, you can give higher weight to the late romanticism style, and get the quantitative score after weighted summation. The style similarity calculation is to calculate the cosine similarity between the feature vectors to quantify the closeness of the works in a certain style feature.

[0100] The analysis results are presented in various visualization forms. Radar chart is used to show the scores of the work in eight feature dimensions, and the characteristics of the work in timbre change, rhythm freedom and harmonic complexity are clearly compared. The style probability bar chart intuitively shows the probability of the work belonging to each era, genre and composer style. The similar work heat map shows the other 19th century works that are most similar to the target work such as Beethoven's Moonlight Sonata in poetic concept in a work library, which can be used for music recommendation or historical research.

[0101] Through the implementation of the embodiment, the method of the application can overcome subjective impression criticism and provide a data-driven, quantifiable and comparable objective analysis tool for the poetic concept research of 19th century Western instrumental works, and deeply reveal the internal rules and aesthetic value of music art in this great historical period.

[0102] Further, the multi-dimensional feature set for quantifying the poetic concept of instrumental music extracted in step S3 specifically includes:

[0103] The timbre variance feature is obtained by calculating the variance of the Mel frequency cepstral coefficient (MFCC), and the specific calculation formula is:

[0104] ;

[0105] Wherein, is the timbre variance feature, is the MFCC coefficient vector of the t-th frame of audio, is the mean vector of all frame MFCC coefficients, is the total number of audio frames;

[0106] The dynamic gradient feature is obtained by calculating the gradient standard deviation of the root mean square (RMS) energy of the audio signal, and the specific calculation formula is:

[0107] ;

[0108] Wherein, is the dynamic gradient feature, is the energy change of adjacent frames, is the average energy change, is the total number of frames;

[0109] The duration offset feature is obtained by calculating the standard deviation of the local time offset between the actual performance audio and the standard rhythm of the score, and the specific calculation formula is:

[0110] ;

[0111] wherein, is the duration offset feature, is the difference between the actual performance duration of the i-th note and the standard duration of the score, is the total number of notes, is the average time offset;

[0112] The interval distribution feature is quantified by calculating the melody interval smoothness index, and the specific calculation formula is:

[0113] ;

[0114] wherein, is the interval distribution feature, is the pitch difference between adjacent notes, is the total number of melody notes;

[0115] The rhythm entropy value feature is quantified by calculating the information entropy of the rhythm pattern, and the specific calculation formula is:

[0116] ;

[0117] wherein, is the rhythm entropy value feature, is the frequency of the j-th rhythm pattern appearing in the music, and R is the total number of rhythm pattern categories;

[0118] The chord distance feature is quantified by calculating the proportion of non-traditional harmony, and the specific calculation formula is:

[0119] ;

[0120] wherein, is the chord distance feature, is the number of times of non-traditional harmony in the work, is the total number of chord performances in the work;

[0121] The emotion score feature is calculated by the following formula:

[0122] ;

[0123] wherein, a sentiment score feature, a text data of the first clause, a sentiment score output by the BERT sentiment analysis model, and L is a total number of texts;

[0124] The vocabulary distribution statistical feature is extracted by calculating the vocabulary distribution-inverse document frequency TF-IDF, and the calculation formula is:

[0125]

[0126] wherein, is a vocabulary distribution statistical feature, is a frequency of occurrence of the word w in the document d, and N is a total number of documents, is a number of documents containing the word w.

[0127] Further, the calculation of the poetic concept feature vector specifically includes:

[0128] The extracted timbre variance feature, intensity gradient feature, duration offset feature, interval distribution feature, rhythm entropy value feature, chord distance feature, sentiment score feature, and vocabulary distribution statistical feature are subjected to z-score standardization to eliminate dimensional differences, and the calculation formula is:

[0129]

[0130] wherein, is an original feature value, is a mean value of the feature in the training set, is a standard deviation of the feature in the training set, is a standardized feature value;

[0131] All the feature values after the standardization processing are spliced in a predetermined order to combine into a high-dimensional and unified mathematical vector, and the vector is defined as a poetic concept feature vector .

[0132] Further, the steps of constructing the poetic concept analysis model specifically include:

[0133] A training set is constructed, and a number of historical instrumental works are selected from a historical instrumental work database as training samples. A poetic concept style label is labeled for each instrumental work in the training set. The poetic concept style label includes an era style label, a genre style label, and a composer style label.

[0134] Feature matrix construction is performed. A multi-dimensional feature set is extracted for each instrumental work in the training set, and all feature values are subjected to z-score standardization processing to construct a feature matrix wherein is a feature dimension, ​​The number of samples;

[0135] A machine learning algorithm is used to train a poetic concept analysis model, with a poetic concept feature vector as input and a poetic concept style label as output target. The machine learning algorithm uses a deep neural network structure.

[0136] The model performance is evaluated using a cross-validation method, and the model accuracy is optimized through hyperparameter tuning. The trained model parameters are saved to form a poetic concept analysis model that can be applied to new instrumental music analysis.

[0137] Further, the deep neural network structure includes an input layer, three hidden layers, and an output layer. The input layer has a node number corresponding to the feature dimension, the hidden layers have 256, 128, and 64 nodes respectively and use ReLU activation functions, and the output layer has a node number corresponding to the label category number and uses a Softmax activation function.

[0138] Further, the small batch gradient descent algorithm is used to train the poetic concept analysis model, with a batch size of 64. The loss function is cross-entropy loss, the optimizer is Adam optimization algorithm, the initial learning rate is set to 0.001, the training number is set to 100 and the early stopping method is used.

[0139] Further, the hyperparameters of the Adam optimizer are set as , , .

[0140] Further, step S5 specifically includes the following steps:

[0141] For the instrumental music to be analyzed, after extracting the feature vector and standardizing it, input the trained poetic concept analysis model, and output the poetic concept style probability distribution , where represents the probability of belonging to the i-th poetic concept style label.

[0142] Based on the poetic concept style probability distribution, calculate the comprehensive poetic concept quantitative score S, and the calculation formula is:

[0143] ;

[0144] where, is the preset weight coefficient of the i-th poetic concept style;

[0145] Based on the poetic concept feature vectors of multiple works, calculate the poetic concept style similarity. The similarity calculation uses the cosine similarity algorithm, and the calculation formula is:

[0146] ;

[0147] where, and are feature vectors of two instrumental works, respectively, is a poetic concept style similarity of the two instrumental works.

[0148] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for analyzing the poetic quality of instrumental music based on historical data, characterized in that, Includes the following steps: S1. The raw data includes audio signal data, instrumental work symbol data, and text evaluation data related to the work; S2. Perform standardization preprocessing on the raw data to form an analysis dataset, specifically including: The acquired raw data is cleaned to remove invalid or damaged data segments; for multiple data formats of the same instrumental work, timeline alignment is performed to ensure the consistency of audio signals, instrumental work symbols, and text evaluation data in terms of time sequence. Audio signal standardization: The audio signal data is denoised, the volume amplitude is normalized, and it is uniformly converted into the same sampling rate and bit depth to generate audio waveform data in a standard format. Instrumental music symbol standardization: The instrumental music symbol data is parsed and converted into a unified internal data representation structure, all tonalities are uniformly converted to C major or A minor to eliminate tonal differences, and all rhythm durations are quantized to the smallest integer multiple of the standard note value to generate standardized symbol sequence data; Text data standardization: The text evaluation data is cleaned to remove stop words and irrelevant symbols; natural language processing technology is used for stemming and word form restoration, and all text is converted into lowercase text with a unified encoding format to generate standardized text corpus data; The preprocessed standardized audio waveform data, standardized symbol sequence data, and standardized text corpus data are associated and stored according to the unique identifier of the work to form a structured analysis dataset; S3. Based on the analysis dataset, extract a multidimensional feature set for quantifying the poetic concept of instrumental music; the multidimensional feature set includes: Audio-level features extracted from audio signal data include timbre variance features, velocity gradient features, and duration offset features; structural-level features extracted from instrumental work symbol data include interval distribution features, rhythm entropy features, and chord distance features; semantic-level features extracted from text evaluation data include sentiment score features and vocabulary distribution statistics. S4. Based on the multi-dimensional feature set, calculate a comprehensive poetic concept feature vector for each instrumental work; using a machine learning algorithm, based on a training set labeled with historical prior knowledge, train the poetic concept feature vector to construct a poetic concept analysis model. S5. Input the instrumental work to be analyzed into the trained poetic concept analysis model, output the quantitative score of the poetic concept of the work; and calculate the similarity of poetic concept style based on the poetic concept feature vectors of multiple works. S6. Output the quantitative scoring of poetic concepts and the results of style similarity analysis in a visual form.

2. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 1, characterized in that, The extraction of the multi-dimensional feature set for quantifying the poetic concept of instrumental music in step S3 specifically includes: The timbre variance feature is obtained by calculating the variance of the Mel-frequency cepstral coefficients (MFCCs), and the specific calculation formula is as follows: ； in, This represents the timbre variance characteristic. Let be the MFCC coefficient vector of the audio in frame t. This is the mean vector of the MFCC coefficients across all frames. This represents the total number of audio frames. The velocity gradient feature is obtained by calculating the standard deviation of the gradient of the root mean square (RMS) energy of the audio signal. The specific calculation formula is as follows: ; in, For intensity gradient features, For energy changes between adjacent frames, The average value of energy change. Total number of frames; The duration offset feature is obtained by calculating the standard deviation of the local time offset between the actual played audio and the standard rhythm of the score. The specific calculation formula is as follows: ； in, For duration offset features, The difference between the actual playing time of the i-th note and the standard playing time of the score. The total number of notes, This represents the average of the time offsets; The interval distribution characteristics are quantified by calculating the melody interval smoothness index, and the specific calculation formula is as follows: ； in, For interval distribution characteristics, The pitch difference between adjacent notes. The total number of melody notes; The rhythm entropy value feature is quantified by calculating the information entropy of the rhythm pattern, and the specific calculation formula is as follows: ； in, For rhythm entropy value features, Let R be the frequency of the j-th rhythmic pattern in the music, and R be the total number of rhythmic patterns. The chordal distance feature is quantified by calculating the proportion of non-traditional harmonics, and the specific calculation formula is as follows: ； in, As a chord distance feature, The number of times non-traditional harmonic progressions appear in the work. The total number of chord progressions in the work; The emotional score feature is calculated using the following formula: ； in, For sentiment score features, For the l-th text data, L represents the sentiment score output by the BERT sentiment analysis model, where L is the total number of texts. The vocabulary distribution statistical features are extracted by calculating the vocabulary distribution-inverse document frequency (TF-IDF), and the calculation formula is as follows: ； in, For the statistical characteristics of word distribution, Let w be the frequency of word w in document d, and N be the total number of documents. This represents the number of documents containing the word w.

3. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 1, characterized in that, The calculation of the poetic concept feature vector specifically includes: The extracted timbre variance features, dynamic gradient features, duration offset features, interval distribution features, rhythm entropy features, chord distance features, emotional score features, and vocabulary distribution statistical features are z-score standardized to eliminate dimensional differences. The calculation formula is as follows: ； in, These are the original eigenvalues. This is the mean of the feature in the training set. The standard deviation of this feature in the training set. These are the standardized eigenvalues; All standardized feature values ​​are concatenated in a predetermined order to form a high-dimensional, unified mathematical vector, which is defined as the poetic concept feature vector. .

4. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 1, characterized in that, The specific steps for constructing the poetic concept analysis model include: A training set is constructed by selecting several instrumental works from a historical instrumental works database as training samples, and labeling each instrumental work in the training set with a poetic concept style tag; the poetic concept style tag includes era style tag, genre style tag, and composer style tag. To construct the feature matrix, the multidimensional feature set is extracted for each instrumental piece in the training set, and all feature values ​​are standardized using z-score to construct the feature matrix. ,in For feature dimension, The number of samples; A machine learning algorithm is used to train a poetic concept analysis model, with poetic concept feature vectors as input and poetic concept style tags as output targets; the machine learning algorithm adopts a deep neural network structure. The model performance was evaluated using cross-validation, and the model accuracy was optimized through hyperparameter tuning. The trained model parameters were saved to form a poetic concept analysis model that can be applied to the analysis of new instrumental works.

5. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 4, characterized in that, The deep neural network structure includes an input layer, three hidden layers, and an output layer. The number of nodes in the input layer corresponds to the feature dimension. The hidden layers have 256, 128, and 64 nodes respectively and use the ReLU activation function. The number of nodes in the output layer corresponds to the number of label categories and uses the Softmax activation function.

6. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 4, characterized in that, The training model for analyzing poetic concepts uses a mini-batch gradient descent algorithm with a batch size of 64. The loss function is cross-entropy loss, the optimizer is the Adam optimization algorithm, the initial learning rate is 0.001, the number of training epochs is 100, and early stopping is used.

7. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 6, characterized in that, The hyperparameters of the Adam optimizer are set to... , , .

8. The method for analyzing the poetic quality of instrumental music based on historical data according to claim 1, characterized in that, Step S5 specifically includes the following steps: For the instrumental works being analyzed, feature vectors are extracted, standardized, and then input into a pre-trained poetic concept analysis model. The output is a probability distribution of poetic concept style. ,in This represents the probability of belonging to the i-th type of poetic concept style label; Based on the probability distribution of the poetic concept style, the comprehensive poetic concept quantitative score S is calculated using the following formula: ； in, For the preset weighting coefficient of the i-th type of poetic conceptual style; Based on the feature vectors of poetic concepts from multiple works, the similarity of poetic concepts and styles is calculated. The similarity calculation uses the cosine similarity algorithm, and the calculation formula is as follows: ； in, and These are the feature vectors of two instrumental works. The similarity in poetic concept and style between the two instrumental works.

Images