Event clue knowledge enhancement generation method and system based on deep learning

Abstract

The application discloses a kind of event clue knowledge enhancement generation method and system based on deep learning, comprising: obtaining detective event text corpus, obtain preprocessed detective event text corpus;Event recognition and clue extraction are carried out to the preprocessed detective event text corpus, form detective event five tuple set and total clue fragment set;Detective event clue knowledge graph is constructed based on detective event five tuple set and total clue fragment set, and detective event clue knowledge graph is generated;Output complete detective event clue subgraph;Form multi-granularity contrast sample set;Get the clue vector representation set optimized after hierarchical contrast learning;Get final event clue knowledge enhancement text.The application achieves high-quality, full-link, traceable event knowledge enhancement text generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of clue generation, in particular to an event clue knowledge enhancement generation method and system based on deep learning. BACKGROUND

[0002] With the continuous development of artificial intelligence and natural language processing technology, event clue automatic generation and knowledge enhancement tasks have attracted widespread attention in the field of investigation and decryption. Existing event clue generation technologies mostly rely on local context modeling or static knowledge bases. For complex events involving multiple documents, long-chain time sequences and multi-causal interweaving, they can only capture part of the event nodes, resulting in fragmented timelines and incomplete causal chains, making it difficult to support event restoration and reasoning from a global perspective.

[0003] Traditional knowledge enhancement text generation methods mostly use single-layer contrast learning or simple clue splicing. When dealing with the alignment of event clue representation and generated text expression, there are often problems such as insufficient fact injection, semantic expression misplacement and logical jumping. Existing technologies lack multi-layer structural expression and dynamic association mechanisms when faced with deep event clue mining across documents and paragraphs. SUMMARY

[0004] One object of the present application is to provide an event clue knowledge enhancement generation method and system based on deep learning. The present application achieves high-quality, full-link, traceable event knowledge enhancement text generation.

[0005] An event clue knowledge enhancement generation method based on deep learning according to an embodiment of the present application includes the following steps:

[0006] An investigation event text corpus is obtained, and the investigation event text corpus is preprocessed by sentence segmentation, word segmentation and noise cleaning to obtain a preprocessed investigation event text corpus;

[0007] Event recognition and clue extraction are performed on the preprocessed investigation event text corpus. Each paragraph is extracted as an investigation event five-tuple, and a clue segment associated with the investigation event five-tuple is simultaneously extracted to form an investigation event five-tuple set and a total clue segment set;

[0008] An investigation event clue knowledge graph is constructed based on the investigation event five-tuple set and the total clue segment set. The investigation event five-tuples are aligned and merged, and the source text identifier and confidence information of each graph node are labeled to generate an investigation event clue knowledge graph;

[0009] An improved multi-hop reasoning module is called on the investigation event clue knowledge graph to output a complete investigation event clue subgraph;

[0010] Based on the complete detective event clue subgraph, word-level, sentence-level, section-level and event-level contrast samples are constructed to form a multi-granularity contrast sample set;

[0011] The multi-granularity contrast sample set is input into a hierarchical contrast learning module, and representation alignment training is performed in the word-level semantic space, the sentence-level semantic space, the section-level semantic space and the event-level semantic space to obtain a clue vector representation set optimized by hierarchical contrast learning;

[0012] Based on the clue vector representation set and the context semantic vector of the preprocessed detective event text corpus, a coherent detective event narrative text is generated, and slot integrity detection and confidence evaluation are performed to obtain a final event clue knowledge enhanced text.

[0013] Optionally, the forming of the detective event five-tuple set and the total clue fragment set comprises:

[0014] Each paragraph in the preprocessed detective event text corpus is extracted in turn to obtain a paragraph set;

[0015] For each paragraph in the paragraph set, a sequence labeling model is used to recognize the subject element, object element, behavior element, time element and location element in the paragraph, and a detective event five-tuple is formed;

[0016] The detective event five-tuples extracted from all paragraphs are sequentially organized to form a detective event five-tuple set;

[0017] For each paragraph, the detective event five-tuple extraction is performed at the same time, and a clue fragment set directly related to the current five-tuple is recognized;

[0018] The clue fragment sets recognized in all paragraphs are sequentially combined to form a total clue fragment set, and the detective event five-tuple set and the corresponding total clue fragment set are output.

[0019] Optionally, the generating of the detective event clue knowledge graph comprises:

[0020] All detective event five-tuples of the detective event five-tuple set are aligned and merged according to the subject consistency principle, the time proximity principle and the semantic similarity principle, and all detective event five-tuples satisfying the merging conditions are grouped into several detective event five-tuple merging groups;

[0021] For each detective event five-tuple merging group, a detective event clue knowledge graph node is constructed based on the contained detective event five-tuple elements;

[0022] For each of the plurality of clue segments in the clue segment set, each clue segment is connected to one or more nodes in the node set by an edge, and if the clue segment complements, explains or supports the corresponding node in the text semantics, a directed edge is established in the detective event clue knowledge graph;

[0023] For each of the nodes in the node set and each of the directed edges in the edge set, a text source identifier is labeled, and a confidence level is configured for each node and each directed edge;

[0024] The detective event clue knowledge graph is outputted, which is composed of the node set, the edge set, the text source identifier and the confidence level.

[0025] Optionally, the subject consistency principle is that if the subject element in the detective event five tuple is equal, it is considered to be consistent in the subject dimension and is divided into the same detective event five tuple merging group;

[0026] The time proximity principle is that if the difference between the time elements in the detective event five tuple is less than or equal to the time proximity threshold value and the dimension of the time proximity threshold value is consistent with the time element, the two are adjacent in the time dimension and are divided into the same detective event five tuple merging group;

[0027] The semantic similarity principle is that the detective event five tuple has a semantic vector representation, and when the cosine similarity between the semantic vectors of any two detective event five tuples is greater than or equal to the semantic similarity threshold value and the semantic similarity threshold value is in the range of zero to one, the two are related in the semantic dimension and are divided into the same detective event five tuple merging group.

[0028] Optionally, the improved multi-hop reasoning module comprises:

[0029] A multi-hop reasoning strategy network based on reinforcement learning is constructed on the detective event clue knowledge graph, and the state space of the multi-hop reasoning strategy network is used to describe the detective event reasoning path representation composed of the current node and the historical node;

[0030] The deep path memory network based on the fusion attention mechanism aggregates the features of the nodes and their semantic information in the historical reasoning path to obtain the context semantic vector on each reasoning path;

[0031] The reward function of multi-constraint fusion is obtained by weighted summation of the semantic similarity score between event nodes, the event time sequence continuity score and the context logical consistency score between event nodes;

[0032] The multi-hop reasoning strategy network calculates, in each step of reasoning, a reward function of multi-constraint fusion for each candidate node in the action space of the current node, simultaneously evaluates the semantic relevance, time continuity and context logical consistency of the current node and the candidate node, selects the optimal action according to the numerical value of the reward function, and recursively jumps to search the most valuable nodes and edges in the detective event clue knowledge graph;

[0033] The multi-hop reasoning strategy network recursively updates the reasoning path based on the reward function, and uses the reward function to guide the reasoning path to dynamically explore and complete the potential implicit event nodes and bridge edges in the detective event clue knowledge graph, and generate a detective event clue subgraph containing reasoning incremental results.

[0034] Optionally, the forming of the multi-granularity contrast sample set comprises:

[0035] The corresponding text content of each node and edge in the detective event clue subgraph is extracted and hierarchically labeled, and the hierarchical labeling comprises four levels of word-level labeling, sentence-level labeling, paragraph-level labeling and event-level labeling;

[0036] The word-level labeling, sentence-level labeling, paragraph-level labeling and event-level labeling are respectively one-to-one corresponding to the nodes or edges in the detective event clue subgraph, and a multi-granularity labeling mapping set is formed;

[0037] For the labeling mapping of each level of the multi-granularity labeling mapping set, a positive sample pair and a negative sample pair are constructed;

[0038] All positive sample pairs and negative sample pairs under each granularity are combined to form a multi-granularity contrast sample set.

[0039] Optionally, the improved hierarchical contrast learning module comprises:

[0040] The multi-granularity contrast sample set is obtained, and the multi-granularity contrast sample set is input into the improved hierarchical contrast learning module;

[0041] In the word-level semantic space, a word-level contrast loss function is calculated, and the word-level contrast loss function is used to enhance the consistency of the word representations of different nodes related to the same detective clue;

[0042] In the sentence-level semantic space, a sentence-level contrast loss function is calculated, and the sentence-level contrast loss function is used to improve the semantic alignment degree between different text sentences in the detective event clue;

[0043] In the paragraph-level semantic space, a paragraph-level contrast loss function is calculated, and the paragraph-level contrast loss function is used to evaluate the logical coherence of the paragraph fragments composed of multiple related sentences in the event context in the detective event clue knowledge;

[0044] In the event-level semantic space, an event-level contrast loss function is calculated, and the event-level contrast loss function is used to overall control the logical consistency of the detective event clue and the completeness of the event description.

[0045] The word-level contrast loss function, the sentence-level contrast loss function, the section-level contrast loss function, and the event-level contrast loss function are weighted and fused to obtain a total-level contrast loss function.

[0046] Based on the total-level contrast loss function, the improved hierarchical contrast learning module is iteratively trained to generate a clue vector representation set optimized by hierarchical contrast learning.

[0047] Optionally, the generated detective event narrative text is subjected to slot completeness detection and confidence evaluation, including:

[0048] In the knowledge injection type text generation model, the clue vector representation set and the context semantic vector are fused through a multi-source gating attention mechanism.

[0049] The fused clue vector representation set and context semantic vector drive the decoder to generate the detective event narrative text.

[0050] The detective event narrative text is subjected to slot completeness detection.

[0051] The generated detective event narrative text is subjected to confidence evaluation, and the confidence of each event clause in the text and the overall text confidence are calculated.

[0052] The quality of the detective event narrative text is evaluated through the slot completeness loss rate and the overall text confidence:

[0053] When the slot completeness loss rate is lower than the preset slot completeness threshold and the overall text confidence is higher than the text confidence threshold, the generated text is identified as meeting the quality requirements; otherwise, the system automatically generates an error feedback signal.

[0054] The error feedback signal is fed back to the multi-hop reasoning module, triggering incremental search and updating of the detective event clue knowledge graph, to obtain an updated detective event clue knowledge graph.

[0055] Based on the updated detective event clue knowledge graph, the text generation process is re-executed iteratively until the slot completeness loss rate is lower than the preset slot completeness threshold and the overall text confidence is higher than the text confidence threshold, and the final event clue knowledge enhanced text meeting the requirements is output.

[0056] An event clue knowledge enhanced generation system based on deep learning, for executing an event clue knowledge enhanced generation method based on deep learning, comprising:

[0057] The text preprocessing and event clue extraction module is used for sentence segmentation, word segmentation, noise filtering, and extraction of event five-tuples and clue fragments on the detective event text corpus;

[0058] The event clue knowledge graph construction module is used for merging the five-tuples and clues according to the principles of subject consistency, time proximity and semantic similarity, and generating a detective event clue knowledge graph;

[0059] The multi-hop reasoning and incremental completion module performs improved multi-hop reasoning driven by reinforcement learning on the detective event clue knowledge graph, and recursively completes implicit nodes and bridge edges;

[0060] The multi-granularity contrast sample generation and hierarchical contrast learning module constructs positive and negative samples at the word, sentence, paragraph and event levels;

[0061] The knowledge injection type text generation and closed loop optimization module generates detective event narrative text by fusing clue vectors and context, automatically detects slot completeness and confidence, and obtains the final event clue knowledge enhanced text.

[0062] The beneficial effects of the present application are:

[0063] (1) The present application introduces an improved multi-hop reasoning strategy network on the detective event clue knowledge graph, combines the deep path memory of the attention mechanism and the reward function of multi-constraint fusion, realizes dynamic fusion and recursive completion of node time sequence, semantics, context logic in multiple dimensions, effectively mines the implicit bridge nodes and causal chains between texts, and significantly improves the completeness of event reasoning and the continuity of event clue network.

[0064] (2) The present application breaks the limitation of traditional hierarchical contrast learning which only optimizes independently at each layer by using a dynamic contrast loss function for cross-layer semantic propagation, realizes collaborative optimization of word-level, sentence-level, paragraph-level and event-level representations, and effectively solves the misalignment and fact omission problems between knowledge clues and text expressions by aggregating fine-grained clue knowledge layer by layer and feeding back coarse-grained semantic guidance layer by layer.

[0065] (3) The present application introduces multi-source gated attention and dynamic confidence detection in the text generation module, which can automatically feed back to the reasoning module and trigger incremental search and graph update when event element missing or expression confidence is insufficient in the generated text, driven by the slot completeness missing rate parameter and the overall text confidence parameter, and finally achieve high-quality, full-link, traceable event knowledge enhanced text generation. BRIEF DESCRIPTION OF DRAWINGS

[0066] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0067] Figure 1 A flowchart of an event clue knowledge enhancement generation method and system based on deep learning is proposed for the present application. DETAILED DESCRIPTION

[0068] Embodiment 1:

[0069] Reference Figure 1 An event clue knowledge enhancement generation method based on deep learning, characterized in that it comprises the following steps:

[0070] An investigator event text corpus is obtained, and the investigator event text corpus is subjected to sentence division, word division and noise cleaning preprocessing to obtain a preprocessed investigator event text corpus;

[0071] Event recognition and clue extraction are performed on the preprocessed investigator event text corpus, the subject element, object element, behavior element, time element and place element in each paragraph are extracted as an investigator event five-tuple, and the clue segment associated with the investigator event five-tuple is simultaneously extracted to form an investigator event five-tuple set and a total clue segment set;

[0072] An investigator event clue knowledge graph is constructed based on the investigator event five-tuple set and the total clue segment set, the investigator event five-tuples are aligned and merged according to the subject consistency principle, time proximity principle and semantic similarity principle, the source text identifier and confidence information of each graph node are labeled, and an investigator event clue knowledge graph is generated;

[0073] An improved multi-hop reasoning module is called on the investigator event clue knowledge graph, and a complete investigator event clue subgraph is output;

[0074] Based on the complete investigator event clue subgraph, word-level, sentence-level, paragraph-level and event-level comparison samples are constructed to form a multi-granularity comparison sample set;

[0075] The multi-granularity comparison sample set is input into a hierarchical contrast learning module, and representation alignment training is performed in the word-level semantic space, sentence-level semantic space, paragraph-level semantic space and event-level semantic space, respectively, to obtain a clue vector representation set optimized by hierarchical contrast learning;

[0076] Based on the clue vector representation set and the context semantic vector of the preprocessed investigator event text corpus, a coherent and consistent investigator event narrative text is generated, slot completeness detection and confidence evaluation are performed on the generated investigator event narrative text, missing event nodes or logically incoherent segments are identified, and the detection results are fed back to the multi-hop reasoning module to trigger incremental search and investigator event clue knowledge graph update until the preset completeness threshold and confidence threshold are met, and a final event clue knowledge enhancement text is obtained.

[0077] In the embodiment, a detective event five tuple set and a total clue fragment set are formed, including:

[0078] Each paragraph in the preprocessed detective event text corpus is extracted in turn to obtain a paragraph set;

[0079] The paragraph set is used for event element and clue fragment extraction, and each paragraph in the paragraph set is identified by a unique index, and the total number of paragraphs is denoted as the number of paragraphs;

[0080] For each paragraph in the paragraph set, a sequence labeling model is used to identify the subject element, object element, action element, time element and location element in the paragraph, and form a detective event five tuple;

[0081] Each detective event five tuple is uniquely determined by the subject element, action element, object element, time element and location element, the subject element is used to indicate the investigation target, the action element is used to indicate the action content, the object element is used to indicate the involved object, the time element is used to indicate the event occurrence time, and the location element is used to indicate the event occurrence position;

[0082] The detective event five tuples extracted from all paragraphs are sequentially organized to form a detective event five tuple set;

[0083] The detective event five tuple set is composed of the detective event five tuples generated in each paragraph.

[0084] For each paragraph, the detective event five tuple extraction is performed at the same time, and a clue fragment set directly related to the current five tuple is identified;

[0085] The clue fragment set is composed of semantic fragments that have causal relationship, parallel relationship, reasoning relationship or context supplement relationship with any element in the detective event five tuple, each clue fragment in the clue fragment set is identified by a unique index, and the number of clue fragment sets in each paragraph is used to represent the richness of the clue information in the paragraph.

[0086] All the clue fragment sets identified in all paragraphs are sequentially combined to form a total clue fragment set;

[0087] The total clue fragment set includes all available clue fragments in the preprocessed detective event text corpus.

[0088] The detective event five tuple set and the corresponding total clue fragment set are output;

[0089] Each five tuple in the detective event five tuple set and each clue fragment in the clue fragment set is labeled with the index of the source paragraph and the specific position index in the text corpus.

[0090] In this embodiment, a detective event clue knowledge graph is generated, including:

[0091] All detective event five-tuples of the detective event five-tuple set are aligned and merged according to the subject consistency principle, the time proximity principle and the semantic similarity principle, and all detective event five-tuples satisfying the merging conditions are grouped into a plurality of detective event five-tuple merging groups.

[0092] The total number of detective event five-tuple merging groups is denoted as the total number of merging groups, and each detective event five-tuple merging group includes a plurality of detective event five-tuples satisfying the rules.

[0093] For each detective event five-tuple merging group, a detective event clue knowledge graph node is constructed based on the included detective event five-tuple elements.

[0094] The set of detective event clue knowledge graph nodes is denoted as the node set, and the total number of detective event five-tuple merging groups represents the total number of the node set. Each node corresponds to one detective event five-tuple merging group.

[0095] For a plurality of clue fragments in the total clue fragment set, each clue fragment is connected to one or more nodes in the node set, and if the clue fragment supplements, explains or supports the corresponding node in the text semantics, a directed edge is established in the detective event clue knowledge graph.

[0096] The directed edge is used to describe the association between the clue and the node, and all directed edges form an edge set.

[0097] Each node in the node set and each directed edge in the edge set is labeled with a text source identifier, and each node and each directed edge is assigned a confidence level.

[0098] The text source identifier is used to indicate the paragraph index and specific text position index of the node or directed edge in the preprocessed detective event text corpus, and the confidence level is used to represent the credibility of the node and directed edge in the detective event clue knowledge graph. The confidence level ranges from zero to one.

[0099] The detective event clue knowledge graph is outputted, which is composed of the node set, the edge set, the text source identifier and the confidence level.

[0100] In this embodiment, the subject consistency principle is that if the subject elements in the detective event five-tuple are equal, it is considered that the subjects are consistent in the subject dimension and are divided into the same detective event five-tuple merging group.

[0101] The time proximity principle is: if the difference between the time elements in the detective event five tuple is less than or equal to the time proximity threshold, and the dimension of the time proximity threshold is consistent with the time element, then the two are adjacent in the time dimension and are divided into the same detective event five tuple merging group;

[0102] The semantic similarity principle is: the detective event five tuple has a semantic vector representation, when the cosine similarity between the semantic vectors of any two detective event five tuples is greater than or equal to the semantic similarity threshold, and the semantic similarity threshold takes a value range of zero to one, then the two are related in the semantic dimension and are divided into the same detective event five tuple merging group.

[0103] In the embodiment, the multi-hop reasoning module is improved, including:

[0104] A multi-hop reasoning strategy network based on reinforcement learning is constructed on the detective event clue knowledge graph;

[0105] The state space of the multi-hop reasoning strategy network is used to describe the detective event reasoning path representation composed of the current node and the historical node, the action space is used to describe the adjacent node set that the current node can reach, and each action is used to indicate the selection of the next detective event clue knowledge graph node.

[0106] The deep path memory network based on the fusion attention mechanism aggregates the features of the nodes and their semantic information in the historical reasoning path, to obtain the reasoning path context semantic vector on each step;

[0107] The reasoning path context semantic vector is used to represent the global logical relationship of the node set on the reasoning path in the context, and provides a path level reference for the multi-hop reasoning strategy network when selecting actions.

[0108] The reward function of multi-constraint fusion is obtained by weighted summation of the semantic similarity score between event nodes, the node event time sequence continuity score and the context logical consistency score between event nodes;

[0109] The semantic similarity score between event nodes is used to measure the relevance of the current node and the candidate node on the semantic vector, the node event time sequence continuity score is used to measure the time continuity between the time elements of the candidate node and the current node, and the context logical consistency score between event nodes is used to describe the global consistency between the reasoning path context semantic vector and the semantic vector of the candidate node.

[0110] The multi-hop reasoning strategy network calculates, in each step of reasoning, a reward function of multi-constraint fusion for each candidate node in the action space of the current node, simultaneously incorporates semantic relevance, time continuity and context logical consistency of the current node and the candidate node into evaluation, selects the optimal action according to the numerical value of the reward function, and recursively jumps to search the most valuable nodes and edges in the detective event clue knowledge graph;

[0111] The multi-hop reasoning strategy network updates the reasoning path based on the reward function recursively;

[0112] The node search process satisfies the node event time sequence constraint and the semantic similarity constraint, that is, the time elements of each node on the reasoning path are in a strictly non-decreasing relationship, and the semantic similarity of any two nodes is higher than a set semantic similarity threshold.

[0113] The multi-hop reasoning strategy network uses the reward function to guide the reasoning path to dynamically explore and complete potential implicit event nodes and bridge edges in the detective event clue knowledge graph, and generate a detective event clue subgraph containing reasoning incremental results;

[0114] The detective event clue subgraph includes full-amount nodes and their connection relationships after reasoning completion.

[0115] In the embodiment, a multi-granularity contrast sample set is formed, including:

[0116] The corresponding text content is extracted for each node and edge in the detective event clue subgraph and hierarchical labeling is performed, including four levels of word-level labeling, sentence-level labeling, paragraph-level labeling and event-level labeling;

[0117] The word-level labeling is used to locate keywords or proper nouns in the text content, the sentence-level labeling is used to identify complete sentences expressing a single fact or clue relationship, the paragraph-level labeling is used to represent a coherent description fragment containing multiple related sentences, and the event-level labeling is used to summarize all context information representing a single detective event.

[0118] The word-level labeling, the sentence-level labeling, the paragraph-level labeling and the event-level labeling are respectively one-to-one corresponding to the nodes or edges in the detective event clue subgraph, forming a multi-granularity labeling mapping set;

[0119] The multi-granularity labeling mapping set is used to realize comprehensive coverage of the detective event clue knowledge graph content at different semantic levels.

[0120] For the labeling mapping of each level of the multi-granularity labeling mapping set, a positive sample pair and a negative sample pair are constructed;

[0121] The positive sample pair is composed of a pair of texts with consistent sources or causal chain relationship, and the negative sample pair is composed of a pair of texts with no direct correlation or obvious semantic separation. The word-level positive sample pair measures the consistency of the expression of the same clue or entity in different nodes, the sentence-level positive sample pair measures the semantic similarity of the same event node under different text descriptions, the paragraph-level positive sample pair measures the contextual coherence of the related event paragraph, the event-level positive sample pair judges the consistency of the expression of the whole event chain, and the negative sample pair corresponds to the contrast relationship with semantic inconsistency or reasoning path inconsistency at each level.

[0122] All positive sample pairs and negative sample pairs at each granularity are combined to form a multi-granularity contrast sample set;

[0123] The multi-granularity contrast sample set includes a word-level contrast sample set, a sentence-level contrast sample set, a paragraph-level contrast sample set, and an event-level contrast sample set, each of which is composed of positive sample pairs and negative sample pairs and is organized in a unified data structure.

[0124] In this embodiment, the improved hierarchical contrast learning module includes:

[0125] The multi-granularity contrast sample set is obtained, and the multi-granularity contrast sample set is input into the improved hierarchical contrast learning module;

[0126] The improved hierarchical contrast learning module is used to construct and optimize the word-level semantic space, the sentence-level semantic space, the paragraph-level semantic space, and the event-level semantic space, to enhance the cross-level alignment capability of the detective event clue knowledge at different granularities, and to unify the fact and logical consistency of the generated detective event text;

[0127] For the detective event clue knowledge enhancement generation task, a dynamic contrast loss function of cross-layer semantic propagation is proposed. The dynamic contrast loss function of cross-layer semantic propagation aggregates fine-grained clue knowledge layer by layer upwards and feeds back coarse-grained semantic guidance layer by layer downwards by fusing the feature representations of adjacent semantic levels, to realize the collaborative optimization of word-level, sentence-level, paragraph-level, and event-level representations;

[0128] In the word-level semantic space, a word-level contrast loss function is calculated;

[0129] The word-level contrast loss function is used to enhance the consistency of the word representation of different nodes involving the same detective clue, and dynamically adjust the weight of the word-level positive and negative sample pair according to the semantic guidance information fed back from the sentence-level semantic space, to optimize the expression accuracy of the word-level information in the whole context in the detective event clue knowledge;

[0130] In the sentence-level semantic space, a sentence-level contrast loss function is calculated;

[0131] The sentence-level contrast loss function is used to improve the semantic alignment between different text sentences in the detective event clue, and to fuse the semantic guidance information from the paragraph-level semantic space. The selection strategy of the positive and negative sample pairs at the sentence level is dynamically corrected according to the context consistency score of the corresponding paragraph level, so as to ensure the logical reasoning coherence of the detective event sentences.

[0132] In the paragraph-level semantic space, a paragraph-level contrast loss function is calculated.

[0133] The paragraph-level contrast loss function is used to evaluate the logical coherence of the paragraph fragment composed of multiple related sentences in the detective event clue knowledge under the event context, and dynamically adjust the weight of the positive and negative sample pairs at the paragraph level according to the high-level event integrity feedback result of the event-level semantic space, so as to strengthen the overall narrative consistency of the detective event clue knowledge at the paragraph level.

[0134] In the event-level semantic space, an event-level contrast loss function is calculated.

[0135] The event-level contrast loss function is used to control the logical consistency and completeness of the event description of the detective event clue as a whole. By fusing the context logical coherence features of the paragraph-level semantic space, the combination mode of the positive and negative sample pairs at the event level is dynamically adjusted in a cross-layer semantic propagation manner, so as to further enhance the high-level semantic alignment capability of the detective event clue knowledge at the event level.

[0136] The word-level contrast loss function, the sentence-level contrast loss function, the paragraph-level contrast loss function and the event-level contrast loss function are weighted and fused to obtain a total-level contrast loss function.

[0137] The dynamic weight in the weighted fusion is adjusted in real time according to the training feedback effect of the current level, so as to strengthen the semantic propagation and cooperation between each granularity, and promote the collaborative optimization of the detective event clue knowledge representation at each level.

[0138] Based on the total-level contrast loss function, the improved hierarchical contrast learning module is iteratively trained to generate a set of clue vector representations optimized by hierarchical contrast learning.

[0139] The set of clue vector representations includes a set of word-level clue vector representations, a set of sentence-level clue vector representations, a set of paragraph-level clue vector representations and a set of event-level clue vector representations. Each set in the set of clue vector representations contains a detective event clue knowledge representation optimized by a cross-layer semantic propagation mechanism, so that the detective event generation text maintains high consistency and coherence in logic and semantics across granularities.

[0140] In this embodiment, the generated detective event narrative text is subjected to slot integrity detection and confidence evaluation, including:

[0141] In the construction of the knowledge injection text generation model, the multi-source gating attention mechanism is used to fuse the set of clue vector representations and the context semantic vector;

[0142] The multi-source gating attention mechanism controls the injection proportion of the set of clue vector representations at different semantic levels in the generation process by introducing gating weights, the gating weights are summed up to equal one at all levels, and the value range of each gating weight is from zero to one.

[0143] The fused set of clue vector representations and the context semantic vector drive the decoder to generate the detective event narrative text;

[0144] The detective event narrative text is used to express the clue structure and logical chain in the detective event clue knowledge graph.

[0145] Slot integrity detection is performed on the detective event narrative text;

[0146] The slot integrity detection identifies whether the subject element, object element, behavior element, time element, and location element key event element are missing in the detective event narrative text through the integrity detection function, the integrity detection function calculates the slot integrity missing rate by taking the difference between the total number of detective event five-tuple sets that should be included and the total number of detective event five-tuple sets actually detected in the generation of the detective event narrative text, taking the total number of detective event five-tuple sets that should be included as the denominator, the slot integrity missing rate is used to measure the proportion of missing elements in the detective event narrative text.

[0147] The confidence of the generated detective event narrative text is evaluated, and the confidence of each event clause in the text and the overall text confidence are calculated;

[0148] The event clause confidence is obtained by comparing the consistency score of each event clause in the detective event narrative text with the semantic vector of the source node, and the overall text confidence is obtained by weightedly averaging all event clause confidences.

[0149] The quality of the detective event narrative text is evaluated by the slot integrity missing rate and the overall text confidence:

[0150] When the slot integrity missing rate is lower than the preset slot integrity threshold and the overall text confidence is higher than the text confidence threshold, the generated text is recognized as meeting the quality standard; otherwise, the system automatically generates an error feedback signal;

[0151] The error feedback signal is returned to the multi-hop reasoning module, triggering incremental search and updating of the detective event clue knowledge graph, obtaining the updated detective event clue knowledge graph;

[0152] The incremental search introduces new jump searches on the basis of reserving original reasoning paths, fills in missing nodes and bridging edges, and obtains an updated detective event clue knowledge graph.

[0153] The generation text process is re-executed based on the updated detective event clue knowledge graph, iteratively executed until the slot completeness missing rate is lower than the preset slot completeness threshold and the overall text confidence is higher than the text confidence threshold, and the final event clue knowledge enhanced text meeting the requirements is output;

[0154] Each statement in the final event clue knowledge enhanced text can be traced back to its corresponding clue vector representation set node or edge, realizing the explainability and coherence of the detective event line.

[0155] An event clue knowledge enhancement generation system based on deep learning, for executing an event clue knowledge enhancement generation method based on deep learning, comprising:

[0156] A text preprocessing and event clue extraction module for sentence segmentation, word segmentation and noise filtering of the detective event text corpus, and extracting event five-tuples and clue fragments;

[0157] An event clue knowledge graph construction module for merging five-tuples and clues according to the principles of subject consistency, time proximity and semantic similarity, generating a detective event clue knowledge graph;

[0158] A multi-hop reasoning and incremental completion module for executing improved multi-hop reasoning driven by reinforcement learning on the detective event clue knowledge graph, recursively completing implicit nodes and bridging edges;

[0159] A multi-granularity contrast sample generation and hierarchical contrast learning module for constructing positive and negative samples at the word, sentence, paragraph and event levels;

[0160] A knowledge injection type text generation and closed loop optimization module for generating detective event narrative text by fusing clue vectors and context, automatically detecting slot completeness and confidence, and obtaining final event clue knowledge enhanced text.

[0161] Embodiment 2:

[0162] A detective investigation office recently received a complex case of entrustment, the goal is to restore the whole process of the lost cultural relics theft case, the case involves multiple clues, multiple event nodes and multiple parties, the clues are distributed in different investigation reports, monitoring logs, scene investigation records, suspect interrogation records, physical evidence detection reports and various social public opinion information. All data are in text form, with various expressions, some information has time gaps and content fragmentation, which needs to be automatically spliced into complete event chains for further analysis and evidence report.

[0163] The detective investigation team imports all recent text evidence related to the theft case into the system, including on-site monitoring summaries, investigation logs, item lists, in-out warehouse records, suspect access tracks, public opinion hot posts, third-party analysis conclusions, and internal information briefings. The original data is about 20,000 pieces, with a text length of about 600,000 words. The system automatically performs text segmentation and noise filtering to obtain about 13,000 preprocessed text segments.

[0164] The system performs event five-tuple and clue fragment extraction on all paragraphs. It automatically identifies the following specific elements in Example 2: the security guard discovers damaged display cases, suspicious figures appear at night, in-out warehouse records change abnormally, fingerprints do not hit the suspect database, and social platform users post that a strange vehicle appeared at the crime scene. It also extracts the following clue fragments: moving shadows appear in monitoring dead angles, in-out records do not match before and after the incident, and there are signs of external force prying on the safe. In total, 9,800 detective event five-tuples and 11,200 clue fragments are formed.

[0165] Based on the above five-tuples and clue fragments, the system automatically constructs a detective event clue knowledge graph according to the rules of subject consistency (Example 2 is the same suspect target or item), time proximity (Example 2 is a continuous night event), and semantic similarity (Example 2 is different descriptions pointing to the same behavior). Each node is labeled with its source and confidence. The knowledge graph automatically merges to form: night abnormal in-out - monitoring abnormal picture - abnormal in-out - safe prying - evidence analysis results, complete event chain.

[0166] In the reasoning phase, the system calls a multi-hop reasoning strategy network driven by reinforcement learning. Starting from the last in-out of cultural relics, it automatically recursively retrieves monitoring dead angles, suspicious vehicle appearance, access control record tampering, witness testimony of strange noises at multiple implicit bridge nodes, and causal edges. The reasoning network prioritizes nodes with high temporal, spatial, semantic, and behavioral correlation based on a reward function that integrates multiple constraints, enabling the reasoning path to link various heterogeneous evidence fragments and complete the fact chain: night vehicle departure - in-out record modification - monitoring data partial loss, which is difficult for humans to trace.

[0167] In the multi-granularity contrast learning phase, the system constructs positive and negative samples at the word, sentence, segment, and event levels. Example 2 is a word-level positive sample for the incident, theft, and disappearance. The sentence-level positive sample is the appearance of suspicious figures near the display case at night. The segment-level positive sample is the high overlap between the monitoring record loss paragraph and the safe prying time. The event-level positive sample is the incident - in-out - disposal - evidence verification closed chain. Negative samples include inconsistent reporting time and monitoring abnormalities, and lack of temporal connection between some evidence. The system dynamically trains contrast loss for cross-layer semantic propagation, aligning vectors at each layer and improving semantic consistency.

[0168] In the text generation link, the system uses the gated multi-source attention to fuse the optimized multi-layer clue vector and the context vector into the generation model, automatically generates a high-integrity narrative text, and example embodiment 2 is:

[0169] According to the analysis of the access control log and the monitoring video on the night of the incident, there is a blurred figure in the west side of the exhibition hall at 23:47, and then the alarm of the showcase in the southwest corner is abnormally triggered. The on-site investigation log shows that there are pry marks on the surface of the safe, the monitoring log shows that the image is temporarily interrupted during this period, the abnormal change of the artifact number is found in the out-of-warehouse record, and it is consistent with the period when the strange vehicle leaves the monitoring area the next morning. The suspect's interrogation record cannot explain the night activity track, and the evidence comparison report points out that the fingerprint is not extracted from the scene but has indirect DNA signal. According to the comprehensive evidence, it is highly related to the night abnormal activity, the change of the out-of-warehouse record, and the missing of the monitoring data.

[0170] The system automatically detects the slot integrity of the above generated text and finds that all the key event five-tuples are covered. The overall text confidence score is 0.97, there is no missing of any main node, and each statement can be traced back to the corresponding original clue node, realizing the explainable, traceable and logically rigorous event analysis.

[0171] Compared with the traditional single-hop retrieval method, the traditional scheme can only connect the showcase damage-alarm-reporting-alteration of out-of-warehouse record-case closing, five-node chain, missing the monitoring dead angle-abnormal entry-abnormal traces on the scene-suspect vehicle, and bridging details. The slot integrity missing rate of the traditional scheme is 0.22, and the overall confidence score is 0.65. It is found that there are 6 fact chain breakpoints that need to be manually supplemented after artificial post-supplementation.

[0172] Table 1 shows the specific comparison data

[0173]

[0174] Training sample comparison examples:

[0175] Traditional positive sample: "showcase damage-reporting-case closing";

[0176] The positive sample of the present application: "monitoring dead angle-abnormal figure-access control tampering-vehicle leaving-evidence report-case closing";

[0177] The negative sample of the present application: "reporting-monitoring dead angle"-both have no direct evidence support, and the system automatically filters.

[0178] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A deep learning-based event clue knowledge enhancement generation method, characterized in that, The method comprises the following steps: acquire a detective event text corpus, perform sentence segmentation, word segmentation and noise cleaning preprocessing on the detective event text corpus, and obtain a preprocessed detective event text corpus; perform event recognition and clue extraction on the preprocessed detective event text corpus, extract each paragraph as a detective event five-tuple, and simultaneously extract a clue segment associated with the detective event five-tuple, to form a detective event five-tuple set and a total clue segment set; construct a detective event clue knowledge graph based on the detective event five-tuple set and the total clue segment set, align and merge the detective event five-tuple, label the source text identifier and confidence information of each graph node, and generate a detective event clue knowledge graph; call an improved multi-hop reasoning module on the detective event clue knowledge graph, and output a complete detective event clue subgraph; based on the complete detective event clue subgraph, construct word-level, sentence-level, paragraph-level and event-level comparison samples to form a multi-granularity comparison sample set; input the multi-granularity comparison sample set into an improved hierarchical contrast learning module, and perform representation alignment training in a word-level semantic space, a sentence-level semantic space, a paragraph-level semantic space and an event-level semantic space to obtain a clue vector representation set optimized by hierarchical contrast learning; based on the clue vector representation set and the context semantic vector of the preprocessed detective event text corpus, generate a coherent detective event narrative text, and perform slot integrity detection and confidence evaluation to obtain a final event clue knowledge enhanced text; The improved multi-hop reasoning module comprises: construct a multi-hop reasoning strategy network based on reinforcement learning on the detective event clue knowledge graph, and the state space of the multi-hop reasoning strategy network is used to describe the detective event reasoning path representation composed of the current node and the historical node; a deep path memory network based on a fusion attention mechanism is used to aggregate the features of the nodes and their semantic information in the historical reasoning path to obtain a context semantic vector on each reasoning path; a reward function of multi-constraint fusion is obtained by weighted summation of the semantic similarity score between event nodes, the time sequence coherence score of nodes and the context logical consistency score between event nodes; in each reasoning process of the multi-hop reasoning strategy network, the reward function of multi-constraint fusion is calculated for each candidate node in the current node and the action space, the semantic relevance, time coherence and context logical consistency of the current node and the candidate node are simultaneously evaluated, the optimal action is selected according to the numerical value of the reward function, and the recursive jump search of the most valuable nodes and edges in the detective event clue knowledge graph is performed; the multi-hop reasoning strategy network recursively updates the reasoning path based on the reward function, and uses the reward function to guide the dynamic exploration and completion of potential implicit event nodes and bridging edges in the detective event clue knowledge graph, to generate a detective event clue subgraph containing reasoning incremental results; The improved hierarchical contrast learning module comprises: acquire a multi-granularity comparison sample set, and input the multi-granularity comparison sample set into an improved hierarchical contrast learning module; In the word-level semantic space, a word-level contrast loss function is calculated, and the word-level contrast loss function is used to enhance the consistency of the word representations of different nodes involved in the same detective clue; In the sentence-level semantic space, a sentence-level contrast loss function is calculated, and the sentence-level contrast loss function is used to improve the semantic alignment degree between different text sentences in the detective event clue; In the paragraph-level semantic space, a paragraph-level contrast loss function is calculated, and the paragraph-level contrast loss function is used to evaluate the logical coherence of the paragraph fragment composed of multiple related sentences in the detective event clue knowledge in the event context; In the event-level semantic space, an event-level contrast loss function is calculated, and the event-level contrast loss function is used to comprehensively control the logical consistency and event description integrity of the detective event clue; The word-level contrast loss function, the sentence-level contrast loss function, the paragraph-level contrast loss function and the event-level contrast loss function are weighted and fused to obtain a total-level contrast loss function; Based on the total-level contrast loss function, the improved hierarchical contrast learning module is iteratively trained to generate a set of clue vector representations optimized by hierarchical contrast learning.

2. The event lead knowledge augmentation generation method based on deep learning according to claim 1, characterized in that, The formation of the detective event five-tuple set and the total clue fragment set comprises: Each paragraph in the preprocessed detective event text corpus is extracted in turn to obtain a paragraph set; For each paragraph in the paragraph set, a sequence labeling model is used to identify the subject element, object element, action element, time element and location element in the paragraph respectively, and a detective event five-tuple is formed; The detective event five-tuples extracted from all paragraphs are sequentially organized to form a detective event five-tuple set; For each paragraph, the detective event five-tuple extraction is performed at the same time, and a clue fragment set directly related to the current five-tuple is identified; All clue fragment sets identified in the paragraphs are combined to form a total clue fragment set, and the detective event five-tuple set and the corresponding total clue fragment set are output.

3. The event lead knowledge augmentation generation method based on deep learning according to claim 1, characterized in that, The generation of the detective event clue knowledge graph comprises: All detective event five-tuples of the detective event five-tuple set are aligned and merged according to the subject consistency principle, time proximity principle and semantic similarity principle, and all detective event five-tuples satisfying the merging conditions are grouped into several detective event five-tuple merging groups; For each detective event five-tuple merging group, a detective event clue knowledge graph node is constructed based on the contained detective event five-tuple elements; For each clue fragment in the total clue fragment set, each clue fragment is connected to one or more nodes in the node set, and if the clue fragment supplements, explains or supports the corresponding node in the text semantics, a directed edge is established in the detective event clue knowledge graph; Each node in the node set and each directed edge in the edge set are labeled with a text source identifier, and each node and each directed edge are assigned a confidence level; The detective event clue knowledge graph composed of the node set, the edge set, the text source identifier and the confidence level is output.

4. The event lead knowledge augmentation generation method based on deep learning according to claim 3, characterized in that, The subject consistency principle is that if the subject elements in the detective event five-tuple are equal, it is considered to be consistent in the subject dimension and is divided into the same detective event five-tuple merging group. The time proximity principle is that if the difference between the time elements in the detective event five tuple is less than or equal to the time proximity threshold value, and the dimension of the time proximity threshold value is consistent with the time element, then the two are adjacent in the time dimension and are divided into the same detective event five tuple merging group. The semantic similarity principle is that the detective event five tuple has a semantic vector representation, and when the cosine similarity between the semantic vectors of any two detective event five tuples is greater than or equal to the semantic similarity threshold value and the semantic similarity threshold value is in the range of zero to one, then the two are related in the semantic dimension and are divided into the same detective event five tuple merging group.

5. The event lead knowledge augmentation generation method based on deep learning according to claim 1, characterized in that, The formation of the multi-granularity contrast sample set includes: Extracting the corresponding text content for each node and edge in the detective event clue subgraph and performing hierarchical labeling, which includes four levels of word-level labeling, sentence-level labeling, section-level labeling, and event-level labeling; Corresponding the word-level labeling, sentence-level labeling, section-level labeling, and event-level labeling to the nodes or edges in the detective event clue subgraph one by one to form a multi-granularity labeling mapping set; For each level of labeling mapping in the multi-granularity labeling mapping set, construct positive sample pairs and negative sample pairs; Combine all positive sample pairs and negative sample pairs under each granularity to form a multi-granularity contrast sample set.

6. The event lead knowledge augmentation generation method based on deep learning according to claim 1, characterized in that, Based on the clue vector representation set and the context semantic vector of the preprocessed detective event text corpus, a coherent and consistent detective event narrative text is generated, and slot integrity detection and confidence evaluation are performed, including: In the knowledge injection type text generation model, the clue vector representation set and the context semantic vector are fused through a multi-source gating attention mechanism; The fused clue vector representation set and context semantic vector drive the decoder to generate the detective event narrative text; Perform slot integrity detection on the detective event narrative text; Perform confidence evaluation on the generated detective event narrative text to calculate the confidence of each event clause and the overall text confidence; Evaluate the quality of the detective event narrative text through the slot integrity missing rate and the overall text confidence: When the slot integrity missing rate is lower than the preset slot integrity threshold and the overall text confidence is higher than the text confidence threshold, the generated text is recognized as meeting the quality requirements; otherwise, the system automatically generates an error feedback signal; The error feedback signal is returned to the multi-hop reasoning module, triggering incremental search and updating of the detective event clue knowledge graph, obtaining an updated detective event clue knowledge graph; Based on the updated detective event clue knowledge graph, the text generation process is re-executed, and the iteration is performed until the slot integrity missing rate is lower than the preset slot integrity threshold and the overall text confidence is higher than the text confidence threshold, and the final event clue knowledge enhanced text that meets the requirements is output.

Images