A Deep Learning-Based Sentiment Classification Method for Social Network Corpora

Abstract

This invention discloses a deep learning-based method for sentiment classification of social network corpora, comprising the following steps: acquiring and preprocessing social network corpora to generate an aligned input set; performing shared semantic encoding to obtain representations of the current text, context, social signals, and user style; performing irony trigger decomposition based on the shared semantic encoding results to form explicit emotions, event contexts, tone modifiers, and social reinforcement segments; generating irony probability, conflict intensity, and literal sentiment results via the irony branch; generating true emotions and emotion cause results via the true emotion branch; and outputting the sentiment classification result and evidence package result through consistency adjudication. This invention uses deep learning and event-constrained inversion methods to achieve true sentiment classification of social network corpora, and has the advantages of accurate irony recognition, interpretable results, and strong traceability.

Description

Technical Field

[0001] This invention relates to the field of emotion recognition technology, and in particular to an emotion classification method for social network corpora based on deep learning. Background Technology

[0002] With the widespread use of social networking platforms, social network corpora such as microblogs, comment sections, forum replies, and forwarded comments have become important data sources for user emotion recognition. Existing technologies commonly employ convolutional neural networks, recurrent neural networks, pre-trained language models, or combinations thereof to vectorize social text and classify its sentiment. Some solutions also combine sentiment dictionaries, contextual semantic information, or dual-channel feature extraction methods to improve the processing capabilities for short texts, colloquial expressions, and non-standard expressions. Building upon this foundation, some technologies are beginning to attempt to detect and analyze satirical texts, ironic texts, or the causes of emotions to enhance the accuracy and interpretability of emotion recognition.

[0003] However, most existing technologies still focus on overall text classification, lacking structured decomposition of explicit emotional segments, event context segments, tone modification segments, and social reinforcement segments, making it difficult to accurately identify the conflict relationship between literal emotion and implicit context. At the same time, existing solutions usually focus on a single link in irony detection or emotion classification, lacking a real emotion recovery mechanism based on event anomaly type and anomaly severity, making it difficult to stably output real emotion results and emotion cause results in irony scenarios. Furthermore, in the final output stage, there is a lack of rule-based adjudication links corresponding to irony probability results, conflict intensity results, literal emotion results, and real emotion results, resulting in insufficient interpretability and traceability of classification results.

[0004] Therefore, how to provide a deep learning-based method for sentiment classification of social network corpora is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for classifying the sentiment of social network corpora. This invention uses deep learning and event-constrained inversion methods to classify the true sentiment of social corpora, and has the advantages of accurate irony recognition, interpretable results, and strong traceability.

[0006] A deep learning-based sentiment classification method for social network corpora according to an embodiment of the present invention includes the following steps: Acquire social network corpora, perform text alignment, context association, social additional signal collection, and user style fragment organization to generate an aligned input set; The aligned input set is fed into the shared semantic encoding layer of the irony-real emotion dual-branch consistent network to generate shared semantic encoding results, including current text representation, context representation, social signal representation, and user style representation; The shared semantic encoding results are input into the irony trigger splitting layer to generate a set of fragment splits, including a set of explicit emotion fragments, a set of event context fragments, a set of tone modification fragments, and a set of social reinforcement fragments; The fragment splitting set and the shared semantic encoding result are input into the irony branch, and literal emotion modeling, implicit context modeling, conflict graph construction and conflict graph reasoning are performed to generate irony probability results, conflict intensity results and literal emotion results. The fragment splitting set and the shared semantic encoding result are input into the real emotion branch, and event constraint inversion and real emotion prototype matching are performed to generate real emotion results and emotion cause results. Based on the results of irony probability, conflict intensity, literal emotion, true emotion, and emotion cause, a decision is made at the consistency decision layer to generate the final emotion classification result and evidence package result.

[0007] Optionally, the generation of the aligned input set specifically includes: The social network corpus is acquired and merged according to the publisher identifier, publication timestamp, conversation identifier, and corpus source identifier to generate the original set of social network corpus. Perform text normalization and text alignment on the original set of social network corpus to generate the current set of text fragments; Context tracing and adjacency association are performed on the original set of social network corpus to generate a set of conversation contexts; Social additional signals are aggregated based on non-textual information in the original set of social network corpus, and a set of social additional signals is generated. Perform historical backtracking, tone pattern extraction, facial expression usage pattern extraction, and evaluation tendency extraction on the original set of social network corpus to generate a set of user style fragments, tone pattern records, facial expression usage pattern records, and evaluation tendency records; The current text fragment set, the conversation context set, the social additional signal set, and the user style fragment set are aligned and encapsulated according to a unified sample primary key to generate an aligned input set.

[0008] Optionally, the generation of the shared semantic encoding result specifically includes: The aligned input set is input into the shared semantic encoding layer of the irony-true emotion dual-branch consistent network, parsed, and a field alignment index is built to generate the shared semantic encoding input set. Perform word-granular segmentation, fragment boundary annotation, and position writing on the current text field in the shared semantic encoded input set to generate the current text encoded input sequence; Perform hierarchical expansion, sequential concatenation, and context role writing on the session context fields in the shared semantic encoded input set to generate a context encoded input sequence; Perform type mapping and position backfilling on the social additional signal fields in the shared semantic encoded input set to generate a social signal encoded input sequence; Based on tone pattern records, facial expression usage pattern records, and evaluation tendency records, historical fragment sorting, style fragment filtering, and aggregate coding preparation are performed on the user style fragment field in the shared semantic coding input set to generate a user style coding input sequence. Hierarchical encoding and cross-source fusion are performed on the current text-encoded input sequence, context-encoded input sequence, social signal-encoded input sequence, and user style-encoded input sequence to generate a shared representation sequence; Based on the shared representation sequence, source aggregation and association encapsulation are performed according to the primary key field of the sample to generate shared semantic encoding results.

[0009] Optionally, the irony-real emotion dual-branch consistency network includes a shared semantic encoding layer, an irony triggering splitting layer, an irony branch, a real emotion branch, and a consistency adjudication layer.

[0010] Optionally, the generation of the fragment split set specifically includes: The shared semantic encoding result is input into the irony trigger splitting layer, and a trigger alignment sequence is established according to the current text representation, context representation, social signal representation and user style representation to generate a trigger input set; Based on the trigger input set, trigger candidate localization is performed, and a trigger candidate set is generated; Based on the trigger candidate set, perform dominant emotion segment determination and boundary merging to generate a dominant emotion segment set; Based on the trigger candidate set, the event context fragment determination and boundary extension are performed to generate the event context fragment set; Based on the trigger candidate set, the tone modification fragment determination and social reinforcement fragment determination are performed, and a tone modification fragment set and a social reinforcement fragment set are generated; Based on the sets of explicit emotion fragments, event context fragments, tone modification fragments, and social reinforcement fragments, the fragment splitting conflict resolution and result encapsulation are performed to generate a fragment splitting set.

[0011] Optionally, the generation of the irony probability result, conflict intensity result, and literal sentiment result specifically includes: The fragment splitting set and the shared semantic encoding result are input into the irony branch to perform fragment alignment and role mapping, generating the irony branch input set; Based on the input set of ironic branches, literal sentiment modeling is performed to generate intermediate records of literal sentiment. Latent context modeling is performed based on the input set of irony branches to generate intermediate records of the latent context; Conflict graphs are constructed based on literal emotion intermediate records, implicit context intermediate records, tone modification fragment sets, and social reinforcement fragment sets; Conflict graph reasoning is performed based on the conflict graph, and conflict reasoning records are generated. Based on the aggregation of literal sentiment intermediate records and conflict inference records, the results of irony probability, conflict intensity, and literal sentiment are generated.

[0012] Optionally, for each explicit emotional fragment node, the conflict graph reasoning searches for event context fragment nodes connected to it via fragment position adjacency edges, semantic transition edges, or contextual response edges. When the emotional direction label of the explicit emotional fragment node is opposite to the context polarity label of the event context fragment node, an opposite direction label is written. When the emotional intensity label of the explicit emotional fragment node is higher than the context severity label of the event context fragment node and there is a semantic transition edge or social reinforcement edge between them, a reinforced conflict label is written. When there are both contextual response edges and style matching edges between the explicit emotional fragment node and the event context fragment node, and the evaluation direction corresponding to the user style representation is inconsistent with the emotional direction label of the current explicit emotional fragment node, a style contrast label is written. The number of opposite direction labels, reinforced conflict labels, and style contrast labels are counted for each explicit emotional fragment node to obtain the conflict reasoning record.

[0013] Optionally, the generation of the actual emotional outcome and the emotional cause outcome specifically includes: The fragment split set and the shared semantic encoding result are input into the real emotion branch to perform inversion alignment, generating the real emotion branch input set; Based on the input set of real emotion branches, the event constraint object is located and a set of event constraint objects is generated. Based on the set of event constraint objects, perform exception type determination and exception severity determination, and generate event constraint records; Based on the event constraint record, perform event constraint inversion to generate a true emotion inversion record; Based on the real emotion inversion record, the real emotion prototype screening, prototype matching and cause location are performed to generate real emotion intermediate record and emotion cause record; Based on the encapsulation of the execution results of the real emotion intermediate records and emotion cause records, the real emotion results and emotion cause results are generated.

[0014] Optionally, the generation of the final emotion classification result and evidence package result specifically includes: The results of irony probability, conflict intensity, literal sentiment, true sentiment, and sentiment cause are input into the consistency adjudication layer to align the execution results and collect samples, generating an adjudication input set. Based on the set of adjudication inputs, an adjudication pattern determination is performed, and an adjudication pattern result is generated. A consistency determination is performed based on the set of adjudication inputs and the adjudication pattern results, and a consistency determination result is generated. Based on the adjudication pattern results and consistency judgment results, the adjudication output is executed to generate the final sentiment classification result; Based on the adjudication input set, adjudication mode result, consistency judgment result, and final emotion intermediate record execution evidence field encapsulation, an evidence package result is generated.

[0015] Optionally, the consistency determination includes reading the emotion direction marker in the literal emotion result, reading the matching prototype identifier in the real emotion result and the corresponding real evaluation direction as the real emotion direction marker, and writing a reverse causal conflict marker when the corresponding segment position in the literal emotion result overlaps with the corresponding segment position in the emotion cause result or has a direct response association through the associated context marker, and the literal emotion direction marker is opposite to the real emotion direction marker. When the literal emotion direction marker is the same as the real emotion direction marker and the corresponding segment position overlaps, a same-direction consistency marker is written. When the literal emotion direction marker is opposite to the real emotion direction marker and there is no position overlap and no direct response association, a reverse causal conflict marker is written.

[0016] The beneficial effects of this invention are: This invention addresses common issues in social network corpora, such as irony, context reversal, emoticon reinforcement, and individual user expression style differences. It constructs a complete processing chain from generating aligned input sets, shared semantic encoding, irony trigger decomposition, irony branch determination, recovery of genuine emotion branches, to consistent judgment output. By refining social network corpora into current text, conversational context, social additional signals, and user style fragments, and further subdividing them into explicit emotion fragments, event context fragments, tone modification fragments, and social reinforcement fragments, the relationship between literal emotion and implicit context is no longer limited to coarse-grained judgment at the overall text level, but can be structurally analyzed at the fragment level. This improves the ability to identify ironic statements, emotion reversal statements, and statements with reinforcement signals.

[0017] Building upon this foundation, this invention outputs irony probability results, conflict intensity results, and literal emotion results through an irony branch. Then, through a true emotion branch, it generates true emotion results and emotion cause results based on the event constraint object, anomaly type, anomaly severity, and true emotion prototype matching. This allows the determination of true emotion to move beyond simply relying on surface-level emotion word direction, instead combining event subject description, event result description, and anomalous facts to complete constraint inversion. Furthermore, this invention integrates irony probability results, conflict intensity results, literal emotion results, true emotion results, and emotion cause results in a rule-based adjudication layer, simultaneously generating an evidence package result. Therefore, it not only improves the accuracy of the final emotion classification result in complex social contexts but also enhances the interpretability, traceability, and stability of the results. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based social network corpus sentiment classification method proposed in this invention; Figure 2 This is a schematic diagram of the irony branch process of a deep learning-based social network corpus sentiment classification method proposed in this invention. Figure 3 This is a schematic diagram of the real emotion branch flow of a deep learning-based social network corpus emotion classification method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A deep learning-based method for sentiment classification of social network corpora includes the following steps: Acquire social network corpora, perform text alignment, context association, social additional signal collection, and user style fragment organization to generate an aligned input set; The aligned input set is fed into the shared semantic encoding layer of the irony-real emotion dual-branch consistent network to generate shared semantic encoding results, including current text representation, context representation, social signal representation, and user style representation; The shared semantic encoding results are input into the irony trigger splitting layer to generate a set of fragment splits, including a set of explicit emotion fragments, a set of event context fragments, a set of tone modification fragments, and a set of social reinforcement fragments; The fragment splitting set and the shared semantic encoding result are input into the irony branch, and literal emotion modeling, implicit context modeling, conflict graph construction and conflict graph reasoning are performed to generate irony probability results, conflict intensity results and literal emotion results. The fragment splitting set and the shared semantic encoding result are input into the real emotion branch, and event constraint inversion and real emotion prototype matching are performed to generate real emotion results and emotion cause results. Based on the results of irony probability, conflict intensity, literal emotion, true emotion, and emotion cause, a decision is made at the consistency decision layer to generate the final emotion classification result and evidence package result.

[0021] In this embodiment, the generation of the aligned input set specifically includes: The social network corpus is acquired and merged according to the publisher identifier, publication timestamp, conversation identifier, and corpus source identifier to generate the original set of social network corpus. The original set of social network corpus includes main post text, comment text, reply text, forwarded comments, emoji sequence, topic tag sequence, interaction object identifier, and source location identifier, and writes a corpus index record for each piece of social network corpus; Perform text normalization and text alignment on the original set of social network corpus to generate the current set of text fragments; The text normalization process includes removing invalid whitespace, unifying repeated punctuation, retaining emoji placeholders, retaining topic tag placeholders, retaining interactive object placeholders, unifying forwarding and reply prefixes, and performing sentence segmentation, sequential backfilling, and missing insertion on the normalized text content according to the corpus index records. Context tracing and adjacency association are performed on the original set of social network corpus to generate a set of conversation contexts; Context tracing and adjacency association include locating the upstream main post, direct reply, previous reply and adjacent comments in the same session corresponding to the current text based on the session identifier, reply pointing relationship and posting timestamp, constructing the context association order according to the reply level, and removing corpus fragments that have no reply relationship with the current text and are outside the preset session window; Social additional signals are aggregated based on non-textual information in the original set of social network corpus, and a set of social additional signals is generated. The collection of social additional signals includes extracting emoji sequences, topic tag sequences, interaction object identifiers, forwarding relationship identifiers, citation relationship identifiers, repeated punctuation identifiers, rhetorical question identifiers, and emphasis identifiers, and mapping them to the corresponding current text segments according to the corpus index records; Perform historical backtracking, tone pattern extraction, facial expression usage pattern extraction, and evaluation tendency extraction on the original set of social network corpus to generate a set of user style fragments, tone pattern records, facial expression usage pattern records, and evaluation tendency records; The process includes historical backtracking, tone pattern extraction, emoji usage pattern extraction, and evaluation tendency extraction. This involves retrieving historical texts within a preset time window from the same publisher based on the publisher's identifier and timestamp. Texts repeatedly forwarded in the current conversation and texts shorter than a preset length are removed. Normalization and cross-text counting are performed on sentence-ending modifiers, rhetorical question markers, exclamation markers, repeated punctuation, degree intensifiers, and transitional words in the retained historical text to generate tone pattern records. Emoji sequence in the retained historical text is mapped by emoji category, marked with position, and counted across text to generate emoji usage pattern records. Evaluation direction merging and cross-text counting are performed on evaluation words, negation words, degree words, emotion words, and event outcome descriptions in the retained historical text to generate evaluation tendency records. Finally, the retained historical texts are organized according to timestamp and tone similarity to obtain a set of user style fragments. The current text fragment set, the conversation context set, the social additional signal set, and the user style fragment set are aligned and encapsulated according to a unified sample primary key to generate an aligned input set.

[0022] In this embodiment, the generation of shared semantic encoding results specifically includes: The aligned input set is input into the shared semantic encoding layer of the irony-true emotion dual-branch consistent network, parsed, and a field alignment index is built to generate the shared semantic encoding input set. The field alignment index includes the current text position index, the session context order index, the social additional signal mapping index, and the user style fragment order index. Based on the field alignment index, null value masking, length normalization, and order correction are performed on each field to obtain a shared semantic encoding input set. Perform word-granular segmentation, fragment boundary annotation, and position writing on the current text field in the shared semantic encoded input set to generate the current text encoded input sequence; The current text encoding input sequence is based on the sentence segmentation results in the current text fragment set. Emoji placeholders, topic tag placeholders, interactive object placeholders, forwarding prefixes, and reply prefixes are written as reserved symbols into a unified sequence, and sentence boundary markers and fragment order markers are written according to the current text position index. Perform hierarchical expansion, sequential concatenation, and context role writing on the session context fields in the shared semantic encoded input set to generate a context encoded input sequence; The hierarchical expansion includes unfolding each context fragment in the order of upstream main post, direct reply, previous reply, and adjacent comments in the same session context set. The context role writing includes writing the main post role tag, direct reply role tag, previous reply role tag, or adjacent comment role tag for each context fragment, and completing the splicing according to the session context order index. Perform type mapping and position backfilling on the social additional signal fields in the shared semantic encoded input set to generate a social signal encoded input sequence; The type mapping includes mapping an emoji sequence to an emoji type marker, mapping a topic tag sequence to a topic tag type marker, and mapping interaction object identifiers, forwarding relationship identifiers, citation relationship identifiers, repeated punctuation identifiers, rhetorical question identifiers, and emphasis identifiers to corresponding signal type markers. The position backfilling includes backfilling each signal type marker to the corresponding current text position or fragment position according to the social additional signal mapping index, to obtain the social signal encoding input sequence. Based on tone pattern records, facial expression usage pattern records, and evaluation tendency records, historical fragment sorting, style fragment filtering, and aggregate coding preparation are performed on the user style fragment field in the shared semantic coding input set to generate a user style coding input sequence. The historical fragment sorting process includes arranging historical text fragments according to the publication timestamp and tone similarity order in the user style fragment set. The style fragment filtering process includes recording the tone expression form in the unified historical text fragments according to tone pattern, recording the expression category and expression position form in the unified historical text fragments according to expression usage pattern, and marking the evaluation direction in the historical text fragments according to evaluation tendency record. The fragment boundaries corresponding to the tone pattern record, expression usage pattern record, and evaluation tendency record are preserved to obtain the user style encoding input sequence. Hierarchical encoding and cross-source fusion are performed on the current text-encoded input sequence, context-encoded input sequence, social signal-encoded input sequence, and user style-encoded input sequence to generate a shared representation sequence; The hierarchical encoding includes performing sequence encoding on the current text encoding input sequence, the context encoding input sequence, the social signal encoding input sequence, and the user style encoding input sequence respectively to obtain the current text initial representation sequence, the context initial representation sequence, the social signal initial representation sequence, and the user style initial representation sequence. The cross-source fusion includes using the current text initial representation sequence as the main sequence, and aligning and fusing the context initial representation sequence, the social signal initial representation sequence, and the user style initial representation sequence to their corresponding positions based on the field alignment index to obtain a shared representation sequence. Based on the shared representation sequence, source aggregation and association encapsulation are performed according to the primary key field of the sample to generate shared semantic encoding results.

[0023] In this embodiment, the irony-real emotion dual-branch consistency network includes a shared semantic encoding layer, an irony triggering splitting layer, an irony branch, a real emotion branch, and a consistency adjudication layer.

[0024] In this embodiment, the generation of the fragment split set specifically includes: The shared semantic encoding result is input into the irony trigger splitting layer, and a trigger alignment sequence is established according to the current text representation, context representation, social signal representation and user style representation to generate a trigger input set; The trigger alignment sequence is based on the text position order corresponding to the current text representation. Alignment mapping is performed on positions in the context representation that have a reply association or adjacent association with the current text. Position backfilling is performed on the emoji type marker, topic tag type marker, interaction object type marker, repeated punctuation type marker, rhetorical question type marker, and emphasis type marker corresponding to the current text position in the social signal representation. Style association is performed on the positions of historical fragments in the user style representation that match the tone expression, emoji category, and evaluation direction of the current text to obtain the trigger input set. Based on the trigger input set, trigger candidate localization is performed, and a trigger candidate set is generated; The trigger candidate localization uses the positions of emotion words, negation words, degree words, transition words, rhetorical question markers, exclamation markers, repeated punctuation marks, emoticons, topic tags, and event result descriptions in the current text representation as fixed splitting entry points. It combines the semantic transition positions in the context representation, the reinforcement signal positions in the social signal representation, and the high-frequency tone positions in the user style representation to extract candidates and retain their order, thus obtaining the trigger candidate set. Based on the trigger candidate set, perform dominant emotion segment determination and boundary merging to generate a dominant emotion segment set; The explicit emotion segment determination and boundary merging process involves taking the position of the emotion word, degree word, negation word, and exclamation mark in the trigger candidate set as the core position, absorbing the degree intensifier, negation word, and tone modifier that are continuously adjacent to it along the current position, absorbing the evaluation object descriptor and exclamation mark that are continuously adjacent to it along the current position, performing boundary merging on overlapping segments, and eliminating candidate segments that only contain function words and do not contain emotional expression, thus obtaining the explicit emotion segment set. Based on the trigger candidate set, the event context fragment determination and boundary extension are performed to generate the event context fragment set; The event context fragment determination and boundary extension include taking the event result description position, time delay description position, quantity anomaly description position, service behavior description position, functional status description position and context semantic transition position in the trigger candidate set as the core positions, extending along the sentence segment boundary to both sides to the complete event expression unit, performing cross-fragment splicing on context fragments that have a direct response relationship with the current event expression unit, and removing candidate fragments that do not contain event subject or result description, to obtain the event context fragment set; Based on the trigger candidate set, the tone modification fragment determination and social reinforcement fragment determination are performed, and a tone modification fragment set and a social reinforcement fragment set are generated; The determination of tone-modifying segments includes extracting rhetorical question markers, transition words, sentence-ending tone words, repeated punctuation, and intensity words located before and after emotion words in the current text, and merging them into tone-modifying segments according to their adjacent positions; The social reinforcement fragment determination includes extracting emoji type markers, topic tag type markers, interaction object type markers, citation relationship markers, forwarding relationship markers, and emphasis type markers corresponding to the current text position, and merging them into social reinforcement fragments according to the current position or fragment position to obtain a tone modification fragment set and a social reinforcement fragment set; Based on the sets of explicit emotion fragments, event context fragments, tone modification fragments, and social reinforcement fragments, the fragment splitting conflict resolution and result encapsulation are performed to generate a fragment splitting set; The conflict resolution process includes adjudicating any overlapping segments according to a fixed priority order, where event context segments take precedence over explicit emotion segments, explicit emotion segments take precedence over tone modification segments, tone modification segments take precedence over social reinforcement segments, low-priority segments that are completely covered by high-priority segments are pruned, and low-priority segments that only partially overlap with high-priority segments retain non-overlapping boundaries. The result encapsulation includes writing fragment type tags, fragment start position, fragment end position, source position, and associated context tags for each fragment, and summarizing them in order of the current text position.

[0025] In this embodiment, the generation of irony probability results, conflict intensity results, and literal sentiment results specifically includes: The fragment splitting set and the shared semantic encoding result are input into the irony branch to perform fragment alignment and role mapping, generating the irony branch input set; The segment alignment and role mapping include aligning the explicit emotion segment set with the current text representation according to the segment start position and segment end position, aligning the event context segment set with the context representation according to the associated context marker, aligning the tone modification segment set and the social reinforcement segment set with the social signal representation and the user style representation according to the source position, and writing segment type marker, position order marker and associated segment marker for each aligned segment; Based on the input set of ironic branches, literal sentiment modeling is performed to generate intermediate records of literal sentiment. In the literal emotion modeling, the position of the emotion word is first located as the core position for each explicit emotion segment. If a negative word appears consecutively before the core position, the emotion direction of the explicit emotion segment is reversed once. If a degree word appears consecutively before and after the core position, the emotion intensity of the explicit emotion segment is increased by one level. If an exclamation mark or repeated punctuation appears consecutively after the core position, the emotion intensity of the explicit emotion segment is increased by another level. If there are transition words in the same explicit emotion segment, the core position is re-determined based on the emotion word position after the transition word. For each explicit emotion segment, the emotion direction mark, emotion intensity mark, and corresponding segment position are output to obtain the intermediate record of literal emotion. Latent context modeling is performed based on the input set of irony branches to generate intermediate records of the latent context; Implicit context modeling locates the event result description position, time delay description position, quantity anomaly description position, service behavior description position, and functional status description position for each event context fragment. When an event context fragment hits any of the time delay description position, quantity anomaly description position, service behavior description position, or functional status description position, the event context fragment is marked as a negative context. When an event context fragment only hits the positive result description position and does not hit any of the aforementioned positions, the event context fragment is marked as a positive context. When the same event context fragment hits both the positive result description position and at least one of the aforementioned positions, the context fragment associated with the direct response is used. If the context fragment associated with the direct response contains a negative word position, a rhetorical question mark position, or a repeated punctuation position, the event context fragment is marked as a negative context. For each event context fragment, the context polarity label, context severity label, and corresponding fragment position are output to obtain the implicit context intermediate record. Conflict graphs are constructed based on literal emotion intermediate records, implicit context intermediate records, tone modification fragment sets, and social reinforcement fragment sets; The conflict graph uses explicit emotion fragments, event context fragments, tone modifier fragments, and social reinforcement fragments as graph nodes. Fragment position adjacency edges are established when fragments are adjacent, semantic transition edges are established when there are transition words, contextual response edges are established when there are direct response associations, social reinforcement edges are established when there are emoji type markers or repeated punctuation type markers, and style matching edges are established when there are tone expression forms or evaluation directions that match the user's style. The emotion direction markers and emotion intensity markers in the literal emotion intermediate records are written into the explicit emotion fragment nodes, and the situation polarity markers and situation severity markers in the implicit situation intermediate records are written into the event context fragment nodes. Conflict graph reasoning is performed based on the conflict graph, and conflict reasoning records are generated. Based on the aggregation of literal sentiment intermediate records and conflict reasoning records, irony probability results, conflict intensity results, and literal sentiment results are generated. The results aggregation includes marking the irony probability result as high when there are at least two opposite direction markers; marking the irony probability result as medium when there is one opposite direction marker and at least one of the number of reinforcement conflict markers and style contrast markers is not zero; marking the irony probability result as low when there are zero opposite direction markers, or when there is one opposite direction marker and both the number of reinforcement conflict markers and style contrast markers are zero; marking the conflict intensity result as strong when at least two of the number of opposite direction markers, reinforcement conflict markers, and style contrast markers are not zero; and marking the conflict intensity result as weak when only one of the three is not zero. The emotional direction marker, emotional intensity marker, and corresponding segment position corresponding to the dominant emotional segment with the highest number of opposite direction markers in the literal emotion intermediate record are encapsulated into the literal emotion result.

[0026] In this embodiment, for each explicit emotional fragment node, the conflict graph reasoning searches for event context fragment nodes connected to it via fragment position adjacency edges, semantic transition edges, or contextual response edges. When the emotional direction label of the explicit emotional fragment node is opposite to the context polarity label of the event context fragment node, an opposite direction label is written. When the emotional intensity label of the explicit emotional fragment node is higher than the context severity label of the event context fragment node and there is a semantic transition edge or social reinforcement edge between them, a reinforced conflict label is written. When there are both contextual response edges and style matching edges between the explicit emotional fragment node and the event context fragment node, and the evaluation direction corresponding to the user style representation is inconsistent with the emotional direction label of the current explicit emotional fragment node, a style contrast label is written. The number of opposite direction labels, reinforced conflict labels, and style contrast labels are counted for each explicit emotional fragment node to obtain the conflict reasoning record.

[0027] In this embodiment, the generation of the actual emotional outcome and the emotional cause outcome specifically includes: The fragment split set and the shared semantic encoding result are input into the real emotion branch to perform inversion alignment, generating the real emotion branch input set; The inversion alignment includes aligning the set of explicit emotion fragments with the current text representation according to the fragment start position and fragment end position, aligning the set of event context fragments with the context representation according to the associated context markers, and aligning the set of tone modification fragments and social reinforcement fragments with the social signal representation and user style representation according to the source position. For each aligned fragment, a fragment type marker, position order marker, associated context marker, and source position marker are written to obtain the set of true emotion branch inputs. Based on the input set of real emotion branches, the event constraint object is located and a set of event constraint objects is generated. The event constraint object localization process includes using event context fragments and explicit emotion fragments as candidate associated fragments. When an event context fragment and an explicit emotion fragment in the same current text have a fragment position adjacency relationship, a semantic transition relationship, or a direct response relationship, a candidate event constraint object is established. When the event context fragment in the candidate event constraint object hits at least one of the following positions: time delay description position, quantity anomaly description position, service behavior description position, functional status description position, or positive result description position, the candidate event constraint object is written into the event subject description and the event result description to obtain the event constraint object set. Based on the set of event constraint objects, perform exception type determination and exception severity determination, and generate event constraint records; The exception type determination includes writing a time-delay exception type when the event constraint object hits the time delay description position, writing a quantity exception type when the event constraint object hits the quantity exception description position, writing a service failure exception type when the event constraint object hits the service behavior description position, writing a function failure exception type when the event constraint object hits the function status description position, and writing a result reversal exception type when the event constraint object hits at least one of the above exception description positions. The severity determination of anomalies includes writing a mild anomaly marker when the event constraint object hits only one type of anomaly and does not contain rhetorical question markers, repeated punctuation marks, or emoticon markers; writing a moderate anomaly marker when the event constraint object hits only one type of anomaly and contains at least one of the rhetorical question markers, repeated punctuation marks, or emoticon markers; and writing a severe anomaly marker when the event constraint object hits two or more types of anomalies, or hits one type of anomaly and contains at least two of the rhetorical question markers, repeated punctuation marks, or emoticon markers, thus obtaining the event constraint record. Based on the event constraint record, perform event constraint inversion to generate a true emotion inversion record; In the event constraint inversion, when the anomaly type is a time-mismatch anomaly, quantity anomaly, service misconduct anomaly, or function failure anomaly, and the anomaly severity is marked as mild anomaly, a negative true evaluation direction is written. When the anomaly type is any of the above anomaly types, and the anomaly severity is marked as moderate or severe anomaly, a strong negative true evaluation direction is written. When the anomaly type is a result reversal anomaly, a true evaluation direction opposite to the emotional direction corresponding to the explicit emotional fragment is written. When the evaluation tendency record corresponding to the user style representation is consistent with the currently written true evaluation direction, the current true evaluation direction is retained. When the evaluation tendency record corresponding to the user style representation is inconsistent with the currently written true evaluation direction, the current true evaluation direction is reduced by one level to obtain the true emotional inversion record. Based on the real emotion inversion record, the real emotion prototype screening, prototype matching and cause location are performed to generate real emotion intermediate record and emotion cause record; The real emotion prototype screening process involves comparing the event subject description, event result description, real evaluation direction, and abnormal severity in the real emotion inversion record with each real emotion prototype in the pre-set real emotion prototype library. First, real emotion prototypes with inconsistent event result descriptions are filtered out, then real emotion prototypes with inconsistent real evaluation directions are filtered out, and real emotion prototypes with consistent event subject descriptions and abnormal severity are retained in the remaining real emotion prototypes to obtain a real emotion prototype candidate set. The prototype matching process involves selecting real emotion prototypes from the real emotion prototype candidate set that are completely consistent with the event subject description, event result description and real evaluation direction in the real emotion inversion record as matching prototypes, and writing the corresponding real emotion category into the real emotion intermediate record. The cause localization includes writing the corresponding emotional cause record based on the abnormal type in the event constraint record; Based on the encapsulation of the execution results of the real emotion intermediate records and emotion cause records, the real emotion results and emotion cause results are generated.

[0028] In this embodiment, the generation of the final emotion classification result and evidence package result specifically includes: The results of irony probability, conflict intensity, literal sentiment, true sentiment, and sentiment cause are input into the consistency adjudication layer to align the execution results and collect samples, generating an adjudication input set. The result alignment and sample aggregation include performing one-to-one correspondence merging of the irony probability results, conflict intensity results, literal sentiment results, true sentiment results, and sentiment cause results according to the primary key field of the samples; performing position alignment between the corresponding fragment positions in the literal sentiment results and the corresponding fragment positions in the true sentiment results; performing context alignment between the cause-corresponding fragment positions in the sentiment cause results and the associated context markers; and writing a decision sequence number marker for each group of merged results. Based on the set of adjudication inputs, an adjudication pattern determination is performed, and an adjudication pattern result is generated. The adjudication mode determination includes marking the current sample as the true emotion priority adjudication mode when the irony probability result is high, marking the current sample as the literal emotion priority adjudication mode when the irony probability result is low and the conflict intensity result is weak, and marking the current sample as the consistency review adjudication mode when the irony probability result is medium, or when the irony probability result is low and the conflict intensity result is strong, thus obtaining the adjudication mode result. A consistency determination is performed based on the set of adjudication inputs and the adjudication pattern results, and a consistency determination result is generated. Based on the adjudication pattern results and consistency judgment results, the adjudication output is executed to generate the final sentiment classification result; The adjudication output includes the following: When the adjudication mode result is the true emotion priority adjudication mode, the true emotion category in the true emotion result is written into the final emotion classification result; when the adjudication mode result is the literal emotion priority adjudication mode, if the emotion direction mark in the literal emotion result is positive, the positive emotion is written into the final emotion classification result; if the emotion direction mark in the literal emotion result is negative, the negative emotion is written into the final emotion classification result; when the adjudication mode result is the consistency review adjudication mode, if there is a reverse causal conflict mark in the consistency judgment result, the true emotion category in the true emotion result is written into the final emotion classification result; if there is a same-direction consistency mark in the consistency judgment result, the true emotion category in the true emotion result is written into the final emotion classification result; if there is a reverse causal conflict mark in the consistency judgment result, the emotion direction corresponding to the literal emotion result is written into the final emotion classification result, and encapsulated with the adjudication source mark and the corresponding segment position. Based on the adjudication input set, adjudication pattern results, consistency judgment results, and final emotional intermediate records, the evidence field is encapsulated to generate an evidence package result. The evidence field encapsulation includes writing the sample primary key field, adjudication pattern result, irony probability result, conflict intensity result, emotion direction marker, emotion intensity marker and corresponding fragment position in the literal emotion result, true emotion category, matching prototype identifier and corresponding fragment position in the true emotion result, cause type, cause corresponding fragment position and associated context marker in the emotion cause result, consistency judgment result, and final emotion intermediate record into the same evidence field set, and encapsulating it with the adjudication sequence number marker and sample primary key field into an evidence package result.

[0029] In this embodiment, consistency determination includes reading the emotion direction marker in the literal emotion result, reading the true evaluation direction corresponding to the matching prototype identifier in the true emotion result as the true emotion direction marker, and writing a reverse causal conflict marker when the corresponding segment position in the literal emotion result overlaps with the corresponding segment position in the emotion cause result or has a direct response association through the associated context marker, and the literal emotion direction marker is opposite to the true emotion direction marker. When the literal emotion direction marker is the same as the true emotion direction marker and the corresponding segment position overlaps, a same-direction consistency marker is written. When the literal emotion direction marker is opposite to the true emotion direction marker, there is no position overlap, and there is no direct response association, a reverse causal conflict marker is written.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a sentiment analysis scenario of product and service review corpus on a social media platform. The application location was selected in the data processing environment of an e-commerce operation center in East China, and the application time covered peak user review periods after several consecutive promotional cycles. In this scenario, a large number of reviews simultaneously contained superficially positive sentiments, contextual complaints, emoji reinforcement, and reply chain transitions. For example, users might use words of praise in the main comment but express concerns about logistics delays, slow customer service response, or functional malfunctions in subsequent replies. Traditional sentiment classification methods easily misclassify such corpus as positive sentiments, making it difficult to identify the true source of dissatisfaction and providing traceable judgment criteria.

[0031] In practical applications, social network corpora are first obtained from the platform's comment section, follow-up comment section, and reply section. The text content, contextual relationships, emoticons, hashtags, and users' historical expression habits are aligned and organized to form an aligned input set. Then, through shared semantic encoding, a unified representation of the current text, context, social signals, and user style is obtained. In the irony triggering decomposition layer, explicit emotional fragments, event context fragments, tone modification fragments, and social reinforcement fragments are extracted. Subsequently, in the irony branch, conflict analysis is performed on literal emotion and implicit context to obtain irony probability results, conflict intensity results, and literal emotion results. In the true emotion branch, event constraint inversion is performed based on the event anomaly type and severity, and true emotion prototype matching is combined to output true emotion results and emotion cause results. Finally, in the consistency adjudication layer, all results are synthesized to complete the adjudication, and an evidence package result is generated simultaneously for operations personnel to review the judgment process.

[0032] In this application, comment text, reply text, emoji sequences, event anomaly descriptions, adjudication records, and manual review records are collected as verification data. After continuous operation, it can be seen that this invention maintains a more stable classification direction for seemingly positive but actually negative ironic comments, complaint comments with rhetorical questions and repeated punctuation, and reply chains where the true emotion can only be identified through context. Simultaneously, the output results can correspond to specific event context fragments and emotion cause records, allowing operators to directly pinpoint the causes such as logistics delays, quantity anomalies, service inadequacies, or functional failures. This demonstrates that this invention can effectively solve the problems of difficult irony identification, difficulty in recovering true emotions, and unclear judgment criteria in social network corpora, possessing good application feasibility and result traceability.

[0033] Table 1. Comparison of Comprehensive Performance of Sentiment Classification in Social Network Corpora

[0034] Overall, the method of this invention achieves the highest levels in accuracy, macro-average precision, macro-average recall, and macro-average F1 score. Specifically, the accuracy reaches 92.48%, representing an improvement of 10.32 percentage points compared to text convolutional classification methods, 8.74 percentage points compared to bidirectional cyclic classification methods, 5.57 percentage points compared to pre-trained semantic classification methods, 4.41 percentage points compared to context-enhanced sentiment classification methods, and 3.14 percentage points compared to methods that concatenate irony detection and sentiment classification. This indicates that in social network corpus scenarios, relying solely on ordinary text features or general context enhancement, while improving sentiment classification performance, still has significant recognition limitations for irony, context reversal, and complex expressions with social reinforcement signals. In contrast, this invention, through layered processing of aligned input sets, shared semantic encoding results, and fragment splitting sets, transforms the classification process from overall text judgment to structured semantic analysis, thus resulting in more stable comprehensive classification performance.

[0035] In terms of the accuracy rates of irony sample recognition and true emotion recovery, the method of this invention achieves 88.93% and 89.71% respectively, both significantly higher than the comparative methods. In particular, compared with the method that concatenates irony detection and emotion classification, the accuracy rate of irony sample recognition is improved by 6.26 percentage points, and the accuracy rate of true emotion recovery is improved by 6.13 percentage points. This indicates that simply concatenating irony detection and emotion classification, while mitigating some misjudgments, still tends to remain within the two-stage processing logic of literal emotion and irony label, making it difficult to effectively recover the true evaluation direction. This invention outputs irony probability results, conflict intensity results, and literal emotion results in the irony branch, and performs event constraint inversion based on the event constraint object set, anomaly type, anomaly severity, and true emotion prototype candidate set in the true emotion branch. This ensures that the solution of true emotion is based on the facts of the event context, rather than remaining on surface expression or single reconstruction results. Therefore, it has a stronger recovery capability for corpora that appear positive but are actually negative, appear derogatory but are actually positive, and rely on context to reconstruct true emotions.

[0036] In terms of the accuracy of emotion cause matching and the traceability consistency rate of evidence packages, the method of this invention achieves 84.26% and 87.35% respectively, which better reflects the structural advantages of this invention. Compared with pre-trained semantic classification methods, the accuracy of emotion cause matching is improved by 13.00 percentage points, and the traceability consistency rate of evidence packages is improved by 13.51 percentage points; compared with context-enhanced emotion classification methods, the accuracy of emotion cause matching is improved by 9.43 percentage points, and the traceability consistency rate of evidence packages is improved by 9.89 percentage points. This result shows that even if ordinary classification methods can provide a good judgment of emotion direction, they often cannot explain the event cause, triggering fragment, and adjudication basis corresponding to the judgment, resulting in insufficient interpretability of the results. This invention generates emotion cause records in the real emotion branch and uniformly writes the irony probability result, conflict intensity result, literal emotion result, real emotion result, emotion cause result, and consistency judgment result into the evidence field set in the consistency adjudication layer. Therefore, it not only improves the accuracy of emotion result judgment but also significantly enhances the correspondence and traceability consistency between the results and the original corpus.

[0037] Further analysis of the data reveals that while the pre-trained semantic classification method achieves a high level of overall classification performance, it still lags behind this invention in terms of accuracy in identifying irony samples and accuracy in recovering true emotions. This indicates that while relying solely on high-quality semantic representation can capture general emotional features, it is insufficient to resolve the conflict between literal emotions and implicit context. The context-enhanced emotion classification method further improves upon the pre-trained semantic classification method, proving that contextual information is indeed effective for social network corpora. However, its improvement is mainly concentrated on overall classification performance, with limited increases in the accuracy of emotion cause matching and the traceability consistency rate of evidence packages. This suggests that its context utilization remains at the level of assisting understanding, without forming a complete chain of fragment-level triggering, event-level constraints, and adjudication-level encapsulation. The irony detection and emotion classification concatenation method has achieved an accuracy of 82.67% in identifying irony samples, but its accuracy in recovering true emotions and matching emotion causes is still significantly lower than this invention. This is because the concatenation structure typically uses the irony recognition result as a single intermediate label, lacking joint constraints on event context fragments, anomaly types, and anomaly severity, making it difficult to stably recover the true evaluation direction.

[0038] In summary, this invention elevates the emotion classification of social network corpora from a simple text polarity judgment to a structured analysis process that is interpretable, traceable, and recoverable of real emotions by constructing a continuous processing link between aligned input sets, shared semantic encoding results, fragment splitting sets, irony probability results, conflict intensity results, literal emotion results, true emotion results, emotion cause results, and evidence package results. This not only more accurately identifies the true evaluation direction in irony scenarios but also simultaneously locates the emotion cause and forms a stable basis for adjudication. Therefore, it exhibits more prominent comprehensive advantages in terms of classification accuracy, true emotion recovery ability, and result traceability in complex social contexts.

[0039] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for sentiment classification of social network corpora, characterized in that, Includes the following steps: Acquire social network corpora, perform text alignment, context association, social additional signal collection, and user style fragment organization to generate an aligned input set; The aligned input set is fed into the shared semantic encoding layer of the irony-real emotion dual-branch consistent network to generate shared semantic encoding results, including current text representation, context representation, social signal representation, and user style representation; The shared semantic encoding results are input into the irony trigger splitting layer to generate a set of fragment splits, including a set of explicit emotion fragments, a set of event context fragments, a set of tone modification fragments, and a set of social reinforcement fragments; The fragment splitting set and the shared semantic encoding result are input into the irony branch, and literal emotion modeling, implicit context modeling, conflict graph construction and conflict graph reasoning are performed to generate irony probability results, conflict intensity results and literal emotion results. The fragment splitting set and the shared semantic encoding result are input into the real emotion branch, and event constraint inversion and real emotion prototype matching are performed to generate real emotion results and emotion cause results. Based on the results of irony probability, conflict intensity, literal emotion, true emotion, and emotion cause, a decision is made at the consistency decision layer to generate the final emotion classification result and evidence package result.

2. The deep learning-based social network corpus sentiment classification method according to claim 1, characterized in that, The generation of the alignment input set specifically includes: Acquire social network corpora and merge them according to the publisher identifier, publication timestamp, conversation identifier, and corpus source identifier to generate the original set of social network corpora; Perform text normalization and text alignment on the original set of social network corpus to generate the current set of text fragments; Context tracing and adjacency association are performed on the original set of social network corpus to generate a set of conversation contexts; Social additional signals are aggregated based on non-textual information in the original set of social network corpus, and a set of social additional signals is generated. Perform historical backtracking, tone pattern extraction, facial expression usage pattern extraction, and evaluation tendency extraction on the original set of social network corpus to generate a set of user style fragments, tone pattern records, facial expression usage pattern records, and evaluation tendency records; The current text fragment set, the conversation context set, the social additional signal set, and the user style fragment set are aligned and encapsulated according to a unified sample primary key to generate an aligned input set.

3. The deep learning-based social network corpus sentiment classification method according to claim 1, characterized in that, The generation of the shared semantic encoding result specifically includes: The aligned input set is input into the shared semantic encoding layer of the irony-true emotion dual-branch consistent network, parsed, and a field alignment index is built to generate the shared semantic encoding input set. Perform word-granular segmentation, fragment boundary annotation, and position writing on the current text field in the shared semantic encoded input set to generate the current text encoded input sequence; Perform hierarchical expansion, sequential concatenation, and context role writing on the session context fields in the shared semantic encoded input set to generate a context encoded input sequence; Perform type mapping and position backfilling on the social additional signal fields in the shared semantic encoded input set to generate a social signal encoded input sequence; Based on tone pattern records, facial expression usage pattern records, and evaluation tendency records, historical fragment sorting, style fragment filtering, and aggregate coding preparation are performed on the user style fragment field in the shared semantic coding input set to generate a user style coding input sequence. Hierarchical encoding and cross-source fusion are performed on the current text-encoded input sequence, context-encoded input sequence, social signal-encoded input sequence, and user style-encoded input sequence to generate a shared representation sequence; Based on the shared representation sequence, source aggregation and association encapsulation are performed according to the primary key field of the sample to generate shared semantic encoding results.

4. The deep learning-based social network corpus sentiment classification method according to claim 3, characterized in that, The irony-real emotion dual-branch consistency network includes a shared semantic encoding layer, an irony triggering splitting layer, an irony branch, a real emotion branch, and a consistency adjudication layer.

5. The deep learning-based social network corpus sentiment classification method according to claim 1, characterized in that, The generation of the fragment split set specifically includes: The shared semantic encoding result is input into the irony trigger splitting layer, and a trigger alignment sequence is established according to the current text representation, context representation, social signal representation and user style representation to generate a trigger input set; Based on the trigger input set, trigger candidate localization is performed, and a trigger candidate set is generated; Based on the trigger candidate set, perform dominant emotion segment determination and boundary merging to generate a dominant emotion segment set; Based on the trigger candidate set, the event context fragment determination and boundary extension are performed to generate the event context fragment set; Based on the trigger candidate set, the tone modification fragment determination and social reinforcement fragment determination are performed, and a tone modification fragment set and a social reinforcement fragment set are generated; Based on the sets of explicit emotion fragments, event context fragments, tone modification fragments, and social reinforcement fragments, the fragment splitting conflict resolution and result encapsulation are performed to generate a fragment splitting set.

6. The deep learning-based social network corpus sentiment classification method according to claim 1, characterized in that, The generation of the irony probability result, conflict intensity result, and literal sentiment result specifically includes: The fragment splitting set and the shared semantic encoding result are input into the irony branch to perform fragment alignment and role mapping, generating the irony branch input set; Based on the input set of ironic branches, literal sentiment modeling is performed to generate intermediate records of literal sentiment. Latent context modeling is performed based on the input set of irony branches to generate intermediate records of the latent context; Conflict graphs are constructed based on literal emotion intermediate records, implicit context intermediate records, tone modification fragment sets, and social reinforcement fragment sets; Conflict graph reasoning is performed based on the conflict graph, and conflict reasoning records are generated. Based on the aggregation of literal sentiment intermediate records and conflict inference records, the results of irony probability, conflict intensity, and literal sentiment are generated.

7. The deep learning-based social network corpus sentiment classification method according to claim 6, characterized in that, For each explicit emotion fragment node, the conflict graph reasoning process searches for event context fragment nodes connected to it via fragment position adjacency edges, semantic transition edges, or contextual response edges. When the emotion direction label of the explicit emotion fragment node is opposite to the context polarity label of the event context fragment node, an opposite direction label is written. When the emotion intensity label of the explicit emotion fragment node is higher than the context severity label of the event context fragment node and there is a semantic transition edge or social reinforcement edge between them, a reinforced conflict label is written. When there are both contextual response edges and style matching edges between the explicit emotion fragment node and the event context fragment node, and the evaluation direction corresponding to the user style representation is inconsistent with the emotion direction label of the current explicit emotion fragment node, a style contrast label is written. The number of opposite direction labels, reinforced conflict labels, and style contrast labels are counted for each explicit emotion fragment node to obtain the conflict reasoning record.

8. The deep learning-based social network corpus sentiment classification method according to claim 1, characterized in that, The generation of the actual emotional outcome and the emotional cause outcome specifically includes: The fragment split set and the shared semantic encoding result are input into the real emotion branch to perform inversion alignment, generating the real emotion branch input set; Based on the input set of real emotion branches, the event constraint object is located and a set of event constraint objects is generated. Based on the set of event constraint objects, perform exception type determination and exception severity determination, and generate event constraint records; Based on the event constraint record, perform event constraint inversion to generate a true emotion inversion record; Based on the real emotion inversion record, the real emotion prototype screening, prototype matching and cause location are performed to generate real emotion intermediate record and emotion cause record; Based on the encapsulation of the execution results of the real emotion intermediate records and emotion cause records, the real emotion results and emotion cause results are generated.

9. The deep learning-based social network corpus sentiment classification method according to claim 1, characterized in that, The generation of the final emotion classification result and evidence package result specifically includes: The results of irony probability, conflict intensity, literal sentiment, true sentiment, and sentiment cause are input into the consistency adjudication layer to align the execution results and collect samples, generating an adjudication input set. Based on the set of adjudication inputs, an adjudication pattern determination is performed, and an adjudication pattern result is generated. A consistency determination is performed based on the set of adjudication inputs and the adjudication pattern results, and a consistency determination result is generated. Based on the adjudication pattern results and consistency judgment results, the adjudication output is executed to generate the final sentiment classification result; Based on the adjudication input set, adjudication mode result, consistency judgment result, and final emotion intermediate record execution evidence field encapsulation, an evidence package result is generated.

10. The deep learning-based social network corpus sentiment classification method according to claim 9, characterized in that, The consistency determination includes reading the emotion direction marker in the literal emotion result, reading the matching prototype identifier in the real emotion result and the corresponding real evaluation direction as the real emotion direction marker. When the corresponding segment position in the literal emotion result overlaps with the corresponding segment position in the emotion cause result or has a direct response association through the associated context marker, and the literal emotion direction marker is opposite to the real emotion direction marker, a reverse causal conflict marker is written. When the literal emotion direction marker is the same as the real emotion direction marker and the corresponding segment position overlaps, a same-direction consistency marker is written. When the literal emotion direction marker is opposite to the real emotion direction marker, there is no position overlap, and there is no direct response association, a reverse causal conflict marker is written.

Images