Event implementation system and event implementation method based on artificial intelligence

Abstract

The invention provides an event implementation system based on artificial intelligence. The event instrumentation system includes an input device, an output device, a graphics processing unit (GPU), and a processor. The processor is connected with the input device, the output device and the GPU. The event instrumentation system receives free text through the input device, and the processor performs event instrumentation on the free text through the GPU. The GPU retrieves and retrieves a plurality of verb-centered events from the free text execution event using semantic parsing, and performs event abstraction and retrieves a plurality of abstract events, and implements the abstract events to a plurality of anchor events of an event-centered KG, and inference a sub-graph through an inference model, and the sub-graph contains the abstract event and the anchor event. The GPU generates a prediction based on the reasoning and provides the prediction through the processor and the output device.

Description

Technical Field

[0001] The present invention generally relates to an event grounding method based on artificial intelligence. More specifically, the present invention relates to a system for event grounding method using a knowledge graph (KG). Background Art

[0002] Reasoning about narratives is a fundamental task in natural language processing (NLP) and has attracted great interest in the NLP community. It is crucial for downstream applications such as text summarization and dialogue generation.

[0003] The most critical challenge in narrative reasoning is modeling the relations between events, which often requires extensive background world knowledge.

[0004] Consider the following story: "Tom was tired and wanted to have fun. He bought a ticket to see Harry Potter." This narrative can be broken down into multiple sub-sentences:

[0005] (E1) Tom is very tired.

[0006] (E2) Tom wants to have fun.

[0007] (E3) He bought a ticket for the Harry Potter movie.

[0008] Each clause represents an event with a verb and one or more arguments. These events convey most of the meaning within their context.

[0009] For humans, understanding these semantic units relies heavily on background world knowledge beyond the immediate context. For example, given E1 and E2, we can infer that Tom has just finished his work. Knowing that watching a movie is a pleasure, it is reasonable to infer that Tom has chosen this activity (from E2 to E3). We can also infer from E3 that Tom will need to arrive at the theater before the movie starts.

[0010] To help machines leverage this type of event knowledge, existing solutions generally fall into two categories:

[0011] Implicit modeling using language models (LM): Some approaches involve pre-training language models using event-aware objectives. While this approach captures event knowledge, it often sacrifices transparency and interpretability of reasoning. Knowledge is embedded in a way that is not easily interpretable.

[0012] Explicit Modeling with Knowledge Graphs (KGs): Other approaches explicitly organize the world knowledge of events into structured, event-centric knowledge graphs. These graphs provide a clear and organized representation of events and their relationships. Despite the existence of these approaches, research on effectively leveraging the symbolic event knowledge in these KGs for reasoning is limited. Existing solutions primarily deal with the constrained format of text (subject-predicate-object) and do not generalize well to free text.

[0013] In recent years, many large-scale knowledge graphs have been constructed, such as ATOMIC, ConceptNet, ASER, and GLUCOSE. However, how to effectively utilize the knowledge in these resources remains a problem. Figure 6 ,Current solutions can be roughly classified into two groups: the knowledge model paradigm and the retrieval and integration paradigm.

[0014] Knowledge Model Paradigm: This approach leverages external KGs by pre-training LMs with carefully designed objectives. Most existing knowledge-enhanced LMs focus on using entity-centric KGs. When it comes to leveraging external event knowledge, this paradigm involves fine-tuning LMs on event-aware KGs through methods such as event-pair relation modeling, complete event recovery / masking, and relevance-based event ranking.

[0015] Retrieval and integration paradigm: In contrast, this approach explicitly retrieves triples or subgraphs from an external KG. Recent research on reasoning with external knowledge bases (KBs) and text has explored reifying entities into KGs in tasks such as open-domain question answering (QA), commonsense QA, and narrative reasoning. However, most of these works focus on entity-centric KGs that contain little or no event knowledge. Although some research on script reasoning has explored event usage, their methods are usually limited to structured text similar to "subject-predicate-object" in narrative chain machine comprehension (MCNC) tasks, making it challenging to extend to general free text.

[0016] In contrast, our solution addresses the more complex problem of externalizing events in free text into an event-centric KG. Given the pressing need for interpretability in artificial intelligence (AI), our approach extends the retrieval and integration paradigm to achieve this goal. By doing so, we enhance narrative reasoning with a more interpretable and efficient framework for externalizing free text into an event-centric KG. Summary of the invention

[0017] To address the above problems, some embodiments of the present invention propose a novel AI-based framework called EvenGround to explicitly materialize free text into an event-centric knowledge graph (KG). Our approach addresses two main challenges: event representation and event sparsity.

[0018] Event Representation Problem: We adopt a semantic parsing-based event extraction approach enhanced with an event normalization module. This module separates events from their context while preserving mutual reference information, ensuring that anaphoric terms such as pronouns remain correctly linked to their antecedents.

[0019] Event sparsity problem: Inspired by human abstract thinking, we introduce a multi-level event abstraction method. This method conceptualizes events into various levels of abstraction by omitting detailed arguments. This abstraction helps to summarize events, making it easier to match them with the incomplete nature of KGs.

[0020] In some embodiments of the present invention, we empirically demonstrate that our solution significantly alleviates the sparsity problem.

[0021] In addition, in some embodiments of the present invention, we materialize the abstracted events into KG to create a joint reasoning sub-graph. Subsequently, we use a model based on a graph neural network (GNN) to perform reasoning.

[0022] Experimental results on three narrative reasoning tasks show that our framework consistently outperforms current state-of-the-art models in some embodiments of the present invention. In addition, we provide qualitative studies showing how our approach provides human-interpretable evidence for model predictions.

[0023] Contributions of the present invention:

[0024] In some embodiments of the present invention, we present an initial formulation of the problem of materializing free text into an event-centric KG.

[0025] In some embodiments of the present invention, we introduce an artificial intelligence-based system approach EvenGround to solve the event representation and sparsity problems, thereby achieving effective narrative reasoning based on embodied information.

[0026] Experimental results indicate that our approach of the embodiments outperforms strong baselines, achieving new state-of-the-art performance on three datasets while providing human-interpretable evidence.

[0027] In some aspects of some embodiments of the present invention, an event concretization method based on artificial intelligence is provided. The event concretization method comprises: performing event acquisition from input free text using semantic parsing by an event concretization system and acquiring multiple verb-centric events; performing event abstraction by the event concretization system and acquiring multiple abstract events; the event concretization system concretizes the abstract events into multiple anchor events of the event-centric KG; the event concretization system infers a subgraph through an inference model; and generating a prediction. The subgraph includes abstract events and anchor events.

[0028] In another aspect of some embodiments of the present invention, an event materialization system based on artificial intelligence is provided. The event materialization system includes an input device, an output device, a graphics processing unit (GPU), and a processor. The processor connects the input device, the output device, and the GPU. The event materialization system receives free text through the input device, and the processor performs event materialization on the free text through the GPU. The GPU uses semantic parsing to perform event acquisition from the free text and obtains multiple verb-centric events, and performs event abstraction and obtains multiple abstract events, and materializes the abstract events to multiple anchor events of the event-centric KG, and infers the subgraph through the reasoning model, and the subgraph includes abstract events and anchor events. The GPU generates predictions based on reasoning and provides predictions through the processor and the output device.

[0029] In an embodiment of the present invention, event acquisition includes: event extraction and event normalization.

[0030] In an embodiment of the present invention, event acquisition includes: an event materialization system extracts verb-centric events from free text. Each verb-centric event includes a triggering verb and an argument set, and each argument has a semantic role.

[0031] In an embodiment of the present invention, event acquisition includes: the event materialization system replaces multiple first tokens in a verb-centric event with multiple second tokens, and each first token refers to a person, and each second token refers to one or more first tokens in the first token that refer to the same person.

[0032] In an embodiment of the present invention, event abstraction includes: the event concretization system discards the arguments of each verb-centered event according to their importance; and obtaining an abstract event.

[0033] In an embodiment of the present invention, the event materialization system materializes the abstract event into multiple nodes in the event-centric KG. The event materialization system obtains anchor events from the nodes, and each abstract event is linked to one or more anchor events.

[0034] In an embodiment of the present invention, the event materialization system obtains a sub-graph, and the sub-graph includes all anchored events, abstract events, and verb-centric events.

[0035] In an embodiment of the present invention, the event materialization system uses a GNN module to perform reasoning on the sub-graph.

[0036] In an embodiment of the present invention, all words in an event are lemmatized.

[0037] In an embodiment of the present invention, when processing the first token and the second token, multiple word fragments are detected through grammatical parsing and animation classification, and the event concretization system uses the mutual reference information between these fragments to normalize all fragments referring to characters and generate a second token. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Embodiments of the present invention are described in more detail below with reference to the accompanying drawings, in which:

[0039] Figure 1 It is an overview of EvenGround, an artificial intelligence-based system method of an embodiment of the present invention;

[0040] Figure 2 is an overview of the sub-graph of an embodiment of the present invention, and the sub-graph is from the SCT-v1.0 dataset. The attention weights of the first 10 nodes are shown in the bar chart. The first 3 nodes are bolded and underlined;

[0041] Figure 3 is another overview of the sub-graph of an embodiment of the present invention. The attention weights of the first 10 nodes are shown in a bar chart. The first 3 nodes are bolded and underlined;

[0042] Figure 4 is another overview of the sub-graph of an embodiment of the present invention. The attention weights of the first 10 nodes are shown in a bar chart. The first 3 nodes are bolded and underlined;

[0043] Figure 5 is a comparison of event materialization performance under different settings related to some embodiments of the present invention. The bar graph (with the y-axis on the left) shows the hit rate percentage of event matching. The line shows the average L2 distance d. We do not normalize for "no extraction";

[0044] Figure 6 is an overview of the knowledge model paradigm (left) and the retrieval and integration paradigm (right). The knowledge model paradigm pre-trains the LM with a specially designed objective, and then further fine-tunes the LM to adapt to the downstream prediction task. The retrieval and integration paradigm retrieves the relevant sub-graph of the story context, and then makes predictions based on the retrieved sub-graph; and

[0045] Figure 7 is a curve of F1 score and threshold value related to some embodiments of the present invention. The curve reflects the event matching performance under different threshold values ​​l. DETAILED DESCRIPTION

[0046] In the following description, artificial intelligence-based methods and systems for event materialization are described as preferred examples. It should be understood by those skilled in the art that modifications, including additions and / or substitutions, may be made without departing from the scope and spirit of the invention. Specific details may be omitted so as not to obscure the invention; however, this disclosure is written to enable those skilled in the art to practice the teachings herein without undue experimentation.

[0047] In some embodiments of the present invention, an artificial intelligence-based event concretization system and event concretization method are provided to solve the unresolved problem of concretizing free text into an event-centric KG.

[0048] In some embodiments of the present invention, one of the main goals of the artificial intelligence-based event materialization system and event materialization method of the present invention includes the design of EvenGround. EvenGround provides a comprehensive framework for materializing free text into event-centric KGs to enhance contextualized narrative reasoning. The present invention addresses key challenges in the field of NLP, especially in the understanding and reasoning of narratives. Specific goals and features of the present invention are as follows:

[0049] Enhanced events represent:

[0050] In some embodiments of the present invention, the system employs an event extraction method based on semantic parsing. This method is equipped with an event normalization module designed to separate events from their context while preserving mutual reference information. This ensures that anaphoric terms such as pronouns remain correctly linked to their antecedents, which is crucial for accurate narrative comprehension.

[0051] Mitigation of event sparsity:

[0052] Inspired by the human abstract thinking process, in some embodiments of the present invention, the system introduces a multi-level event abstraction method. This method conceptualizes events into various abstraction levels by omitting detailed arguments. This abstraction helps to generalize events, making it easier to match the events with the incomplete nature of KGs, and significantly reduces the failure rate of event materialization.

[0053] Event reification and joint reasoning:

[0054] In some embodiments of the present invention, the abstracted events are materialized into KG to create joint reasoning sub-graphs. The system of the embodiment uses a GNN-based model to perform reasoning on these sub-graphs. This combination of event abstraction and GNN-based reasoning enhances the system's ability to effectively interpret and predict narrative outcomes.

[0055] Human-interpretable evidence:

[0056] In some embodiments of the present invention, this framework provides human-interpretable evidence for model predictions, thereby improving the transparency and explainability of the reasoning process. This feature is critical for building trust in AI systems, especially for applications that require clear justification of model decisions.

[0057] In other words, in some embodiments of the present invention, systems and methods including EvenGround are provided, which relate to methods and systems for materializing free text into event-centric KGs for the purpose of narrative reasoning. These systems and methods address the challenge of leveraging structured world knowledge in NLP tasks, particularly in the field of narrative understanding and prediction.

[0058] One of the goals of the present invention is to provide a comprehensive framework that can effectively materialize events from free-form text to event-centric knowledge graphs, thereby enhancing the reasoning capabilities of natural language processing systems. In some embodiments of the present invention, EvenGround aims to overcome two key problems in the art: the event representation problem and the event sparsity problem.

[0059] A key feature of the present invention is its ability to handle free text input, which distinguishes the present invention from previous methods that are restricted to specific structured formats. In some embodiments of the present invention, the EvenGround method incorporates novel techniques for event extraction, normalization, and multi-level abstraction that together address the challenges of representing events from unstructured text and mitigating the sparsity of events in natural language.

[0060] Another important feature of the present invention is that it uses GNN to reason about the constructed joint knowledge-enhanced subgraph. This method allows combining contextual information from the input text with relevant background knowledge from the event-centric knowledge graph, thereby achieving more sophisticated and accurate narrative reasoning.

[0061] In the following description, multiple systems of multiple embodiments will be used to explain the artificial intelligence-based event materialization system and event materialization method of the present invention, and the system executes or utilizes the event materialization method presented in each embodiment, which will not be described separately in some of the following embodiments. However, it is obvious to a person skilled in the art that the following embodiments can all be used as the system or method claimed at the end of this document.

[0062] In an embodiment of the present invention, an event materialization system based on artificial intelligence is provided. The event materialization system has an input device, an output device, a GPU and a processor, and the processor connects the input device, the output device and the graphics processing unit. For example, the input device may include a keyboard, a mouse, a touchpad, a touch module, a microphone or any other device that can be used for text input in any form, and the output device may include a display, a projector, a speaker, a printer or any other device that can be used for text output in any form. The processor may include a central processing unit (CPU).

[0063] For some embodiments of the present invention, the GPU may be a data center GPU designed to accelerate AI, high performance computing (HPC), data science, and graphics. In some embodiments, the GPU may feature tensor cores dedicated to deep learning tasks. For example, the GPU may include an NVIDIA Tesla V100.

[0064] In some other embodiments, the system may include semantic parsing tools. In these embodiments, semantic parsing tools accurately extract events from free text input. These tools should support semantic role labeling (SRL) to identify verbs and their corresponding arguments in the text.

[0065] In some other embodiments, the system may include an event normalization module. The event normalization module reduces event sparsity while maintaining inter-referencing information. This module should be able to detect tokens that refer to entities and replace them with standardized referents.

[0066] In some other embodiments, the system must implement a GNN-based reasoning model, such as a relational graph convolutional network (RGCN), to perform reasoning on joint subgraphs. This model should be able to leverage the relational information within the KG for effective narrative reasoning.

[0067] In the system of some embodiments of the present invention, it is necessary to access comprehensive event-centric KGs, such as ASER, ATOMIC or GLUCOSE. These KGs should be normalized to ensure compatibility with the event abstraction and materialization process.

[0068] In some other embodiments of the present invention, the system should include computing resources. The system requires a large amount of computing resources, including further more high-performance GPUs, to handle the complex tasks of event extraction, normalization, abstraction, and GNN-based reasoning. Sufficient memory and processing power are essential for running large-scale experiments and achieving optimal performance.

[0069] In some other embodiments, the system should include a natural language processing framework. It is suggested to integrate with advanced NLP frameworks such as Huggingface Transformer and Deep Graph Library (DGL) to facilitate the implementation of language models and graph-based reasoning algorithms.

[0070] In other words, in some embodiments of the present invention, a computer system with sufficient processing power and memory is required. The system should be able to run natural language processing tools, graph processing algorithms, and neural network models. In addition, access to large-scale event-centric knowledge graphs is necessary for the knowledge retrieval and integration steps of the method.

[0071] In an embodiment of the present invention, the event materialization system receives free text through an input device, and the processor performs event materialization on the free text through a GPU.

[0072] Figure 1 An overview of the event materialization process of the system in this embodiment is shown. Figure 1 In the event materialization system of the embodiment, the GPU uses semantic parsing to perform event acquisition from free text and obtains multiple verb-centric events (i.e., extraction). After extraction, the GPU performs event abstraction and obtains multiple abstract events (i.e., conceptualization). Subsequently, the GPU materializes the abstract events to multiple anchor events of the event-centric KG and infers the subgraph through the inference model (i.e., event linking and subgraph retrieval). Finally, the GPU generates predictions based on reasoning and provides the predictions through the processor and output device (i.e., reasoning).

[0073] In this embodiment, the system appropriately extracts events from free text. Through event abstraction, the system improves the efficiency of materialization or linking between events and KGs to provide reliable reasoning and prediction of free text. This method uses advanced NLP techniques and graph-based techniques to convert free text input into structured reasoning output by materializing it in an event-centric knowledge graph. Each step is critical to ensuring accurate event representation, handling sparsity, and achieving effective narrative reasoning.

[0074] Specifically, in this embodiment, the method includes event acquisition from free text. The method begins by acquiring events from free text input using semantic parsing techniques. This involves identifying and extracting verb-centric events, where each event is represented by a verb and its associated arguments. The process ensures that multiple events are accurately captured from the text.

[0075] In some embodiments of the present invention, event acquisition includes: event extraction and event normalization. In order to solve the event representation problem, the system of this embodiment is equipped with event extraction based on semantic parsing and event normalization modules to separate events from contexts while retaining the mutual reference information of their arguments.

[0076] In these embodiments, event extraction is a basic component of NLP, which involves identifying and capturing specific events within a given text. This process focuses on identifying verbs (actions) and their associated arguments (such as subjects, objects, and other related entities). Event extraction is intended to systematically decompose sentences into discrete meaningful information units. By isolating these events, the system can better understand the narrative structure and context of the text. This ability is crucial for applications such as information retrieval, text summarization, and narrative reasoning, where understanding the sequence and relationship between events is essential. Effective event extraction enables automatic identification of key actions and interactions, thereby promoting a deeper understanding of the text content.

[0077] Additionally, event normalization is a complementary process to event extraction that aims to standardize the representation of identified events. This involves resolving mutual references (e.g., determining that "he" refers to "Tom") and converting synonymous phrases to a consistent format (e.g., "purchase a ticket" and "buy a ticket" both become "buy a ticket"). The goal of event normalization is to ensure that different expressions of the same event are treated uniformly, thereby reducing the variability and ambiguity of the data. This standardization is critical for integrating the extracted events into a structured knowledge base or knowledge graph, where the consistency of representation enhances the accuracy and reliability of data retrieval and reasoning tasks. By normalizing events, the system can better generalize from specific examples, thereby improving its ability to handle diverse and complex narratives with greater coherence and clarity.

[0078] Specifically, in these embodiments, event acquisition includes: the GPU extracts verb-centric events from free text. In this step, each verb-centric event includes a triggering verb and an argument set, and each argument has a semantic role.

[0079] In this process of these embodiments, the system utilizes a semantic parsing-based method to extract the event from the context of the event. For example, for a text segment s having n sentences, that is: s=[s 1 ,s 2 ,…,s n ]. The system performs semantic role labeling (SRL) on the text s to extract a series of events (P) centered on the verb, and each event (p i ) has a trigger verb (verb i ) and the argument set (A i ).Right now: P = {p 1 ,p 2 ,…,p m}, and p i =(verb i ,A i ).

[0080] In these embodiments, each argument in the set has a semantic role. A semantic role (also called a subject role) is the potential relationship of an argument (noun phrase) to a main verb in a clause. Semantic roles provide information about how entities involved in an event participate in the event. They help identify the function of each argument in the context of the event described by the verb.

[0081] Common semantic roles: Agent: The entity that performs the action. Example: Tom in “Tom kicked the ball”. Patient (or Subject): The entity that experiences or is affected by the action. Example: the ball in “Tom kicked the ball”. Experiencer: The entity that experiences or perceives an event. Example: John in “John feels happy”. Means: The way in which an action is performed. Example: with a hammer in “She used a hammer to hit the nail.” Beneficiary (or recipient): The entity for whose benefit the action is performed. Example: for Mary in “He baked a cake for Mary”. Location: The place where the action takes place. Example: The park in “They were playing in the park”. Source: The starting point of an action. Example: from New York in “she flew from New York”. Goal: The end point of the action. Example: to Paris in “He sent the letter to Paris”.

[0082] The features of semantic roles in these embodiments include: understanding and disambiguation. Semantic roles help understand the meaning of a sentence by clarifying the function of each argument. They disambiguate sentences by providing context about the relationship between different elements in the sentence. For example: in the sentence "John gave Mary a book.", their semantic roles indicate that John (the agent) performed the giving; Mary (the beneficiary) received the book; and the book (the subject) is the item given.

[0083] The features of semantic roles in these embodiments also include: information extraction and retrieval. In NLP applications, semantic roles are used to extract structured information from unstructured text. This is useful in tasks such as information retrieval, question answering, and summarization. For example, extracting events from news articles to identify who did what to whom, when, and where.

[0084] The features of semantic roles in these embodiments also include: machine learning and AI. In these embodiments, semantic roles can be used to train machine learning models for various NLP tasks. They provide features that help models understand sentence structure and meaning, thereby improving the performance of tasks such as translation, sentiment analysis, and text generation. For example: in machine translation, understanding roles can help translate sentences more accurately by maintaining relationships between elements.

[0085] Features of semantic roles in these embodiments also include: Knowledge graph construction. When building a knowledge graph, semantic roles help link entities with the correct relationships. This structured knowledge representation supports better query and reasoning. For example: building a knowledge graph of scientific literature, in which entities such as researchers, institutions, and discoveries are linked to appropriate roles.

[0086] In these embodiments, semantic roles are crucial to understanding the interactions and relationships between different components of a sentence. They enhance the accuracy of various NLP applications by providing a deeper understanding of sentence structure and meaning.

[0087] Specifically, in these embodiments, the semantic roles of arguments are based on the PropBank annotation guidelines. The semantic roles include: ARG 0 : Usually indicates the agent or performer of an action. This is usually the subject of a sentence in an active voice structure. For example, in "John ate an apple" "John" would be the ARG 0 . ARG 1 : Usually indicates the patient or subject of an action. This is usually the direct object of a verb or the subject in a passive structure. In "John ate an apple" "apple" would be the ARG 1 .

[0088] In these embodiments, the semantic role further comprises: ARG 2 , ARG 3 , ARG 4 These are used for less common roles that can vary depending on the verb. They usually mean: ARG 2 : means, beneficiary or attributive ARG 3 : starting point, beneficiary or attributive ARG 4 :end

[0089] In these embodiments, the semantic role further comprises: ARG M . This indicates a modifier argument. These provide additional information about the event but are not central to the meaning of the verb. Examples include: ARG M-TMP :Time information (when) ARG M-LOC :Location information (where) ARG M-MNR :Method information (method) ARG M-CAU : Reason or cause.

[0090] In the context of the EvenGround framework of these embodiments, these semantic roles are important for several reasons:

[0091] Event Extraction: Semantic role labeling helps identify the core components of an event (who did what to whom, when, where, how, etc.).

[0092] Event Normalization: Understanding roles helps retain important information during the normalization process.

[0093] Multi-level abstraction: The framework uses the importance of these roles to decide the order of argument discarding during abstraction. Specifically, the framework discards arguments in the following order: ARGM, then ARG2 / 3 / 4, then ARG1, and finally ARG0.

[0094] This structured understanding of event arguments allows the system to create more meaningful and hierarchical event representations, which is crucial for the efficient materialization of knowledge graphs and subsequent reasoning tasks.

[0095] In some embodiments, event acquisition includes event normalization. The GPU replaces a plurality of first tokens in a verb-centric event with a plurality of second tokens, wherein each first token refers to a person; and each second token refers to one or more first tokens in the first tokens that refer to the same person.

[0096] For example, in an embodiment of the present invention, three events are extracted from the text: 1.The general took some wine at the party. 2.He felt sleepy. 3.He said goodbye to them.

[0097] In these events, "General" and "He" refer to the same person, while "They" refers to another group of people. During event normalization, the first token in this text contains "General", "He", and "They".

[0098] After the replacement, the events become as follows: 1.[P0]Took some wine at the party. 2. [P0] Feeling sleepy. 3. [P0] said goodbye to [P1].

[0099] In these events, [P0] and [P1] are the second tokens. [P0] refers to "general" and "he", and [P1] refers to "they". Normalization in this embodiment helps reduce event sparsity by removing details in character words. This increases the probability that the event is successfully materialized into the KG.

[0100] In some embodiments of the present invention, event abstraction includes: the GPU discarding arguments of each verb-centric event according to their importance; and obtaining abstract events.

[0101] Specifically, the abstraction process is based on the importance of event arguments in semantic role labeling (i.e., semantic roles). 0 and ARG 1 are the most important because they usually specify the subject and object. In contrast, the modifier argument ARG M Conveys the least information, as it usually defines additional constraints on the predicate, such as when and where the event occurred.

[0102] In an embodiment, the system discards arguments in the following order: 1.ARG M ; 2.ARG 2 , ARG 3 , ARG 4 ; 3.ARG 1 ;as well as 4.ARG 0 .

[0103] For example, the abstraction of the series of events (P) centered on the verb includes the following steps: generating a new abstract event set P based on the series of events P abs, where P abs = {q 1 ,q 2 ,…,q m}. Each q i is the event corresponding to p i An abstract sequence of events.

[0104] For example, a sequence q contains abstract events q 0 ,q 1 ,q 2 ,q 3 ,in q 0 ={ARG 0 : [P0], V: Evacuated, ARG 2 : To relatives' house, ARG M :last night}; q 1 ={ARG 0 : [P0], V: Evacuated, ARG 2 : to a relative's house}; q 2 ={ARG 0 :[P0], V:evacuated}; q 3 ={V: evacuated}.

[0105] Each time an argument is dropped, the level of abstraction of the event increases. At the same time, events at higher levels of abstraction (e.g., 2 、p 3 ) are more likely to be recorded in the KG, which alleviates the sparsity problem.

[0106] In other words, in this embodiment, each time an argument is discarded, a new event of a higher level of abstraction is created. This produces a series of increasingly abstract representations of the original event.

[0107] In this embodiment, this process aims to address the sparsity problem in event materialization. More abstract versions of an event are more likely to match entries in the KG, thereby increasing the chance of successful materialization.

[0108] The system of an embodiment preserves all levels of abstraction, allowing it to use the most appropriate level for matching in a KG.

[0109] In this embodiment, this approach strikes a balance between preserving event details (at a more specific level) and increasing the likelihood of finding a match in the KG (at a more abstract level).

[0110] Multi-level event abstraction is a key step in EvenGround because it significantly improves event materialization performance by increasing the hit rate in the KG while still maintaining the ability to access more detailed event information when needed.

[0111] In this embodiment, the method includes event abstraction. After extracting verb-centric events, the method performs event abstraction to create multiple abstract events. This step involves conceptualizing the verb-centric events by simplifying the verb-centric events into a more general form, omitting specific details while retaining core semantic information. This abstraction process helps deal with the sparsity of events and enhances generalization.

[0112] Table 1 shows the impact of event extraction, normalization, and abstraction. The mean and standard deviation of the accuracy on SCT-v1.0 are reported, where “RB” and “BB” refer to the RoBERTa-base version and BERT-base version.

[0113] Table 1

[0114] As shown in Table 1, the system of the embodiment eliminates event extraction ("no extraction"), event normalization ("no normalization"), and multi-granularity event abstraction ("no abstraction" and "-ARGX"), respectively. Specifically, when the event extraction module is eliminated, we use the complete statement for event concretization instead. When the event normalization part is eliminated, we skip the normalization step and use the original event for concretization. For multi-level event abstraction, we discard event arguments in the order that the highest level ("-ARG1") contains all abstract events in the previous levels. The baseline ("no knowledge") shows the results of the ordinary language model, which does not utilize any external knowledge.

[0115] In this example, the event extraction and normalization steps are necessary. When removed, performance relative to the baseline is not improved, and even degrades. In addition, the event abstraction step is also crucial. By taking only the first level of abstraction (removing the modifier arguments), we have seen considerable performance gains. The model reaches its best performance after discarding ARG1.

[0116] On the other hand, reference Figure 5 We perform the analysis using the following automatic metrics: (1) the average L2 distance d among event matches, and (2) the percentage of events that are considered successful matches (i.e., with an L2 distance below l = 0.65) (the hit rate).

[0117] Refer to Table 2 and Figure 7We evaluate the matching results through human annotation. Three domain experts are asked to annotate whether the event matching is successful for 50 stories (~500 events) randomly sampled from the validation set of SCTv1.0. Fleiss's κ value is 0.7414. We obtain the ground truth labels by majority voting and present the accuracy of different event matching methods in Table 2. To explore the impact of the threshold l used in Section 3.2.1, we Figure 7 The F1 scores at different thresholds are visualized in .

[0118] Table 2

[0119] We can observe that: 1) The performance of directly matching sentences to KGs (without extraction) is quite low, which requires an event extraction stage. 2) The event normalization step significantly improves the matching performance. Removing the normalization step can reduce the accuracy by up to 76.7%. 3) In general, the matching performance gradually increases with the increase of abstraction level. 4) The Pearson's r value between the automatic evaluation results and the human evaluation results is 0.8977, indicating that thresholding on the L2 distance is a reasonable way to automatically filter out poorly matched events.

[0120] In this embodiment, after acquiring the abstract event, the GPU materializes the abstract event. Specifically, in this embodiment, the system materializes the event through the GPU. Subsequently, the abstract event is materialized into multiple anchor events within the event-centric knowledge graph (KG). This step involves linking the abstract event to the corresponding event or node in the KG. The system obtains the anchor event from the node, and each abstract event is linked to one of the anchor events, thereby ensuring that the event is appropriately contextualized within the larger framework of the KG.

[0121] In this embodiment, the goal of event matching is to find the most semantically similar node (called "anchor event") for each extracted and abstracted event in the event-centric KG. This process involves the following steps:

[0122] Step 1: Define the matching target

[0123] For the abstracted event set P abs For each event p in , we aim to find the node v in the knowledge graph G = (V, E) that minimizes the distance d(p, v). This can be expressed as:

[0124]

[0125] where V is the set of nodes in the graph, and d(·,·) represents the distance between events.

[0126] Step 2: Choose a semantic similarity measure

[0127] The system of an embodiment chooses a semantic similarity approach instead of using a token-level similarity metric such as TF-IDF or BM25 that may not capture the semantic meaning of an event. The system uses a sentence transformer to encode both the extracted event p and the knowledge graph node v.

[0128] Step 3: Calculate semantic similarity

[0129] We use the L2 distance between the encoded representations to calculate the semantic similarity:

[0130] d(p,v)=||SBERT(text(p))-SBERT(text(v))|| 2

[0131] Where SBERT() represents the sentence transformer encoding function, and text() retrieves the textual representation of the event.

[0132] Step 4: Apply matching threshold

[0133] To filter out bad matches, the system empirically sets a threshold l on the distance d(p,v). If the distance exceeds this threshold, the match is considered unsuccessful.

[0134] Step 5: Collect matching anchor events

[0135] For P abs For each abstract event in , we collect the set The anchor events that are successfully matched in is the abstracted event sequence from the corresponding Matched sequence of anchor events.

[0136] Step 6: Handle unmatched events

[0137] Events that do not find a match within the threshold are not included in the set of anchor events. This helps ensure that only events with reliable matches are used in subsequent inference steps.

[0138] This matching process allows the system to materialize events extracted and abstracted from the input text to the most semantically similar events in the knowledge graph. By using semantic similarity rather than token-level matching, the system can better handle variations in event descriptions and capture the underlying meaning of events. The use of thresholds helps filter out bad matches, thereby improving the quality of materialized events used in later reasoning stages. Multiple levels of abstraction from previous steps increase the likelihood of finding a good match, as more abstract event representations are more likely to have counterparts in the knowledge graph.

[0139] In this embodiment, after acquiring the anchor event, the GPU acquires the sub-graph, and the sub-graph contains all anchor events, abstract events, and verb-centric events.

[0140] Specifically, in this embodiment, the method includes reasoning subgraph creation. The method proceeds to reasoning about the subgraph by creating a reasoning model. This involves retrieving relevant subgraphs from the KG that are connected to the materialized events. The reasoning subgraph includes interconnected events and entities, forming a coherent structure for narrative understanding and reasoning.

[0141] In the EvenGround framework, from G sub =(V sub ,E sub ) starts, the system searches The shortest path within γ hops between each event pair in G. For any path obtained, add the nodes and edges along the path to G sub . Subgraph G sub Acts as a bridge between the input narrative and the large amount of knowledge stored in the event-centric KG, allowing the system to leverage relevant background knowledge to improve reasoning. The system further constructs a joint knowledge-enhanced subgraph G for reasoning 联合 =(V 联合 ,E 联合 ). 联合 Contains G sub In addition, we add the context events in P as nodes to G 联合 , where the materialized relationship between the context event and the anchor event in C and the context relationship are added as edges.

[0142] This submap G 联合 They play a vital role in the EvenGround framework by providing a centralized, relevant, and structured representation of the background knowledge required for effective narrative reasoning.

[0143] Specifically, the characteristics of the sub-graph include:

[0144] Relevance: The subgraph contains the events and relations that are most relevant to the input narrative, filtered from the larger knowledge graph.

[0145] Compactness: A subgraph is a much smaller and more manageable structure than the full KG, thus focusing computational resources on the most relevant information.

[0146] Connectivity: Subgraphs keep events connected, thus maintaining relational information that is critical for reasoning.

[0147] Path-based: A subgraph is constructed by finding the shortest path between matching anchor events, thus ensuring logical connections between related events.

[0148] Limited size: Subgraph retrieval is limited to paths within γ hops, thus controlling the size and complexity of the retrieved information.

[0149] Contextual enrichment: Subgraphs capture not only individual events but also the surrounding context in the knowledge graph, thus providing a richer basis for reasoning.

[0150] Interpretability: The subgraph structure allows for more interpretable reasoning processes, as the system can trace its reasoning through the graph.

[0151] Flexibility: The subgraph can adapt to different levels of event abstraction, from specific to more general representations.

[0152] Integration with input: The subgraphs, when combined with contextual events from the input text, form a joint subgraph (G 联合 )’s basis.

[0153] Reasoning Basis: The subgraph serves as the main knowledge structure against which the graph neural network performs its reasoning operations.

[0154] In this embodiment, after acquiring the sub-graph, the GPU uses a GNN module to perform reasoning on the sub-graph. Specifically, the embodiment uses a GNN-based reasoning model to process the joint knowledge-enhanced sub-graph and generate predictions for narrative reasoning tasks. The operation of this model is as follows:

[0155] First, the system uses the language model representation to classify the input text s and the joint subgraph V 联合 Each node v in is encoded with two:

[0156] v=f LM (text(v)),s=f LM (s)

[0157] where f LM represents a language model function, and text(v) retrieves the textual content of node v.

[0158] Then, the system applies the GNN module to the joint sub-graph G 联合 In this embodiment, the system uses a relational graph convolutional network (RGCN) to efficiently 联合 RGCN updates V of each layer l in L-layer GNN according to the following equation 联合 The representation of each node i in is:

[0159]

[0160] in:

[0161] R stands for E 联合 The set of edge types in ;

[0162] N r (i) represents the neighborhood of node I with relation r;

[0163] σ(·) is a nonlinear activation function;

[0164] W r is the learnable weight matrix for each relation type r;

[0165] After the L layer updates, the system obtains G by pooling the hidden node embeddings from the final layer. 联合 The vector representation of is:

[0166] g = pooling ({h i L :i∈V 联合})

[0167] Finally, the system uses a multi-layer perceptron (MLP) module to produce the predicted probabilities:

[0168] p(s)∝MLP(s+g)

[0169] Where s represents the encoded input text, and g represents the pooled graph representation.

[0170] This GNN-based reasoning model enables the system to effectively integrate information from both the input text and the knowledge graph, thereby capturing complex relationships between events and facilitating sophisticated narrative reasoning. The use of a relational GNN architecture allows the model to exploit the rich structural information present in the event-centric knowledge graph, thereby enhancing its reasoning capabilities.

[0171] The model's ability to process both textual and graph-structured inputs in a unified framework represents a significant advance in narrative understanding and prediction tasks. By combining these diverse sources of information, the system can make more informed and contextually relevant predictions, improving performance on a wide range of narrative reasoning benchmarks.

[0172] Specifically, in this embodiment, the method includes prediction generation. Ultimately, the method generates predictions based on the reasoning performed on the subgraph. This involves using a reasoning model such as a GNN to analyze the subgraph and make informed predictions. The results are then processed and provided through an output device, completing the reasoning cycle.

[0173] In an embodiment, the output prediction can be used as a story completion. For example, the input free text shows "Caroline is a medical student. Caroline studies hard to get good grades.", and the prediction can be "She did very well."

[0174] refer to Figure 2 , in another embodiment, for example, the input free text shows "Caroline is a medical student. Caroline studies very hard to get good grades. One day Caroline failed the test by one point. Caroline was very frustrated, but she continued to study hard.", and the prediction can be "Later, she passed the test."

[0175] refer to Figure 3 , in another embodiment, for example, the input free text shows "Ava needs to go shopping with her two-year-old son. But she couldn't find his shoes even though she looked everywhere! She had no choice but to decide to buy him new shoes. She took him into the store to pick out a new pair of shoes.", and the prediction may be "Ava takes good care of her son.".

[0176] refer to Figure 4 , in another embodiment, for example, the input free text shows "The children were playing at home when they heard music. They ran to their mother and asked for change. She handed them a few dollars. They ran outside.", and the prediction may be "The children excitedly bought ice cream cones.".

[0177] In another embodiment, for example, the input free text reads "After finishing his work, Tom decided to relax. He went to the cinema and bought a movie ticket", and the prediction may be "Tom will enjoy watching the movie and will feel more relaxed afterwards".

[0178] In another embodiment, the output prediction can be used as a next event prediction in healthcare. For example, the input free text shows "Patient A was diagnosed with hypertension. The doctor prescribed medication and recommended lifestyle changes.", and the prediction can be "The patient is likely to have a follow-up visit in three months to monitor the effectiveness of the medication and lifestyle changes.".

[0179] In another embodiment, the output predictions can be used as customer support automation. For example, the input free text shows "The customer reported that their internet connection was down. They reset their router, but the problem persisted." and the prediction can be "The customer's next step is to contact technical support to further troubleshoot or schedule a technician to come out."

[0180] In another embodiment, the output prediction can be used for financial news analysis. For example, the input free text shows "Company X announced a sharp drop in quarterly earnings due to a drop in sales in its main market.", and the prediction can be "The company's stock price is likely to fall in the next trading day as investors react to the earnings report.".

[0181] In another embodiment, the output predictions can be used as a personal assistant scheduler. For example, input free text showing "John has a meeting with his project team at 10 AM. He needs to prepare a presentation in advance", and the prediction can be "John should start preparing his presentation at 8 AM to ensure he is ready for the 10 AM meeting.".

[0182] In some embodiments of the invention, all words in an event are lemmatized.

[0183] During the event extraction and representation process of the present invention, all words within the extracted events undergo lemmatization. Lemmatization is a key preprocessing step that simplifies the inflected forms of words to their base or dictionary forms (called lemmas). This process helps standardize word representations in different events, reduces variability, and improves the system's ability to match semantically similar events. For example, verbs in different tenses (e.g., "running," "ran," "runs") are all simplified to their base form ("run"), and plural nouns are converted to their singular forms.

[0184] The lemmatization process significantly helps to address the event sparsity problem. By normalizing word forms, it increases the likelihood of finding matches between extracted events and events stored in the knowledge graph. This normalization is particularly beneficial when performing event abstraction and matching, as it allows the system to identify and group semantically equivalent events that can be expressed using different inflected forms in the original text.

[0185] In some embodiments of the present invention, when processing the first token and the second token, multiple word fragments are detected through grammatical parsing and animation classification, and the event concretization system uses the mutual reference information between these fragments to normalize all fragments referring to characters and generate a second token.

[0186] The event reification system uses sophisticated natural language processing techniques to process multi-word expressions and entity pronouns within events. Specifically, the system uses grammatical parsing to identify grammatical structures and word dependencies within the text. At the same time, the system applies animacy classification to determine which word fragments can refer to animate entities, especially people. These processes allow the system to detect and isolate meaningful multi-word fragments that represent coherent concepts or entities within events.

[0187] Once these segments are identified, the system uses cross-reference information to track references to the same entity across different events or sentences. This cross-reference solution enables the system to normalize all segments that refer to people and replace them with standardized tokens (e.g., "[P0]", "[P1]"). This normalization process is critical to maintaining consistency in event representation and facilitating more accurate event matching and reasoning. By abstracting away specific entity names while retaining their roles and relationships within events, the system can more effectively externalize events to the knowledge graph and reason about narrative structure independent of specific character identities.

[0188] The functional units and modules of the system according to the embodiments disclosed herein can be implemented using a computing device, a computer processor or an electronic circuit system, including but not limited to an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a microcontroller and other programmable logic devices configured or programmed according to the teachings of the present disclosure. The computer instructions or software codes running in the computing device, computer processor or programmable logic device can be easily made by practitioners in the field of software or electronic technology based on the teachings of the present disclosure.

[0189] All or part of the method according to the embodiment may be executed in one or more computing devices including a server computer, a personal computer, a laptop computer, a mobile computing device (eg, a smart phone), and a tablet computer.

[0190] Embodiments may include computer storage media, transitory and non-transitory memory devices having stored therein computer instructions or software codes that may be used to program or configure a computing device, a computer processor, or an electronic circuit system to perform any of the processes of the present invention. Storage media, transitory and non-transitory memory devices may include, but are not limited to, floppy disks, optical disks, Blu-ray disks, DVDs, CD-ROMs and magneto-optical disks, ROMs, RAMs, flash memory devices, or any type of media or device suitable for storing instructions, codes, and / or data.

[0191] Each of the functional units and modules according to various embodiments may also be implemented in a distributed computing environment and / or a cloud computing environment, where all or part of a machine instruction is executed in a distributed manner by one or more processing devices interconnected by a communication network, such as an intranet, a wide area network (WAN), a local area network (LAN), the Internet, and other forms of data transmission media.

[0192] The foregoing description of the present invention has been provided for the purpose of illustration and description. The present invention is not intended to be exhaustive or to limit the present invention to the precise form disclosed. Many modifications and variations will be apparent to those skilled in the art.

[0193] The embodiment was chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications as are suited to the particular use contemplated.

Claims

1. An event materialization method based on artificial intelligence, characterized in that: include: The event materialization system uses semantic parsing to obtain events from input free text and obtain multiple verb-centric events; Abstracting events and obtaining multiple abstract events through the event materialization system; The event materialization system materializes the abstract event into multiple anchor events of an event-centric knowledge graph (KG); The event materialization system infers the sub-graph through the inference model; and Generate predictions, The sub-graph includes the abstract event and the anchor event.

2. The event materialization method according to claim 1, characterized in that: The event acquisition includes: Event extraction and event normalization.

3. The event materialization method according to claim 1, characterized in that: The event acquisition includes: The event materialization system extracts the verb-centric events from the free text, And each verb-centric event includes a triggering verb and an argument set, and each of the arguments has a semantic role.

4. The event materialization method according to claim 1, characterized in that: The event acquisition includes: The event reification system replaces a plurality of first tokens in the verb-centric event with a plurality of second tokens, and each first token refers to a person, And each second token refers to one or more first tokens among the first tokens that refer to the same person.

5. The event materialization method according to claim 3, characterized in that: The event abstraction includes: The event reification system discards the arguments according to the importance of the arguments for each verb-centric event; and Gets the abstract event.

6. The event materialization method according to claim 1, characterized in that: The event materialization system materializes the abstract event into multiple nodes in the event-centric KG. And the event materialization system obtains the anchor event from the node, and each abstract event is linked to one or more anchor events.

7. The event materialization method according to claim 1, characterized in that: The event concretization system obtains the sub-graph, and the sub-graph includes all the anchor events, the abstract events, and the verb-centric events.

8. The event materialization method according to claim 1, characterized in that: The event materialization system uses a GNN module to perform reasoning on the sub-graph.

9. The event materialization method according to claim 1, characterized in that: All words in the event are lemmatized.

10. The event materialization method according to claim 1, characterized in that: When processing the first token and the second token, multiple word fragments are detected through grammatical parsing and animation classification, and the event concretization system uses the mutual reference information between these fragments to normalize all fragments referring to characters and generate the second token.

11. An event materialization system based on artificial intelligence, characterized in that: include: Input device; Output device; Graphics processing unit (GPU); as well as a processor connected to the input device, the output device and the graphics processing unit, wherein the event materialization system receives free text via the input device, and the processor performs event materialization on the free text via the GPU, and the GPU performs event acquisition from the free text using semantic parsing and acquires a plurality of verb-centric events, and the GPU performs event abstraction and obtains a plurality of abstract events, And the GPU materializes the abstract event into multiple anchor events of the event-centric KG, The GPU infers the sub-graph through the inference model, and the sub-graph includes the abstract event and the anchor event, And the GPU generates a prediction based on the inference and provides the prediction through the processor and the output device.

12. The event materialization system according to claim 11, characterized in that: The event acquisition includes: Event extraction and event normalization.

13. The event materialization system according to claim 11, characterized in that: The event acquisition includes: The GPU extracts the verb-centric events from the free text, And each verb-centric event includes a triggering verb and an argument set, and each of the arguments has a semantic role.

14. The event materialization system according to claim 11, characterized in that: The event acquisition includes: The GPU replaces a plurality of first tokens in the verb-centric event with a plurality of second tokens, and each first token refers to a character, And each second token refers to one or more first tokens among the first tokens that refer to the same person.

15. The event materialization system according to claim 13, characterized in that: The event abstraction includes: The GPU discards the arguments according to the importance of the arguments for each verb-centric event; and Gets the abstract event.

16. The event materialization system according to claim 11, characterized in that: The GPU materializes the abstract event into multiple nodes in the event-centric KG, The GPU obtains the anchor events from the nodes, and each abstract event is linked to one or more anchor events.

17. The event materialization system according to claim 11, characterized in that: The GPU obtains the sub-graph, and the sub-graph includes all the anchor events, the abstract events, and the verb-centric events.

18. The event materialization system according to claim 11, characterized in that: The GPU uses a GNN module to perform reasoning on the sub-graph.

19. The event materialization system according to claim 11, characterized in that: All words in the event are lemmatized.

20. The event materialization system according to claim 11, characterized in that: When processing the first token and the second token, multiple word fragments are detected through grammatical parsing and animation classification, and the event concretization system uses the mutual reference information between these fragments to normalize all fragments referring to characters and generate the second token.

Images