Method and server for processing a remote interactive session based on an ops microcomputer

Abstract

The application provides a remote interactive session processing method and server based on an OPS microcomputer, which comprises the following steps: on the OPS terminal side, the semantic unit cohesion relationship of a remote session flow is analyzed, the potential cohesion relationship between the speech segment units of a sending end and a receiving end is extracted, and a session promotion semantic chain is generated; when a transmission interruption event occurs, self-regression deduction is performed according to a semantic trajectory, a transitional session segment is generated, and a receiving end presentation link is embedded; the potential intention pointing trend is determined based on the speech prosody fluctuation characteristics and the semantic turning vector of the receiving end, and a visual semantic guide anchor point is established and pushed to a session interface; the session rhythm generation adjustment strategy is determined through the interaction behavior data of the anchor point of the receiving end, the generation frequency of the transitional session segment and the anchor point pushing time are adjusted, the overall structure modeling of the session intention trajectory is realized, the transitional content which is coherent with the context semantic is generated when the transmission is interrupted, the potential intention direction is perceived, and the dynamically adapted guide strategy is provided.

Description

Technical Field

[0001] This invention relates to the field of computer interaction technology, specifically to a remote interactive session processing method and server based on an OPS microcomputer. Background Technology

[0002] In remote interactive session systems, the OPS microcomputer, acting as the terminal device, typically runs the session transport protocol stack and real-time communication applications. It is responsible for encapsulating the sender's voice or text messages into a remote session stream and transmitting it to the receiver for presentation. To enhance the interactive experience, existing systems typically feature voice activity detection and jitter buffer management, maintaining audio continuity by adjusting playback latency or packet loss compensation during network fluctuations. Some technologies also integrate speech transcription and semantic analysis modules, enabling real-time transcription and keyword extraction of session content, which is then displayed as subtitles or floating labels on the receiver's interface. Additionally, some technologies attempt to predict a list of phrases the user might reply to based on historical session content, allowing the receiver to quickly input them. However, such predictions are usually independent of the deeper intent of the current session and cannot cope with content loss caused by transmission interruptions.

[0003] However, existing systems lack overall structural modeling of the session intent trajectory, making it difficult to generate transitional content that is semantically coherent with the context when transmission is interrupted, and also unable to proactively perceive the potential intent direction of the receiving end and provide dynamically adapted guidance strategies. Summary of the Invention

[0004] The purpose of this invention is to provide a remote interactive session processing method and server based on an OPS microcomputer to solve the problems mentioned in the background art.

[0005] This invention provides a remote interactive session processing method based on an OPS microcomputer, comprising: On the OPS terminal side, semantic unit connection relationship parsing processing is performed on the remote session stream to extract the potential semantic unit connection relationship between the sending end speech fragment unit and the receiving end speech fragment unit in the remote session stream, and generate a session advancement semantic chain. The session advancement semantic chain is used to characterize the semantic trajectory of the intention direction in the remote session stream. When a transmission interruption event occurs in the remote session stream, the sender speech segment unit is subjected to autoregressive inference processing based on the semantic trajectory to generate a transitional session segment and embed it into the timing position in the receiving end presentation link corresponding to the transmission interruption event. The potential intent direction trend of the receiver conversation in the remote conversation stream is determined based on the prosodic fluctuation features and semantic transition vectors of the receiver speech fragment units. Based on the potential intent pointing trend, a visual semantic guidance anchor is established, and the visual semantic guidance anchor is pushed to the conversation interface of the receiving end. The visual semantic guidance anchor is used to guide the response direction of the conversation of the receiving end. The receiving end collects interactive behavior data on the visual semantic guidance anchor point, and determines a conversation rhythm generation and adjustment strategy based on the interactive behavior data. The conversation rhythm generation and adjustment strategy is used to adjust the generation frequency of the transitional conversation segments and the push timing of the visual semantic guidance anchor point.

[0006] This invention provides a remote interactive session processing server, comprising: A processor; a storage device on which a computer program is stored; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the above-described remote interactive session processing method based on an OPS microcomputer.

[0007] The present invention provides a readable storage medium on which a program or instruction is stored, and when the program or instruction is executed by a processor, it implements the above-described remote interactive session processing method based on an OPS microcomputer.

[0008] Compared with existing technologies, the beneficial effects of this invention are: It establishes a processing logic on the OPS terminal side encompassing semantic unit connection parsing, transmission interruption compensation, latent intent perception, guidance anchor point generation, and rhythm adaptive adjustment. By using a conversation-progressive semantic chain to structurally represent the intent trajectory of interleaved utterance fragments, it enables autoregressive deduction based on this semantic trajectory during transmission interruptions, generating transitional conversation fragments consistent with the contextual intent direction to fill conversation gaps. Simultaneously, by fusing the prosodic fluctuations of the receiver's utterances with semantic transition vectors, it mines latent intent trends. The resulting visualized semantic guidance anchor points can make implicit intent trends explicit as directional guidance. Furthermore, by collecting data on the receiver's interaction with the anchor points, it dynamically adjusts the conversation rhythm generation strategy, achieving real-time adaptation between guidance intensity and receiver response patterns. This realizes a systematic functional leap in remote conversations from passive transmission to active semantic compensation and intelligent guidance. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1A flowchart illustrating a remote interactive session processing method based on an OPS microcomputer, provided as an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of three-terminal interaction for remote interactive session processing based on an OPS microcomputer, provided as an embodiment of the present invention.

[0012] Figure 3 This is an exemplary interface diagram of the receiving end provided in an embodiment of the present invention. Detailed Implementation

[0013] 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, and 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.

[0014] Please see Figure 1 , Figure 1 This is a flowchart of a remote interactive session processing method based on an OPS microcomputer provided by an embodiment of the present invention. The method can be executed by a remote interactive session processing server and includes steps 110-150.

[0015] The embodiments of the present invention can realize the overall structural modeling of the session intent trajectory, thereby generating transitional content that is consistent with the context semantics when the transmission is interrupted, and then actively sensing the potential intent direction of the receiving end and providing a dynamically adapted guidance strategy.

[0016] Step 110: On the OPS terminal side, perform semantic unit connection relationship parsing on the remote session stream, extract the potential semantic unit connection relationship between the sending end speech fragment unit and the receiving end speech fragment unit in the remote session stream, and generate a session advancement semantic chain.

[0017] In this step, the OPS microcomputer acts as the terminal device for the remote interactive session. The remote session stream it receives contains multiple sender speech fragment units generated by the sender and multiple receiver speech fragment units generated by the receiver. These speech fragment units are arranged interleaved in the session stream according to their respective timestamps. The purpose of parsing the semantic unit connection relationship of the remote session stream is to identify which sender speech fragment units and which receiver speech fragment units have semantic connection or response relationships from the interleaved speech fragment units, and to organize these relationships into a semantic trajectory structure that can characterize the intention direction in the remote session stream.

[0018] Step 1101: Perform semantic unit segmentation on the sender speech fragment units in the remote session stream to obtain a sender semantic unit sequence. The sender semantic unit sequence contains multiple sender semantic units with semantic independence.

[0019] Specifically, the text content data of each utterance segment unit from the sending end is first acquired. This text content data can be obtained by transcribing the speech input from the sending end using a speech recognition engine, or extracted directly from the text message input from the sending end. Then, the text content data is segmented according to the principle of semantic integrity. The principle of semantic integrity uses sentence boundaries and punctuation marks as the primary segmentation criteria, and at the same time, it combines the results of dependency parsing to correct the segmentation boundaries, ensuring that each semantic unit from the sending end obtained by segmentation has the ability to independently express a complete semantic proposition.

[0020] Dependency parsing determines the boundaries of semantic units by identifying the dependency relationships between the core predicate and the arguments it governs in the text content data. When all the arguments governed by the core predicate appear, a complete semantic proposition is formed, and the text span corresponding to this semantic proposition is a sender semantic unit. All the segmented sender semantic units are arranged according to their order of appearance in the sender discourse segment units, and combined with the timestamp of each sender discourse segment unit in the remote conversation stream, corresponding time position information is assigned to each sender semantic unit, thus forming a sequence of sender semantic units.

[0021] Step 1102: Perform semantic unit segmentation on the receiving speech fragment units in the remote session stream to obtain a receiving semantic unit sequence. The receiving semantic unit sequence contains multiple semantically independent receiving semantic units.

[0022] Extract multiple speech fragments generated by the receiver from the remote session stream, and obtain the text content data of each speech fragment. This text content data can be obtained by transcribing the voice input of the receiver through a speech recognition engine, or directly extracted from the text message input by the receiver.

[0023] The text content data is segmented according to the same semantic integrity principle as in step 1101. First, sentence boundaries and punctuation are used as the initial segmentation criteria. Then, the segmentation boundaries are corrected based on the results of dependency parsing to ensure that each segmented receiver semantic unit has the ability to independently express a complete semantic proposition. All segmented receiver semantic units are arranged according to their order of appearance in the receiver discourse segment units. Combined with the timestamp of each receiver discourse segment unit in the remote conversation stream, corresponding time position information is assigned to each receiver semantic unit, thus forming a sequence of receiver semantic units.

[0024] Step 1103: Calculate the semantic association strength parameter between each semantic unit in the sending semantic unit sequence and each semantic unit in the receiving semantic unit sequence.

[0025] Each sending-end semantic unit in the sending-end semantic unit sequence is paired with each receiving-end semantic unit in the receiving-end semantic unit sequence to obtain multiple sending-end-receiving semantic unit pairs. For each semantic unit pair, a semantic association strength parameter is calculated. For any semantic unit pair consisting of any sending-end semantic unit and any receiving-end semantic unit, the sending-end semantic unit is first subjected to word extraction processing, removing stop words to obtain the sending-end core vocabulary set. Simultaneously, the receiving-end semantic unit is subjected to word extraction processing, removing stop words to obtain the receiving-end core vocabulary set. The number of words in the intersection between the sending-end and receiving-end core vocabulary sets is counted. This number of words in the intersection is divided by the geometric mean of the number of words in the sending-end and receiving-end core vocabulary sets to obtain the word co-occurrence index.

[0026] Simultaneously, the sending semantic unit is input into a pre-trained semantic embedding model. This model employs a dual-tower structure based on a Transformer encoder. The encoder layer contains multiple self-attention layers and feedforward network layers, which stack the input text to map it into a fixed-dimensional semantic embedding vector, resulting in the sending semantic embedding vector. Similarly, the receiving semantic unit is input into the same semantic embedding model to obtain the receiving semantic embedding vector. The cosine similarity between the sending and receiving semantic embedding vectors is calculated as the semantic distance metric. The word co-occurrence metric and the semantic distance metric are multiplied by the first and second weighting coefficients, respectively, and then summed to obtain the semantic association strength parameter between the sending and receiving semantic units.

[0027] Step 1104: Create semantic connection pairs between the sending semantic unit and the receiving semantic unit based on the semantic association strength parameter. Each semantic connection pair contains a sending semantic unit, a receiving semantic unit, and the semantic association strength parameter between them.

[0028] For each transmitting semantic unit, iterate through all receiving semantic units in the receiving semantic unit sequence, obtain the semantic association strength parameter between the transmitting semantic unit and each receiving semantic unit, and select the receiving semantic unit with the largest semantic association strength parameter from all receiving semantic units as the target receiving semantic unit. Combine the transmitting semantic unit, the target receiving semantic unit, and the semantic association strength parameter between them into a semantic connection pair.

[0029] The above pairing operation is performed on all sending semantic units to obtain multiple semantic connection pairs equal to the number of sending semantic units. Each semantic connection pair represents a direct semantic connection between the intent content of a sending semantic unit and the reverse intent content of a receiving semantic unit. The semantic association strength parameter is used to quantify the tightness of this connection.

[0030] Step 1105: Arrange the semantic connection pairs according to the timestamp order of the sending end speech fragment unit and the receiving end speech fragment unit to obtain the semantic connection pair sequence.

[0031] For each semantic connection pair formed in step 1104, the time position information corresponding to the sending semantic unit and the receiving semantic unit in the semantic connection pair are extracted. The order of their timestamps is compared, and the later timestamp is taken as the attribution time of the semantic connection pair in the remote session stream. All semantic connection pairs are sorted according to the order of their attribution times, with the semantic connection pairs with earlier attribution times arranged first and the semantic connection pairs with later attribution times arranged last, forming a semantic connection pair sequence. Each position in the semantic connection pair sequence corresponds to a semantic connection event in the remote session stream, and the overall arrangement of the sequence reflects the temporal evolution of the intended interaction between the sending and receiving ends in the remote session stream.

[0032] Step 1106: Perform semantic trajectory continuity analysis on adjacent semantic connection pairs in the semantic connection pair sequence to generate semantic transition direction vectors between adjacent semantic connection pairs.

[0033] Step 11061: Extract the sending semantic unit and receiving semantic unit of the current semantic connection pair in the semantic connection pair sequence, as well as the sending semantic unit and receiving semantic unit of the next semantic connection pair.

[0034] Specifically, any two adjacent semantic connection pairs are selected from the semantic connection pair sequence. The semantic connection pair that appears earlier in the sequence is designated as the current semantic connection pair, and the semantic connection pair that appears later in the sequence is designated as the next semantic connection pair. The sending-end semantic units and receiving-end semantic units contained in the current semantic connection pair are extracted, and the sending-end semantic units and receiving-end semantic units contained in the next semantic connection pair are extracted, providing semantic material for subsequent continuity analysis.

[0035] Step 11062: Calculate the semantic continuity tightness between the receiving semantic unit of the current semantic connection pair and the sending semantic unit of the next semantic connection pair, and determine the semantic trajectory breakage index between the current semantic connection pair and the next semantic connection pair based on the semantic continuity tightness.

[0036] For example, the semantic unit at the receiving end of the current semantic connection pair is taken as the starting point, and the semantic unit at the sending end of the next semantic connection pair is taken as the ending point. The semantic connection tightness between the starting point and the ending point is calculated. The semantic connection tightness is calculated as follows: First, the starting point and the ending point are processed by word extraction to obtain their respective core word sets, and the number of words in the intersection of the two core word sets is counted; then, the starting point and the ending point are input into the pre-trained semantic embedding model to obtain their respective semantic embedding vectors, and the cosine similarity between the two semantic embedding vectors is calculated; the number of words in the intersection and the cosine similarity are weighted and summed according to the preset fusion weight to obtain the semantic connection tightness.

[0037] The semantic connection tightness is compared with a preset connection threshold. If the semantic connection tightness is greater than or equal to the connection threshold, the semantic trajectory breakage index is set to the first value, indicating that the semantic connection between the two semantic connection pairs is relatively smooth. If the semantic connection tightness is less than the connection threshold, the semantic trajectory breakage index is set to the second value. If the second value is greater than the first value, it indicates that there is a significant semantic trajectory break between the two semantic connection pairs.

[0038] Step 11063: Extract the discourse intent category features of the receiving semantic unit in the current semantic connection pair and the discourse intent category features of the sending semantic unit in the next semantic connection pair. Based on the discourse intent category features of the receiving semantic unit in the current semantic connection pair and the discourse intent category features of the sending semantic unit in the next semantic connection pair, calculate the intent transition direction angle parameter.

[0039] For example, the receiving semantic unit in the current semantic connection pair is input into a pre-trained intent classifier. The intent classifier adopts a classification structure based on a multilayer perceptron, which includes an input layer, multiple hidden layers and an output layer. The output layer uses a softmax activation function to output the probability distribution of the receiving semantic unit on each preset intent category. The intent category vector corresponding to the intent category with the highest probability is used as the discourse intent category feature of the receiving semantic unit in the current semantic connection pair.

[0040] Similarly, the sending semantic unit in the next semantic connection pair is input into the same intent classifier to obtain the discourse intent category features of the sending semantic unit in the next semantic connection pair. In the preset intent category relationship graph, each intent category corresponds to a node in the graph, the edges between nodes represent the transition relationship between intent categories, and the direction attribute on the edge represents the direction of transition. Based on the discourse intent category features of the receiving semantic unit in the current semantic connection pair, the first node in the intent category relationship graph is located, and based on the discourse intent category features of the sending semantic unit in the next semantic connection pair, the second node in the intent category relationship graph is located. The direction angle of the directed edge from the first node to the second node is calculated as the intent transition direction angle parameter.

[0041] Step 11064: Input the semantic trajectory breakage degree index and the intention jump direction angle parameter into the semantic jump direction generation function to generate a semantic jump direction vector between adjacent semantic connection pairs. The semantic jump direction vector contains both the continuity information of the semantic trajectory and the direction information of the intention jump.

[0042] For example, the semantic trajectory breakage index obtained in step 11062 and the intention transition direction angle parameter obtained in step 11063 are used as input variables and processed by the semantic transition direction generation function. The semantic transition direction generation function first maps the semantic trajectory breakage index into the component values ​​of a vector in the first dimension, and then decomposes the intention transition direction angle parameter into the component values ​​of the vector in the other dimensions.

[0043] Specifically, when the intentional transition direction angle parameter is expressed in polar coordinates, it includes the angle magnitude and the direction deflection. The angle magnitude is mapped to the component value of the vector in the second dimension, and the direction deflection is decomposed into the component values ​​of the vector in the third and fourth dimensions. The combination of these component values ​​forms a complete semantic transition direction vector. Steps 11061 to 11064 are performed on each pair of adjacent semantic connection pairs in the semantic connection pair sequence to generate the corresponding semantic transition direction vector, thereby forming a semantic transition direction vector sequence.

[0044] Step 1107: Create a session advancement semantic chain for the remote session stream based on the semantic transition direction vector and the semantic connection pair sequence. The session advancement semantic chain has the sending semantic unit and the receiving semantic unit as nodes and the semantic transition direction vector as edges.

[0045] Step 11071: Map each semantic connection pair in the semantic connection pair sequence to a semantic node in the session advancement semantic chain. The semantic node contains the sending semantic unit, the receiving semantic unit, and the semantic association strength parameter between the two of the semantic connection pair.

[0046] Iterate through each semantic join pair in the semantic join pair sequence, creating a corresponding semantic node for each semantic join pair in the session advancement semantic chain. Each semantic node's data structure contains three fields: the first field stores the sending semantic unit in the semantic join pair, the second field stores the receiving semantic unit in the semantic join pair, and the third field stores the semantic association strength parameter in the semantic join pair. All created semantic nodes are organized according to the order of the semantic join pairs in the semantic join pair sequence to form a semantic node sequence.

[0047] Step 11072: Determine the order of semantic nodes in the session advancement semantic chain based on the timestamp order of the semantic connection pair sequence, and map the semantic transition direction vector between adjacent semantic connection pairs to the directed edges between adjacent semantic nodes in the session advancement semantic chain.

[0048] It can be understood that the order of semantic nodes in the conversation advancement semantic chain is consistent with the order of semantic connector pairs in the semantic connector pair sequence. That is, the semantic node corresponding to the i-th semantic connector pair in the semantic connector pair sequence is arranged at the i-th position in the conversation advancement semantic chain. For any two adjacent semantic nodes in the semantic node sequence, i.e., the i-th semantic node and the (i+1)-th semantic node, the corresponding i-th semantic transition direction vector is extracted from the semantic transition direction vector sequence, and this semantic transition direction vector is mapped to a directed edge from the i-th semantic node to the (i+1)-th semantic node.

[0049] Step 11073: Extract the intention transition direction angle parameter contained in each semantic transition direction vector, assign the intention transition direction angle parameter as the direction attribute of the directed edge, extract the semantic trajectory breakage index contained in each semantic transition direction vector, and assign the semantic trajectory breakage index as the weight attribute of the directed edge.

[0050] For each semantic transition direction vector, the intention transition direction angle parameter is reconstructed from the component values ​​of the second, third, and fourth dimensions of the vector, and the intention transition direction angle parameter is assigned to the direction attribute of the corresponding directed edge.

[0051] Simultaneously, a semantic trajectory breakage index is extracted from the first dimension component value of the vector, and this semantic trajectory breakage index is assigned to the weight attribute of the corresponding directed edge. After the above assignment process, each directed edge in the conversation advancement semantic chain carries both direction information and weight information. The direction information represents the direction of intent transfer, and the weight information represents the smoothness of the transfer.

[0052] Step 11074: Perform statistical processing on the local density distribution of semantic nodes in the session advancement semantic chain to generate the boundary of the clustering region of semantic nodes in the session advancement semantic chain.

[0053] Each semantic node in the conversation advancement semantic chain is mapped to a two-dimensional semantic coordinate space, where the horizontal axis represents the position index of the semantic node in the semantic node sequence, and the vertical axis represents the value of the semantic association strength parameter of that semantic node. In this two-dimensional semantic coordinate space, a kernel density estimation algorithm is used to estimate the distribution density of the semantic nodes. The algorithm applies a Gaussian kernel function centered on each semantic node, and the contributions of each Gaussian kernel function are superimposed to obtain the density distribution surface over the entire coordinate space.

[0054] On the density distribution surface, a density threshold is set, and continuous regions with density values ​​greater than the threshold are marked as high-density clustered regions, while continuous regions with density values ​​less than or equal to the threshold are marked as low-density sparse regions. The boundary coordinates of each high-density clustered region are extracted as the clustered region boundary of the semantic node.

[0055] Step 11075: Divide the session advance semantic chain into multiple semantic intent blocks according to the aggregation region boundary of the semantic nodes. Each semantic intent block corresponds to an intent topic unit in the remote session stream.

[0056] Based on the boundary coordinates of the clustered region obtained in step 11074, the semantic nodes are grouped in the semantic node sequence: all semantic nodes located within the boundary of the same high-density clustered region belong to the same semantic intent block, and semantic nodes located in the low-density sparse region are assigned to the semantic intent block corresponding to the nearest high-density clustered region according to their distance from the adjacent high-density clustered region.

[0057] After grouping, all semantic nodes in the session advancement semantic chain are divided into multiple semantic intent blocks. The semantic nodes contained in each semantic intent block have a high degree of consistency in semantic content, corresponding to a relatively independent intent topic unit in the remote session stream.

[0058] Step 11076: Generate an intent topic description tag for each semantic intent block and attach the intent topic description tag to the corresponding set of semantic nodes in the session advancement semantic chain.

[0059] For each semantic intent block, the sending and receiving semantic units contained in all semantic nodes within the block are collected. All collected semantic units are concatenated and input into a pre-trained topic labeling model. This model employs a sequence-to-sequence structure based on an attention mechanism. The encoder encodes the input semantic unit sequence to obtain a context vector, and the decoder generates a sequence of topic descriptive words progressively based on the context vector. This generated sequence of topic descriptive words serves as the intent topic description label for the semantic intent block. The generated intent topic description label is then appended to the additional attribute fields of all semantic nodes within the semantic intent block, thus establishing the association between intent topic information and semantic nodes.

[0060] Step 120: When a transmission interruption event occurs in the remote session stream, perform autoregressive deduction processing on the speech fragment unit at the sending end based on the semantic trajectory, generate a transitional session fragment, and embed it into the timing position in the receiving end presentation link corresponding to the transmission interruption event.

[0061] In this step, the remote session stream may experience transmission interruptions due to network fluctuations, device processing delays, or brief periods of silence at the sending end. These interruptions are characterized by arrival times of speech fragments at the sending end significantly exceeding the average interval under normal session rhythm. When a transmission interruption is detected, the semantic trajectory information in the established session progression semantic chain is used to infer the speech content that the sending end might generate during the interruption, thereby generating transitional session fragments to fill the session gaps. Step 1201: Monitor the arrival time interval of the speech segment units sent by the remote session stream. When the arrival time interval exceeds the preset interruption judgment threshold, determine that a transmission interruption event has occurred in the remote session stream and record the start time of the transmission interruption event.

[0062] The system continuously monitors the received timestamp of each newly arriving sender speech segment unit in the remote session stream. It calculates the difference between the received timestamp of the currently arriving sender speech segment unit and the received timestamp of the previously arriving sender speech segment unit to obtain the arrival time interval. This arrival time interval is compared with a preset interruption threshold, which is set based on the statistical distribution of historical arrival time intervals in the remote session stream under normal transmission conditions. Specifically, it is the average of historical arrival time intervals plus three times the standard deviation. If the arrival time interval exceeds the interruption threshold, a transmission interruption event is determined to have occurred in the remote session stream, and the received timestamp of the currently arriving sender speech segment unit is recorded as the start time of the transmission interruption event.

[0063] Step 1202: Extract the latest generated semantic transition direction vector in the session advancement semantic chain before the transmission interruption event occurs, and determine the extension direction of the current semantic trajectory based on the latest generated semantic transition direction vector.

[0064] When a transmission interruption event is determined to have occurred, the last generated semantic transition direction vector in the session advancement semantic chain is obtained. This semantic transition direction vector corresponds to the intent transfer relationship between the last adjacent semantic connection pair before the transmission interruption event occurred. The directional attribute component is extracted from this semantic transition direction vector, that is, the intent transition direction angle parameter is reconstructed from the second, third, and fourth dimension component values ​​of the vector. This intent transition direction angle parameter is used as the extension direction of the current semantic trajectory.

[0065] Step 1203: Extract the intent topic description label corresponding to the extension direction of the current semantic trajectory from the session advancement semantic chain to determine the intent topic context of the current session.

[0066] Based on the extension direction of the current semantic trajectory determined in step 1202, the intent topic description tags of all semantic intent blocks in the session advancement semantic chain are traversed. The relevance between each intent topic description tag and the extension direction of the current semantic trajectory is calculated. The relevance is calculated by mapping the intent topic description tag to a tag vector and then calculating the inner product with the extension direction vector of the current semantic trajectory. The intent topic description tag with the highest relevance is selected as the intent topic context of the current session. This intent topic context represents the core topic being discussed in the remote session stream when the transmission interruption event occurs.

[0067] Step 1204: Extract the last transmitting semantic unit and the last receiving semantic unit before the transmission interruption event occurs, and use the last transmitting semantic unit and the last receiving semantic unit as the starting semantic state for autoregressive inference.

[0068] The last semantic connection pair generated before the start of the transmission interruption event is located in the session progression semantic chain. The sending semantic unit is extracted from this semantic connection pair as the last sending semantic unit, and the receiving semantic unit is extracted as the last receiving semantic unit. The last sending semantic unit and the last receiving semantic unit are used together as the starting semantic state for autoregressive inference. This starting semantic state contains the latest semantic content of the session before the transmission interruption event occurred.

[0069] Step 1205: Starting from the initial semantic state, and combining the extension direction of the current semantic trajectory and the intent topic context of the current session, perform autoregressive inference processing to generate multiple candidate transitional semantic units.

[0070] Step 12051: Use the last receiving semantic unit as the initial input state for autoregressive inference, and use the last sending semantic unit as the context information of the initial input state.

[0071] At the initial moment of the autoregressive inference process, the last receiving semantic unit extracted in step 1204 is set as the initial input state, which serves as the input to the autoregressive inference model at the current time step. The last sending semantic unit extracted in step 1204 is set as the context information, which serves as the auxiliary reference information for the autoregressive inference model at the current time step.

[0072] Step 12052: Use the pre-built conversational semantic inference model to perform semantic state encoding on the initial input state and the preceding context information to generate an initial semantic state vector.

[0073] The conversational semantic inference model employs an autoregressive language model architecture based on a Transformer decoder. The main body of the model comprises a decoder module consisting of multiple masked self-attention layers, cross-attention layers, and feedforward network layers stacked together. The initial input state and preceding context information are encoded using word embeddings and positional encoding, respectively, and then concatenated to form the encoder input sequence. This sequence is input to the encoder part of the conversational semantic inference model for semantic state encoding. The encoder part utilizes multi-layer bidirectional self-attention computation to ensure that the representation at each position incorporates the contextual information of the entire input sequence. After multi-layer encoding processing, a fixed-dimensional initial semantic state vector is generated.

[0074] Step 12053: The extension direction of the current semantic trajectory is transformed into a direction constraint vector. The direction constraint vector is used to limit the generation direction of semantic units in the autoregressive deduction process.

[0075] The extension direction of the current semantic trajectory determined in step 1202, i.e., the intention transition direction angle parameter, is mapped to the same vector space as the semantic embedding vector. A linear transformation matrix is ​​then used to convert the intention transition direction angle parameter into a direction constraint vector with the same dimension as the semantic state vector. This direction constraint vector serves as a constraint condition during the autoregressive deduction process, guiding the generated semantic units to shift in a specific direction within the semantic space.

[0076] Step 12054: Transform the intent topic context of the current session into an intent constraint vector. The intent constraint vector is used to limit the range of generated content of semantic units during the autoregressive inference process.

[0077] The intent topic description labels for the current session, determined in step 1203, are input into a pre-trained topic embedding model. This model employs a dual-tower encoder with the same structure as the semantic embedding model, mapping the intent topic description labels to topic embedding vectors. The topic embedding vectors are then projected onto a dimensional space with the same dimensions as the semantic state vectors through a linear transformation, yielding intent constraint vectors. These intent constraint vectors serve as content constraints during the autoregressive inference process, ensuring that the generated semantic units maintain semantic content consistency with the current intent topic.

[0078] Step 12055: The initial semantic state vector, direction constraint vector, and intention constraint vector are fused to generate the inference constraint vector. The inference constraint vector is then input into the semantic unit generation module of the session semantic inference model to generate candidate transition semantic units for the current time step.

[0079] For example, the initial semantic state vector, direction constraint vector, and intent constraint vector are weighted and summed according to preset fusion weights to obtain the inference constraint vector. The inference constraint vector serves as the initial hidden state input to the decoder of the conversation semantic inference model. Based on this initial hidden state and the context representation output by the encoder, the decoder generates output words one by one through an autoregressive approach.

[0080] During the generation process, the probability distribution of the words output by the decoder at each step is calculated using the softmax function. The word with the highest probability from the probability distribution is selected as the output word for the current time step. Generation of the current time step stops when a complete semantic unit is generated and marked as the end. All generated words are then combined into candidate transition semantic units for the current time step.

[0081] Step 12056: Feed back the candidate transition semantic units of the current time step to the initial input state, update the initial input state to the candidate transition semantic units of the current time step, and iteratively execute the steps of semantic state encoding processing, deduction of constraint vector generation and candidate transition semantic unit generation until a candidate transition semantic unit sequence that meets the preset number requirement is generated.

[0082] The candidate transition semantic units generated in step 12055 are used as the input state for the next time step, replacing the original initial input state. Maintaining the preceding context information, steps 12052 to 12055 are repeated to generate candidate transition semantic units for the next time step. One candidate transition semantic unit is generated in each iteration, and these units are arranged sequentially to form a candidate transition semantic unit sequence. Iteration stops when the number of candidate transition semantic units in the sequence reaches a preset threshold. This preset threshold is determined based on the duration of the transmission interruption event; a longer duration results in a larger threshold.

[0083] Step 1206: Calculate the semantic trajectory consistency parameter between each candidate transition semantic unit and the receiving semantic unit in the initial semantic state, and select the target transition semantic unit from multiple candidate transition semantic units based on the semantic trajectory consistency parameter.

[0084] It is understood that for each candidate transition semantic unit generated in step 1205, the candidate transition semantic unit is paired with the receiving semantic unit in the initial semantic state to form a semantic unit pair. The semantic association strength of the semantic unit pair is calculated according to the method for calculating the semantic association strength parameter in step 1103, and the semantic association strength is used as the semantic trajectory consistency parameter corresponding to the candidate transition semantic unit. The semantic trajectory consistency parameters of all candidate transition semantic units are compared, and the candidate transition semantic unit with the largest semantic trajectory consistency parameter is selected as the target transition semantic unit. This target transition semantic unit has the highest continuity consistency with the session content before the transmission interruption event in terms of semantic content.

[0085] Step 1207: The target transitional semantic unit is serialized and organized according to the extension direction of the current semantic trajectory to generate a transitional session fragment. The semantic content of the transitional session fragment maintains semantic trajectory continuity with the session advancement semantic chain before the transmission interruption event.

[0086] Furthermore, the target transitional semantic units selected in step 1206 are arranged according to their generation time order, and the arrangement order is fine-tuned according to the extension direction of the current semantic trajectory to ensure that the semantic flow between the target transitional semantic units is consistent with the extension direction of the current semantic trajectory. All the fine-tuned target transitional semantic units are combined into a complete session fragment to obtain a transitional session fragment. The semantic content of this transitional session fragment continues the semantic trajectory represented by the session progression semantic chain before the transmission interruption event in terms of intent direction.

[0087] Step 1208: Determine the timing position of the transmission interruption event in the receiving end presentation link and embed it into the timing position in the receiving end presentation link.

[0088] Based on the start time of the transmission interruption event recorded in step 1201, the corresponding position of this start time on the timeline of the receiving end presentation link is calculated. The timeline of the receiving end presentation link takes the moment when the receiving device begins to present the remote session stream as zero and progresses at normal playback speed. The transitional session fragment generated in step 1207 is inserted into the position corresponding to this start time in the receiving end presentation link, so that when the receiving end presents the remote session stream, the transitional session fragment can appear at the time position where the transmission interruption event occurred, filling the session gap caused by the delayed arrival of the sending end speech fragment unit.

[0089] Step 130: Determine the potential intent direction trend of the receiver conversation in the remote conversation stream based on the prosodic fluctuation features and semantic transition vectors of the receiver's speech fragment units.

[0090] In this step, in order to achieve more accurate conversation guidance at the receiving end, it is necessary to conduct in-depth intent mining of the recipient's discourse. By comprehensively analyzing the prosodic features and semantic transition features in the recipient's discourse segments, we can identify the potential but not yet explicitly stated intent of the recipient during the conversation, thus providing a basis for generating visual guidance anchors in the future.

[0091] Step 1301: Perform prosodic feature extraction processing on the speech segment unit at the receiving end to generate the speech prosodic fluctuation features of the speech segment unit at the receiving end. The speech prosodic fluctuation features include the tone rise and fall sequence and the speech rate change sequence.

[0092] Audio signal data of speech segments from the receiving end is acquired from the remote session stream. This audio signal data is recorded by the audio acquisition module of the OPS microcomputer during the session at the receiving end. The audio signal data is then processed by frame segmentation, dividing the audio signal into multiple overlapping audio frames with a fixed frame shift step size. The duration of each audio frame is set to a first duration value. Fundamental frequency features and energy features are extracted from each audio frame. The fundamental frequency features are calculated using the autocorrelation method, and the energy features are obtained by calculating the sum of squares of the amplitudes of the sampling points within the audio frame. The fundamental frequency features of each audio frame are arranged in chronological order to form a fundamental frequency time series. Difference calculation is performed on this fundamental frequency time series to obtain a pitch rise and fall sequence. The value at each position in the pitch rise and fall sequence represents the difference in fundamental frequency between two adjacent audio frames.

[0093] Simultaneously, the energy features of each audio frame are arranged in chronological order to form an energy time series. The ratio of energy features between adjacent audio frames is calculated as a local speech rate estimate. The local speech rate estimates are then arranged in chronological order to obtain a speech rate variation sequence.

[0094] Step 1302: Perform semantic transition analysis on the received speech fragment unit to generate a semantic transition vector for the received speech fragment unit. The semantic transition vector includes the direction angle parameter and the magnitude parameter of the semantic transition.

[0095] The text content data of the received speech fragment unit is segmented into multiple received semantic units according to the method in step 1102. Each received semantic unit is sequentially input into a pre-trained semantic embedding model to obtain a semantic embedding vector corresponding to each received semantic unit. The semantic embedding vectors corresponding to adjacent received semantic units are differentially calculated to obtain multiple difference vectors. Each difference vector represents the displacement between two adjacent semantic units in the semantic space. The direction angle of each difference vector is calculated as the direction angle parameter of the semantic transition, and the magnitude of each difference vector is calculated as the magnitude parameter of the semantic transition. All difference vectors are arranged in the temporal order of the corresponding semantic units to form a semantic transition vector sequence. Each semantic transition vector contains both the direction angle parameter and the magnitude parameter.

[0096] Step 1303: Calculate the distribution information of intonation inflection points of the received speech segment units based on the intonation rise and fall sequence, and determine the position of the emotional tendency transition of the received speech based on the intonation inflection point distribution information.

[0097] The second-order difference calculation is performed on the intonation rise and fall sequence to obtain the intonation acceleration sequence. The zero-crossing points where the intonation acceleration sequence changes from positive to negative are identified as the inflection points from intonation rise to fall, and the zero-crossing points where the intonation acceleration sequence changes from negative to positive are identified as the inflection points from intonation fall to rise. All identified inflection point positions are marked on the time axis of the received speech segment unit, forming intonation inflection point distribution information. The time corresponding to each intonation inflection point is the position of the change in emotional tendency of the received speech.

[0098] Step 1304: Calculate the distribution information of speech rate change points of the receiving end speech segment unit based on the speech rate change sequence, and determine the emphasis position of the receiving end speech based on the speech rate change point distribution information.

[0099] Abrupt change detection is performed on the speech rate variation sequence. A sliding window method is used to calculate the standard deviation of the speech rate variation value within each local window. When the standard deviation at a certain position exceeds a preset abrupt change threshold, that position is marked as a speech rate abrupt change point. All identified speech rate abrupt change points are marked on the time axis of the speech segment unit at the receiving end, forming speech rate abrupt change point distribution information. The time corresponding to each speech rate abrupt change point is the emphasis position of the speech at the receiving end.

[0100] Step 1305: The position of the emotional tendency transition and the position of the emphasis are associated and matched with the semantic transition direction angle parameter in the semantic transition vector to generate the intention pointing marker of the receiving end discourse.

[0101] It is understood that for each sentiment shift position determined in step 1303 and each emphasis position determined in step 1304, the semantic shift vector that is closest to that position in time is found on the timeline, and the direction angle parameter in the semantic shift vector is associated and bound to the currently processed sentiment shift position or emphasis position. The sentiment shift positions and emphasis positions bound to the direction angle parameter are collectively referred to as intention pointing markers, and each intention pointing marker contains its position information on the timeline and the associated direction angle parameter.

[0102] Step 1306: Quantify the intent pointing intensity of the intent pointing marker based on the semantic transition amplitude parameter in the semantic transition vector to generate the intent pointing confidence parameter of the intent pointing marker.

[0103] For each intent-pointing marker, obtain the magnitude parameter of the semantic transition vector associated with it, and map the magnitude parameter to a continuous value range between zero and one using the sigmoid function. The value obtained after mapping is used as the intent-pointing confidence parameter of the intent-pointing marker. The larger the intent-pointing confidence parameter, the higher the clarity of the intent expressed by the receiver at the marker.

[0104] Step 1307: Based on the intent pointing marker and the intent pointing confidence parameter of the intent pointing marker, determine the potential intent pointing trend of the receiving end session. The potential intent pointing trend includes the directional distribution characteristics and intensity distribution characteristics of intent pointing.

[0105] All intention-pointing markers are grouped according to their direction angle parameters. Intention-pointing markers whose direction angle parameters fall within the same angle interval are grouped together. The number of intention-pointing markers in each angle interval and the sum of the intention-pointing confidence parameters of each marker are counted to obtain the intention-pointing intensity in that direction. The intention-pointing intensity corresponding to each angle interval is normalized to obtain the directional distribution characteristics of intention-pointing. At the same time, the intention-pointing confidence parameters of each intention-pointing marker are arranged in chronological order to form the intensity distribution characteristics of intention-pointing.

[0106] Step 140: Establish visual semantic guidance anchors based on potential intent pointing trends, and push the visual semantic guidance anchors to the receiving end's conversation interface. The visual semantic guidance anchors are used to guide the response direction of the receiving end's conversation.

[0107] After determining the potential intent direction trend of the receiving end session, this trend information needs to be transformed into a guidance form that the receiving end user can intuitively perceive. This is achieved by establishing visual semantic guidance anchors in the receiving end session interface, providing clear directional guidance for the receiving end user. A visual semantic guidance anchor is a graphical interface element that integrates directional indication and semantic prompts. Its presentation and content are adaptively configured according to the characteristics of the potential intent direction trend. Step 140 further includes: Step 1401: Analyze the directional distribution features of intent pointing contained in the potential intent pointing trend, and extract the intent pointing direction with the highest intent pointing confidence parameter in the directional distribution features as the dominant intent direction.

[0108] The directional distribution features of intent obtained in step 1307 are traversed, and the intent confidence parameter corresponding to each direction in the directional distribution features is found. The direction with the largest intent confidence parameter is selected as the dominant intent direction. This direction represents the direction corresponding to the intent content that the receiver is most likely to express but has not yet explicitly stated in the current session.

[0109] Step 1402: Based on the dominant intent direction, retrieve candidate intent semantic templates that match the dominant intent direction from the preset intent semantic template library. The candidate intent semantic templates include guiding semantic content fragments and guiding semantic presentation styles.

[0110] The pre-defined intent semantic template library contains multiple intent semantic templates, each pre-associated with a direction angle range. The dominant intent direction is matched against the direction angle range associated with each intent semantic template, and intent semantic templates whose dominant intent direction falls within their respective direction angle ranges are selected as candidate intent semantic templates. Each candidate intent semantic template contains a guiding semantic content fragment, which is a pre-defined prompt text used to guide the receiver to respond along the dominant intent direction. It also includes a guiding semantic presentation style, which defines the display format of the guiding semantic content fragment in the conversation interface.

[0111] Step 1403: Extract the intensity distribution features of intent pointing contained in the potential intent pointing trend, and determine the presentation intensity level parameters of the candidate intent semantic template based on the intensity distribution features.

[0112] Extract the intention direction confidence parameter corresponding to the dominant intention direction from the intensity distribution features of the intention direction obtained in step 1307, and map the intention direction confidence parameter to one of multiple discrete presentation intensity levels. The mapping method is to divide the numerical range from zero to one into multiple sub-intervals, each sub-interval corresponding to a presentation intensity level. After determining the sub-interval in which the intention direction confidence parameter falls, take the presentation intensity level corresponding to the sub-interval as the presentation intensity level parameter of the candidate intention semantic template.

[0113] Step 1404: Input the dominant intent direction, candidate intent semantic template and presentation intensity level parameters into the anchor point generation function to generate a visual semantic guidance anchor point. The visual semantic guidance anchor point includes guiding semantic text content and visual presentation attribute parameters.

[0114] The anchor point generation function first generates guiding semantic text content based on guiding semantic content fragments in the candidate intent semantic template. Then, it determines the direction indication attribute in the visualization presentation attribute parameters based on the dominant intent direction, and determines the prominence attribute in the visualization presentation attribute parameters based on the presentation intensity level parameter. Finally, it combines the guiding semantic text content and the visualization presentation attribute parameters to generate a complete visual semantic guiding anchor point.

[0115] Step 1405: Render the visual semantic guidance anchor point to the preset anchor point display area in the conversation interface of the receiving end according to the visual presentation attribute parameters.

[0116] The receiving end's session interface pre-defined an anchor point display area, located on the side or bottom of the session interface. The position offset of the visual semantic guide anchor point within the anchor point display area is determined based on the direction indication attribute defined in the visualization presentation attribute parameters. The display brightness and size of the visual semantic guide anchor point are determined based on the prominence attribute. The visual semantic guide anchor point is then rendered into the anchor point display area according to the determined position offset, display brightness, and size.

[0117] Step 150: Collect interactive behavior data of the receiving end for the visual semantic guidance anchor point, and determine the conversation rhythm generation adjustment strategy based on the interactive behavior data. The conversation rhythm generation adjustment strategy is used to adjust the generation frequency of transitional conversation segments and the timing of pushing visual semantic guidance anchor points.

[0118] After the visual semantic guidance anchor is pushed to the receiving end's conversation interface, the receiving end user will exhibit various interactive behaviors such as gazing, clicking, and ignoring it. By collecting the above interactive behavior data, the receiving end user's acceptance of the guidance content and response pattern can be quantitatively evaluated, thereby dynamically adjusting the rhythm strategy of subsequent conversation guidance to match the guidance method with the receiving end user's interaction habits. In specific implementation, the collected interactive behavior data includes anchor gaze duration, anchor click trigger time, and anchor ignoring count. Based on the above data, response latency parameters and adoption rate parameters are calculated. Combined with the degree of matching between the actual input response content after the receiving end triggers the anchor and the anchor's intent, a conversation rhythm profile of the receiving end is constructed. Based on this, the direction of adjusting the generation frequency of transitional conversation segments and the direction of adjusting the timing of anchor push are determined, and finally, a conversation rhythm generation adjustment strategy is generated to guide the subsequent generation of transitional conversation segments and the pushing operation of visual semantic guidance anchors.

[0119] In one embodiment, step 150 includes: Step 1501: Monitor the presentation status of the visual semantic guidance anchors in the receiver's session interface, and collect the interaction behavior data of the receiver to the visual semantic guidance anchors. The interaction behavior data includes anchor gaze duration parameters, anchor click trigger time parameters, and anchor ignore count parameters.

[0120] After the visual semantic guidance anchor is pushed to the receiving end's session interface, the front-end monitoring script on the receiving end device continuously records the user's interaction with the anchor. The monitoring script uses the receiving end device's camera to collect the position of the user's gaze on the screen. Timing starts when the position falls within the display area of ​​the visual semantic guidance anchor and stops when the position moves out of the display area, thus obtaining the anchor gaze duration parameter.

[0121] When a user clicks or touches a visual semantic guidance anchor point, the time difference between the click trigger time and the anchor point push time is recorded as the anchor point click trigger time parameter. If the visual semantic guidance anchor point remains displayed in the display area for a preset expiration time without any interaction from the user, the anchor point is marked as ignored, and the number of ignored anchor points within a preset statistical period is counted as the anchor point ignore count parameter.

[0122] Step 1502: Calculate the response delay parameter of the receiver to the visual semantic guidance anchor point based on the anchor point gaze duration parameter and the anchor point click trigger time parameter.

[0123] The anchor point gaze duration parameter and the anchor point click trigger time parameter are weighted and summed according to a preset weight coefficient. Specifically, the anchor point click trigger time parameter is multiplied by the first weight value to obtain the first product, the anchor point gaze duration parameter is multiplied by the second weight value to obtain the second product, and the first product and the second product are added together to obtain the response delay parameter.

[0124] Step 1503: Calculate the acceptance rate parameter of the visual semantic guidance anchor at the receiving end based on the anchor ignore count parameter and the total number of times the visual semantic guidance anchor is pushed.

[0125] Within a preset statistical period, the total number of pushes for the visual semantic guidance anchor is counted. The ignore rate is obtained by dividing the anchor ignore count parameter by the total number of pushes. The adoption rate parameter is obtained by subtracting the ignore rate from the value.

[0126] Step 1504: Extract the response utterance fragment unit input by the receiving end after triggering the visual semantic guidance anchor point, and calculate the intent matching degree parameter between the response utterance fragment unit and the potential intent pointing trend corresponding to the visual semantic guidance anchor point.

[0127] When the receiving user triggers the visual semantic guidance anchor point, the text content of the reply utterance fragment unit subsequently input by the receiving end is collected. The semantic association strength between the reply utterance fragment unit and the potential intention pointing trend corresponding to the visual semantic guidance anchor point is calculated according to the semantic association strength parameter calculation method in step 1103. The semantic association strength is used as the intention matching degree parameter.

[0128] Step 1505: Construct a receiver session rhythm profile based on response latency parameters, acceptance rate parameters, and intent matching degree parameters. The receiver session rhythm profile includes the receiver's response speed characteristics to session guidance and the receiver's acceptance degree characteristics to session guidance.

[0129] The response latency parameter, acceptance rate parameter, and intent matching degree parameter are combined into a three-dimensional feature vector. The response latency parameter, as the first dimension component, represents the response speed of the receiver to the session guidance. The smaller the response latency parameter, the faster the response speed. The acceptance rate parameter and intent matching degree parameter are the second and third dimension components, respectively. Together, they represent the degree of acceptance of the receiver to the session guidance.

[0130] Step 1506: Based on the receiver's conversation rhythm profile, determine the direction of frequency adjustment for generating transitional conversation segments and the direction of timing adjustment for pushing visual semantic guidance anchors.

[0131] The response speed characteristics in the receiver's conversation rhythm profile are compared with a preset response speed threshold. If the response speed characteristics are less than the response speed threshold, the generation frequency adjustment direction for transitional conversation segments is determined to be decreasing, and the timing adjustment direction for pushing visual semantic guidance anchors is determined to be delaying. If the response speed characteristics are greater than or equal to the response speed threshold, the generation frequency adjustment direction for transitional conversation segments is determined to be maintaining, and the timing adjustment direction for pushing visual semantic guidance anchors is determined to be maintaining.

[0132] Simultaneously, the acceptance level feature is compared with the preset acceptance level threshold. If the acceptance level feature is less than the acceptance level threshold, the generation frequency of the transitional conversation segment is adjusted to decrease, and the timing of the push of the visual semantic guidance anchor is adjusted to delay.

[0133] Step 1507: Adjust the generation frequency parameter of the transitional session segment according to the generation frequency adjustment direction, and adjust the push timing parameter of the visual semantic guidance anchor point according to the push timing adjustment direction.

[0134] If the generation frequency adjustment direction is decreasing, the generation frequency parameter of the transitional session segment is reduced by a preset step size. If the generation frequency adjustment direction is maintaining, the generation frequency parameter of the transitional session segment remains unchanged. If the push timing adjustment direction is delaying or postponing, the delay time value in the push timing parameter of the visual semantic guidance anchor is increased by a preset step size. If the push timing adjustment direction is maintaining, the push timing parameter of the visual semantic guidance anchor remains unchanged.

[0135] Step 1508: Combine the adjusted frequency parameters of the transitional conversation segments and the timing parameters of the push of the visual semantic guidance anchor points into a conversation rhythm generation adjustment strategy.

[0136] Furthermore, the adjusted generation frequency parameters of the transitional session segments and the adjusted push timing parameters of the visual semantic guidance anchors from step 1507 are packaged into a policy data package. This policy data package contains two fields: the first field stores the adjusted generation frequency parameters, and the second field stores the adjusted push timing parameters. This policy data package is the session rhythm generation adjustment policy. The session rhythm generation adjustment policy is output to the policy execution module of the remote interactive session processing system, which controls the subsequent transitional session segment generation operations and visual semantic guidance anchor push operations according to the policy content.

[0137] In another alternative embodiment, the method further includes: Step 161: Obtain the semantic response topology of the receiving speech fragment unit corresponding to the transmission interruption event in the remote session stream. By comparing the distribution offset between the semantic response topology and the intention direction of the visualized semantic guidance anchor, generate the guidance anchor offset feedback parameter.

[0138] After determining the conversation rhythm generation and adjustment strategy based on interaction behavior data, in order to further improve the guidance accuracy of the visualized semantic guidance anchors, it is necessary to evaluate the performance of the guidance anchors in the actual conversation process. Specifically, the receiving end speech fragment units corresponding to the transmission interruption event in time are extracted from the remote conversation stream. After semantic unit segmentation of the receiving end speech fragment units, the semantic response topology of the receiving end speech fragment units is established in accordance with the method of creating the conversation advancement semantic chain in step 1107. This topology uses the receiving end semantic units as nodes and the semantic associations between semantic units as edges, representing the overall semantic response pattern of the receiving end after the transmission interruption event.

[0139] Then, the overall semantic center vector of the semantic response topology is extracted, and the intention direction vector of the visual semantic guidance anchor is extracted. The vector difference between the overall semantic center vector and the intention direction vector is calculated to obtain the distribution offset. Based on the magnitude and direction of the distribution offset, a guidance anchor offset feedback parameter is generated. When the magnitude exceeds a preset offset threshold, the guidance anchor offset feedback parameter indicates that direction correction is required.

[0140] Step 162: Based on the guidance anchor bias feedback parameter, perform direction correction processing on the intention direction of the visual semantic guidance anchor, generate the corrected visual semantic guidance anchor, and synchronize the guidance anchor bias feedback parameter to the conversation advancement semantic chain, and update the direction attribute of the semantic transition direction vector in the conversation advancement semantic chain.

[0141] If the guide anchor point offset feedback parameter generated in step 161 requires direction correction, then the offset direction angle and offset magnitude value in the guide anchor point offset feedback parameter are extracted. The intention pointing direction vector of the current visual semantic guide anchor point is rotated and scaled according to the opposite direction of the offset direction angle and the offset magnitude value to obtain the corrected intention pointing direction vector. The visual semantic guide anchor point is regenerated using the corrected intention pointing direction vector to obtain the corrected visual semantic guide anchor point.

[0142] Simultaneously, the bias direction angle and bias magnitude values ​​in the guide anchor bias feedback parameters are transmitted to the session advancement semantic chain management module. This module traverses all semantic transition direction vectors in the session advancement semantic chain, performs corresponding angle offsets on the direction attribute components of each semantic transition direction vector according to the bias direction angle, and performs corresponding scaling on the weight attribute components of each semantic transition direction vector according to the bias magnitude value, thereby realizing the online update of the session advancement semantic chain.

[0143] Step 163: Map the corrected visual semantic guidance anchor points to preset multimodal guidance presentation channels. Perform modal decomposition processing on the corrected visual semantic guidance anchor points according to the perceived stimulus intensity weights of each multimodal guidance presentation channel to generate multimodal guidance segment signals that correspond one-to-one with the multimodal guidance presentation channels.

[0144] Specifically, the preset multimodal guidance presentation channels include a visual text channel, a visual graphics channel, and an auditory cue channel. The guiding semantic text content in the corrected visual semantic guidance anchor points is mapped to the visual text channel to obtain the text channel signal; the directional indication attributes in the corrected visual semantic guidance anchor points are mapped to the visual graphics channel to generate a graphics channel signal containing directional arrows; and the guiding semantic text content in the corrected visual semantic guidance anchor points is converted into an auditory cue signal through a speech synthesis engine to obtain the auditory channel signal. The power of the signals in the above three channels is adjusted according to the perceptual stimulus intensity weights of each multimodal guidance presentation channel. The perceptual stimulus intensity weights are determined by the historical interaction preferences of the receiving user, with channels with higher preference levels receiving higher power gains.

[0145] Step 164: Perform time synchronization fusion processing on the multimodal guidance fragment signal and the transitional session fragment in the receiver presentation link to generate a multimodal collaborative guidance session stream to enhance the guidance of the response direction of the receiver session.

[0146] Specifically, the text channel signal, graphic channel signal, and auditory channel signal generated in step 163 are time-aligned according to their respective presentation delay parameters to ensure that the signals from the three channels arrive at the corresponding presentation device at the receiving end at the same time. The time-aligned multimodal guidance segment signal and the transitional session segments in the presentation link at the receiving end are interleaved on the time axis according to their respective timestamps to form a multimodal collaborative guidance session stream. When this multimodal collaborative guidance session stream is presented at the receiving end, the transitional session segments are displayed in the main session area in the form of text, the text channel signal is displayed in the anchor point display area in the form of prompt text, the graphic channel signal is displayed in the anchor point display area in the form of directional indicator graphics, and the auditory channel signal is played through the receiving end speaker in the form of voice prompts.

[0147] As another optional embodiment, the method further includes: Step 171: Perform conversation topology evolution tracking on the visual semantic guidance anchor triggered by the receiving end speech fragment unit to determine the propagation dynamic trajectory of the visual semantic guidance anchor in the conversation advancement semantic chain. The propagation dynamic trajectory represents the diffusion path and influence decay mode of the visual semantic guidance anchor's intention among subsequent semantic nodes.

[0148] When a user triggers a visual semantic guidance anchor point, the triggering event is recorded as the starting node of the propagation. During subsequent sessions, newly added semantic nodes in the session progression semantic chain are continuously monitored. The semantic distance between the new semantic node and the intent direction corresponding to the triggering anchor point is calculated. When the semantic distance is less than a preset propagation threshold, the new semantic node is marked as a successor node in the propagation chain. Using the propagation starting node as the origin, the semantic distance of each successor node as the propagation depth, and the directional deviation of each successor node relative to its predecessor node as the propagation direction, a propagation trajectory path is constructed. Simultaneously, the attenuation ratio of the semantic association strength parameter at each successor node is recorded to obtain the influence attenuation pattern.

[0149] Step 172: Calculate the potential semantic interference strength between the visualized semantic guidance anchor and other untriggered visualized semantic guidance anchors based on the propagation dynamics trajectory, and select the sequence of anchors to be suppressed from other visualized semantic guidance anchors that interfere with the dominant intention direction based on the potential semantic interference strength.

[0150] For each untriggered visual semantic guidance anchor point, obtain the intention direction vector of that anchor point, calculate the cross-correlation between the intention direction vector and the directional deflection of each direction in the propagation trajectory path of triggered anchor points, and take the maximum value of the cross-correlation as the potential semantic interference strength. Untriggered anchor points with potential semantic interference strength greater than a preset interference threshold are marked as anchor points that interfere with the dominant intention direction, and arranged in descending order of interference strength to form a sequence of anchor points to be suppressed.

[0151] Step 173: Apply semantic annihilation processing operation to each visual semantic guidance anchor in the anchor sequence to be suppressed. Generate anchor silence instruction by iteratively decaying the intention pointing intensity of the anchor to be suppressed to a preset silence threshold.

[0152] For each anchor in the suppression anchor sequence, the number of iterative decay steps is determined according to its sorting position in the sequence, with earlier-sorted anchors requiring more decay steps. In each decay step, the anchor's intention-pointing confidence parameter is multiplied by a decay factor less than one, and the anchor's display transparency is gradually increased according to the same decay ratio. After a preset number of iterative decay steps, the anchor's intention-pointing confidence parameter decreases below the silencing threshold, and the anchor's display transparency reaches a near-completely transparent level. At this point, a silencing command is generated for that anchor.

[0153] Step 174: Update the presentation layout of the visual semantic guidance anchor in the receiver's session interface based on the anchor silence instruction to suppress intent fork disturbances in the remote session stream.

[0154] The anchor point silencing instructions generated in step 173 are sent to the session interface rendering module at the receiving end. Based on the anchor point identifier and transparency parameter indicated in the silencing instructions, this module adjusts the transparency of the corresponding visual semantic guidance anchor point in the session interface to the specified value, and simultaneously switches the interactive state of the anchor point to a non-interactive state. After the above update process, all visual semantic guidance anchor points presented in the session interface are anchor points consistent with the current dominant intent direction. Potential interfering anchor points on the propagation path of triggered anchor points have been semantically annihilated, thereby effectively suppressing intent forking disturbances in the remote session stream.

[0155] In the above technical solution, for the pre-trained semantic embedding model used in steps 1103, 11061, and 1302, the model adopts a dual-tower architecture based on the Transformer encoder, specifically a pre-trained Sentence-BERT (SBERT) network structure, using "all-MiniLM-L6-v2" as the basic backbone network. This network contains 6 Transformer encoder layers, each of which contains a 12-head self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by a normalization layer and a residual connection, outputting a fixed-length semantic vector with a dimension of 384. During inference, the input text is first segmented and padded to the maximum sequence length of 128. After being fed into the model, the mean pooling vector output by the pooling layer is taken as the semantic embedding vector. No additional fine-tuning is required, and it is directly used for cosine similarity calculation.

[0156] For the pre-trained intent classifier used in step 11063, the classifier adopts a multilayer perceptron (MLP) architecture, which includes an input layer (384-dimensional, consistent with the semantic embedding vector dimension), two hidden layers (dimensions of 256 and 128 respectively, with ReLU activation function), and an output layer (output dimension of 12, corresponding to 12 preset intent categories, such as inquiry, confirmation, rebuttal, supplementation, summary, transfer, emphasis, questioning, suggestion, evaluation, request, and refusal). The output layer is followed by a softmax function to generate a probability distribution. During the training phase, the classifier uses an internal corpus containing 50,000 labeled remote conversation utterances. Each data point consists of utterance text and its corresponding intent category label. The input format is a 384-dimensional vector of text encoded by SBERT, and the output is a 12-dimensional probability distribution. During training, the Adam optimizer was used with an initial learning rate of 0.0005, a batch size of 64, and 50 training epochs. The cross-entropy loss function was used as the optimization objective, and an early stopping strategy was employed. The macro-average F1 score on the validation set was used as the early stopping metric; training stopped when the F1 score failed to improve for five consecutive epochs. In application, the input semantic unit text was first encoded into a 384-dimensional vector using SBERT before being fed into the MLP input layer. The category with the highest probability in the output probability distribution was taken as the intent category. Simultaneously, the row vectors corresponding to the category in the weight matrix of the output layer were used as the intent category feature vectors for subsequent direction and angle calculations.

[0157] For the pre-trained topic labeling model used in step 11076, the model adopts a sequence-to-sequence (Seq2Seq) structure based on an attention mechanism. The encoder is a 6-layer bidirectional LSTM with a hidden layer dimension of 512, the decoder is a 6-layer unidirectional LSTM with a hidden layer dimension of 512, the attention mechanism is additive attention (Bahdanau attention), the output layer is a linear layer followed by softmax, and the vocabulary size is 30,000. During training, the model uses a parallel corpus containing 100,000 sets of conversational semantic unit sequences and their corresponding topic description tags. The input is a concatenated text of multiple semantic units (connected by a special delimiter, with a maximum sequence length of 256), and the output is a sequence of topic description words (maximum length 20). Training uses the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, 80 training epochs, and cross-entropy loss. Planned sampling (with a planned sampling rate of 0.5) is used to assist training. In application, the input semantic units are concatenated and then segmented and serialized. The resulting sequence is then fed into the encoder to obtain the context vector. The decoder generates a topic description word sequence word by word in an autoregressive manner until an end tag is generated or the maximum length is reached. The generated word sequence is then directly used as the topic description tag.

[0158] For the pre-built conversational semantic inference model used in steps 12052 and 12055, the model adopts an autoregressive language model architecture based on the Transformer decoder, specifically a small version of GPT-2 (124M parameters), containing 12 decoder layers. Each layer contains a masked multi-head self-attention sublayer (12 heads) and a feedforward network sublayer (768 dimensions), and uses layer normalization and residual connections. In the pre-training stage, the model is self-supervised pre-trained using a general domain Chinese corpus (approximately 20GB of text data). The pre-training task is standard language modeling (predicting the next word), using the Adam optimizer, with an initial learning rate of 5e-5, a batch size of 8, and 3 training epochs. Building upon this foundation, fine-tuning was performed using 50,000 pairs of "context-receiver response" dialogues in a remote conversation scenario. During fine-tuning, the input format was a concatenated sequence of "context information + initial input state," and the output was the next response. The optimizer remained Adam, with the learning rate adjusted to 1e-5, batch size to 4, and training epochs to 10. Evaluation metrics included perplexity and BLEU-4 score. In application, the initial input state and context information were segmented and positionally encoded before being concatenated and fed into the decoder. Simultaneously, directional constraint vectors and intent constraint vectors were injected into each layer of the decoder via a cross-attention mechanism (as additional key-value pairs participating in cross-attention calculation). The decoder generated output words sequentially using an autoregressive approach. The generation process employed top-k sampling (k=50) and a temperature parameter of 0.9 to control generation diversity until a generation end marker was displayed. The generated word sequences were combined to form candidate transitional semantic units.

[0159] For the anchor point generation function in step 1404, a lightweight rule generation module is used internally instead of a neural network model. This module maps the dominant intent direction to directional words (such as "turn left", "turn right", "keep") based on preset template filling rules. It combines text fragments in the candidate intent semantic template with presentation intensity level parameters (three levels: low, medium, and high) to generate guiding statements of corresponding lengths (the higher the intensity, the longer the statement and the more definitive the vocabulary). This module does not require training and directly calls the predefined mapping table and string concatenation logic to complete the generation. When all the above models are deployed on the OPS microcomputer, INT8 quantization compression is used to reduce memory usage. Inference is accelerated based on ONNXRuntime. The input text of each model is uniformly preprocessed by converting full-width characters to half-width characters, removing redundant whitespace characters, and word segmentation (using the jieba word segmentation library). The output results are directly used in the original format (vector, category label, text sequence) for subsequent steps without additional post-processing.

[0160] Please refer to the following: Figure 2This illustration shows a schematic diagram of the three-terminal interaction between the remote interactive session processing server, the OPS terminal side, and the receiving end in an embodiment of the present invention. Figure 2 In this system, the data flow and processing logic between each terminal are divided according to time sequence and functional modules.

[0161] On the OPS terminal side, after the remote session stream is collected, it first undergoes semantic unit potential connection relationship parsing to extract the semantic association structure between the sending and receiving speech fragment units. This structure is then aggregated on the server side to generate a session progression semantic chain. This semantic chain represents the semantic trajectory of intent direction in the remote session stream in the form of a directed graph. This semantic chain is continuously updated to track the topic evolution of the current session. On the other hand, when a transmission interruption event occurs, it is sent to the autoregressive inference module (as illustrated by the RNN in the figure). Combining the last receiving speech unit and the current intent topic context, it infers transitional session fragments and backfills them into the temporal position corresponding to the interruption event in the receiving end presentation link.

[0162] On the receiving end, the utterance segments are extracted and uploaded to the server simultaneously with their presentation. The server determines the potential intent direction of the receiving conversation based on the utterance's prosodic features and semantic transition vectors. This trend is fed into an anchor generator, which, combined with the dominant intent direction and a pre-defined semantic intent template, generates visual semantic guidance anchors, which are then pushed to the receiving end's conversation interface. User interaction data with these anchors (gaze, click, ignore, etc.) is fed back to the server. After conversation rhythm profiling, a PID-like adjustment strategy is generated, which in turn adjusts the frequency of transitional conversation segments and the timing of the visual semantic guidance anchor pushes, thus forming a complete closed loop from perception to compensation, and from guidance to adaptive adjustment.

[0163] Please refer to the following: Figure 3 This illustrates a schematic diagram of the interface session on the receiving end. Figure 3 In the main conversation area, the interactive text between the sender and receiver is displayed in a bubble format. The sender's question, "What do you think of this solution?", and the receiver's reply, "I think the interface could be further optimized," are arranged chronologically. When a transmission interruption occurs, the system inserts a transitional conversation segment at the corresponding time point in the transmission chain on the receiver's display. This is illustrated in the image, where the text "Please wait, network fluctuations, connecting the context for you..." is displayed with a gradient background and italics. This segment visually informs the user of the current transmission delay while maintaining the continuity of the conversation and avoiding interruptions in interaction caused by blank waiting areas.

[0164] Visual semantic guidance anchors are displayed in a fixed area on the side or bottom of the conversation interface (labeled "AI Assistance" in the image). The example anchor in the image is "Suggested Reply Interface Parameters," which is generated based on the potential intent direction determined by the receiver's speech rhythm and semantic transition vector, and presented as a clickable button. When the user clicks this anchor, the system automatically fills in the corresponding guiding semantic text "Regarding interface parameters, I suggest..." into the input box, which the user can then edit or send directly. Below the anchor, three auxiliary operation options are provided: "More Detailed," "More Concise," and "Change Direction," used to adjust the level of detail in the guidance content or switch the intent direction, respectively. The "Current Assistance Frequency: Low / Medium / High" indicator in the upper right corner of the interface displays the execution status of the conversation rhythm generation adjustment strategy. This status is dynamically adjusted based on the receiver's interaction data with the anchor (such as gaze duration, click trigger time, and number of ignores), allowing the user to intuitively perceive the real-time changes in the system's assistance intensity, thereby achieving dynamic adaptation between guidance density and user response patterns.

[0165] In summary, this invention establishes a processing logic on the OPS terminal side encompassing semantic unit connection parsing, transmission interruption compensation, latent intent perception, guidance anchor point generation, and rhythm adaptive adjustment. By using a conversational advancement semantic chain to structurally represent the intent trajectory of interleaved utterance fragments, autoregressive deduction can be performed based on this semantic trajectory during transmission interruptions, generating transitional conversation fragments consistent with the contextual intent direction to fill conversation gaps. Simultaneously, by fusing the prosodic fluctuations of the receiver's utterances with semantic transition vectors, latent intent trends are mined. The resulting visualized semantic guidance anchor points can make implicit intent trends explicit as directional guidance. Furthermore, by collecting data on the receiver's interaction with the anchor points, the conversation rhythm generation strategy is dynamically adjusted, achieving real-time adaptation between guidance intensity and receiver response patterns. This realizes a systematic functional leap in remote conversations from passive transmission to active semantic compensation and intelligent guidance.

[0166] This invention also provides a remote interactive session processing server: a processor; a storage device storing a computer program thereon; and a network interface for providing network communication functions; when the computer program is executed by the processor, the processor implements any of the remote interactive session processing methods based on an OPS microcomputer.

[0167] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0168] Furthermore, it should be noted that this embodiment of the invention also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of the remote interactive session processing server reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the remote interactive session processing server to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this invention, please refer to the description of the method embodiments of this invention.

[0169] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A remote interactive session processing method based on an OPS microcomputer, characterized in that, include: On the OPS terminal side, semantic unit connection relationship parsing processing is performed on the remote session stream to extract the potential semantic unit connection relationship between the sending end speech fragment unit and the receiving end speech fragment unit in the remote session stream, and generate a session advancement semantic chain. The session advancement semantic chain is used to characterize the semantic trajectory of the intention direction in the remote session stream. When a transmission interruption event occurs in the remote session stream, the sender speech segment unit is subjected to autoregressive inference processing based on the semantic trajectory to generate a transitional session segment and embed it into the timing position in the receiving end presentation link corresponding to the transmission interruption event. The potential intent direction trend of the receiver conversation in the remote conversation stream is determined based on the prosodic fluctuation features and semantic transition vectors of the receiver speech fragment units. Based on the potential intent pointing trend, a visual semantic guidance anchor is established, and the visual semantic guidance anchor is pushed to the conversation interface of the receiving end. The visual semantic guidance anchor is used to guide the response direction of the conversation of the receiving end. The receiving end collects interactive behavior data on the visual semantic guidance anchor point, and determines a conversation rhythm generation and adjustment strategy based on the interactive behavior data. The conversation rhythm generation and adjustment strategy is used to adjust the generation frequency of the transitional conversation segments and the push timing of the visual semantic guidance anchor point.

2. The method according to claim 1, characterized in that, The step of parsing semantic unit connections in the remote session stream on the OPS terminal side, extracting the potential semantic unit connections between the sending and receiving speech fragment units in the remote session stream, and generating a session advancement semantic chain includes: The sending end speech segment unit in the remote session stream is subjected to semantic unit segmentation processing to obtain a sending end semantic unit sequence, which contains multiple semantically independent sending end semantic units; The receiving end speech segment unit in the remote session stream is subjected to semantic unit segmentation processing to obtain a receiving end semantic unit sequence, which contains multiple semantically independent receiving end semantic units. Calculate the semantic association strength parameter between each transmitting semantic unit in the transmitting semantic unit sequence and each receiving semantic unit in the receiving semantic unit sequence; A semantic connection pair between the sending semantic unit and the receiving semantic unit is created based on the semantic association strength parameter. Each semantic connection pair includes a sending semantic unit, a receiving semantic unit, and the semantic association strength parameter between them. The semantic connection pairs are arranged according to the timestamp order of the sending speech fragment unit and the receiving speech fragment unit to obtain a semantic connection pair sequence; Semantic trajectory continuity analysis is performed on adjacent semantic connection pairs in the semantic connection pair sequence to generate semantic transition direction vectors between adjacent semantic connection pairs; A session advancement semantic chain for the remote session stream is created based on the semantic transition direction vector and the semantic connection pair sequence, wherein the session advancement semantic chain has the sending semantic unit and the receiving semantic unit as nodes and the semantic transition direction vector as edges.

3. The method according to claim 2, characterized in that, The step of performing semantic trajectory continuity analysis on adjacent semantic connection pairs in the semantic connection pair sequence to generate semantic transition direction vectors between adjacent semantic connection pairs includes: Extract the sending semantic unit and receiving semantic unit of the current semantic connection pair in the semantic connection pair sequence, as well as the sending semantic unit and receiving semantic unit of the next semantic connection pair; Calculate the semantic continuity tightness between the receiving semantic unit of the current semantic connection pair and the sending semantic unit of the next semantic connection pair, and determine the semantic trajectory breakage index between the current semantic connection pair and the next semantic connection pair based on the semantic continuity tightness. Extract the discourse intent category features of the receiving semantic unit in the current semantic connection pair and the discourse intent category features of the sending semantic unit in the next semantic connection pair. Calculate the intent transition direction angle parameter based on the discourse intent category features of the receiving semantic unit in the current semantic connection pair and the discourse intent category features of the sending semantic unit in the next semantic connection pair. The semantic trajectory breakage index and the intention transition direction angle parameter are input into the semantic transition direction generation function to generate a semantic transition direction vector between adjacent semantic connection pairs. The semantic transition direction vector contains both the continuity information of the semantic trajectory and the direction information of the intention transition.

4. The method according to claim 2, characterized in that, The step of creating the session advancement semantic chain of the remote session stream based on the semantic transition direction vector and the semantic connection sequence includes: Each semantic connection pair in the semantic connection pair sequence is mapped to a semantic node in the session advancement semantic chain. The semantic node contains the sending semantic unit, the receiving semantic unit, and the semantic association strength parameter between the two of the semantic connection pair. The semantic nodes in the session advancement semantic chain are arranged according to the timestamp order of the semantic connection pair sequence, and the semantic transition direction vector between adjacent semantic connection pairs is mapped to the directed edge between adjacent semantic nodes in the session advancement semantic chain. Extract the intention transition direction angle parameter contained in each semantic transition direction vector, and assign the intention transition direction angle parameter as the direction attribute of the directed edge. Extract the semantic trajectory breakage index contained in each semantic transition direction vector, and assign the semantic trajectory breakage index as the weight attribute of the directed edge. Statistical processing is performed on the local density distribution of the semantic nodes in the session advancement semantic chain to generate the clustering region boundary of the semantic nodes in the session advancement semantic chain; The session advance semantic chain is divided into multiple semantic intent blocks according to the aggregation region boundary of the semantic nodes, and each semantic intent block corresponds to an intent topic unit in the remote session stream; Generate an intent topic description tag for each semantic intent block, and attach the intent topic description tag to the corresponding set of semantic nodes in the session advancement semantic chain.

5. The method according to claim 1, characterized in that, When a transmission interruption event occurs in the remote session stream, the process of performing autoregressive deduction processing on the sending end speech segment unit based on the semantic trajectory to generate a transitional session segment and embedding it into the receiving end presentation link at the timing position corresponding to the transmission interruption event includes: Monitor the arrival time interval of the speech segment units sent by the remote session stream. When the arrival time interval exceeds a preset interruption determination threshold, determine that a transmission interruption event has occurred in the remote session stream and record the start time of the transmission interruption event. Extract the latest generated semantic transition direction vector in the session advancement semantic chain before the transmission interruption event occurs, and determine the extension direction of the current semantic trajectory based on the latest generated semantic transition direction vector; Extract intent topic description tags corresponding to the extension direction of the current semantic trajectory from the session advancement semantic chain to determine the intent topic context of the current session; Extract the last transmitting semantic unit and the last receiving semantic unit before the transmission interruption event occurs, and use the last transmitting semantic unit and the last receiving semantic unit as the starting semantic state for autoregressive inference; Starting from the initial semantic state, and combining the extension direction of the current semantic trajectory and the intent topic context of the current session, autoregressive inference processing is performed to generate multiple candidate transition semantic units. Calculate the semantic trajectory consistency parameter between each candidate transition semantic unit and the receiving semantic unit in the initial semantic state, and select the target transition semantic unit from the plurality of candidate transition semantic units based on the semantic trajectory consistency parameter; The target transitional semantic unit is serialized and organized according to the extension direction of the current semantic trajectory to generate a transitional session fragment. The semantic content of the transitional session fragment maintains semantic trajectory continuity with the session advancement semantic chain before the transmission interruption event. Determine the timing position of the transmission interruption event in the receiving end presentation link and embed the timing position in the receiving end presentation link.

6. The method according to claim 5, characterized in that, Starting from the initial semantic state, and combining the extension direction of the current semantic trajectory and the intent topic context of the current session, autoregressive inference processing is performed to generate multiple candidate transitional semantic units, including: The last receiving semantic unit is used as the initial input state for autoregressive inference, and the last sending semantic unit is used as the contextual information of the initial input state. The initial input state and the preceding context information are semantically encoded using a pre-built conversational semantic inference model to generate an initial semantic state vector. The extension direction of the current semantic trajectory is transformed into a direction constraint vector, which is used to limit the generation direction of semantic units during the autoregressive inference process; The intent topic context of the current session is transformed into an intent constraint vector, which is used to limit the range of semantic unit generated content during the autoregressive inference process; The initial semantic state vector, the direction constraint vector, and the intention constraint vector are fused to generate an inference constraint vector. The inference constraint vector is then input into the semantic unit generation module of the conversation semantic inference model to generate candidate transition semantic units for the current time step. The candidate transition semantic units of the current time step are fed back to the initial input state, the initial input state is updated to the candidate transition semantic units of the current time step, and the steps of semantic state encoding processing, deduction constraint vector generation and candidate transition semantic unit generation are iteratively executed until a candidate transition semantic unit sequence that meets the preset number requirement is generated.

7. The method according to claim 1, characterized in that, The determination of the potential intent orientation trend of the receiver's conversation in the remote conversation stream based on the prosodic fluctuation features and semantic transition vectors of the receiver's speech fragment units includes: The receiving end speech segment unit is subjected to prosodic feature extraction processing to generate the speech prosodic fluctuation feature of the receiving end speech segment unit. The speech prosodic fluctuation feature includes the intonation rise and fall sequence and the speech rate change sequence. The receiving end speech segment unit is subjected to semantic transition analysis processing to generate a semantic transition vector of the receiving end speech segment unit. The semantic transition vector includes the direction angle parameter and the magnitude parameter of the semantic transition. The distribution information of intonation inflection points of the received speech segment unit is calculated based on the intonation rise and fall sequence, and the position of the emotional tendency transition of the received speech is determined based on the intonation inflection point distribution information. The distribution information of speech rate change points of the receiving end speech segment unit is calculated based on the speech rate change sequence, and the emphasis position of the receiving end speech is determined based on the speech rate change point distribution information. The emotional tendency transition position and the emphasis position are associated and matched with the semantic transition direction angle parameter in the semantic transition vector to generate the intention pointing marker point of the receiving end utterance; The intent pointing intensity of the intent pointing marker is quantified based on the semantic transition amplitude parameter in the semantic transition vector to generate the intent pointing confidence parameter of the intent pointing marker. Based on the intent pointing marker and the intent pointing confidence parameter of the intent pointing marker, the potential intent pointing trend of the receiving end session is determined. The potential intent pointing trend includes the directional distribution characteristics and the intensity distribution characteristics of intent pointing.

8. The method according to claim 1, characterized in that, The establishment of visual semantic guidance anchors based on the potential intent-oriented trends includes: The directional distribution features of intent pointing contained in the potential intent pointing trend are analyzed, and the intent pointing direction with the highest intent pointing confidence parameter in the directional distribution features is extracted as the dominant intent direction; Based on the dominant intent direction, a candidate intent semantic template matching the dominant intent direction is retrieved from a preset intent semantic template library. The candidate intent semantic template includes guiding semantic content fragments and guiding semantic presentation styles. Extract the intensity distribution features of intent pointing contained in the potential intent pointing trend, and determine the presentation intensity level parameter of the candidate intent semantic template based on the intensity distribution features; The dominant intent direction, the candidate intent semantic template, and the presentation intensity level parameter are input into the anchor point generation function to generate a visual semantic guidance anchor point. The visual semantic guidance anchor point includes guiding semantic text content and visual presentation attribute parameters. The visual semantic guidance anchor is rendered to the preset anchor display area in the conversation interface of the receiving end according to the visual presentation attribute parameters.

9. The method according to claim 1, characterized in that, The data acquisition and receiving end, based on the interactive behavior data of the visual semantic guidance anchor point, determines a session rhythm generation and adjustment strategy, including: Monitor the presentation status of the visual semantic guidance anchor in the receiving terminal's session interface, and collect the interactive behavior data of the receiving terminal to the visual semantic guidance anchor. The interactive behavior data includes anchor gaze duration parameter, anchor click trigger time parameter, and anchor ignore count parameter. Calculate the response delay parameter of the receiving end to the visual semantic guidance anchor point based on the anchor point gaze duration parameter and the anchor point click trigger time parameter; The acceptance rate parameter of the visual semantic guidance anchor point is calculated based on the anchor point ignore count parameter and the total number of times the visual semantic guidance anchor point is pushed. Extract the response utterance fragment unit input by the receiving end after triggering the visual semantic guidance anchor point, and calculate the intent matching degree parameter between the response utterance fragment unit and the potential intent pointing trend corresponding to the visual semantic guidance anchor point; A receiver conversation rhythm profile is constructed based on the response delay parameter, the acceptance rate parameter, and the intent matching degree parameter. The receiver conversation rhythm profile includes the receiver's response speed characteristics to conversation guidance and the receiver's acceptance degree characteristics to conversation guidance. Based on the receiver's conversation rhythm profile, determine the frequency adjustment direction for generating the transitional conversation segment and the timing adjustment direction for pushing the visual semantic guidance anchor point; The generation frequency parameter of the transitional session segment is adjusted according to the generation frequency adjustment direction of the transitional session segment, and the push timing parameter of the visual semantic guidance anchor point is adjusted according to the push timing adjustment direction of the visual semantic guidance anchor point. The adjusted frequency parameters for generating the transitional conversation segments and the adjusted timing parameters for pushing the visual semantic guidance anchors are combined to form a conversation rhythm generation and adjustment strategy.

10. A remote interactive session processing server, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the remote interactive session processing method based on an OPS microcomputer as described in any one of claims 1-9.

Images