A suicide risk assessment method based on emotional perception

Abstract

The application discloses a suicide risk assessment method based on emotion perception, and obtains the embedded representation of an input post through an emotion coding model to obtain an emotion vector of a suicidal post and an emotion vector of a non-suicidal post, which are used to generate an initialization vector of the suicidal post and an initialization vector of the non-suicidal post, so that the fine-grained emotion expressed in the post can be captured; in a T-LSTM layer, a time-aware T-LSTM model is used to simulate the dynamic change of the emotion in the non-suicidal post, and the time interval information is fully utilized; finally, the probability distribution of each suicide risk level is obtained by combining the context representation vector of the suicidal post, the context representation vector of the non-suicidal post, an LIWC feature vector and a post time feature vector, so that the accuracy of the suicide risk assessment is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a suicide risk assessment method based on emotion perception. BACKGROUND

[0002] With the popularity of social media, people increasingly disclose their emotions and mental states on social media, which provides extensive practical data for suicide risk assessment. Suicide risk assessment of social media users is a psychological assessment task of predicting the suicide risk level of users by analyzing and utilizing the posts published by users on social media, which is beneficial to the rational allocation of medical resources and the provision of targeted psychological intervention for high-risk users.

[0003] Early studies mainly considered feature engineering methods, using online behaviors, psycholinguistic lexical LIWC and emotional dictionary features for mental problem identification on social media. These methods work well, but consume a lot of manpower and time. In recent years, researchers have begun to use deep neural network models such as CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory) to explore user context such as user emotional changes, social networks and historical posts to better understand the formation process of user psychological crisis and thus improve the level of suicide risk assessment. However, most of these methods are used to identify the suicide risk of a single post, that is, to judge whether a post contains suicidal ideation, and the user-level suicide risk assessment has not been fully explored.

[0004] User-level suicide risk assessment is a user-oriented prediction task, and further refines the suicide risk level. Most of the existing user-level suicide risk assessment methods use all posts of the user, and this method does not distinguish between suicide-related posts and other posts, which may affect the recognition ability of the model. To solve this problem, the prior art divides user posts into SuicideWatch (SW) posts and Non-SuicideWatch (NSW) posts, and models these two types of posts respectively, wherein SuicideWatch posts refer to posts published on the SuicideWatch section of the Reddit forum, and Non-SuicideWatch represents posts published on other sections of Reddit. However, this user-level suicide risk assessment method still has certain limitations.

[0005] First, the emotional information contained in the Non-SuicideWatch posts is not fully utilized. It is usually difficult to determine the suicide risk level of a user according to the content of the Non-SuicideWatch posts. However, the large amount of content related to mental health problems or daily trivia contained in the Non-SuicideWatch posts can provide some help for the suicide risk assessment task, especially the emotional information. For example, the emotional words "hate", "anxious" and "isolated" appear in a Non-SuicideWatch post, indicating that the overall mood of the user is relatively negative, and therefore the probability of predicting that the user is at risk of suicide is increased.

[0006] Second, the emotional changes of the user over time are not considered. Emotions are time-sensitive, and the importance of emotional information changes with the size of the time interval. The time interval information should be considered when predicting the suicide risk level, and the above problems will result in low accuracy of suicide risk assessment. SUMMARY

[0007] To this end, an embodiment of the present application proposes a suicide risk assessment method based on emotion perception to solve the problem that the prior art does not fully utilize the emotional information contained in the posts and does not consider the time interval information.

[0008] The suicide risk assessment method based on emotion perception according to an embodiment of the present application is applied to a suicide risk assessment model based on emotion perception, and the suicide risk assessment model includes a post embedding layer, a Bi-LSTM layer, a T-LSTM layer, a user feature extraction layer and a classification layer.

[0009] The method comprises:

[0010] The post embedding layer obtains the embedding representation of the input post through a semantic encoding model to obtain a semantic encoding vector of the suicide-tendency post and a semantic encoding vector of the non-suicide-tendency post, and obtains the embedding representation of the input post through an emotional encoding model to obtain an emotional vector of the suicide-tendency post and an emotional vector of the non-suicide-tendency post, and then obtains an initialization vector of the suicide-tendency post according to the semantic encoding vector of the suicide-tendency post and the emotional vector of the suicide-tendency post, and obtains an initialization vector of the non-suicide-tendency post according to the semantic encoding vector of the non-suicide-tendency post and the emotional vector of the non-suicide-tendency post;

[0011] The initialization vector of the suicide-tendency post is input into the Bi-LSTM layer to obtain a hidden state vector of the suicide-tendency post, and a context representation vector of the suicide-tendency post is obtained according to the hidden state vector of the suicide-tendency post;

[0012] The initialization vector of the non-suicidal post is input into the T-LSTM layer to obtain the hidden state vector of the non-suicidal post, and the context representation vector of the non-suicidal post is obtained based on the hidden state vector of the non-suicidal post.

[0013] LIWC feature vector and posting time feature vector are obtained through the user feature extraction layer;

[0014] In the classification layer, the context representation vectors of suicidal posts, the context representation vectors of non-suicidal posts, the LIWC feature vectors, and the posting time feature vectors are concatenated. The concatenated vectors are then fed into a fully connected network, and the probability distribution of each suicide risk level is calculated using a normalized exponential function.

[0015] According to the emotion-based suicide risk assessment method of this invention, in the post embedding layer, the embedded representation of the input post is obtained through an emotion coding model to obtain the emotion vectors of suicidal tendencies and non-suicidal tendencies. These vectors are used to generate the initialization vectors of suicidal tendencies and non-suicidal tendencies, which can capture the fine-grained emotions expressed in the post. In the T-LSTM layer, the dynamic changes of emotions in non-suicidal tendencies are simulated through a time-aware T-LSTM model, making full use of time interval information. Finally, the probability distribution of each suicide risk level is obtained by combining the context representation vectors of suicidal tendencies, non-suicidal tendencies, LIWC feature vectors, and posting time feature vectors, thereby improving the accuracy of suicide risk assessment.

[0016] In addition, the suicide risk assessment method based on emotion perception according to the above embodiments of the present invention may also have the following additional technical features:

[0017] Furthermore, the semantic encoding vector of suicidal posts is expressed as follows:

[0018] ;

[0019] in, Indicates user u i The j A suicidal post. Indicates and The semantic encoding vector of the corresponding suicidal post; SentenceBERT represents the semantic encoding function.

[0020] The semantic encoding vector of non-suicidal posts is expressed as follows:

[0021] ;

[0022] in, Indicates user u i The j A post with non-suicidal tendencies Indicates and The semantic encoding vector of the corresponding non-suicidal posts.

[0023] Furthermore, the sentiment encoding model is the PlutchikBERT model obtained by fine-tuning the pre-trained BERT model using the Plutchik-2 dataset.

[0024] Furthermore, the expression for the sentiment vector of suicidal posts is as follows:

[0025] ;

[0026] in, Indicates and The sentiment vector of the corresponding suicidal post; PlutchikBERT represents the sentiment encoding function.

[0027] The expression for the sentiment vector of non-suicidal posts is as follows:

[0028] ;

[0029] in, Indicates and The sentiment vector of the corresponding non-suicidal posts.

[0030] Furthermore, in the step of obtaining the initialization vector of a suicidal post based on its semantic encoding vector and sentiment vector, the expression for the initialization vector of the suicidal post is as follows:

[0031] ;

[0032] in, Indicates and The initialization vector for the corresponding suicidal post. Indicates a splicing operation;

[0033] In the step of obtaining the initialization vector of a non-suicidal post based on its semantic encoding vector and sentiment vector, the expression for the initialization vector of the non-suicidal post is as follows:

[0034] ;

[0035] in, Indicates and The initialization vector for the corresponding suicidal post.

[0036] Furthermore, in the step of inputting the initialization vector of the suicidal post into the Bi-LSTM layer to obtain the hidden state vector of the suicidal post, and obtaining the context representation vector of the suicidal post based on the hidden state vector, the expression for the hidden state vector of the suicidal post is as follows:

[0037] ;

[0038] ;

[0039] ;

[0040] in, Indicates user u i The j The hidden state vector of a suicidal post. Indicates user u i The j The forward hidden vector of a suicidal post. Indicates user u i The j The backward hidden vector of a suicidal post. Indicates user u i The j -1 forward hidden vector of suicidal posts, Indicates user u i The j -1 backward hidden vector of suicidal posts;

[0041] The expression for the context representation vector of suicidal posts is as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] in, Indicates user u i The contextual representation vector of posts with suicidal tendencies. Indicates the first j The attention coefficient of a suicidal post Indicates the first jThe attention weight of a suicidal post after normalization. n For the total number of suicidal posts, e xp Based on the natural constant e An exponential function with base 0. , , Trainable parameters for calculating attention to suicidal posts. T Indicates transpose. tanh This represents the hyperbolic tangent function.

[0046] Furthermore, in the step of inputting the initialization vector of the non-suicidal post into the T-LSTM layer to obtain the hidden state vector of the non-suicidal post, and obtaining the context representation vector of the non-suicidal post based on the hidden state vector of the non-suicidal post, the expression of the hidden state vector of the non-suicidal post is as follows:

[0047] ;

[0048] in, Indicates user u i The j The hidden state vector of a non-suicidal post, T-LSTM represents a time-aware T-LSTM function;

[0049] The expression for the context representation vector of non-suicidal posts is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] in, Indicates user u i The contextual representation vector of non-suicidal posts. Indicates the first j The attention score of a non-suicidal post. Indicates the first j The attention weight of non-suicidal posts after normalization. m The total number of posts that are not suicidal. , , Trainable parameters for calculating attention to non-suicidal posts.

[0054] Furthermore, in the step of calculating the probability distribution of each suicide risk level using the normalized exponential function, the calculation formula is as follows:

[0055] ;

[0056] in, The probability distribution of suicide risk levels is represented by Softmax, which represents the normalized exponential function. , These are the trainable parameters in the normalized exponential function. Indicates user u i LIWC eigenvectors, Indicates user u i The posting time feature vector.

[0057] Furthermore, the T-LSTM layer satisfies the following condition:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] in, express t The short-term memory unit at time -1, where tanh represents the hyperbolic tangent function. , For trainable network parameters, express t The memory unit at time -1 Indicates user u i The time interval between two consecutive non-suicidal posts. Represents the decay function. express t Short-term memory units decaying at time -1 express t Long-term memory units at time -1 Indicates the adjusted t The memory unit at time -1. Attached Figure Description

[0064] The above and / or additional aspects and advantages of the embodiments of the present invention will become apparent and readily understood from the description of the embodiments in conjunction with the following drawings, wherein:

[0065] Figure 1 This is a schematic diagram of the framework of a suicide risk assessment model based on emotion perception according to an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] This invention proposes a suicide risk assessment method based on emotion perception, which is applied to an emotion perception-based suicide risk assessment model. Please refer to [link / reference]. Figure 1 The suicide risk assessment model includes a post embedding layer, a Bi-LSTM layer, a T-LSTM layer, a user feature extraction layer, and a classification layer. The post embedding layer initializes the embedding representation of the input posts; the Bi-LSTM layer processes features in posts with suicidal tendencies; the T-LSTM layer models the sentiment changes in posts without suicidal tendencies; the user feature extraction layer extracts the user's LIWC word frequency and posting behavior features; and the classification layer predicts the user's suicide risk level.

[0068] The method includes steps 1 to 5:

[0069] Step 1: In the post embedding layer, obtain the embedding representation of the input post through the semantic encoding model to obtain the semantic encoding vector of the suicidal tendency post and the semantic encoding vector of the non-suicidal tendency post. In addition, obtain the embedding representation of the input post through the sentiment encoding model to obtain the sentiment vector of the suicidal tendency post and the sentiment vector of the non-suicidal tendency post. Then, obtain the initialization vector of the suicidal tendency post based on the semantic encoding vector and the sentiment vector of the suicidal tendency post. Obtain the initialization vector of the non-suicidal tendency post based on the semantic encoding vector and the sentiment vector of the non-suicidal tendency post.

[0070] Wherein, given the first i individual users u i The definition of a suicidal post ,in A collection of posts expressing suicidal tendencies. , , They represent Article 1 and Article 2 respectively. j Article, No. n A suicidal post. , , Is with , , The corresponding posting time. Similarly, Posts representing non-suicidal tendencies, among which A collection of posts indicating non-suicidal tendencies. , , They represent Article 1 and Article 2 respectively. j Article, No. m A post with non-suicidal tendencies , , Is with , , The corresponding posting time.

[0071] The language style of social media users can reflect their mental state. Therefore, the semantic features of posts are important for understanding users' suicidal thoughts. Given the excellent performance of the pre-trained language model BERT on a large number of NLP tasks, SentenceBERT is used to obtain the embedding representation of the input post.

[0072] The semantic encoding vector of suicidal posts is expressed as follows:

[0073] ;

[0074] in, Indicates user u i The j A suicidal post. Indicates and The semantic encoding vector of the corresponding suicidal post; SentenceBERT represents the semantic encoding function.

[0075] The semantic encoding vector of non-suicidal posts is expressed as follows:

[0076] ;

[0077] in, Indicates user u i The j A post with non-suicidal tendencies Indicates and The semantic encoding vector of the corresponding non-suicidal posts.

[0078] While conventional pre-trained models can semantically model text, they cannot capture the fine-grained emotions expressed in social media posts. To capture these fine-grained emotions, the Plutchik-2 emotion wheel, composed of eight emotions, was used: joy, sadness, surprise, anticipation, anger, fear, trust, and disgust.

[0079] In this embodiment, the sentiment encoding model is the PlutchikBERT model obtained by fine-tuning the pre-trained BERT model using the Plutchik-2 dataset. It can extract fine-grained sentiment encoding from Reddit posts.

[0080] The expression for the sentiment vector of suicidal posts is as follows:

[0081] ;

[0082] in, Indicates and The sentiment vector of the corresponding suicidal post; PlutchikBERT represents the sentiment encoding function.

[0083] The expression for the sentiment vector of non-suicidal posts is as follows:

[0084] ;

[0085] in, Indicates and The sentiment vector of the corresponding non-suicidal posts.

[0086] The expression for the initialization vector of suicidal posts is as follows:

[0087] ;

[0088] in, Indicates and The initialization vector for the corresponding suicidal post. Indicates a splicing operation;

[0089] In the step of obtaining the initialization vector of a non-suicidal post based on its semantic encoding vector and sentiment vector, the expression for the initialization vector of the non-suicidal post is as follows:

[0090] ;

[0091] in, Indicates and The initialization vector for the corresponding suicidal post.

[0092] Step 2: Input the initialization vector of the suicidal post into the Bi-LSTM layer to obtain the hidden state vector of the suicidal post, and obtain the context representation vector of the suicidal post based on the hidden state vector of the suicidal post.

[0093] Users with suicidal thoughts or behaviors frequently post on Reddit's SuicideWatch subreddit, revealing their painful experiences and seeking help, so every post with suicidal tendencies may indicate a suicide risk. Modeling the suicide context of users can provide deeper insights into the generation of suicidal ideation and how their mental state changes over time. Given that Long Short-Term Memory (LSTM) networks are well-suited for capturing long-term dependencies, Bi-LSTM is employed to encode users' changing mental states over time.

[0094] The expression for the latent state vector of suicidal posts is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] in, Indicates user u i The j The hidden state vector of a suicidal post. Indicates user u i The j The forward hidden vector of a suicidal post. Indicates user u i The j The backward hidden vector of a suicidal post. Indicates user u i The j -1 forward hidden vector of suicidal posts, Indicates user u i The j -1 backward hidden vector of suicidal posts.

[0099] In addition, this embodiment also employs an attention mechanism to demonstrate the importance of different suicidal posts for users' suicide risk assessment.

[0100] Specifically, the expression for the context representation vector of suicidal posts is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] in, Indicates user u i The contextual representation vector of posts with suicidal tendencies. Indicates the first j The attention coefficient of a suicidal post Indicates the first j The attention weight of a suicidal post after normalization. n The total number of posts with suicidal tendencies exp Based on the natural constant e An exponential function with base 0. , , Trainable parameters for calculating attention to suicidal posts. T Indicates transpose. tanh This represents the hyperbolic tangent function.

[0105] Step 3: Input the initialization vector of the non-suicidal post into the time-aware T-LSTM layer to obtain the hidden state vector of the non-suicidal post, and obtain the context representation vector of the non-suicidal post based on the hidden state vector of the non-suicidal post.

[0106] In the real world, users' psychological states change over time, so it's necessary to incorporate posting intervals to simulate this process as closely as possible. Clearly, the longer the time interval between the previous and current post, the smaller the impact of the previous post on the user's current psychological state. However, existing LSTM models assume that the time intervals between consecutive posts are equal, failing to consider the significant variations in these intervals, ranging from seconds to days. To simulate changes in user emotional states from a large number of non-suicidal posts, a time-aware T-LSTM layer is employed. This layer incorporates the time intervals between consecutive posts into the LSTM to capture the irregularities in historical posting times.

[0107] T-LSTM layers apply time interval information to the decay of memory cells. Long-term memory cannot be completely discarded, while short-term memory should be... and The time interval between two posts published at two different times is adjusted proportionally, i.e., the short-term memory units are weighted. T-LSTM applies a monotonically decreasing function to the elapsed time interval, converting time into weights.

[0108] Specifically, the T-LSTM layer satisfies the following condition:

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] in, express t Short-term memory units at time -1 tanh Represents the hyperbolic tangent function. , For trainable network parameters, express t The memory unit at time -1 Indicates user u i The time interval between two consecutive non-suicidal posts. Represents the decay function. express t Short-term memory units decaying at time -1 express t Long-term memory units at time -1 Indicates the adjusted t The memory unit at time -1.

[0115] In this embodiment, an attention mechanism is also used to learn the adaptive weights of the hidden states of each post. The final representation of non-suicidal posts is the weighted sum of the hidden states of all posts and the attention weights.

[0116] Specifically, the expression for the latent state vector of non-suicidal posts is as follows:

[0117] ;

[0118] in, Indicates user u i The j The hidden state vector of a non-suicidal post, T-LSTM represents a time-aware T-LSTM function;

[0119] The expression for the context representation vector of non-suicidal posts is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] in, Indicates user u i The contextual representation vector of non-suicidal posts. Indicates the first j The attention score of a non-suicidal post. Indicates the first j The attention weight of non-suicidal posts after normalization. m The total number of posts that are not suicidal. , , Trainable parameters for calculating attention to non-suicidal posts.

[0124] Step 4: Obtain the LIWC feature vector and posting time feature vector through the user feature extraction layer.

[0125] Individuals in different psychological states exhibit significant differences in language expression and wording style. Therefore, this embodiment uses the 2007 version of the LIWC dictionary to calculate the wording differences in suicidal tendencies in posts, in order to help detect users' mental health status. Specifically, a 64-dimensional LIWC feature vector is defined. Corresponding to the 64 word classes in the LIWC dictionary, and then the user u i By combining the words in all suicidal posts, the frequency of words in each category is calculated, which is the proportion of words in that category to the total number of words in suicidal posts. The word frequency of this category is used as the feature value of the corresponding dimension.

[0126] Furthermore, users' posting behavior (such as posting time) can also reflect differences in their psychological state. To encode this information, a 4-dimensional posting time feature vector is defined. Each dimension represents the proportion of posts published by users within the time periods (0:00-5:59), (6:00-11:59), (12:00-17:59), and (18:00-23:59) of the day to all posts.

[0127] Step 5: In the classification layer, the context representation vectors of suicidal posts, the context representation vectors of non-suicidal posts, the LIWC feature vectors, and the posting time feature vectors are concatenated. The concatenated vectors are then fed into a fully connected network, and the probability distribution of each suicide risk level is calculated using a normalized exponential function.

[0128] In this embodiment, users will be divided into four suicide risk levels: NR (no risk), LR (low risk), MR (medium risk), and SR (high risk). The specific formula for calculating the probability distribution is as follows:

[0129] ;

[0130] in, The probability distribution of suicide risk levels is represented by Softmax, which represents the normalized exponential function. , These are the trainable parameters in the normalized exponential function. Indicates user u i LIWC eigenvectors, Indicates user u i The posting time feature vector.

[0131] In summary, according to the suicide risk assessment method based on emotion perception provided in this embodiment, in the post embedding layer, the embedded representation of the input post is obtained through an emotion coding model, resulting in the emotion vectors of suicidal tendencies and non-suicidal tendencies. These vectors are used to generate the initialization vectors of suicidal tendencies and non-suicidal tendencies, thus capturing the fine-grained emotions expressed in the posts. In the T-LSTM layer, the dynamic changes of emotions in non-suicidal tendencies are simulated through a time-aware T-LSTM model, making full use of time interval information. Finally, the probability distribution of each suicide risk level is obtained by combining the context representation vectors of suicidal tendencies, non-suicidal tendencies, LIWC feature vectors, and posting time feature vectors, thereby improving the accuracy of suicide risk assessment.

[0132] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0133] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A suicide risk assessment method based on emotion perception, characterized in that, This is applied to an emotion-based suicide risk assessment model, which includes a post embedding layer, a Bi-LSTM layer, a T-LSTM layer, a user feature extraction layer, and a classification layer. The method includes: In the post embedding layer, the embedding representation of the input post is obtained through the semantic encoding model, resulting in the semantic encoding vector of suicidal tendencies and the semantic encoding vector of non-suicidal tendencies. In addition, the embedding representation of the input post is obtained through the sentiment encoding model, resulting in the sentiment vector of suicidal tendencies and the sentiment vector of non-suicidal tendencies. Then, the initialization vector of suicidal tendencies is obtained based on the semantic encoding vector and the sentiment vector of suicidal tendencies. The initialization vector of non-suicidal tendencies is obtained based on the semantic encoding vector and the sentiment vector of non-suicidal tendencies. The initialization vector of the suicidal post is input into the Bi-LSTM layer to obtain the hidden state vector of the suicidal post, and the context representation vector of the suicidal post is obtained based on the hidden state vector of the suicidal post. The initialization vector of the non-suicidal post is input into the time-aware T-LSTM layer to obtain the hidden state vector of the non-suicidal post, and the context representation vector of the non-suicidal post is obtained based on the hidden state vector of the non-suicidal post. LIWC feature vector and posting time feature vector are obtained through the user feature extraction layer; In the classification layer, the context representation vectors of suicidal posts, the context representation vectors of non-suicidal posts, the LIWC feature vectors, and the posting time feature vectors are concatenated. The concatenated vectors are then fed into a fully connected network, and the probability distribution of each suicide risk level is calculated using a normalized exponential function.

2. The suicide risk assessment method based on emotion perception according to claim 1, characterized in that, The semantic encoding vector of suicidal posts is expressed as follows: in, Indicates user u i The jth post about suicidal tendencies Indicates and The semantic encoding vector of the corresponding suicidal post; SentenceBERT represents the semantic encoding function. The semantic encoding vector of non-suicidal posts is expressed as follows: in, Indicates user u i The jth post, which is not suicidal. Indicates and The semantic encoding vector of the corresponding non-suicidal posts.

3. The suicide risk assessment method based on emotion perception according to claim 2, characterized in that, The emotion encoding model is the PlutchikBERT model obtained by fine-tuning the pre-trained BERT model using the Plutchik-2 dataset.

4. The suicide risk assessment method based on emotion perception according to claim 3, characterized in that, The expression for the sentiment vector of suicidal posts is as follows: in, Indicates and The sentiment vector of the corresponding suicidal post; PlutchikBERT represents the sentiment encoding function. The expression for the sentiment vector of non-suicidal posts is as follows: in, Indicates and The sentiment vector of the corresponding non-suicidal posts.

5. The suicide risk assessment method based on emotion perception according to claim 4, characterized in that, In the step of obtaining the initialization vector of a suicidal post based on its semantic encoding vector and sentiment vector, the expression for the initialization vector of the suicidal post is as follows: in, Indicates and The initialization vector for the corresponding suicidal post. Indicates a splicing operation; In the step of obtaining the initialization vector of a non-suicidal post based on its semantic encoding vector and sentiment vector, the expression for the initialization vector of the non-suicidal post is as follows: in, Indicates and The initialization vector for the corresponding suicidal post.

6. The suicide risk assessment method based on emotion perception according to claim 5, characterized in that, In the step of inputting the initialization vector of suicidal posts into a Bi-LSTM layer to obtain the hidden state vector of suicidal posts, and then obtaining the context representation vector of suicidal posts based on the hidden state vector, the expression for the hidden state vector of suicidal posts is as follows: in, Indicates user u i The hidden state vector of the j-th suicidal post, Indicates user u i The forward hidden vector of the j-th suicidal post, Indicates user u i The backward hidden vector of the j-th suicidal post, Indicates user u i The forward hidden vector of the (j-1)th suicidal post, Indicates user u i The backward hidden vector of the (j-1)th suicidal post; The expression for the context representation vector of suicidal posts is as follows: in, Indicates user u i The contextual representation vector of posts with suicidal tendencies. Let α represent the attention coefficient of the j-th suicidal post. j Let represent the normalized attention weight of the j-th suicidal post, n be the total number of suicidal posts, exp be an exponential function with base e, and c be an exponential function. r W r b r Trainable parameters are calculated for attention to suicidal posts, where T represents the transpose and tanh represents the hyperbolic tangent function.

7. The suicide risk assessment method based on emotion perception according to claim 6, characterized in that, In the step of inputting the initialization vector of non-suicidal posts into a T-LSTM layer to obtain the hidden state vector of non-suicidal posts, and then obtaining the context representation vector of non-suicidal posts based on the hidden state vector, the expression for the hidden state vector of non-suicidal posts is as follows: in, Indicates user u i The hidden state vector of the j-th non-suicidal post, where T-LSTM represents a time-aware T-LSTM function; The expression for the context representation vector of non-suicidal posts is as follows: in, Indicates user u i The contextual representation vector of non-suicidal posts. β represents the attention coefficient of the j-th non-suicidal post. j Let represent the normalized attention weight of the j-th non-suicidal post, m be the total number of non-suicidal posts, and c be the total number of posts. u W u b u Trainable parameters for calculating attention to non-suicidal posts.

8. The suicide risk assessment method based on emotion perception according to claim 7, characterized in that, The formula for calculating the probability distribution of each suicide risk level using the normalized exponential function is as follows: in, This represents the probability distribution of suicide risk levels, where Softmax represents the normalized exponential function, and W... t b t These are the trainable parameters in the normalized exponential function. Indicates user u i LIWC eigenvectors, Indicates user u i The posting time feature vector.

9. The suicide risk assessment method based on emotion perception according to claim 1, characterized in that, The T-LSTM layer satisfies the following condition: g(D t )=1 / D t in, Let W represent the short-term memory unit at time t-1, tanh represent the hyperbolic tangent function, and W... k b k For trainable network parameters, C t-1 Δ represents the memory cell at time t-1. t Indicates user u i The time interval between two consecutive non-suicidal posts, where g represents the decay function. This represents the short-term memory unit after decay at time t-1. This represents the long-term memory unit at time t-1. This represents the memory cell at time t-1 after adjustment.

Images