Information processing method

Abstract

The invention discloses an information processing method, which comprises the following steps of: acquiring original information text data from subscribed terminal data streams, and preprocessing and outputting the original information text data; a double-layer BiLSTM-CRF neural network structure is adopted to recognize diversified time expressions, and explicit time expressions are extracted; implicit time information is captured through a multi-head self-attention mechanism, and the implicit time information is extracted in combination with a context understanding model; the relative time is converted into absolute time through time sequence graph reasoning; the reminding frequency and mode are dynamically adjusted based on the time urgency and information importance two-dimensional score; the expired information is automatically classified and archived, and the storage space is released; according to the method disclosed by the invention, the problems that text information has context ambiguity influence and depends on dates can be further solved, fuzzy expression cannot be solved, and the time extraction precision under complex expressions including nesting time entities and the like can also be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to an information processing method. BACKGROUND

[0002] In the existing various terminals, a large amount of information, emails, instant messages or APP push can be generated every day, and these generated information are various time-sensitive information, that is, bank notifications, operator packages, merchant promotions in the information, subscription information, bill information, activity invitations in the email, group notifications in the instant message, time-limited activities shared by friends, and coupon information, member day reminder information pushed by APP. All of them contain explicit or implicit "validity period", but the validity period in these information is usually mixed in the text, which needs to be carefully read by the user to find out. Among them, the information expression such as "before a certain month a certain day", "validity period to a certain month", "before the end of the month", "last three days" is non-uniform and standardized. When the user quickly browses the information, it is easy to miss these key time points.

[0003] In the prior art, the time expression is mainly extracted from the information by natural language processing technology, the time information is extracted from various messages, and the time information is converted into standardized time data, but the recognition accuracy of the complex time-sensitive information is not high, and the recognition error often occurs. Therefore, by further adopting a deep learning and rule hybrid architecture, combining a pre-training language model and an attention mechanism to extract time expressions, the recognition accuracy of the time can be improved under a single natural language processing technology, but the judgment standards of time urgency and importance are different in different scenarios, the diversity and complexity of time expressions, that is, the conversion accuracy is not high in the process of converting the multi-style expression into a standard time format, and the missed conversion often leads to missed reminders. SUMMARY

[0004] The purpose of the present application is to provide an information processing method to solve the problems raised in the background art.

[0005] The specific technical solution provided by the present application is as follows: an information processing method, comprising the following operation steps: S1: obtaining original information text data from each terminal data stream subscribed, and pre-processing and outputting.

[0006] Preferably, from the terminal data stream of the existing subscription, the original information text with timestamp and terminal ID metadata is pulled, a set data delay variable is introduced, and the pulling frequency is adjusted according to the variable when pulling the original information text; at the same time, hierarchical preprocessing is performed using a rule engine, and the byte stream, text structure and context rationality are processed from the three layers of characters, syntax and semantics; noise data is removed through preprocessing, and unified encoding is performed based on byte entropy value using an encoding detection algorithm; after unified encoding, the temporary buffer area of the memory queue is written, and a timeliness weight factor is introduced in the buffer area to prioritize the output of emergency data.

[0007] S2: Based on the preprocessed information text, a double-layer BiLSTM-CRF neural network structure is used to recognize diversified time expressions and extract explicit time expressions.

[0008] Preferably, the double-layer BiLSTM-CRF neural network structure includes a first layer BiLSTM and a second layer BiLSTM, the first layer BiLSTM is used to capture word-level features and output hidden state sequences based on forward and backward LSTM processing sequences; the second layer BiLSTM is used to calculate enhanced hidden states.

[0009] Preferably, for the original information text output from the buffer area, a rule model is used for priority matching: a pre-defined time regular library is matched, the time regular library includes definitions of basic time, absolute time, short-term relative time and extended week expressions; based on the defined time regular library, the original information text is scanned to lock the time expression candidate interval in the information text, generate a protection mark, convert the time expression to a placeholder expression, and store the original expression in a mapping table; the remaining text is input into a BiLSTM word segmentation model to output basic word segmentation results, and then the fragments are recombined, i.e. the placeholder is restored to the original expression and inserted into the corresponding position to form a final word segmentation sequence; the final word segmentation sequence is divided into

[0010] ​Preferably, the first layer BiLSTM is used to process local features within the node, and the second layer BiLSTM is used to enhance the hidden state by cross-node collaboration; the hidden state sequence output by the second layer BiLSTM is calculated to calculate the scores of all labels of the first word, and a recursive calculation is performed for each position and label; when calculating the scores of all possible label sequences, the Viterbi algorithm dynamic programming is further used to find the optimal label sequence, and the basic information of each extracted time expression is extracted; then, a rule and statistical comprehensive classifier is used for automatic classification of types, and the classified information is compared with the placeholder position in the mapping table; if the output position deviates from the rule engine position by more than a set character threshold, error correction is triggered, otherwise the extracted display time expression is output.

[0011] S3: Capture implicit time information through multi-head self-attention mechanism, and extract implicit time information in combination with context understanding model.

[0012] Preferably, the extraction of implicit time information is performed by fusing multi-head self-attention and context understanding model, using multi-head mechanism to focus on multiple context points in parallel, and using context model to learn global dependencies based on Transformer for keywords of a certain event.

[0013] Preferably, the input text word vector is linearly projected into Q, K, and V corresponding to the query, key, and value matrices; the input is divided into h heads by multi-head self-attention, and the attention weight of each head is calculated, that is, the attention score of the trigger word and all event nouns is calculated, the event with the highest score is selected as the reference event, and the inter-word correlation weight is obtained; the weighted sum value vector is output, and the multi-head splicing vector is output; the key context is focused by linear projection, the pre-trained language model BERT is fine-tuned by context understanding model, the multi-head splicing vector is input into the Transformer layer based on multi-layer attention and feedforward network, the context vector is generated through residual connection and layer normalization; the event relationship in the context vector is analyzed, and the implicit time expression is output.

[0014] Preferably, the output implicit time expression specifically includes: identifying the trigger time offset keyword from the context vector, identifying the offset numerical value and unit through regular expression matching, and converting the numerical value using the pre-trained number analysis model when the regular expression matching fails; identifying the event entity through the context understanding model Transformer, analyzing the event relationship in the generated context vector, and determining which event is the reference of the current time offset, outputting the time offset relative to a certain reference event, and obtaining the corresponding implicit time expression: {“time offset”: “numerical value with unit”; “reference event”: “explicitly mentioned event”} S4: Convert relative time to absolute time by using time sequence atlas reasoning.

[0015] Preferably, the extracted time expressions of display time and implicit time are parsed into relative time offsets and reference events, identifying the reference time including the message timestamp ; construct the timing graph, create nodes containing each event time variable, add edges containing relative time offsets and relationship types; set the node with known absolute time as the anchor node, take the sending time of the message as the initial anchor point for the initialization of the timing graph, aggregate neighbor information for each node using GNN, calculate the neighbor time weighted average, update the absolute time with timestamp , where is the correction term to solve the inconsistency of the timing graph, which is updated iteratively, that is, for each non-anchor node, collect all neighbor estimates, perform weighted fusion, calculate the correction term, and converge when the time difference changes less than the set threshold, output the final absolute timestamp.

[0016] S5: Based on the two-dimensional score of time urgency and information importance, dynamically adjust the reminding frequency and method.

[0017] Preferably, read the event absolute time, get the content from the original text, and calculate the urgency; extract keywords, calculate the information importance based on keywords and sentiment using a text classification model; calculate the comprehensive score based on urgency and information importance ; dynamically adjust the feedback according to the comparison of the comprehensive score with the set threshold; mark the notification information based on the dynamic gradient of the HSL color space during the adjustment process.

[0018] S6: Automatically classify and archive expired information to release storage space.

[0019] S7: Synchronize offline recognition using a lightweight architecture, and integrate with existing terminals that generate each message.

[0020] Compared with the prior art, the present application has the beneficial effects that: by constructing a double-layer BiLSTM-CRF architecture, the first layer BiLSTM is used to capture word-level features, and the second layer is used to enhance context dependence, which can further process the problems of context ambiguity and dependence on dates in text information, and improve the time extraction accuracy under complex expressions such as nested time entities. Meanwhile, the multi-head self-attention and context understanding model are fused, the multi-head mechanism is used to focus on multiple context points including event keywords in parallel, the context model learns global dependence based on Transformer, solves ambiguous expressions, and improves the implicit time recall rate. Finally, aiming at the problem that the existing rule engine or simple offset calculation cannot process complex time dependence, an event-time graph structure is constructed, and the conflict is solved by graph propagation algorithm, which not only supports multi-event chain conversion, but also improves the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a step flow diagram of an information processing method provided by an embodiment of the present application. DETAILED DESCRIPTION The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the present application.

[0022] Embodiment 1 As shown in Figure 1 , the information processing method described in this embodiment includes the following steps: S1: Obtain original information text data from each terminal data stream subscribed.

[0023] In this embodiment, by pulling the original information text with timestamp and terminal ID metadata from each terminal data stream subscribed, a set data delay variable is introduced, and the pulling frequency is adjusted according to the variable when pulling the original information text. When the variable is within the set real-time pulling threshold, real-time pulling processing is performed; when the variable is within the set batch pulling threshold, batch pulling processing is performed; when the variable is within the set delay pulling threshold, delay pulling processing is performed. Meanwhile, the rule engine is used for hierarchical preprocessing, and the byte stream, text structure and context rationality are processed from the three layers of characters, syntax and semantics, that is, ASCII control character filtering and byte-by-byte filtering invisible characters, using regular expressions to match the garbled code mode through the predetermined garbled code mode library, and detecting the text repetition rate to delete the repeated fragments. Through preprocessing to remove noise data, and using encoding detection algorithm based on byte entropy value to unify the encoding, after unified encoding, write to the temporary buffer area of the memory queue, introduce time limit weight factor in the buffer area, output the emergency data preferentially.

[0024] S2: Adopting double-layer BiLSTM-CRF neural network structure to identify diversified time expressions and extract explicit time expressions.

[0025] In the embodiment, since the existing single-layer LSTM or rule matching is in the process of diversified time expressions, when there is "next Monday" or "July 30, 2025", it is easy to be affected by the context ambiguity, that is, when there is the text information "tomorrow", it depends on the date, therefore, in the application, the double-layer BiLSTM-CRF architecture is constructed, the first layer BiLSTM is used to capture the word-level features, the second layer is used to enhance the context dependence, and the CRF is used to solve the label conflict. Compared with the traditional single-layer model, it can further process complex expressions such as nested time entities.

[0026] Exemplarily, for the original information text output from the buffer area, the rule model is used to match preferentially: the pre-defined matching time regular library includes the definition of the basic time, absolute time, short-term relative time and extended week expression, and the basic time regular expression is as follows: wherein, "r"..." " represents the original string, and no additional escape is required when matching, " represents the matching of the number expression of [0-9], " represents the matching of 1 to 2 digits, that is, the hour in the time can be represented by 1 or 2 digits, that is, the first part of the time regular expression matches the absolute time representation of "time: minute" format, that is, the time representation of "9:30" or "23:45"; the second part "next week [Monday Tuesday Wednesday Thursday Friday Saturday Sunday]" matches the relative time representation of "next week " expression, and the vertical line " | " represents the logical or relationship. The short-term relative time is defined as "(tomorrow | day after tomorrow)", and the extended week expression is defined as "(last week | this week | next week) [Monday Tuesday Wednesday Thursday Friday Saturday Sunday]. Through the defined time regular expression, not only the explicit time expression can be quickly locked, but also the subsequent deep model error segmentation can be avoided.

[0027] Exemplarily, based on the defined time regular library, the original information text is scanned, and after scanning the original information text, the time expression candidate interval in the information text is locked, a protection mark is generated, and the time expression is converted into a placeholder "##TIME##". "Expression, where time indicates that the replaced content belongs to the category of time expressions, represents a continuous and indivisible text segment, i represents the unique serial number of the current time expression in the text, and finally the original expression is stored in the mapping table. For example, convert the time expression "next Monday 9:00 - 11:30" in the original information text "The meeting is scheduled for next Monday 9:00 - 11:30, please attend on time" into a placeholder representation " ", where represents the segment ID. Therefore, the mapping table representation of the above original information text "The meeting is scheduled for next Monday 9:00 - 11:30, please attend on time" is , Further input the remaining text into the BiLSTM word segmentation model to output the basic word segmentation result. For example: The remaining text "The meeting is scheduled for, please attend on time" is segmented into ["The meeting", "is scheduled for", ",", "please", "attend on time"], and then perform segment recombination, that is, restore the placeholder to the original expression and insert it into the corresponding position to form the final word segmentation sequence: ["The meeting", "is scheduled for", "next Monday 14:00 - 16:00", ",", "please", "attend on time"].

[0028] Divide the segmented sequence into segments, allocate each segment to an independent node for processing, calculate the multi-head attention of each node to its own segment, and output a local attention vector containing the corresponding dimension and segment length. At the same time, the master node aggregates all local attention vectors, filters key information through the gating mechanism, and outputs a global attention vector, that is, the master node is responsible for coordinating and transmitting cross-segment attention vectors.

[0029] After word segmentation, use the first layer of BiLSTM to process local features within the node, that is, learn context features by calculating the forward LSTM and backward LSTM, splice the word vectors of the segmented sequence and the vectors at the corresponding positions in the global attention vector, and output the hidden state. Use the second layer of BiLSTM to enhance the hidden state through cross-node collaboration, that is, input the vector spliced by the first layer and the boundary hidden state of the adjacent node, and use the current node to receive the last hidden state of the previous node and the first hidden state of the next node to enhance the input of the first and last words of the segment.

[0030] The hidden state sequence output by the second BiLSTM layer is calculated, all label scores of the first word are calculated, and a recursive calculation is performed for each position and label. When calculating the scores of all possible label sequences, the Viterbi algorithm dynamic programming is further used to find the optimal label sequence. The basic information of each extracted time expression is extracted, and an automatic classification of the type is performed based on a rule and a statistical comprehensive classifier. The classified information is compared with the placeholder position in the mapping table. If the output position deviates from the rule engine position by more than a set character threshold, error correction is triggered. Otherwise, the extracted display time expression is output.

[0031] S3: Capture implicit time information through multi-head self-attention mechanism, and extract implicit time information in combination with context understanding model.

[0032] In the present embodiment, after the display time expression is proposed as described above, the existing simple attention or rule library technical means cannot capture the implicit time for the implicit time expression such as "one hour after the meeting" depending on a certain event relationship, that is, it is difficult to capture the implicit time. Therefore, the present application uses the multi-head mechanism to focus on multiple context points such as the keywords of a certain event in parallel, uses the context model, learns the global dependency based on the Transformer, adjusts the existing fuzzy expression problem, and improves the recall rate of the implicit time by fusing the multi-head self-attention and the context understanding model.

[0033] For example, the input text word vector is linearly projected into Q, K, and V corresponding to the query, key, and value matrices. The multi-head self-attention is used to divide the input into h heads, and the attention weight of each head is calculated. The attention score between the trigger word and all event nouns is calculated, the event with the highest score is selected as the reference event, the inter-word correlation weight is obtained, and the weighted context vector is output. Finally, the multi-head splicing vector is output, and the key context such as the relationship between "meeting" and "after" is focused by linear projection.

[0034] The reuse context understanding model fine-tunes the pre-trained language model BERT, uses multi-layer attention and feedforward network, inputs the multi-head splicing vector to the Transformer layer, generates the context vector through residual connection and layer normalization, identifies the trigger time offset keywords such as “after”, “before”, “within” and the like from the context vector, identifies the offset numerical value and unit through regular expression matching, for example, the numerical value “one” represents 1, “three” represents 3, the unit “h” represents hour, “min” represents minute and the like, and at the same time when the regular expression matching fails, a pre-trained number analysis model is used for numerical conversion, “three” is converted to 3, and “half” is converted to 0.5, that is, the event noun most related to the trigger word is found through the attention mechanism, and finally the reference event to which the offset is attached is determined, and an event sequence representation is output to infer the time, for example, “after an hour” is offset based on the event; When the multi-head self-attention is fused with the context understanding model, the model outputs the implicit time vector, that is, the implicit time is extracted, the event entity is recognized through the context understanding model Transformer, the event relationship in the generated context vector is analyzed, and it is determined which event is the reference of the current time offset, and the time offset relative to a certain reference event is output, and the corresponding implicit time expression is obtained: {“time offset”: “numerical value with unit”; “reference event”: “explicitly mentioned event”}, for example, there is a relative time expression “after an hour” in natural language, and the relative time corresponds to the implicit time expression {“time offset”: “+1h”; “reference event”: “meeting”}, and subsequent relative time and absolute time conversion is performed. “After” means positive offset, and the output structure of the implicit time expression corresponding to the relative time based on this relative time completely describes the implicit time relationship semantics in the text.

[0035] S4: convert the relative time to absolute time by using the time sequence graph reasoning.

[0036] In the present embodiment, the existing rule engine or simple offset calculation cannot process complex time dependencies, such as “two days after the meeting next week”. Therefore, the present application constructs an event-time graph structure, the nodes represent time points, and the edges represent relationships such as before and after, and solves conflicts through graph propagation algorithm, which not only supports multi-event chain conversion, but also improves the accuracy.

[0037] For example, the time expressions of the display time and the implicit time extracted in steps S2 and S3 are analyzed, and are analyzed into a relative time offset and a reference event, and a reference time including a message timestamp is identified . A time sequence graph is constructed, nodes containing time variables of each event are created, and relative time offsets and relationship type edges. By setting the nodes with known absolute time as anchor nodes, the sending time of the message is taken as the initial anchor for the initialization of the timing diagram. Then the GNN is used to aggregate neighbor information for each node, calculate the neighbor time weighted average, and update the absolute time with timestamp , where To solve the correction term of the inconsistency of the timing diagram, iterative update is performed, that is, for each non-anchor node, the estimates of all neighbors are collected, weighted fusion is performed, the correction term is calculated, and when the iteration converges to a time difference change less than a set threshold, the final absolute timestamp is output.

[0038] S5: Based on the two-dimensional score of time urgency and information importance, dynamically adjust the reminding frequency and method.

[0039] In this embodiment, the absolute time of the reading event is read, the content is obtained from the original text, and the urgency :

[0040] wherein the urgency value range [0, 1], is the absolute time of the event, is the current system time, the smaller the time difference, the greater the value, the more urgent, when the event is about to occur, U tends to 1.

[0041] Then extract keywords including "urgent, deadline, expiration" and the like, and use a text classification model based on keywords and sentiment to calculate information importance :

[0042] wherein, is the TF-IDF weight coefficient, is the sentiment weight coefficient, , is the relative importance calculation of the keywords extracted from the text using the importance algorithm of the keywords, is the sentiment polarity value from text sentiment analysis, , , indicates that the text sentiment analysis result is in the range of strong negative to strong positive, wherein, is the weight matrix, is the text semantic vector, is the bias vector, is the positive sentiment probability output by the classifier.

[0043] Normalized to the value range [0, 1].

[0044] Based on urgency And information importance Calculate the comprehensive score :

[0045] Wherein, is the urgency weight coefficient, is the information importance weight coefficient, according to the comparison of the comprehensive score and the set threshold, the dynamic adjustment feedback is carried out, when the comprehensive score S is greater than the set high frequency threshold, the notification information is marked and pushed by using the high frequency reminding mode, when the comprehensive score S is in the range of the set high frequency threshold and the medium frequency threshold, the notification information is marked and pushed by using the medium frequency reminding mode, when the comprehensive score S is less than the set medium frequency threshold, the notification information is marked and pushed by using the low frequency reminding mode, the notification information is marked based on the dynamic gradient of HSL color space, the information remaining validity period is mapped as the visual gradient from the safe green to the warning yellow and then to the emergency red, and the remaining time countdown label is superimposed on the information thumbnail, the user can identify the time-sensitive information within the set time; and according to the information importance and the urgency, the 5-level reminding priority is automatically distributed, the exponential backoff algorithm is used to dynamically adjust the reminding frequency, the initial interval is 2 hours, the interval is halved for each increase of the urgency level, and the best reminding time is predicted through user behavior analysis. And adaptive learning feedback is carried out, through recording the response mode and processing behavior of the user to the reminding, and combining the time change or new information, dynamic updating is carried out, and the time-sensitive judgment is continuously optimized.

[0046] S6: The expired information is automatically classified and archived, and the storage space is released.

[0047] In the embodiment, the database is automatically queried according to the set time, the expiration time of each information is checked according to the converted event absolute time, and the expiration flag is calculated, when > + expiration threshold, the expiration is marked, for the expired information, the text feature vector is extracted, that is, the TF-IDF vector is extracted, the Naive Bayes classifier is applied to calculate the class probability to which the current document belongs, and the highest probability class is selected. For example, “archive_work”, the high importance class is compressed and moved to the archive storage, the low importance class is directly deleted, and the released space size is recorded, for example, the “garbage” class is directly deleted, and finally the storage saving rate is calculated, triggering the next scan.

[0048] S7: Finally, a lightweight architecture is used for offline identification, and is integrated with existing terminals for generating each message, so as to effectively reduce the occupation of expired information and improve the information retrieval efficiency. In the embodiment, the model is compressed into a micro version through model distillation and quantization, terminal offline operation is supported, and seamless integration is realized.

[0049] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0050] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will still be able to modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An information processing method, characterized in that: Includes the following steps: S1: Obtain raw information text data from the data streams of each subscribed terminal, preprocess and output it; S2: Based on the preprocessed output information text, a two-layer BiLSTM-CRF neural network structure is used to identify diverse time expressions and extract explicit time expressions; The dual-layer BiLSTM-CRF neural network structure includes: a first layer BiLSTM and a second layer BiLSTM. The first layer BiLSTM is used to capture word-level features and outputs a hidden state sequence based on forward and backward LSTM processing sequences. The second layer BiLSTM is used to calculate the enhanced hidden state. S3: Capture latent temporal information through a multi-head self-attention mechanism and extract latent temporal information by combining it with a contextual understanding model; The extraction of implicit temporal information is achieved by fusing multi-head self-attention and contextual understanding models. The multi-head mechanism is used to pay attention to multiple contextual points in parallel. For keywords of a certain event, the contextual model is used to learn global dependencies based on Transformer. S4: Use time-series graph reasoning to convert relative time into absolute time; S5: Based on a dual-dimensional scoring system of time urgency and information importance, dynamically adjust the frequency and method of reminders; S6: Automatically categorizes and archives expired information to free up storage space; S7: Simultaneously adopts a lightweight architecture for offline recognition and integrates with existing terminals that generate various messages.

2. The information processing method according to claim 1, characterized in that: The process of obtaining raw information text data includes: pulling raw information text with timestamps and terminal ID metadata from the data streams of existing subscribed terminals; introducing a set data delay variable and adjusting the pulling frequency based on this variable when pulling raw information text; simultaneously using a rule engine for layered preprocessing, processing byte streams, text structure, and contextual rationality from three layers: character, syntax, and semantics; removing noise data through preprocessing; and performing unified encoding based on byte entropy values ​​using an encoding detection algorithm. After unified encoding, the data is written to a temporary buffer in a memory queue, and a timeliness weight factor is introduced into the buffer to prioritize the output of urgent data.

3. The information processing method according to claim 2, characterized in that: Step S2 specifically includes: For the raw information text output from the buffer, a rule model is used for priority matching: a predefined time regularity library for matching, which includes expression definitions for base time, absolute time, short-term relative time and extended period; Based on the defined time regularity library, the original information text is scanned, the candidate interval of time expression in the information text is locked, a protection tag is generated, the time expression is converted into a placeholder expression, and the original expression is stored in the mapping table. The remaining text is input into the BiLSTM word segmentation model, which outputs the basic word segmentation results. Then, the fragments are reassembled, that is, the placeholders are restored to the original expressions and inserted into the corresponding positions to form the final word segmentation sequence. The final word segmentation sequence is divided into Each segment is assigned to an independent node for processing. The multi-head attention of each node to its own segment is calculated, and the local attention vector containing the corresponding dimension and segment length is output. At the same time, the master node aggregates all local attention vectors, filters key information through a gating mechanism, and outputs a global attention vector. The first layer of BiLSTM is used to process local features within a node, and the second layer of BiLSTM is used to enhance the hidden state through cross-node cooperation. The hidden state sequence output by the second layer BiLSTM is calculated. All label scores of the first word are calculated, and recursive calculation is performed for each position and label. When the scores of all possible label sequences are calculated, the Viterbi algorithm is used to dynamically program and find the optimal label sequence. Basic information is extracted for each extracted time expression. Then, an integrated classifier based on rules and statistics is used to automatically classify the types. The classified information is compared with the placeholder positions in the mapping table. If the deviation between the output position and the position of the rule engine is greater than the set character threshold, error correction is triggered; otherwise, the extracted display time expression is output.

4. The information processing method according to claim 3, characterized in that: The specific steps of converting the time expression into a placeholder expression are: converting the time expression into a placeholder. The expression "time" indicates that the content being replaced belongs to the time expression category. This represents a continuous, indivisible text segment, where i represents the unique index of the current time expression within the text.

5. The information processing method according to claim 4, characterized in that: The first layer BiLSTM processes local features within a node by: learning contextual features by computing forward LSTM and backward LSTM, concatenating the word vectors of the segmented sequence with the vectors at corresponding positions in the global attention vector, and outputting the hidden state; The second layer of BiLSTM enhances the hidden state through cross-node collaboration, including: inputting the concatenated vector from the first layer and the boundary hidden states of neighboring nodes, and utilizing the current node... Receive from the previous node The last hidden state and the next node The first hidden state is used to enhance the input of the first and last words of the fragment.

6. The information processing method according to claim 5, characterized in that: The fusion model of multi-head self-attention and contextual understanding specifically includes: The input text word vectors are linearly projected into the query, key, and value matrices corresponding to Q, K, and V; Using multi-head self-attention segmentation with h heads as input, the attention weight of each head is calculated, that is, the attention score between the trigger word and all event nouns is calculated, and the event with the highest score is selected as the benchmark event to obtain the inter-word relevance weight. The weighted summation vector is output as a concatenated vector with multiple heads. By leveraging linear projection to focus on key contexts, and employing a context understanding model to fine-tune the pre-trained language model BERT, a multi-layer attention and feedforward network is used to input multi-head concatenation vectors into the Transformer layer. Through residual connections and layer normalization, context vectors are generated. Parse the event relationships in the context vector and output the implicit time expression.

7. The information processing method according to claim 6, characterized in that: The output implicit time expression specifically includes: identifying keywords that trigger time offsets from the context vector; identifying the offset value and unit through regular expression matching; when regular expression matching fails, using a pre-trained numerical parsing model to convert the value; identifying event entities through the Transformer context understanding model; parsing the event relationships in the generated context vector; determining which event is the reference for the current time offset; and outputting the time offset relative to a certain reference event to obtain the corresponding implicit time expression: {"time offset": "value with unit"; "reference event": "explicitly mentioned event"}.

8. The information processing method according to claim 7, characterized in that: The conversion of relative time to absolute time includes: The extracted explicit and implicit time expressions are parsed and converted into relative time offsets. And reference events, identify reference times including message timestamps. ; Construct a time series graph, create nodes containing the time variables for each event, and add nodes containing relative time offsets. Edges of relational type; Nodes with known absolute times are set as anchor nodes, and the message sending time is used as the initial anchor point for initializing the time sequence graph. A GNN is used to aggregate neighbor information for each node, calculate a time-weighted average of neighbor times, and update the timestamped absolute time. , ,in To address the inconsistency in the time series graph, an iterative update is performed. Specifically, for each non-anchor node, estimates from all its neighbors are collected, weighted, and fused to calculate the correction term. The iteration converges when the change in time difference is less than a set threshold, and the final absolute timestamp is output.

9. The information processing method according to claim 8, characterized in that: Step S5 specifically includes: Read the absolute time of the event, extract the content from the raw text, and calculate the urgency. Extract keywords, and use a text classification model based on keywords and sentiment to calculate the importance of information; Based on urgency And the importance of information Calculate the overall score; The system dynamically adjusts and provides feedback based on a comparison between the overall score and the set threshold. During the adjustment process, notification information is marked based on the dynamic gradient of the HSL color space.

10. The information processing method according to claim 9, characterized in that: Step S6 includes: automatically querying the database according to the set time, checking the expiration time of each piece of information according to the converted absolute time of the event, calculating the expiration flag, extracting text from the expired information to generate feature vectors, applying a Naive Bayes classifier to calculate the probability of the current document belonging to the category, selecting the category with the highest probability, compressing the high-importance categories and moving them to archive storage, and directly deleting the low-importance categories.

Images