A knowledge acquisition method based on neural-symbol hybrid retrieval enhancement

Abstract

The application discloses a knowledge acquisition method based on neural symbol hybrid retrieval enhancement, comprising the following steps: acquiring an event corpus, a semantic graph, and a symbolized knowledge base related to an event extraction task; mapping each corpus text in the event corpus and an event type in the symbolized knowledge base according to the semantic graph by a weak supervision mode to construct a retrieval knowledge base; when event extraction is performed on a target text, retrieving a plurality of candidate corpus texts related to the target text from the retrieval knowledge base, retrieving an optimal event type for the target text according to a mapping relationship between the candidate corpus texts and the event type, and extracting an event subgraph corresponding to the optimal event type and an optimal corpus text as prompt knowledge; combining the prompt knowledge converted into a prompt text with the target text as an input text, performing knowledge acquisition on the input text by using a language model, and obtaining an event extraction result based on the target text. The method enhances the accuracy and efficiency of knowledge acquisition.

Description

Technical Field

[0001] This invention belongs to the field of knowledge acquisition technology based on natural processing, and specifically relates to a knowledge acquisition method based on neural symbol hybrid retrieval enhancement. Background Technology

[0002] Knowledge acquisition technology has always been a crucial technique in natural language processing. While pre-trained language models possess powerful text understanding capabilities, external knowledge can significantly aid existing models in solving certain retrieval tasks. Therefore, many retrieval-enhancing methods have been proposed. However, existing techniques for acquiring task-related knowledge still face several challenges:

[0003] Question 1: Many event extraction tasks rely on single text or graphs as knowledge sources. For example, the paper "Generating disentangled arguments with prompts: a simple event extraction framework that works" publicly uses event description text as the knowledge source. Another example is "OntoED: Low-resource Event Detection with Ontology Embedding," which publicly utilizes only event-related ontology graphs for enhancement. These methods cannot effectively utilize external knowledge from different sources. Furthermore, single external knowledge often suffers from sparsity. For instance, while symbolic knowledge graphs represent high-density, well-constructed knowledge, they are incomplete. Additionally, different types and granularities of knowledge always exist in different forms, such as triples and entity sets. These different forms of knowledge are difficult to represent uniformly and are not suitable as external knowledge for event retrieval tasks.

[0004] Question 2: While existing event retrieval methods can retrieve task-related external knowledge, they also introduce irrelevant knowledge noise. This noise can negatively impact the retrieval performance of language models. One reason for this noise is that while the retrieved knowledge is relevant to the input, it may not be applicable to the retrieval task itself. For example, in event extraction tasks, text relevant to the input may not be very similar to the event itself, but rather relevant in other aspects, such as scene descriptions or weather backgrounds. This makes the text less useful.

[0005] Therefore, providing a more effective method for knowledge acquisition has greater practical significance. Summary of the Invention

[0006] In view of the above, the purpose of this invention is to provide a knowledge acquisition method based on neural symbol hybrid retrieval enhancement, which can improve the ability to acquire knowledge, enhance the accuracy and efficiency of knowledge acquisition, and reduce computational overhead.

[0007] To achieve the above-mentioned objectives, an embodiment provides a knowledge acquisition method based on neural symbol hybrid retrieval enhancement, comprising the following steps:

[0008] Acquire event corpora, semantic graphs, and symbolic knowledge bases related to the event extraction task;

[0009] Based on semantic graphs, each text in the event corpus is mapped to the event type in the symbolic knowledge base using a weakly supervised approach to construct a retrieval knowledge base;

[0010] When performing event extraction for target text, multiple candidate texts related to the target text are retrieved from the retrieval knowledge base. Based on the mapping relationship between candidate texts and event types, the optimal event type is retrieved for the target text, and the event subgraph and the optimal text corresponding to the optimal event type are extracted as cue knowledge.

[0011] After converting the cue knowledge into cue text, it is combined with the target text as input text. A language model is then used to acquire knowledge from the input text, resulting in event extraction results based on the target text.

[0012] Preferably, the step of mapping each text in the event corpus to the event type in the symbolic knowledge base using a weakly supervised approach based on the semantic graph to construct the retrieval knowledge base includes:

[0013] Each text in the event corpus is segmented to obtain each word and its corresponding part of speech, and each word is mapped to a semantic graph to obtain the semantics of each word;

[0014] Based on part-of-speech, words with the target part-of-speech are selected as candidate words from all words contained in each corpus text. The target part-of-speech refers to the part-of-speech that can determine the type of event.

[0015] The system retrieves event trigger words from the symbolic knowledge base that match candidate words. Based on the retrieved event trigger words, it associates the event types connected to the event trigger words with the corpus text corresponding to the candidate words, thereby realizing the mapping between corpus text and event types and constructing a retrieval knowledge base.

[0016] Preferably, in the retrieval knowledge base, the corpus text and its mapped event subgraph containing event types form a mapping pair, wherein the corpus text serves as the key of the mapping and the event subgraph serves as the value of the mapping.

[0017] Preferably, retrieving multiple candidate texts related to the target text from the retrieval knowledge base includes:

[0018] The BM25 retrieval method retrieves multiple similar texts from the retrieval knowledge base as candidate texts for the target text.

[0019] Preferably, the step of retrieving the optimal event type for the target text based on the mapping relationship between candidate corpus text and event types includes:

[0020] The candidate text and event type are input into a binary classification model. Based on the probability value of the predicted candidate text as the input event type, the event type with the highest probability value is selected as the optimal event type.

[0021] Preferably, the step of extracting the event subgraph corresponding to the optimal event type and the optimal corpus text as cue knowledge includes:

[0022] Based on the optimal event type, an event subgraph is extracted from the knowledge graph. The event subgraph includes the parent and child event types associated with the optimal event type, event trigger words, and argument words.

[0023] Based on the mapping relationship between candidate corpus texts and event types, the candidate corpus text corresponding to the optimal event type is extracted as the optimal corpus text.

[0024] Preferably, the step of converting the prompt knowledge into prompt text and combining it with the target text as input text includes:

[0025] Extract the optimal event type and its associated superior and subordinate event types from the event subgraph to form structured event structure information;

[0026] Extract the optimal event type and its associated event trigger words from the event subgraph to form structured trigger word information;

[0027] Extract the optimal event type and its associated argument words from the event subgraph to form structured argument template information;

[0028] The event structure information, trigger word information, and argument template information are converted into text information, which, along with the optimal corpus text, serves as the prompt text.

[0029] Preferably, the symbolic knowledge base includes a schema graph or a knowledge graph.

[0030] Preferably, Standardford NLP is used to segment each corpus text.

[0031] Preferably, the semantic graph is a WordNet semantic graph.

[0032] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0033] By constructing a retrieval knowledge base, external knowledge represented by text and external knowledge represented by symbols are hybridized and mapped. This retrieval knowledge base, being relevant to the event extraction task, contains less noisy knowledge and is more suitable as cue knowledge. Based on this retrieval knowledge base, a hybrid retrieval of candidate text and event types can automatically filter out cue knowledge relevant to the event extraction task. Since the cue knowledge is obtained from the retrieval knowledge base, the amount of data computation is reduced, lowering computational costs. The extracted cue knowledge serves as part of the input to the language model, enhancing the language model's ability to extract knowledge from target text and improving the accuracy and efficiency of event extraction results. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the knowledge acquisition method based on neural symbol hybrid retrieval enhancement provided in the embodiment. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0037] Figure 1 This is a flowchart illustrating the knowledge acquisition method based on neural symbol hybrid retrieval enhancement provided in the embodiment. For example... Figure 1

[0038] As shown, the knowledge acquisition method provided in this embodiment includes the following steps:

[0039] Step 1: Obtain the event corpus, semantic graph, and symbolic knowledge base related to the event extraction task.

[0040] In this embodiment, the event extraction task refers to extracting words or phrases related to events from the input language text using a language model. The extracted words or phrases constitute an event. The event corpus refers to a corpus describing events, containing multiple event data. This event corpus is unstructured data, represented by text, i.e., the event corpus text. The semantic graph refers to a graph composed of the semantics of words, which can be a WordNet semantic graph, used to obtain the semantics of words. The symbolic knowledge base is structured knowledge data, represented by a graph, which can be a schema graph or a knowledge graph. In the graph, nodes include event types, people, time, location, etc., related to the event, hierarchical categories of the event type, and event trigger words that trigger the event.

[0041] Step 2: Based on the semantic graph, map each text in the event corpus to the event type in the symbolic knowledge base using a weakly supervised method to construct a retrieval knowledge base.

[0042] In this embodiment, when constructing the retrieval knowledge base, firstly, each event corpus text in the event corpus is segmented to obtain each word and its corresponding part-of-speech tag. Each word is then mapped to a semantic graph to obtain its semantic meaning. Specifically, Stanford NLP can be used to segment the event corpus text to obtain each word and its corresponding part-of-speech tag. Each word is then input into the WordNet semantic graph to obtain its meaning.

[0043] Then, words with the target part of speech (PPS) are selected as candidate words from all words in each corpus text based on their PPS. The target PPS refers to the PPS that can determine the event type. Specifically, each word in each corpus text is iterated through, and words are selected based on their PPS. A word is considered a candidate word if its PPS can determine the event type. It should be noted that the target PPS is related to the event type; the target PPS differs depending on the event type. Generally, verbs or nouns are more likely to be target PPSs that can determine the event type, while prepositions, pronouns, quantifiers, adjectives, and adverbs are less likely to be target PPSs.

[0044] Next, event trigger words from the symbolic knowledge base are retrieved to match candidate words. Based on the retrieved event trigger words, the event types connected to the event trigger words are associated with the corresponding corpus text of the candidate words, thus mapping the corpus text to event types and constructing the retrieval knowledge base. Specifically, by calculating the relevance between candidate words and event trigger words in the symbolic knowledge base, the most relevant event trigger words are selected as the retrieved event trigger words, and the event types connected to the event trigger words are associated with the corresponding corpus text of the candidate words. In this way, weakly supervised labels are matched to the corpus text in a weakly supervised manner, that is, the associated event types serve as the weakly supervised labels of the corpus text. At the same time, a pointer from the corpus text to the event types is constructed between the corpus text and the event types to realize the mapping between the corpus text and the event types, thereby constructing the retrieval knowledge base.

[0045] In the retrieval knowledge base, a mapping is formed between the corpus text and its mapped event subgraph containing event types. The corpus text serves as the key, and the event subgraph serves as the value. This method manages the retrieval knowledge base, which simultaneously contains symbolic and textual knowledge. In this embodiment, an event subgraph for an event type refers to an event subgraph containing events related to that event type. The event subgraph includes hierarchical event types associated with the event type, event trigger words, and argument words. Argument words include people, places, times, and tools related to the event type.

[0046] Step 3: When extracting events from the target text, retrieve multiple relevant candidate texts from the retrieval knowledge base. Based on the mapping relationship between the candidate texts and event types, retrieve the optimal event type for the target text and extract the event subgraph and the optimal text corresponding to the optimal event type as cue knowledge.

[0047] In this embodiment, when using a language model to extract events from the target text, the cue knowledge of the target text is first extracted from the retrieval knowledge base, specifically including:

[0048] First, the BM25 retrieval method is used to retrieve multiple similar corpus texts from the retrieval knowledge base based on the key as the target text, serving as candidate corpus texts. However, the results obtained solely from corpus text retrieval often contain a lot of noisy knowledge, and the retrieved corpus texts may not be suitable for the event extraction task. Therefore, it is also necessary to perform a retrieval of symbolic event subgraphs.

[0049] Then, after obtaining the candidate text corpus, the optimal event type is retrieved for the target text based on the mapping relationship between the candidate text corpus and event types. In specific implementation, the candidate text corpus and the event types in the retrieval knowledge base are input into a binary classification model. Based on the probability value of the predicted candidate text corpus being the input event type, the event type with the highest probability value is selected as the optimal event type.

[0050] Finally, the event subgraph corresponding to the optimal event type and the optimal corpus text are extracted as cue knowledge. Specifically, an event subgraph is extracted from the knowledge graph based on the optimal event type. This event subgraph includes the hierarchical event types associated with the optimal event type, event trigger words, and argument words. Argument words include people, locations, times, and tools related to the optimal event type. Candidate corpus text corresponding to the optimal event type is extracted as the optimal corpus text based on the mapping relationship between candidate corpus text and event types. The resulting event subgraph and optimal corpus text together serve as cue knowledge. This cue knowledge obtained through hybrid detection is more consistent with event extraction.

[0051] For example, such as Figure 1 As shown, for the target text "Hariri submitted his resignation during a 10-minute meeting with the head of state at the Baabda presidential palace", three candidate texts were obtained through BM25 retrieval. Each candidate text was then binary classified with the event types End-position, Meet, and Phone-Write to select the optimal event type as Meet, and the event subgraph corresponding to the Meet type was extracted.

[0052] Step 4: After converting the cue knowledge into cue text, combine it with the target text as input text, and use a language model to acquire knowledge from the input text to obtain event extraction results based on the target text.

[0053] In this embodiment, after retrieving the prompt knowledge, it is necessary to convert the prompt knowledge into prompt text to facilitate its integration into the language model. Specifically, this includes: extracting the optimal event type and its associated hierarchical event types from the event subgraph to form structured event structure information; extracting the optimal event type and its associated event trigger words from the event subgraph to form structured trigger word information; extracting the optimal event type and its associated argument words from the event subgraph to form structured argument template information; and converting the event structure information, trigger word information, and argument template information into text information, which, along with the optimal corpus text, serves as the prompt text.

[0054] In this embodiment, the converted prompt text and the target text are used together as input text. A language model is used to acquire knowledge from the input text to obtain event extraction results based on the target text. This prompt text can guide the extraction of event information described in the target text within the language model.

[0055] For example, such as Figure 1 As shown, the event structure information includes structured "Meet" and "Contact," the trigger word information includes structured "Meet," "met," and "meeting," and the argument template information includes structured "Meet," "PER (person)," and "LOC (location)." This information is converted into text and used as cue text along with the optimal corpus text. The resulting event extraction includes: event trigger word: Meeting; participant I: Hariri; participant II: head of state; location: Palace.

[0056] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge acquisition method based on neural symbol hybrid retrieval enhancement, characterized in that, Includes the following steps: Acquire event corpora, semantic graphs, and symbolic knowledge bases related to the event extraction task; Based on semantic graphs, a weakly supervised mapping method is used to map each text in the event corpus to event types in a symbolic knowledge base to construct a retrieval knowledge base. This includes: segmenting each text in the event corpus to obtain each word and its corresponding part of speech; mapping each word to a semantic graph to obtain the semantics of each word; selecting words with target parts of speech as candidate words from all words contained in each text, where the target part of speech refers to the part of speech that can determine the event type; retrieving event trigger words from the symbolic knowledge base to match the candidate words; and associating the event types connected to the event trigger words with the corresponding text in the text, thereby mapping the text in the corpus to the event types and constructing the retrieval knowledge base. When performing event extraction for target text, multiple candidate texts related to the target text are retrieved from the retrieval knowledge base. Based on the mapping relationship between candidate texts and event types, the optimal event type is retrieved for the target text, and the event subgraph and the optimal text corresponding to the optimal event type are extracted as cue knowledge. After converting the cue knowledge into cue text, it is combined with the target text as input text. A language model is then used to acquire knowledge from the input text, resulting in event extraction results based on the target text.

2. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, In the retrieval knowledge base, a mapping is formed between the corpus text and its mapped event subgraph containing event types, where the corpus text serves as the key of the mapping and the event subgraph serves as the value of the mapping.

3. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, The process of retrieving multiple candidate texts related to the target text from a retrieval knowledge base includes: The BM25 retrieval method retrieves multiple similar texts from the retrieval knowledge base as candidate texts for the target text.

4. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, The method of retrieving the optimal event type for the target text based on the mapping relationship between candidate corpus text and event types includes: The candidate text and event type are input into a binary classification model. Based on the probability value of the predicted candidate text as the input event type, the event type with the highest probability value is selected as the optimal event type.

5. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, The extraction of the event subgraph corresponding to the optimal event type and the optimal corpus text as cue knowledge includes: Based on the optimal event type, an event subgraph is extracted from the knowledge graph. The event subgraph includes the parent and child event types associated with the optimal event type, event trigger words, and argument words. Based on the mapping relationship between candidate corpus texts and event types, the candidate corpus text corresponding to the optimal event type is extracted as the optimal corpus text.

6. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 4, characterized in that, The process of converting the prompt knowledge into prompt text and combining it with the target text as input text includes: Extract the optimal event type and its associated superior and subordinate event types from the event subgraph to form structured event structure information; Extract the optimal event type and its associated event trigger words from the event subgraph to form structured trigger word information; Extract the optimal event type and its associated argument words from the event subgraph to form structured argument template information; The event structure information, trigger word information, and argument template information are converted into text information, which, along with the optimal corpus text, serves as the prompt text.

7. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, The symbolic knowledge base includes a schema graph or knowledge graph.

8. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, Standford NLP was used to segment each text in the corpus.

9. The knowledge acquisition method based on neural symbol hybrid retrieval enhancement according to claim 1, characterized in that, The semantic graph is the WordNet semantic graph.

Images