Family storm risk identification method, system and device based on artificial intelligence and medium

Abstract

The invention relates to a home violation risk identification method, system and device based on artificial intelligence and a medium. The method comprises the following steps: preprocessing audio data to obtain a voice segment; performing automatic voice recognition on the voice segments to obtain a dialogue text sequence, and performing acoustic feature extraction to obtain a time sequence acoustic feature sequence; performing key feature extraction based on context semantics and risk knowledge on the dialogue text sequence to generate a semantic feature vector; performing deep emotion mode learning on the time sequence acoustic feature sequence to generate an acoustic feature vector; performing multi-modal fusion on the semantic feature vector and the acoustic feature vector to obtain a fusion feature vector; and carrying out collaborative risk judgment on the fused feature vector to generate a result containing high, medium and low risk levels and corresponding judgment confidence coefficients. By adopting the method, the limitation of single modal analysis can be overcome, and the home violence risk can be identified more comprehensively and accurately.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method, system, device and medium for identifying domestic violence risks based on artificial intelligence. Background Technology

[0002] With the popularization of smart home and IoT technologies, utilizing audio acquisition devices deployed in the home environment (such as smart speakers and security system microphones) to monitor abnormal situations and thus identify the risk of domestic violence has become an important technological development direction in the field of digital security. Related technologies aim to replace or assist traditional human intervention through automated analysis, achieving early warning of potential dangers.

[0003] In existing technologies, audio-based domestic violence risk identification mainly follows two technical paths. The first path focuses on acoustic feature analysis, which directly extracts the physical characteristics of audio signals (such as pitch, volume, and spectrum) and uses pattern classification algorithms to determine whether the sound event belongs to categories such as "argument," "fight," or "cry for help." The second path focuses on semantic content analysis, which first converts speech into text and then uses natural language processing techniques to detect the presence of specific keywords related to violence or threats or to determine sentiment polarity in the text.

[0004] However, the aforementioned existing technical approaches all have significant limitations, resulting in insufficient accuracy and low reliability in practical applications. Specifically, methods relying solely on acoustic features are highly susceptible to false positives due to common non-violent high-decibel sounds in the home environment (such as heated debates, children crying, and movie sound effects); simultaneously, they struggle to identify forms of violence with inconspicuous acoustic features, such as verbal coercion and emotional abuse, leading to missed detections. Methods relying solely on semantic analysis are limited by the accuracy of speech-to-text conversion, and when faced with subtle, ironic, or context-dependent threatening language, the lack of understanding of paralinguistic information such as tone and emotion also leads to serious misjudgments and missed detections. Summary of the Invention

[0005] Therefore, it is necessary to provide an AI-based method, system, device, and medium for identifying domestic violence risks that can comprehensively understand the content of the dialogue and the emotional state, thereby more accurately and reliably identifying the risks.

[0006] Firstly, this application provides an artificial intelligence-based method for identifying domestic violence risks, including:

[0007] S1. Preprocess the acquired raw audio data to obtain the speech segment to be analyzed;

[0008] S2. Perform automatic speech recognition on the speech segment to obtain the corresponding dialogue text sequence;

[0009] S3. Extract key features from the dialogue text sequence based on contextual semantics and risk knowledge to generate semantic feature vectors;

[0010] S4. Extract acoustic features from the speech segment to obtain a temporal acoustic feature sequence;

[0011] S5. Perform deep emotion pattern learning on the temporal acoustic feature sequence to generate acoustic feature vectors;

[0012] S6. Perform multimodal fusion on the semantic feature vector and the acoustic feature vector to obtain the fused feature vector;

[0013] S7. Perform collaborative risk assessment on the fused feature vectors to generate domestic violence risk levels and corresponding assessment confidence levels; domestic violence risk levels include high risk, medium risk and low risk.

[0014] In one embodiment, S3 includes:

[0015] S11. Based on a pre-built domestic violence risk knowledge base, risk words and context rules are matched to the dialogue text sequence to obtain labeled text with risk type tags.

[0016] S12. Perform context semantic embedding modeling on the annotated text to obtain the context vector representation of each word;

[0017] S13. Using sentences and phrases from the dialogue text sequence as nodes, calculate edge weights based on the semantic similarity of each node and the number of shared risk words to construct a text semantic graph; the edge weights between nodes are calculated using the following formula:

[0018]

[0019] in, The edge weights between nodes. For semantic similarity between nodes, The normalized number of risk terms shared among nodes. This is the balance coefficient;

[0020] S14. Iteratively calculate the importance score of each node based on the text semantic graph; the update formula for the node importance score is:

[0021]

[0022] in, For the updated node importance score, The damping coefficient is... Pointing to a node The set of nodes, Let the edge weights be those between nodes j and i. Let J be the edge weight between nodes j and k. For nodes The set of nodes it points to;

[0023] S15. Sort and filter according to importance score to obtain a set of key risk semantic segments;

[0024] S16. The vector representation of the set of key semantic segments of risk, the risk type distribution features of the labeled text, the global sentiment features of the dialogue, and the statistical features of the conversation interaction are spliced ​​and fused to generate a semantic feature vector.

[0025] In one embodiment, S5 includes:

[0026] S21. Perform statistical aggregation on the temporal acoustic feature sequence to generate a statistical acoustic feature vector;

[0027] S22. Reassemble the temporal acoustic feature sequence into a two-dimensional feature map, and input the two-dimensional feature map into a pre-trained deep convolutional neural network for forward propagation processing to obtain a deep acoustic feature vector.

[0028] S23. Concatenate the statistical acoustic feature vector and the depth acoustic feature vector to generate an acoustic feature vector.

[0029] In one embodiment, S6 includes:

[0030] S31. Evaluate the confidence of the semantic feature vector to obtain the semantic confidence.

[0031] S32. Evaluate the confidence level of the acoustic feature vectors to obtain the acoustic confidence level;

[0032] S33. Based on semantic confidence and acoustic confidence, the fusion weights of the semantic feature vectors and acoustic feature vectors are calculated; the fusion weights of the semantic feature vectors are calculated using the following formula:

[0033]

[0034] in, The fusion weights for semantic feature vectors Scaling factor For semantic confidence, Acoustic confidence level;

[0035] S34. Based on the fusion weights of the semantic feature vector and the acoustic feature vector, the semantic feature vector and the acoustic feature vector are weighted and summed to obtain the fused feature vector.

[0036] In one embodiment, S31 includes:

[0037] S311. Calculate the average recognition confidence of the dialogue text sequence based on the posterior probability of the word units output by the speech recognition process.

[0038] S312. Calculate the dispersion of the importance scores of each segment in the set of key risk semantic segments to obtain the dispersion of the score set.

[0039] S313. Calculate the information entropy of the risk type distribution characteristics of the annotated text to obtain the risk information entropy; the formula for calculating the risk information entropy is:

[0040]

[0041] in, For risk information entropy, For the number of risk types, For the first Normalized frequency of risk-related labels;

[0042] S314. The semantic confidence is obtained by weighted summation of the average identification confidence, the reciprocal of the score set dispersion, and the negative correlation mapping value of the risk information entropy.

[0043] In one embodiment, after S7, a step of issuing an early warning response based on the level of domestic violence risk is further included, including:

[0044] S41. In response to the domestic violence risk level being high and the confidence level being higher than the high risk threshold, generate a key evidence report containing key semantic fragments of risk and acoustic emotion analysis conclusions, and send the key evidence report to the preset security service platform.

[0045] S42. In response to the domestic violence risk level being medium risk and the confidence level being higher than the medium risk threshold, send an anonymous attention notification to a pre-set trusted contact.

[0046] S43. In response to the domestic violence risk level being low, perform anonymized log recording operations, and when the judgment confidence level is lower than the low confidence threshold, add a low confidence mark to the low confidence judgment result record.

[0047] Secondly, this application also provides an artificial intelligence-based domestic violence risk identification system, including:

[0048] The speech preprocessing module is used to preprocess the acquired raw audio data to obtain the speech segments to be analyzed.

[0049] The automatic speech recognition module is used to automatically recognize speech segments and obtain the corresponding dialogue text sequence.

[0050] The semantic feature extraction module is used to extract key features from the dialogue text sequence based on contextual semantics and risk knowledge, and generate semantic feature vectors.

[0051] The acoustic feature extraction module is used to extract acoustic features from speech segments to obtain a temporal acoustic feature sequence.

[0052] The emotion pattern learning module is used to perform deep emotion pattern learning on temporal acoustic feature sequences and generate acoustic feature vectors.

[0053] The multimodal fusion module is used to perform multimodal fusion of semantic feature vectors and acoustic feature vectors to obtain fused feature vectors;

[0054] The domestic violence risk assessment module is used to collaboratively assess the risk of fused feature vectors and generate domestic violence risk levels and corresponding assessment confidence levels; domestic violence risk levels include high risk, medium risk and low risk.

[0055] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the artificial intelligence-based domestic violence risk identification method as described in the first aspect.

[0056] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based domestic violence risk identification method of the first aspect.

[0057] The aforementioned AI-based method, system, device, and medium for identifying domestic violence risks preprocesses raw audio data to obtain speech segments. These segments are then subjected to automatic speech recognition to obtain a dialogue text sequence, and acoustic feature extraction to obtain a temporal acoustic feature sequence. Subsequently, key feature extraction based on contextual semantics and risk knowledge is performed on the dialogue text sequence to generate a semantic feature vector. Simultaneously, deep emotion pattern learning is performed on the temporal acoustic feature sequence to generate an acoustic feature vector. The semantic and acoustic feature vectors are then fused in a multimodal manner to obtain a fused feature vector. Finally, the fused feature vector is used for collaborative risk assessment to generate domestic violence risk levels (high, medium, and low) and corresponding confidence levels. Through the parallel extraction, fusion, and collaborative assessment of semantic and acoustic dual-modal features, a more comprehensive and accurate identification and assessment of domestic violence risks is achieved. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart illustrating an artificial intelligence-based method for identifying domestic violence risks provided by this invention;

[0060] Figure 2 A flowchart illustrating a method for generating an early warning response in one optional embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the structure of an artificial intelligence-based domestic violence risk identification system provided by the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based method for identifying domestic violence risks is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps S1 to S7:

[0064] S1. Preprocess the acquired raw audio data to obtain the speech segment to be analyzed.

[0065] Optionally, an adaptive filtering algorithm is used to remove noise from the original audio data, with Least Mean Square (LMS) filtering being preferred. The principle is to iteratively adjust the filter coefficients to minimize the mean square error between the output signal and the desired signal, thereby canceling background noise in the home environment. Subsequently, format standardization is performed, converting the original audio data with different sampling rates and bit depths into a preset standard format. Then, an endpoint detection method combining short-time energy and zero-crossing rate is used to segment the effective speech segments. Specifically, the short-time energy of the audio signal is calculated to reflect signal strength, the zero-crossing rate is statistically analyzed to reflect signal frequency changes, and dual thresholds are set to distinguish speech and non-speech regions. Finally, continuous speech regions are extracted as the speech segments to be analyzed.

[0066] S2. Perform automatic speech recognition on the speech segment to obtain the corresponding dialogue text sequence.

[0067] Optionally, an automatic speech recognition (ASR) model based on the Transformer architecture is used to implement the conversion from speech to text. This model captures the long-range dependencies of the speech sequence through the self-attention mechanism. Its core principle is to use the multi-head attention module to calculate the feature correlations in different dimensions in parallel, and combine the feed-forward neural network to perform non-linear transformation on the features. In specific implementation, first perform frame windowing on the speech segment, use the Hanning window to reduce the discontinuity of the inter-frame signal, and then extract the Mel Frequency Cepstral Coefficients (MFCC) as the input feature. This feature is obtained by performing discrete cosine transform after mapping the audio spectrum to the Mel scale that conforms to the human auditory characteristics. Input the MFCC feature sequence into the trained ASR (Automatic Speech Recognition) model. The ASR model outputs a character probability distribution sequence, and the initial text is obtained through Viterbi decoding. Then, correct homophone errors and remove duplicate characters through post-processing based on the language model, and finally generate the corresponding dialogue text sequence.

[0068] S3. Extract key features based on context semantics and risk knowledge from the dialogue text sequence to generate semantic feature vectors.

[0069] Optionally, first perform word segmentation and stop word removal on the dialogue text sequence. Use the Jieba word segmentation tool to implement Chinese word segmentation, and filter out the words without actual semantics through a preset stop word list. Exemplarily, the stop word list includes function words such as "de", "le", "a", etc. Subsequently, construct a domestic violence risk knowledge dictionary, which is generated based on a large number of domestic violence case corpora, covering risk-related words such as explicit violence words, threatening expressions, and emotion tendency words, and at the same time annotate the risk weights of each word. Convert the word-segmented text sequence into word embedding vectors, which are obtained by training with the Word2Vec model. This Word2Vec model learns the distributed representation of words through the context window, making words with similar semantics have similar vector space distributions. Input the word embedding vectors into the Bidirectional Long Short-Term Memory (BiLSTM). This bidirectional long short-term memory network captures the context semantic information of the text through forward and backward propagation, and outputs the context-related vectors of each word. Match the context-related vectors with the risk knowledge dictionary, calculate the risk semantic contribution degree of each word, and finally generate semantic feature vectors with a fixed dimension through the max pooling operation to aggregate the sequence features.

[0070] S4. Perform acoustic feature extraction on the speech segment to obtain the temporal acoustic feature sequence.

[0071] Optionally, the speech segment is framed and windowed using a Hanning window. The frame length and frame shift are set according to conventional speech signal processing methods to ensure signal continuity. Multi-dimensional acoustic features are extracted from each frame, including MFCC and its first-order and second-order differences, short-time energy, short-time zero-crossing rate, and fundamental frequency (F0). The MFCC is obtained by converting the frame signal into a spectrum using a Fast Fourier Transform (FFT), followed by Mel filter bank filtering, logarithmic operations, and discrete cosine transform. The first-order and second-order differences are calculated from the MFCC of adjacent frames, reflecting the temporal trend of feature changes. The short-time energy is calculated from the sum of squares of the frame signals, reflecting signal strength. The short-time zero-crossing rate is obtained by counting the number of times the signal crosses the zero axis within a frame, reflecting the signal frequency characteristics. The fundamental frequency is extracted using the autocorrelation method, reflecting the speaker's pitch variation. The above features of each frame are sequentially concatenated to form a frame-level acoustic feature vector. The feature vectors of all frames are arranged in chronological order to obtain a temporal acoustic feature sequence.

[0072] S5. Perform deep emotion pattern learning on the temporal acoustic feature sequence to generate acoustic feature vectors.

[0073] Optionally, a deep emotion pattern learning model is constructed using a Gated Recurrent Unit (GRU). This model regulates information transmission and forgetting through update and reset gates, effectively capturing long-distance dependencies in temporal acoustic features. Compared to Long Short-Term Memory (LSTM) networks, it has a simpler network structure and higher computational efficiency. Specifically, the temporal acoustic feature sequence is used as the model input, mapped to a hidden layer space of a specified dimension through the input layer. After iterative learning via a multi-layer GRU network, the model automatically extracts emotion-related patterns from the sequence, such as the acoustic feature changes corresponding to emotions like anger and fear. During training, the cross-entropy loss function is used to optimize the model parameters, enabling the model to accurately learn the acoustic feature patterns corresponding to different emotional states. The output of the deep emotion pattern learning model is an emotion feature vector at each time step. Global mean pooling is used to aggregate temporal features, generating a fixed-dimensional acoustic feature vector. This acoustic feature vector contains emotional information and acoustic characteristics from the speech signal.

[0074] S6. Perform multimodal fusion of semantic feature vector and acoustic feature vector to obtain fused feature vector.

[0075] Optionally, an attention mechanism can be employed to achieve multimodal fusion of semantic and acoustic feature vectors. This attention mechanism calculates the correlation weights between the two modal features, highlighting the more critical feature information for risk identification. Specifically, the semantic feature vector is used as the query vector, and the acoustic feature vector as the key vector. An attention score is calculated using a dot product operation, and the weight distribution of the two features is obtained by normalization using a softmax function. The normalized weights are multiplied by their corresponding feature vectors to obtain weighted semantic and acoustic features, which are then added element-wise to obtain the fused feature vector. Alternatively, a concatenation fusion method can be used, directly concatenating the semantic and acoustic feature vectors end-to-end, and then mapping them to a unified-dimensional fused feature vector through a fully connected layer. The preferred attention mechanism fusion method fully utilizes the complementary information of the two modal features.

[0076] S7. Perform collaborative risk assessment on the fused feature vectors to generate domestic violence risk levels and corresponding assessment confidence levels; domestic violence risk levels include high risk, medium risk and low risk.

[0077] Optionally, a multilayer perceptron (MLP) is used to construct a collaborative risk assessment model. This model consists of an input layer, hidden layers, and an output layer. The hidden layers employ the ReLU activation function to introduce non-linear feature transformation. Predefined rules for determining domestic violence risk levels are established. The model is trained based on a large number of fused feature samples labeled with domestic violence risk levels. After the fused feature vector is input into the model, it undergoes multiple feature transformations in the hidden layers. The output layer outputs the probability distributions for high-risk, medium-risk, and low-risk categories using a softmax function. The category with the highest probability value is the final domestic violence risk level, and this highest probability value is the corresponding confidence level. During model training, cross-validation is used to optimize parameters such as the number of network layers and the number of neurons in the hidden layers to ensure the model has good generalization ability. Accurate determination of domestic violence risk is achieved by collaboratively utilizing semantic and acoustic fusion features.

[0078] In the aforementioned AI-based method for identifying domestic violence risks, effective speech segments are screened through audio preprocessing, and a dialogue text sequence is obtained through automatic speech recognition. Semantic features containing contextual semantics and risk knowledge, and acoustic features containing emotional patterns are extracted separately. Multimodal fusion is used to achieve complementary enhancement of the two types of features, and finally, accurate risk levels and confidence scores are generated through collaborative risk assessment. This method overcomes the limitations of single acoustic features or semantic analysis, reduces false alarms caused by non-violent interference and false negatives caused by subtle violence, significantly improves the accuracy and reliability of domestic violence risk identification, and provides practical technical support for early warning of domestic violence risks.

[0079] In one embodiment, S3 includes:

[0080] S11. Based on a pre-built knowledge base of domestic violence risks, risk words and context rules are matched to the dialogue text sequence to obtain labeled text with risk type tags.

[0081] Optionally, the pre-built domestic violence risk knowledge base comprises two parts: a risk vocabulary database and a context rule database. The risk vocabulary database is generated by compiling a large corpus of domestic violence cases, covering core vocabulary corresponding to different risk types such as physical violence, verbal threats, and psychological oppression, with each vocabulary labeled with its corresponding risk type. The context rule database is constructed based on the semantic logic of domestic violence scenarios, including rules such as effective collocation patterns of risk vocabulary and sentence structures that trigger risk judgments. For example, the sentence structure "I'm going to break your legs" ("action + target of harm") is included in the verbal threat context rule. First, the dialogue text sequence is segmented, and then the segmentation results are precisely matched with the risk vocabulary database to obtain initially matched risk vocabulary and corresponding candidate risk types. Further, the context of the matched risk vocabulary is verified in conjunction with the context rule database to determine whether the vocabulary is in a valid risk context, eliminating vocabulary appearing in non-risk scenarios due to ambiguity, quotations, etc. Finally, the text fragments that meet the rules are labeled with the corresponding risk type tags, generating labeled text with risk type tags.

[0082] S12. Perform context semantic embedding modeling on the annotated text to obtain the context vector representation of each word.

[0083] Optionally, a Bidirectional Encoder Representations from Transformers (BERT) model is used to model contextual semantic embedding. This model captures the contextual dependencies of words through a multi-layer bidirectional Transformer encoder. Its core principle is to use a self-attention mechanism to simultaneously focus on the text information to the left and right of a word, so that the generated vector can accurately reflect the semantic meaning of the word in a specific context. In specific implementation, the labeled text is first segmented to generate a word sequence, and special words [CLS] are added as sequence start identifiers and [SEP] as sentence separator identifiers. A positional encoding is assigned to each word to preserve the temporal information of the sequence. The word embedding and the positional encoding are added to obtain the initial input vector. The initial input vector is input into the pre-trained BERT model. The model iteratively transforms the features through a multi-layer self-attention module and a feedforward neural network, and finally outputs the hidden state vector corresponding to each word. This hidden state vector is the context vector representation of each word.

[0084] S13. Using sentences and phrases from the dialogue text sequence as nodes, calculate edge weights based on the semantic similarity of each node and the number of shared risk words to construct a text semantic graph; the edge weights between nodes are calculated using the following formula:

[0085]

[0086] in, The edge weights between nodes. For semantic similarity between nodes, The normalized number of risk terms shared among nodes. This is the balance coefficient.

[0087] Optionally, the dialogue text sequence is first divided into nodes, with punctuation marks used as delimiters to divide sentence units. Dependency parsing is then used to extract phrase units. For example, by analyzing subject-verb and verb-object dependencies between words, noun phrases and verb phrases such as "threaten you" and "don't let you go out" are extracted as phrase units. Sentence units and phrase units together constitute the nodes of the text semantic graph. Calculating the edge weights between nodes requires obtaining two core parameters, one of which is the semantic similarity between nodes. The cosine similarity is calculated by dividing the vector dot product by the product of the magnitudes of the two vectors. This cosine similarity effectively measures the semantic similarity between the two nodes. The normalized number of shared risk terms between nodes is also included. The normalization of the number of risk words shared by two nodes is achieved by dividing the result by the maximum total number of risk words in both nodes. The balance coefficient λ, ranging from 0 to 1, adjusts the contribution ratio between semantic similarity and the number of shared risk words. The optimal value can be determined based on the validation results of the annotated corpus. , Substituting λ into the formula for calculating the edge weight between nodes, the edge weight between every two nodes is calculated, and finally, a text semantic graph is constructed based on the node and edge weights.

[0088] S14. Iteratively calculate the importance score of each node based on the text semantic graph; the update formula for the node importance score is:

[0089]

[0090] in, For the updated node importance score, The damping coefficient is... Pointing to a node The set of nodes, Let the edge weights be those between nodes j and i. Let J be the edge weight between nodes j and k. For nodes The set of nodes it points to.

[0091] Optionally, the node importance score is calculated using a graph-based iterative sorting algorithm. The core principle is that a node's importance is determined by the importance of other nodes pointing to that node, as well as the edge weights. First, all nodes are initialized with the same importance score and damping coefficient. The value ranges from 0 to 1, representing the probability of a random jump, used to prevent the iterative sorting algorithm from getting trapped in local optima. During the iterative calculation process, for each node... First, determine the set of nodes that point to that node. Iterate through each node in the collection ,calculate right Importance contribution value, this value The current importance score multiplied by the edge weight Divide by The set of all nodes pointed to The sum of the middle edge weights is obtained. Towards The importance score of the transfer. All of them. Sum the contributions of the nodes in the set and multiply by the damping coefficient. , plus (1− Get the node Updated importance score Repeat the above iterative process until the change in importance scores of all nodes is less than a preset threshold, achieving convergence. The scores obtained at this point are the final importance scores of each node.

[0092] S15. Sort and filter according to importance score to obtain a set of key semantic segments of risk.

[0093] Optionally, all nodes are sorted in descending order of their final importance scores. After sorting, a combination of threshold filtering and quantity filtering is used to extract risk semantic key fragments. First, an importance score threshold is set, and nodes with scores higher than the threshold are filtered out. If the number of nodes after filtering is too large, a maximum number of nodes to be filtered is set, and the top-ranked nodes are selected. If the number of nodes after filtering is too small, the threshold can be appropriately lowered and the filtering can be repeated. The sentence units or phrase units corresponding to the filtered nodes are then chronologically arranged to maintain their order in the original dialogue text sequence, while duplicate or semantically highly overlapping fragments are removed, ultimately resulting in a set of risk semantic key fragments.

[0094] S16. The vector representation of the set of key semantic segments of risk, the risk type distribution features of the labeled text, the global sentiment features of the dialogue, and the statistical features of the conversation interaction are spliced ​​and fused to generate a semantic feature vector.

[0095] Optionally, for each segment in the risk semantic key segment set, its context vector representation is extracted, and aggregated using mean pooling to obtain the vector representation of the entire set. Mean pooling is achieved by calculating the average of the corresponding dimensions of all segment vectors, generating a fixed-dimensional segment set vector. The risk type distribution features of the labeled text are obtained by statistically analyzing the proportion of each risk type label in the labeled text, and converted into a fixed-dimensional vector using one-hot encoding. The global sentiment features of the dialogue are extracted using the sentiment analysis tool VADER (valence aware dictionary and sEntimentreasoner). This tool calculates the positive, negative, and neutral sentiment scores of the text using a pre-defined sentiment dictionary and grammatical rules, forming a global sentiment feature vector. The statistical features of the conversation interaction are obtained by statistically analyzing the number of turns, the number of speaker switching, the length of each turn, and other interaction information in the dialogue. The statistical results are standardized and converted into vector form. Finally, the above four feature vectors are concatenated in sequence to obtain a semantic feature vector that integrates multi-dimensional semantic information.

[0096] In the above embodiments, accurate risk labeling of text is achieved through a domestic violence risk knowledge base, context-related semantic embeddings are generated using the BERT model, key risk segments are screened by constructing a text semantic graph and iteratively calculating node importance, and finally, multi-dimensional semantically related features are fused to generate a semantic feature vector. This embodiment fully explores the contextual semantics and risk association information of the text, improves the targeting and effectiveness of semantic features, and can more accurately capture obscure and complex semantic signals of domestic violence risks.

[0097] In one embodiment, S5 includes:

[0098] S21. Perform statistical aggregation on the temporal acoustic feature sequence to generate a statistical acoustic feature vector.

[0099] Optionally, global statistics for each dimension of the temporal acoustic feature sequence are selected as the aggregation target, including mean, variance, maximum, minimum, kurtosis, skewness, and quartiles. In practice, each feature dimension of the temporal acoustic feature sequence is traversed, and the above statistics are calculated along the time axis. The mean reflects the overall average level of the feature, the variance reflects the dispersion of the feature, the maximum and minimum values ​​define the range of feature values, the kurtosis reflects the steepness of the feature distribution, the skewness reflects the symmetry of the feature distribution, and the quartiles reflect the distribution interval of the feature. All statistics corresponding to each dimension are arranged in a preset order, and a fixed-length vector is formed by concatenating the dimensions. This vector is the statistical acoustic feature vector, which can comprehensively characterize the global distribution characteristics of the temporal acoustic features.

[0100] S22. Reassemble the temporal acoustic feature sequence into a two-dimensional feature map, and input the two-dimensional feature map into a pre-trained deep convolutional neural network for forward propagation processing to obtain a deep acoustic feature vector.

[0101] Optionally, the temporal acoustic feature sequence is first re-dimensioned using a sliding window partitioning method. This divides the continuous temporal features into multiple feature subsequences within a fixed-length time window. Each feature subsequence serves as a row in a two-dimensional feature map. All subsequences are arranged chronologically to form a two-dimensional feature matrix. This matrix is ​​then expanded to a two-dimensional feature map format that meets the input requirements of a deep convolutional neural network. For example, if the feature subsequence length is m and the number of windows is n, then the dimension of the two-dimensional feature map is n×m. The pre-trained deep convolutional neural network preferably uses a residual network (ResNet). Its core principle is to directly transfer shallow features to deep layers through residual connection structures across multiple convolutional layers, effectively alleviating the gradient vanishing problem in deep networks. The two-dimensional feature map is input into the network for forward propagation, and then iteratively processed through convolutional layers, batch normalization, activation functions, and pooling layers. The convolutional layers extract local spatial features by sliding the convolutional kernels, and batch normalization standardizes the features to accelerate training convergence. The activation function uses a rectified linear unit (ReLU) to introduce non-linear feature transformation. The pooling layers retain key features and reduce dimensionality through downsampling. Finally, the high-dimensional features are mapped to a fixed-dimensional vector through a fully connected layer, which is the deep acoustic feature vector.

[0102] S23. Concatenate the statistical acoustic feature vector and the depth acoustic feature vector to generate an acoustic feature vector.

[0103] Optionally, the statistical acoustic feature vector and the depth acoustic feature vector are first checked for dimensionality compatibility to ensure that they are consistent in vector format and have no dimensional conflicts. A first-end concatenation fusion method is used, arranging all elements of the statistical acoustic feature vector in their original order, followed by all elements of the depth acoustic feature vector. During the concatenation process, an ordered combination of elements is achieved through a vector dimension concatenation function, without the need for additional feature transformation or weight adjustment. After concatenation, the vectors are standardized to ensure that the value range of each element in the vector is consistent, ultimately generating an acoustic feature vector that possesses both global statistical properties and local deep abstraction characteristics.

[0104] In the above embodiment, global distribution information of temporal acoustic features is extracted through statistical aggregation, and then local deep abstract information of the features is mined using a pre-trained deep convolutional neural network. Finally, the two types of features are spliced ​​and fused. This embodiment achieves complementary fusion of global and local information of acoustic features, improves the comprehensiveness and accuracy of acoustic feature representation, and can more sensitively capture the difference signals of different types of sounds (including those with significant and insignificant acoustic features) in domestic violence scenarios.

[0105] In one embodiment, S6 includes:

[0106] S31. Evaluate the confidence of the semantic feature vector to obtain the semantic confidence.

[0107] Optionally, semantic confidence assessment can be achieved by combining feature quality evaluation and classification confidence. Specifically, a semantic confidence evaluator is constructed using a trained multilayer perceptron. The input to this semantic confidence evaluator is a semantic feature vector, and the output is a semantic confidence score between 0 and 1. The core principle is to quantify the reliability of semantic features for risk identification by learning the correlation between semantic feature vectors and real risk labels. During training, a dataset of semantic feature vectors labeled with real risk labels is used, and the model parameters are optimized using the mean squared error loss function. This allows the model to learn the mapping relationship between quality indicators such as the discriminative power and completeness of feature vectors and their confidence scores. During evaluation, the semantic feature vectors are input into the trained semantic confidence evaluator, and the output value is obtained through forward propagation. This value is the semantic confidence score; the closer the value is to 1, the higher the reliability of the semantic features.

[0108] S32. Evaluate the confidence level of the acoustic feature vector to obtain the acoustic confidence level.

[0109] Optionally, an acoustic confidence estimator can be constructed using a multilayer perceptron trained independently, employing a technical framework derived from semantic confidence assessment. This ensures consistency in evaluation criteria while adapting to the characteristics of acoustic features. The acoustic confidence estimator takes acoustic feature vectors as input and outputs acoustic confidence scores between 0 and 1. Its core principle is to learn the correlation between acoustic feature vectors and real risk labels, quantifying the reliability of acoustic features in risk identification. The training dataset uses acoustic feature vector datasets labeled with real risk labels. The model parameters are optimized using the mean squared error loss function, enabling the model to capture the mapping relationship between acoustic feature quality indicators such as discriminability and stability and confidence scores. During evaluation, the acoustic feature vectors are input into the trained acoustic confidence estimator, and the output value is obtained through forward propagation. This value is the acoustic confidence score; the closer the value is to 1, the higher the reliability of the acoustic feature.

[0110] S33. Based on semantic confidence and acoustic confidence, the fusion weights of the semantic feature vectors and acoustic feature vectors are calculated; the fusion weights of the semantic feature vectors are calculated using the following formula:

[0111]

[0112] in, The fusion weights for semantic feature vectors Scaling factor For semantic confidence, Acoustic confidence level.

[0113] Optionally, scaling factor This value, ranging from 0 to 1, is used to adjust the impact of confidence differences on the fusion weights. The optimal value is determined using cross-validation on the labeled fusion feature dataset to ensure the highest accuracy in subsequent risk assessment. The obtained semantic confidence is acquired before calculation. Harmony and acoustic confidence Calculate separately and , and The product of the two features is used. A natural exponential function (exp) is employed to map the product results. This function amplifies the difference in confidence levels, making the exponential values ​​corresponding to high-confidence features more advantageous. The exponential value corresponding to semantic confidence is used as the numerator, and the sum of the numerator and the exponential value corresponding to acoustic confidence is used as the denominator. The ratio of the two is the fusion weight of the semantic feature vector. Fusion weights of acoustic feature vectors Subtract 1 This ensures that the sum of the fusion weights of the two features is 1, thus achieving a normalized weight allocation.

[0114] S34. Based on the fusion weights of the semantic feature vector and the acoustic feature vector, the semantic feature vector and the acoustic feature vector are weighted and summed to obtain the fused feature vector.

[0115] Optionally, the semantic feature vector and acoustic feature vector are first checked for dimensionality consistency to ensure they have the same feature dimension. If a dimensionality difference exists, a fully connected layer is used to map the low-dimensional feature vector to the dimension of the high-dimensional feature vector. During the mapping process, a trained linear transformation matrix is ​​used to achieve dimensionality enhancement. After dimensionality consistency is achieved, element-wise weighted summation is used for fusion. Specifically, each element in the semantic feature vector is summed with its corresponding fusion weight. Multiply to obtain a weighted semantic feature vector; then multiply each element of the acoustic feature vector with its corresponding fusion weight. Multiplying them yields a weighted acoustic feature vector. Adding the two weighted feature vectors element-wise gives the fused feature vector, which fully integrates highly reliable semantic and acoustic information.

[0116] In the above embodiments, the reliability of semantic and acoustic features is quantified by an independent confidence evaluator. Dynamic fusion weights are calculated using an exponential function mapping based on the confidence level, and feature fusion is achieved through weighted summation. This embodiment achieves adaptive allocation of fusion weights, allowing high-reliability features to contribute a higher proportion to the fusion result, effectively suppressing the interference of low-quality features, and improving the representational ability and reliability of the fused feature vector.

[0117] In one embodiment, S31 includes:

[0118] S311. Calculate the average recognition confidence of the dialogue text sequence based on the posterior probability of the word units output by the speech recognition process.

[0119] Optionally, the posterior probability of a word is the recognition probability corresponding to each word in the ASR model output, representing the reliability of the word being correctly recognized. In specific implementation, the posterior probabilities of all words corresponding to the dialogue text sequence are first extracted from the output of the ASR model. Special words used for sequence identification are then removed; for example, the probability values ​​corresponding to the [CLS] start identifier and the [SEP] separator identifier are removed. The arithmetic mean method is used to calculate the average of the remaining word posterior probabilities. This average is the average recognition confidence of the dialogue text sequence. Its core principle is to obtain the overall recognition quality of the entire text sequence by aggregating the recognition reliability of individual words. The higher the average probability, the higher the recognition accuracy of the text sequence.

[0120] S312. Calculate the dispersion of the importance scores of each segment in the set of key risk semantic segments to obtain the dispersion of the score set.

[0121] Optionally, the dispersion uses standard deviation as a quantitative indicator to characterize the degree of dispersion of importance scores among segments in the set of key semantic segments for risk. Specifically, the importance scores of all segments in the set are first obtained, and the arithmetic mean of these scores is calculated. Then, for each score, the squared deviation from the mean is calculated. The sum of all squared deviations is divided by the total number of scores minus one to obtain the variance. Finally, the square root of the Leps variance is taken to obtain the standard deviation, which is the dispersion of the score set. The core principle is that the smaller the dispersion, the more concentrated the importance of the key segments and the more focused the semantic risk information; the larger the dispersion, the greater the difference in importance among the key segments and the more dispersed the semantic risk information. This indicator can quantify the consistency of key semantic information.

[0122] S313. Calculate the information entropy of the risk type distribution characteristics of the annotated text to obtain the risk information entropy; the formula for calculating the risk information entropy is:

[0123]

[0124] in, For risk information entropy, For the number of risk types, For the first Normalized frequency of risk-related labels.

[0125] Optionally, the risk type distribution feature is the proportion of each type of risk label in the labeled text. Information entropy is used to measure the uncertainty of this distribution. A higher entropy value indicates that the risk types are more dispersed and the semantic risks are more complex; a lower entropy value indicates that the risk types are more concentrated and the semantic risks are more explicit. In specific implementation, the total number of risk types contained in the labeled text is first counted. Then, count the occurrence frequency of each type of risk label separately, and divide the occurrence frequency of each type by the total occurrence frequency of all risk labels to obtain the normalized frequency of the i-th type of risk label. .Will Substituting into the information entropy calculation formula, where the logarithm operation is base 2, if a certain type of risk label... If the value is 0, a minimum value needs to be introduced to avoid errors in logarithmic calculations. Through this calculation process, the degree of disorder in the distribution of risk types can be accurately quantified, providing dimensional support for semantic reliability assessment.

[0126] S314. The semantic confidence is obtained by weighted summation of the average identification confidence, the reciprocal of the score set dispersion, and the negative correlation mapping value of the risk information entropy.

[0127] Optionally, the three input indicators are first normalized, mapping the average recognition confidence, the reciprocal of the score set dispersion, and the negative correlation mapping value of risk information entropy to a range of 0 to 1. The reciprocal of the score set dispersion is used to achieve a positive correlation with semantic reliability; that is, the smaller the dispersion, the larger the reciprocal, and the more focused the corresponding semantic information. The negative correlation mapping of risk information entropy is achieved using 1 / (1+H). The risk information entropy is defined as follows: the smaller the entropy value, the larger the mapping value, and the clearer the corresponding risk distribution. Preset weight coefficients for the three indicators. , , The weight coefficients were determined on the labeled corpus dataset through cross-validation and satisfy the following conditions: The three normalized indicators are multiplied by their corresponding weight coefficients, and the sum of the products is obtained as the semantic confidence score. The core principle is to comprehensively quantify the reliability of semantic features by integrating three dimensions: text recognition quality, key segment consistency, and risk distribution certainty.

[0128] In the above embodiments, text recognition confidence is obtained by posterior probability aggregation of lexical units, the importance distribution of key segments is quantified by standard deviation, the uncertainty of risk type distribution is characterized by risk information entropy, and finally, the semantic confidence is obtained by weighted fusion of multi-dimensional indicators. This embodiment realizes a multi-dimensional comprehensive evaluation of semantic confidence, avoids the one-sidedness of single indicator evaluation, and can more accurately and comprehensively quantify the reliability of semantic features, thereby improving the overall accuracy of domestic violence risk identification.

[0129] In an optional embodiment, such as Figure 2 As shown, following S7, there are also steps for early warning and response based on the level of domestic violence risk, including:

[0130] S41. In response to the domestic violence risk level being high and the confidence level being higher than the high risk threshold, generate a key evidence report containing key semantic fragments of risk and acoustic sentiment analysis conclusions, and send the key evidence report to the preset security service platform.

[0131] Optionally, a trigger condition check is first performed, comparing the domestic violence risk level with the confidence level based on the conditional judgment statement. The high-risk threshold is predefined based on a large corpus of domestic violence cases and the intervention needs of the safety service platform. Subsequent operations are triggered only when the risk level is clearly high and the confidence level is strictly higher than the threshold. The generation of the key evidence report needs to integrate two types of core information: the risk semantic key fragment text is directly extracted from the previously obtained risk semantic key fragment set, sorted according to the original dialogue sequence, and then presented in a formatted manner; the acoustic emotion analysis conclusion is extracted from the deep emotion pattern learning results, including emotion type, emotion intensity, and corresponding acoustic feature basis. For example, if the acoustic feature vector matches the anger emotion pattern, the conclusion is marked "There is strong anger emotion, and the corresponding acoustic feature is severe fundamental frequency fluctuation and short-term energy surge". The report adopts a predefined structured format, including modules such as report number, generation time, risk level, confidence level, key evidence text and analysis conclusions. After generation, it establishes communication with the preset security service platform through the application programming interface (API), and uses Hypertext Transfer Protocol Secure (HTTPS) for encrypted data transmission. After the transmission is completed, it receives a reception confirmation signal from the platform to ensure that the report is delivered.

[0132] S42. In response to a domestic violence risk level of medium risk and a confidence level higher than the medium risk threshold, send an anonymous alert to a pre-set trusted contact.

[0133] Optionally, a trigger condition determination is first performed. Logical judgment verifies whether the domestic violence risk level is medium risk, while simultaneously confirming that the confidence level is higher than a predefined medium risk threshold. The medium risk threshold is determined based on a balance between risk intervention costs and early warning accuracy. The notification sending process is initiated only when both conditions are met simultaneously. The trusted contact list is predefined structured data stored in a local encrypted storage module, including contact name, contact information (phone number or email), and preferred notification method. Storage is performed using Advanced Encryption Standard (AES) to ensure data security. Anonymization is achieved by stripping away all information that can be associated with the user's identity, including hiding device identifiers, geolocation, and other privacy data. The notification content is generated using a predefined template and includes non-privacy information such as "Attention alert: Potential family risk detected; please pay attention to the safety of relevant personnel." Based on the contact's preferred notification method, the corresponding communication interface is called to send the notification. If it's a mobile phone number, the Short Message Service (SMS) interface is called to send an SMS; if it's an email address, the Simple Mail Transfer Protocol (SMTP) interface is called to send an email. After sending, the sending time and status are recorded.

[0134] S43. In response to the domestic violence risk level being low, perform anonymized log recording operations, and when the judgment confidence level is lower than the low confidence threshold, add a low confidence mark to the low confidence judgment result record.

[0135] Optionally, when the domestic violence risk level is determined to be low, anonymized logging is directly triggered. The log content adopts a structured format, including the record number, judgment time, risk level, and judgment confidence level. All user identity information and raw audio / text data are removed, retaining only the core indicators related to the risk assessment. Logs are stored in Extensible Markup Language (XML) format in a locally specified encrypted directory, with the storage path predefined in a configuration file. Further, a conditional statement compares the judgment confidence level with a predefined low-confidence threshold. This threshold is used to differentiate the reliability of the low-risk judgment result. If the judgment confidence level is lower than the low-confidence threshold, a preset low-confidence flag field is added to the log record. The flag field content is "Low Confidence Flag: The reliability of this low-risk judgment is low; it is recommended to reassess based on subsequent data," ensuring that low-reliability judgment results can be traced and verified later.

[0136] In the above embodiments, differentiated response strategies are implemented based on the level and confidence level of domestic violence risk. For high-risk situations, a complete evidence report is generated and linked to the safety service platform; for medium-risk situations, anonymous alerts are sent to trusted contacts; and for low-risk situations, anonymous log entries are recorded and confidence levels are differentiated. This embodiment achieves precise risk response grading, balancing emergency handling of high-risk events, early warning of medium-risk events, and traceability management of low-risk events. While protecting user privacy, it improves the timeliness and targeting of risk intervention, effectively connecting technical identification with actual risk handling.

[0137] In the aforementioned AI-based methods, systems, devices, and media for identifying domestic violence risks, the original audio from the home environment is used as input. First, the audio segment to be analyzed is obtained through preprocessing, and then a dialogue text sequence is generated through automatic speech recognition. For the text sequence, risk labeling is completed based on a domestic violence risk knowledge base. Key risk segments are extracted by combining semantic embedding modeling, text semantic graph construction, and node importance ranking. Multi-dimensional text features are fused to generate semantic feature vectors. For the audio segment, acoustic features are extracted through statistical aggregation and deep convolutional networks to generate acoustic feature vectors. Dynamic fusion weights are calculated based on semantic and acoustic confidence levels to achieve multi-modal feature fusion and risk level determination. Finally, differentiated early warning responses are implemented based on risk level and confidence level. This technical solution, through multi-modal feature fusion and adaptive weighting of confidence levels, overcomes the limitations of existing methods such as simple acoustic feature analysis (prone to false positives and false negatives) and simple semantic analysis (limited by transcription accuracy and lack of paralinguistic information understanding). It effectively reduces the false positive and false negative rates of domestic violence risk identification, improves the accuracy and reliability of identification, and achieves effective connection between technical identification and actual risk management through differentiated response strategies.

[0138] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0139] Based on the same inventive concept, this application also provides a system for implementing the aforementioned artificial intelligence-based domestic violence risk identification method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations of one or more embodiments of the artificial intelligence-based domestic violence risk identification system provided below can be found in the limitations of the artificial intelligence-based domestic violence risk identification method described above, and will not be repeated here.

[0140] In one exemplary embodiment, such as Figure 3 As shown, an artificial intelligence-based domestic violence risk identification system 10 is provided, including:

[0141] The speech preprocessing module 11 is used to preprocess the acquired raw audio data to obtain the speech segment to be analyzed.

[0142] Automatic speech recognition module 12 is used to automatically recognize speech segments and obtain the corresponding dialogue text sequence;

[0143] Semantic feature extraction module 13 is used to extract key features from the dialogue text sequence based on contextual semantics and risk knowledge, and generate semantic feature vectors;

[0144] Acoustic feature extraction module 14 is used to extract acoustic features from speech segments to obtain a temporal acoustic feature sequence;

[0145] The emotion pattern learning module 15 is used to perform deep emotion pattern learning on temporal acoustic feature sequences and generate acoustic feature vectors.

[0146] The multimodal fusion module 16 is used to perform multimodal fusion of semantic feature vectors and acoustic feature vectors to obtain fused feature vectors;

[0147] The domestic violence risk assessment module 17 is used to perform collaborative risk assessment on the fused feature vectors and generate domestic violence risk levels and corresponding assessment confidence levels; the domestic violence risk levels include high risk, medium risk and low risk.

[0148] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the artificial intelligence-based domestic violence risk identification method as described above.

[0149] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the artificial intelligence-based domestic violence risk identification method described above.

[0150] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0151] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for identifying domestic violence risks based on artificial intelligence, characterized in that, The method includes: S1. Preprocess the acquired raw audio data to obtain the speech segment to be analyzed; S2. Perform automatic speech recognition on the speech segment to obtain the corresponding dialogue text sequence; S3. Extract key features from the dialogue text sequence based on contextual semantics and risk knowledge to generate a semantic feature vector; S4. Perform acoustic feature extraction on the speech segment to obtain a temporal acoustic feature sequence; S5. Perform deep emotion pattern learning on the temporal acoustic feature sequence to generate acoustic feature vectors; S6. Perform multimodal fusion on the semantic feature vector and the acoustic feature vector to obtain a fused feature vector; S7. Perform collaborative risk assessment on the fused feature vector to generate domestic violence risk levels and corresponding assessment confidence levels; the domestic violence risk levels include high risk, medium risk and low risk.

2. The method according to claim 1, characterized in that, S3 includes: S11. Based on a pre-built domestic violence risk knowledge base, the dialogue text sequence is matched with risk words and context rules to obtain labeled text with risk type tags. S12. Perform context semantic embedding modeling on the annotated text to obtain the context vector representation of each word; S13. Using the sentence and phrase units of the dialogue text sequence as nodes, calculate the edge weights based on the semantic similarity of each node and the number of shared risk words to construct a text semantic graph; wherein the edge weights between nodes are calculated using the following formula: in, The edge weights between nodes. For semantic similarity between nodes, The normalized number of risk terms shared among nodes. This is the balance coefficient; S14. Based on the text semantic graph, iteratively calculate the node importance score to obtain the importance score of each node; the update formula for the node importance score is: in, For the updated node importance score, The damping coefficient is... Pointing to a node The set of nodes, Let the edge weights be those between nodes j and i. Let J be the edge weight between nodes j and k. For nodes The set of nodes it points to; S15. Sort and filter according to the importance scores to obtain a set of key risk semantic segments; S16. The vector representation of the risk semantic key fragment set, the risk type distribution features of the labeled text, the global sentiment features of the dialogue, and the statistical features of the conversation interaction are spliced ​​and fused to generate the semantic feature vector.

3. The method according to claim 1, characterized in that, S5 includes: S21. Perform statistical aggregation on the temporal acoustic feature sequence to generate a statistical acoustic feature vector; S22. The temporal acoustic feature sequence is reconstructed into a two-dimensional feature map, and the two-dimensional feature map is input into a pre-trained deep convolutional neural network for forward propagation processing to obtain a deep acoustic feature vector. S23. The statistical acoustic feature vector and the depth acoustic feature vector are concatenated to generate the acoustic feature vector.

4. The method according to claim 2, characterized in that, S6 includes: S31. Calculate the confidence level of the semantic feature vector to obtain the semantic confidence level; S32. Calculate the confidence level of the acoustic feature vector to obtain the acoustic confidence level; S33. Based on the semantic confidence and the acoustic confidence, calculate the fusion weight of the semantic feature vector and the fusion weight of the acoustic feature vector; the fusion weight of the semantic feature vector is calculated using the following formula: in, The fusion weights for semantic feature vectors Scaling factor The semantic confidence level, The acoustic confidence level; S34. Based on the fusion weights of the semantic feature vector and the acoustic feature vector, the semantic feature vector and the acoustic feature vector are weighted and summed to obtain the fused feature vector.

5. The method according to claim 4, characterized in that, S31 includes: S311. Calculate the average recognition confidence of the dialogue text sequence based on the posterior probability of the word units output by the speech recognition process; S312. Calculate the dispersion of the importance score of each segment in the set of key risk semantic segments to obtain the dispersion of the score set. S313. Calculate the information entropy of the risk type distribution characteristics of the labeled text to obtain the risk information entropy; the formula for calculating the risk information entropy is: in, For risk information entropy, For the number of risk types, For the first Normalized frequency of risk-related labels; S314. The semantic confidence is obtained by weighted summation of the average identification confidence, the reciprocal of the discreteness of the score set, and the negative correlation mapping value of the risk information entropy.

6. The method according to claim 1, characterized in that, Following S7, the method further includes a step of issuing an early warning response based on the domestic violence risk level, including: S41. In response to the domestic violence risk level being high and the confidence level being higher than the high risk threshold, a key evidence report containing key semantic fragments of risk and acoustic emotion analysis conclusions is generated, and the key evidence report is sent to a preset security service platform. S42. In response to the domestic violence risk level being medium risk and the determination confidence level being higher than the medium risk threshold, send an anonymous attention prompt to a preset trusted contact. S43. In response to the domestic violence risk level being low, perform anonymization log recording operation, and when the determination confidence level is lower than the low confidence threshold, add a low confidence mark to the low confidence determination result record.

7. A domestic violence risk identification system based on artificial intelligence, characterized in that, The system includes: The speech preprocessing module is used to preprocess the acquired raw audio data to obtain the speech segments to be analyzed. An automatic speech recognition module is used to automatically recognize the speech segment and obtain the corresponding dialogue text sequence. The semantic feature extraction module is used to extract key features from the dialogue text sequence based on contextual semantics and risk knowledge, and generate a semantic feature vector. An acoustic feature extraction module is used to extract acoustic features from the speech segment to obtain a temporal acoustic feature sequence. The emotion pattern learning module is used to perform deep emotion pattern learning on the temporal acoustic feature sequence to generate acoustic feature vectors. A multimodal fusion module is used to perform multimodal fusion on the semantic feature vector and the acoustic feature vector to obtain a fused feature vector; The domestic violence risk assessment module is used to perform collaborative risk assessment on the fused feature vector to generate a domestic violence risk level and a corresponding assessment confidence level; the domestic violence risk level includes high risk, medium risk and low risk.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images