Method for constructing prospecting prediction question and answer recommendation model

Abstract

The invention relates to the technical field of question and answer model construction, and particularly discloses a prospecting prediction question and answer recommendation model construction method, which comprises modules for constructing a prospecting prediction question and answer recommendation model, a question text input module, a question text receiving and analysis answering module, a question text analysis answering module, a question text analysis answering module and a question text analysis answering module, the big language model is injected into a question and answer database composed of a geological knowledge graph, receives a question text and obtains a question and answer recommendation text based on the geological knowledge graph, and the question and answer result output module receives the question and answer recommendation text and outputs the question and answer recommendation text corresponding to the question text. The problem that related knowledge is difficult to integrate and infer effectively when complex geological data is processed by a retrieval enhancement generation model carried in a traditional large language model is solved.

Description

Technical Field

[0001] The present application relates to the technical field of question-answering model construction, and specifically discloses a method for constructing a mineral prospecting prediction question-answering recommendation model. Background Art

[0002] Question-and-answer recommendations can help users quickly access the information they need through natural language interaction, and building a knowledge base is a common method for building question-and-answer recommendation models. In the field of geology, mineral prospecting and prediction questions involve complex domain knowledge such as ore deposit characteristics, geological structure, and mineralization. To effectively provide accurate knowledge information, structured knowledge graphs are often used to build knowledge bases.

[0003] In existing technologies, knowledge graph construction employs rule-based methods for named entity recognition and relationship extraction, and supervised classification using deep neural networks has become the mainstream approach for geological knowledge graph construction. With the increasing application of large language models in natural language processing tasks, retrieval-enhanced generative models combined with large language models are now a common technique.

[0004] In the field of geology, keyword-based retrieval techniques for querying, searching, and generating structured knowledge bases often result in retrieval cores approaching frequently repeated words. In the context of mineral exploration and prediction, frequency is not the primary factor determining key elements. While semantic-based retrieval methods can effectively map natural language interactions, traditional retrieval-enhanced generative models are often applied to unstructured document data. Due to incomplete data cleaning and the presence of significant noise within documents, they struggle to effectively integrate and infer relevant knowledge when processing complex geological data, compromising the accuracy and comprehensiveness of answers.

[0005] The present invention provides a method for constructing a prospecting prediction question-answer recommendation model to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to solve the problem that the retrieval enhancement generation model carried by the traditional large language model is difficult to effectively integrate and reason about relevant knowledge when processing complex geological data.

[0007] To achieve the above objectives, the basic solution of the present invention provides a method for constructing a prospecting prediction question-answer recommendation model, including constructing the following modules constituting the prospecting prediction question-answer recommendation model:

[0008] Question text input module: used to receive question text;

[0009] Question text reception and parsing answering module: equipped with a large language model, which is injected with a question and answer database composed of a geological knowledge graph. The large language model receives the question text and obtains the question and answer recommendation text based on the geological knowledge graph;

[0010] Question and answer result output module: receives question and answer recommendation text and outputs the corresponding question text.

[0011] Furthermore, the geological knowledge graph includes an ontology layer and a data layer, and the geological knowledge graph is expressed in triple form based on the data layer under the constraints of the ontology layer.

[0012] Furthermore, the construction and representation of the ontology layer are as follows:

[0013]

[0014] In the formula, Geo unit Represents geological field elements, The basic concept domain interpretation of geological elements, Represents the semantic relationship, attribute relationship and spatial relationship between geological elements, Represents an instance used to describe a geological feature.

[0015] Furthermore, the construction of the data layer includes the following steps:

[0016] Obtaining original geological data dataset in text format and preprocessing the original geological data dataset;

[0017] Based on the preprocessed geological data set, sentence segmentation is performed to obtain rough sentence segmentation results;

[0018] The current sentence segmentation result is corrected through regular expressions to obtain the final sentence segmentation result and complete the construction of the data layer.

[0019] Furthermore, the large language model is equipped with a triple extraction prompt word generation framework based on ontology layer constraints. The triple extraction prompt word generation framework is used as a triple prompt word template. The expression of the triple extraction prompt word generation framework is as follows:

[0020] KJ=f(Prompt, Geo unit ,Text)

[0021] In the formula, Prompt is the set of prompt words designed, Text is the geological text information, Geo unit It is the ontology layer, used as a constraint.

[0022] Furthermore, the large language model is also equipped with a question-answering framework based on a retrieval-enhanced generation model, which is used to receive question text and split and parse the question text, and then connect the parsed data to the question text and input it into the large language model to generate prompt words.

[0023] Furthermore, the steps of splitting and parsing the question text are as follows:

[0024] Convert the problem text input by the user into a problem text feature vector, and convert the geological knowledge graph into a knowledge graph feature vector representation;

[0025] Based on the cosine similarity calculation of the problem text feature vector and the knowledge graph feature vector for similarity evaluation, retrieve all triples in sequence, sort the retrieval results according to the similarity evaluation score, and select the most relevant triples in the geological knowledge graph.

[0026] Furthermore, the expression of the retrieval enhancement generation model is as follows:

[0027]

[0028] In the formula, P(a|q) represents the probability that the retrieval enhancement generation model selects knowledge d from the document set D, and P(a|q,d) represents the conditional probability that the generation model generates answer a according to the input q and the retrieved document d;

[0029] The similarity evaluation is obtained based on the relevance calculation, and the formula for the relevance calculation is as follows:

[0030]

[0031] In the formula, sim(q,d i ) represents the similarity between the query q and the text d i , N represents the number of texts in the text set, and i = 1, 2, 3,......

[0032] Furthermore, the Q&A recommendation text Output output by the large language model is obtained by guiding the chain-of-thought reasoning of the large language model based on the triple prompt word template, and the combined expression of the Q&A recommendation text Output is as follows:

[0033] Output = f(Prompt_ans, Query, Knowledge graph, Examples)

[0034] In the formula, f represents the large language model mapping function, Prompt_ans represents the Q&A reasoning guidance prompt word, Query represents the problem text, Knowledge graph represents the supplementary knowledge graph, Examples represents the few-shot example answer content, and the few-shot is made based on the geological ontology constraints and geological data statements of geological expert knowledge.

[0035] The principle and effect of this solution are as follows:

[0036] Compared with the existing technology, the present invention stores professional knowledge in the geological field in a structured manner through a knowledge graph, combines the retrieval-enhanced generation model with the large language model, and effectively combines the two through prompt words. Finally, the natural language processing capabilities of the large language model are used to provide accurate answers to professional questions in the field of mineral exploration and prediction, solving the problem that the retrieval-enhanced generation model carried by the traditional large language model is difficult to effectively integrate and infer relevant knowledge when processing complex geological data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0038] Figure 1 A schematic diagram of a method for constructing a prospecting prediction question-answering recommendation model proposed in an embodiment of the present application is shown;

[0039] Figure 2 A schematic diagram of a mineral prospecting prediction question-answering recommendation model proposed in an embodiment of the present application is shown;

[0040] Figure 3 A schematic diagram showing the ontology layer of the geological knowledge graph proposed in an embodiment of the present application is shown;

[0041] Figure 4 A schematic diagram of the recommendation model answer based on the geological knowledge map of the study area proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0042] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0043] A method for constructing a prospecting prediction question-answering recommendation model, implementing Figure 1 As shown, it includes building the following modules that constitute the prospecting prediction question-answering recommendation model:

[0044] Question text input module: used to receive question text;

[0045] Question text reception and parsing answering module: equipped with a large language model, which is injected with a question and answer database composed of a geological knowledge graph. The large language model receives the question text and obtains the question and answer recommendation text based on the geological knowledge graph;

[0046] Question and answer result output module: receives question and answer recommendation text and outputs the corresponding question text.

[0047] Finally, a prospecting prediction Q&A recommendation model is constructed by a Q&A text input module, a question text receiving, parsing and answering module, and a Q&A result output module, as Figure 2 shown.

[0048] Among them, the large language model carried is the ChatGLM-4 model.

[0049] Among them, the geological knowledge graph that composes the Q&A database is a standard semantic web that represents the knowledge related to geological ore-forming prediction in the form of triples. The geological knowledge graph includes an ontology layer and a data layer. The construction of the geological knowledge graph includes the construction of the ontology layer and the data layer, and under the constraint of the ontology layer, it is expressed in the form of triples based on the data layer. The ontology is a geological ontology, which is a formal and standardized description of geological knowledge concepts and their relationships, including monographs on ore-forming prediction and ore-forming prediction standards. The ontology layer of the geological knowledge graph is constructed using the geological ontology, and it provides a theoretical framework for the attribute decomposition of instances in the data layer during the construction of triples.

[0050] The ontology layer of the geological knowledge graph consists of concept nodes and relationship edges, representing the concepts and their mutual relationships in the geological field. The ontology layer is expressed as:

[0051]

[0052] In the formula, Geo unit represents the elements in the geological field, represents the basic concept domain interpretation of geological elements, represents the semantic relationship, attribute relationship, and spatial relationship between geological elements, represents the instances used to describe geological elements. As Figure 3 shown, in this embodiment, for 8 types of elements in the geological field, such as geography, stratigraphy, rock mass, structure, geological age, mineralization, geophysics, and geochemistry, the hierarchical relationships of each element are divided through prior knowledge, and the attribute relationships and semantic relationships are defined.

[0053] The construction of the data layer of the geological knowledge graph includes the following steps:

[0054] Obtain the original geological data dataset in text format, and preprocess the original geological data dataset. The preprocessing operations include converting the original geological data dataset into unstructured data, performing graphic, text, and table segmentation, using the original unstructured data as the coordinate system, and calculating the position according to the length and width of the document. The calculation formula is as follows:

[0055] loc = (x, y), x ∈ [0, W], y ∈ [0, H]

[0056] Wherein, W and H are the page width and height, and (x, y) are the horizontal and vertical position coordinates of the content element.

[0057] And perform threshold screening according to the coordinates of the content element to convert the original geological data in the original geological data set to obtain the undivided text string data. The expression is as follows:

[0058] |y next -y current |<ε y

[0059] |x next -x current |<ε x

[0060] Wherein, y next and x next respectively represent the string extension coordinates of the vertical text line and the horizontal paragraph, and y current and x current respectively represent the current coordinates of the string of the vertical text line and the horizontal paragraph, and ε y and ε x respectively represent the spacing thresholds in the vertical and horizontal directions.

[0061] And based on the geological data set of the undivided text string data obtained after preprocessing, perform sentence segmentation to ensure semantic integrity.

[0062] Specifically, for the text string T, each word t i is assigned a label s i , i = 1, 2, 3......, used to mark whether the word is the boundary of a sentence. The expression is as follows:

[0063]

[0064] Wherein, when s i = 1, it means that t i is the end point of the sentence, that is, the boundary of the sentence. When s i = 0, it means that t i is not the end point of the sentence, that is, not the boundary of the sentence.

[0065] And based on the following three types of rules: punctuation rule R(t i ), dependency rule D(t i ), and machine learning probability model rule P(t i ), perform constraints and threshold determination to obtain a rough sentence segmentation result:

[0066]

[0067] In the formula, Root indicates the end of the subject-predicate structure of the sentence, punct indicates that the end punctuation mark is attached to the main clause, and dep(t i ) represents the judgment of the complementary dependency relationship of the t i word, and head(t i ) represents the judgment of the head node in the syntactic dependency of the t i word.

[0068] P(t i ) = σ(W·φ(t i )) + b

[0069] In the formula, W represents the weight vector, b represents the bias, σ represents the activation function, and φ(t i ) represents the part-of-speech tagging of the word t i .

[0070] When the values of the punctuation rule R(t i ), the dependency relationship rule D(t i ), and the machine learning probability model rule P(t i ) meet the sentence marker through the logical relationship or judgment, sentence segmentation is performed.

[0071] After obtaining the rough sentence segmentation result, the current sentence segmentation result is corrected through regular expressions. Regular expressions can accurately match and operate on strings. Using the text pattern composed of ordinary characters and metacharacters, it can quickly automatically search and match the error character part in the common error pattern to ensure semantic integrity and perform text cleaning to eliminate the influence of common Chinese parentheses, digital characters, irrelevant symbols, and spaces in the text in the above rule judgment, and obtain the final sentence segmentation result.

[0072] When constructing the triple, first, fuzzy keyword matching needs to be performed under the ontology layer constraint to match keywords and the geological text information related to the description of geological information. Specifically, for fuzzy keyword matching, it is calculated based on the Levenshtein distance. By performing fuzzy similarity matching between the sentence segmentation result and the geological concepts constrained by the ontology layer to obtain the context of relevant knowledge content, the calculation formula is as follows:

[0073]

[0074] In the formula, d(i, j) represents the edit distance between the first i characters of the string s1 and the first j characters of s2, j = 1, 2, 3......, and δ(sr[i], s2[j]) is used to represent whether two characters are equal. If they are equal, it is 1, and if they are not equal, it is 0.

[0075] After that, a triple prompt template is established, and the knowledge paragraphs matched by keywords are injected into the triple prompt template to effectively guide the large language model to extract triples during the question-answering process.

[0076] The large language model is equipped with a triple extraction prompt generation framework based on ontology layer constraints as the triple prompt template. The construction of the geological knowledge graph is a function mapping of geological text information in the large language model. Through the function mapping f as the triple extraction prompt generation framework, the expression is as follows:

[0077] KJ = f(Prompt, Geo unit , Text)

[0078] In the formula, Prompt is the designed set of prompts, Text is the geological text information, and Geo unit is the ontology layer, which is used for constraint.

[0079] Inject the geological text information matched by keywords into the triple prompt template to effectively guide the large language model to extract triples and form a geological knowledge graph. During the triple extraction process, the GLM-4-9B large language model is used to process long text information and quantize it using BF16 precision. The final results are shown in Table 1:

[0080] Table 1 Partial triple data of the geological knowledge graph

[0081]

[0082]

[0083] In each triple, the geological entities are represented by nodes, and the relationships between concepts are connected by edges. Multiple triples are connected by relationships, and finally a geological knowledge network and a geological knowledge graph are formed. This network is represented as:

[0084] ZSWL = (V, E, T)

[0085] In the formula, V represents the set of nodes, E represents the set of edges, and T represents the set of triples composed of nodes and edges.

[0086] The large language model is also equipped with a question-answering framework based on the retrieval-augmented generation model, which is used to receive the question text, split and parse the question text, and input the parsed data together with the question text into the large language model to generate prompts, helping the large language model determine the structure, role, and output limitations of the answer.

[0087] The retrieval-augmented generation model combines the text embedding model and the generation model, and uses the constructed geological knowledge graph as an external knowledge base to supplement the implicit knowledge of the large language model.

[0088] The specific steps for splitting and analyzing the problem text are as follows:

[0089] Convert the problem text input by the user into a problem text feature vector through a text embedding model, and convert the geological knowledge graph into a knowledge graph feature vector;

[0090] In the embedding vector space, based on the calculation of the cosine similarity between the problem text feature vector and the knowledge graph feature vector, perform similarity evaluation, retrieve all triples in sequence, sort the retrieval results according to the similarity evaluation score, and select the most relevant triples in the geological knowledge graph.

[0091] The expression of the retrieval-enhanced generation model is as follows:

[0092]

[0093] In the formula, P(a|q) represents the probability that the retrieval-enhanced generation model selects the knowledge text d from the document set D, and P(a|q,d) represents the conditional probability that the generation model generates the answer text a based on the input problem text q and the retrieved knowledge text d.

[0094] In this implementation, through the embedding text embedding model, perform vector embedding and correlation calculation of the problem text feature vector and the knowledge graph feature vector, perform similarity evaluation, and the calculation formula is as follows:

[0095]

[0096] In the formula, sim(q,d i ) represents measuring the similarity between the query q and the text d i , N represents the number of texts in the text set, and i = 1, 2, 3......

[0097] And sort all the triples after retrieval, and select the top k most relevant triples as the retrieval results.

[0098] In the process of prospecting prediction Q&A, based on the most relevant triples retrieved, and combined with the triple prompt word template, guide the chain of thought reasoning of the large language model, quickly locate the knowledge text position by matching the geological ontology constraint concepts and geological data texts, expand paragraphs based on the context of the obtained relevant knowledge content to ensure the semantic integrity of the context, and set the token length to be less than or equal to 8000 to ensure the rapid iterative reasoning of the large language model.

[0099] During the process of answering questions for prospecting prediction, the chain of thought reasoning of the large language model is guided by the triple prompt template, and finally the combined expression of the Q&A recommended text output by the large language model is as follows:

[0100] Output=f(Prompt_ans,Query,Knowledge graph,Examples)

[0101] In the formula, Output is the Q&A recommended text output by the large language model, f represents the mapping function of the large language model, Prompt_ans represents the prompt for guiding Q&A reasoning, Query represents the question text, Knowledge graph represents the supplementary knowledge graph, and Examples represents the content of few-shot example answers.

[0102] Specifically, guiding the chain of thought reasoning of the large language model by using the triple prompt template can be divided into three steps, including:

[0103] Step 1: Describe the task and define the role, and clarify the task goal and role division. Specifically, use the question text about prospecting prediction proposed by the user as {Query} and the role definition description to enable the large language model to call geological background knowledge.

[0104] Step 2: Use the retrieval-enhanced generation model to obtain triple entities and relationships, including providing semantic matching by reconstructing the question text, obtaining an optimized question expression, and obtaining highly relevant fact triples based on this. The specific steps are as follows:

[0105] Question reconstruction: Tokenize, remove stop words, calculate word frequencies and sort the question text {Query} to generate an optimized question expression. The reconstruction process expression is as follows:

[0106] Query'=Top(Count(Tokenize(Query)\S))

[0107] In the formula, Tokenize represents the function of tokenizing words, that is, dividing words according to the pre-trained tool, S is the set of stop words, Count is the frequency calculation, that is, calculating the number of occurrences of each tokenized word, and Top is the sorting calculation, that is, counting the most frequently occurring tokens.

[0108] Triple acquisition and knowledge graph injection. Based on the optimized question formulation, the similarity between the question formulation and all triples is calculated through the text embedding model embedded in the retrieval-augmented generation model to obtain fact triples highly relevant to the question formulation. And based on the nodes and relationships of the fact triples, I rounds of multi-round jump relationship expansion are set. The fact triples that meet the thresholds of different rounds and the constraints of the geological ontology and avoid the relationship thresholds of the nodes are converted through natural language and then added to the prompt parameter {Knowledge graph}. The process expression is as follows:

[0109] score = max(sim(Query′, s), sim(Query′, r), sim(Query′, o))

[0110] T = {(s, r, o)|socre i >thresholds i}

[0111] In the formula, score represents the scoring calculated by the text embedding model, sim represents the similarity calculation, Query' represents the reconstructed question, s, r, and o respectively represent the subject, relationship, and object of the fact triple, T represents the knowledge graph to be injected, and thresholds i represents the threshold judgment rule set for the i-th round of jump.

[0112] The content of the knowledge graph to be injected, which has been screened and jump-expanded, covers the knowledge domain required by the question text {Query} and is injected into the prompt to ensure that the large language model can access the most relevant geological knowledge.

[0113] And few-shot examples are pre-constructed based on the constraints of the geological ontology and geological data statements based on geological expert knowledge. Based on the few-shot examples, the large language model is guided to answer the question text, and the few-shot examples are added to the prompt parameter {Examples} to guide the answer to related concepts.

[0114] Finally, the large language model outputs the text content with paragraph expansion as the Q&A recommended text from the perspective of a geological expert.

[0115] In this embodiment, based on the geological exploration data of a certain tungsten ore area, a knowledge graph in the large language model is constructed, and the geological ontology design is determined according to the experience of geological experts, and the Q&A process is carried out as Figure 4 shown:

[0116] Question: What strata exist in the Maowan area and what are the lithologies of these strata?

[0117] After the role definition, problem reconstruction, triple acquisition, knowledge graph injection, and few-shot example sentence filling are determined by the large language model, the question-and-answer recommendation text as shown in Figure 4 is obtained.

[0118] The present invention uses a large language model to replace the supervised classification algorithm of the traditional deep neural network, effectively avoiding the manual labeling of the inefficient original data set, and guiding the generation work of the large language model in the vertical field through prompt engineering, solving the selection of data and model for the construction of the geological knowledge graph.

[0119] The present invention uses fuzzy keyword matching and a geological ontology constraint framework for the auxiliary design of prompt engineering. Through the geological ontology constraint, the large language model can effectively focus on the text, thereby effectively improving the interference of redundant information and enhancing the understanding of geological knowledge. At the same time, using the content defined by the geological ontology for fuzzy keyword matching can effectively reduce the text content input into the large language model, which not only reduces the difficulty of text analysis by the large language model but also retains the key geological data text information.

[0120] The present invention uses a prompt generation framework composed of expert experience to enhance the logical analysis ability of the large language model in the field of geological information extraction. By splitting the prompt engineering for the construction of the geological knowledge graph, the defect of insufficient reasoning of the large language model in the geological field is effectively avoided.

[0121] The present invention structurally stores the professional knowledge in the geological field through a knowledge graph, combines a retrieval-augmented generation model and a large language model, effectively combines the two through prompts, and finally uses the natural language processing ability of the large language model to provide accurate answers to professional problems in the field of prospecting prediction.

[0122] The above is only a preferred embodiment of the present invention and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content within the scope of the technical solution of the present invention to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for constructing a prospecting prediction question-and-answer recommendation model, characterized in that Including modules for constructing a prospecting prediction Q&A recommendation model as follows: Question text input module: Used to receive question text; Question text receiving and parsing answer module: Equipped with a large language model, and the large language model is injected with a Q&A database composed of a geological knowledge graph. The large language model receives the question text and obtains Q&A recommendation text based on the geological knowledge graph; Q&A result output module: Receives the Q&A recommendation text and outputs it corresponding to the question text.

2. The method for constructing a prospecting prediction Q&A recommendation model according to claim 1, wherein, The geological knowledge graph includes an ontology layer and a data layer, and the geological knowledge graph is expressed in the form of triples based on the data layer under the constraint of the ontology layer.

3. A method for constructing a prospecting prediction Q&A recommendation model according to claim 2, characterized in that, The construction and representation of the ontology layer are as follows: In the formula, Geo unit represents elements in the geological field, represents the interpretation of the basic conceptual domain of geological elements, represents the semantic, attribute, and spatial relationships between geological elements, represents instances used to describe geological elements.

4. A method for constructing a prospecting prediction Q&A recommendation model according to claim 2, characterized in that, The construction of the data layer includes the following steps: Obtain the original geological data dataset in text format and preprocess the original geological data dataset; Based on the preprocessed geological data dataset, perform sentence segmentation to obtain a rough sentence segmentation result; Correct the current sentence segmentation result through regular expressions to obtain the final sentence segmentation result, completing the construction of the data layer.

5. A method for constructing a prospecting prediction Q&A recommendation model according to claim 2, characterized in that, The large language model is equipped with a triple extraction prompt word generation framework based on the constraints of the ontology layer. The triple extraction prompt word generation framework is used as a triple prompt word template, and the expression of the triple extraction prompt word generation framework is as follows: KJ = f(Prompt, Geo unit , Text) where Prompt is the set of design prompt words, Text is the geological text information, and Geo unit is the ontology layer and is used as a constraint.

6. A method for constructing a prospecting prediction Q&A recommendation model according to claim 2, characterized in that The large language model is also equipped with a Q&A framework based on the retrieval enhanced generation model, which is used to receive the question text, split and parse the question text, and input the parsed data together with the question text into the large language model to generate prompt words.

7. A method for constructing a prospecting prediction Q&A recommendation model according to claim 6, characterized in that, The splitting and parsing steps of the question text are as follows: Convert the question text input by the user into a question text feature vector, and convert the geological knowledge graph into a knowledge graph feature vector representation; Based on the cosine similarity calculation of the question text feature vector and the knowledge graph feature vector, perform similarity evaluation, retrieve all triples in sequence, sort the retrieval results according to the similarity evaluation score, and select the most relevant triples in the geological knowledge graph.

8. A method for constructing a prospecting prediction Q&A recommendation model according to claim 7, characterized in that The expression of the retrieval enhanced generation model is as follows: In the formula, P(a|q) represents the probability that the retrieval enhanced generation model selects knowledge d from the document set D, and P(a|q,d) represents the conditional probability that the generation model generates answer a according to the input q and the retrieved document d; The similarity evaluation is obtained based on the correlation calculation, and the correlation calculation formula is as follows: where sim(q, d i ) represents measuring the similarity between the query q and the text d i , N represents the number of texts in the text set, and i = 1, 2, 3,......

9. A method for constructing a prospecting prediction Q&A recommendation model according to claim 5, characterized in that, The Q&A recommendation text Output output by the large language model is obtained by guiding the chain of thought reasoning of the large language model based on the triple prompt word template. The combined expression of the Q&A recommendation text Output is as follows: Output = f(Prompt_ans, Query, Knowledge graph, Examples) Wherein, f represents the mapping function of the large language model, Prompt_ans represents the guiding prompt for question and answer reasoning, Query represents the question text, Knowledge graph represents the supplementary knowledge graph, and Examples represents the example answer content of few-shot, and the few-shot is obtained by making the geological ontology constraints and geological data statements based on geological expert knowledge.

Images