Intelligent place name translation method and system based on Lora fine tuning and knowledge graph

Abstract

The invention discloses an intelligent geographical name translation method based on Lora fine tuning and a knowledge graph, and relates to the technical field of intelligent geographical name translation. The method comprises the following steps: inputting an English geographical name, and obtaining attribute and relation information through a geographical name translation knowledge graph; selecting a cue word template according to the attribute information, and filling associated information to generate an instantiated cue word; and inputting the instantiated cue word into a geographical name translation big language model, and outputting a Chinese geographical name translation result through context understanding by the model through geographical name translation data set training and Lora parameter fine tuning. According to the method, by constructing the cue word template, the translation process is standardized, and modular translation is achieved; by constructing the knowledge graph, the problem that a large language model lacks geographical name knowledge is solved; and efficient and accurate translation of the place name is realized through Lora fine tuning.

Description

Technical Field

[0001] This application relates to the technical field of intelligent place name translation, and particularly to an intelligent place name translation method based on LoRA fine-tuning and knowledge graph. Background Art

[0002] With the increasing demand for globalized geographical information interaction and bilingual map production, place name translation, as the core link of geographical information standardization, its accuracy and efficiency directly affect the usability of cross-language geographical data.

[0003] Currently, place name translation technologies are mainly divided into two categories: traditional pipeline methods and preliminary applications based on large language models. Both types of technologies have significant bottlenecks in practical applications and are difficult to meet the high-precision and high-efficiency requirements of professional scenarios (such as bilingual map production and global geographical information resource integration). For example, the traditional pipeline place name translation method misidentifies "Palms" in "Palms Restaurant" as a common name, resulting in a translation that deviates from the correct result of "Palms Restaurant"; at the same time, this method has weak processing ability for derived place names and cannot effectively capture the semantic associations of derived place names. For example, for "Jackson Park" and "Jackson Park Road", the former is translated as "Jackson Park" and the latter as "Jackson Road", destroying the关联性 between geographical entities and causing confusion in scenarios such as bilingual maps. In addition, traditional methods rely on manually designed rules and a single dictionary and lack the ability to semantically constrain common names with multiple meanings (such as "Course" can mean "route" or "venue"), and are prone to free translation errors when there is no clear geographical category information, further reducing translation accuracy.

[0004] In recent years, with the rapid development of large language models (such as GPT and ChatGLM) in the field of natural language processing, some studies have tried to directly apply them to place name translation tasks in order to improve translation effects by using the context understanding ability of the models. However, unoptimized large language models have two core defects: one is that there are still translation deviations when processing professional place names. For example, it is unable to accurately judge the meaning of "venue" of "Course" in "Golf Course", and there are still common name translation deviations; the other is that the model training cost is high and the inference efficiency is difficult to meet the batch processing requirements of a large amount of place name data. In this context, there is an urgent need for an English-Chinese place name translation technology with "low error, high efficiency, and strong adaptability" in scenarios such as bilingual map production and global geographical information standardization. There is an urgent need for a translation method that can avoid error propagation, integrate place name professional knowledge, and balance performance and cost to solve the multiple pain points of existing technologies in terms of accuracy, efficiency, and scenario adaptability, and promote the upgrade of English-Chinese place name translation from "stage-by-stage processing" to "joint intelligent translation", providing technical support for cross-language geographical information interaction. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for intelligent translation of place names based on LoRa fine-tuning and knowledge graphs, which can achieve efficient and accurate translation of place names.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] Firstly, this application provides a method for intelligent translation of place names based on LoRa fine-tuning and knowledge graphs, including the following steps.

[0008] Receives English place names input by the user.

[0009] The English place name is sent to the place name translation knowledge graph to retrieve the attribute information and relationship information corresponding to the English place name.

[0010] Select the corresponding prompt word template based on the attribute information corresponding to the English place name, and generate an instantiated prompt word based on the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template; the prompt word template is an integrated template of proper noun transliteration and common noun translation.

[0011] The instantiated prompt words are input into a large-scale language model for place name translation, and the Chinese place name translation result is output through contextual understanding and semantic mapping. The large-scale language model for place name translation is a large-scale language model trained on a general domain based on a place name translation training dataset, and the Lora parameters are fine-tuned. The Chinese place name translation result is a structured intermediate translation result, which includes at least the final translated name field, and further includes at least one of the proper name transliteration field and the generic name semantic translation field.

[0012] Determine whether the Chinese place name translation result meets the consistency verification standard to obtain the first judgment result; if the first judgment result is yes, output the structured intermediate translation result as the Chinese place name translation result; if the first judgment result is no, jump to the step "select the corresponding prompt word template according to the attribute information corresponding to the English place name, and generate an instantiated prompt word according to the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template".

[0013] Secondly, this application provides a place name intelligent translation system based on LoRa fine-tuning and knowledge graph, including the following functional modules.

[0014] The English place name receiving module is used to receive English place names input by the user.

[0015] The knowledge graph retrieval module is used to send the English place names to the place name translation knowledge graph to retrieve the attribute information and relationship information corresponding to the English place names.

[0016] The prompt word generation module is used to select the corresponding prompt word template according to the attribute information corresponding to the English place name, and generate an instantiated prompt word based on the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template; the prompt word template is an integrated template of proper noun transliteration and common noun translation.

[0017] The English place name translation module is used to input the instantiated prompt words into the place name translation language model, and output the Chinese place name translation result through context understanding and semantic mapping. The place name translation language model is a large language model trained on a general domain based on a place name translation training dataset, and the Lora parameters are fine-tuned. The Chinese place name translation result is a structured intermediate translation result, which includes at least the final translated name field, and further includes at least one of the proper name transliteration field and the generic name semantic translation field.

[0018] The consistency verification module is used to determine whether the Chinese place name translation result meets the consistency verification standard and obtain a first judgment result; when the first judgment result is yes, the structured intermediate translation result is output as the Chinese place name translation result; when the first judgment result is no, the prompt word generation module is called.

[0019] According to the specific embodiments provided in this application, this application has the following technical effects.

[0020] This application provides an intelligent English-Chinese place name translation method based on LoRa fine-tuning and semantic enhancement. By constructing an integrated template for proper noun transliteration and common noun semantic translation through place name translation prompt word templates and place name translation knowledge graphs, it fundamentally solves the error propagation problem of traditional technologies and achieves a significant improvement in the accuracy of English-Chinese place name translation. Through LoRa fine-tuning of a large language model, it solves the problem of excessive time consumption for additional data cleaning and information retrieval, and achieves efficient connection of the overall process. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is an application environment diagram of a place name intelligent translation method based on LoRa fine-tuning and knowledge graph in one embodiment of this application.

[0023] Figure 2This is a flowchart illustrating a place name intelligent translation method based on LoRa fine-tuning and knowledge graph, provided as an embodiment of this application.

[0024] Figure 3 This is a schematic diagram of the functional modules of a place name intelligent translation system based on LoRa fine-tuning and knowledge graph, provided as an embodiment of this application.

[0025] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] The place name intelligent translation method based on LoRa fine-tuning and knowledge graph provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 allows users to input English place names and send them to server 104. Upon receiving the input English place name, server 104 sends it to a place name translation knowledge graph. It retrieves the corresponding attribute and relationship information of the English place name from the knowledge graph. Based on the attribute information, server 104 selects a corresponding prompt word template and generates an instantiated prompt word based on the attribute and relationship information of the English place name, combined with the corresponding prompt word template. The prompt word template is an integrated template for proper noun transliteration and generic noun translation. The instantiated prompt word is input into a place name translation language model, which outputs the Chinese place name translation result through contextual understanding and semantic mapping. The place name translation language model is a large language model trained on a general domain based on a place name translation training dataset, with fine-tuning of the LoRa parameters. Server 104 can then feed back the obtained Chinese place name translation result to terminal 102. In addition, in some embodiments, the place name intelligent translation method based on LoRa fine-tuning and knowledge graph can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly perform corresponding operations on the English place names input by the user, or the server 104 can obtain the English place names input by the user from the data storage system and perform corresponding operations on the input English place names.

[0029] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0030] In one exemplary embodiment, such as Figure 2 As shown, a method for intelligent place name translation based on LoRa fine-tuning and knowledge graph is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 205.

[0031] Step 201: Receive the English place name entered by the user.

[0032] Step 202: Send the English place name to the place name translation knowledge graph to retrieve the attribute information and relationship information corresponding to the English place name.

[0033] Step 203: Select the corresponding prompt word template based on the attribute information corresponding to the English place name, and generate an instantiated prompt word based on the attribute and relationship information corresponding to the English place name, combined with the corresponding prompt word template; the prompt word template is an integrated template for proper noun transliteration and common noun translation. In this embodiment, the prompt word template is a structured generation instruction template for constructing a place name translation task.

[0034] Step 204: Input the instantiated prompt words into the place name translation language model, and output the Chinese place name translation result through contextual understanding and semantic mapping; the place name translation language model is a large language model obtained by training a general domain language model based on the place name translation training dataset and fine-tuning the Lora parameters. The Chinese place name translation result is a structured intermediate translation result, which includes at least the final translated name field, and further includes at least one of the proper name transliteration field and the generic name semantic translation field.

[0035] Step 205: Determine whether the Chinese place name translation result meets the consistency verification standard to obtain a first judgment result; if the first judgment result is yes, output the structured intermediate translation result as the Chinese place name translation result; if the first judgment result is no, proceed to step 203. In this embodiment, the consistency verification standard includes one or more of the following verification items: field integrity verification; consistency verification between the generic name translation field and the geographical category; consistency verification between the proper name transliteration field and the preset transliteration rules or transliteration comparison table; consistency verification between the final translated name field and the combination rules of the proper name transliteration field and the generic name translation field; and matching or similarity threshold verification between the final translated name field and the target language translation / alias record in the place name translation knowledge graph.

[0036] Implementing steps 201 to 205 effectively avoids the error propagation and accumulation problems commonly found in traditional translation techniques, and significantly improves accuracy and consistency in place name translation, thereby significantly optimizing the overall accuracy of English-Chinese place name translation. Furthermore, this application can further enhance terminology consistency and contextual adaptability in multilingual scenarios, thus providing more reliable technical support for cross-lingual geographic information processing.

[0037] In another exemplary embodiment of this application, before step 204 described above, the following steps 301 to 304 are also included.

[0038] Step 301: Obtain the English-Chinese place name translation supervision dataset; the English-Chinese place name translation supervision dataset includes English place names and Chinese reference translations; the English place names are common place names or derived place names.

[0039] Step 302: Preprocess and divide the English-Chinese place name translation supervision dataset to obtain a place name translation training dataset, a place name translation test dataset, and a place name translation verification dataset. The place name translation training dataset includes several training samples. Each training sample includes an English place name, its corresponding attribute information, and a Chinese reference translation. When the English place name is a common place name, the attribute information corresponding to the English place name includes a geographical category. When the English place name is a derived place name, the attribute information corresponding to the English place name includes a geographical category, the derived part, and the original geographical category.

[0040] Step 303: Based on the translation characteristics of the common place name and the derived place name, construct prompt word templates respectively, and configure structured output specifications for the prompt word templates, so that the place name translation big language model generates structured intermediate translation results according to preset fields; the structured output specifications define at least the final translation field, and further define at least one of the proper name transliteration field, common name semantic translation field, and derived element translation field.

[0041] Step 304: Based on the place name translation training dataset, combined with the prompt word template and the structured output specification, train the general domain large language model, and fine-tune the Lora parameters to obtain the place name translation large language model.

[0042] In another exemplary embodiment of this application, step 302 is replaced by steps 401 to 402.

[0043] Step 401: Remove invalid data using regular expressions and manually verify the accuracy of place name category labeling.

[0044] Step 402: Convert all data into a structure of "English place name - attribute information - Chinese reference translation", and divide it into place name translation training dataset, place name translation test dataset and place name translation verification dataset according to preset division rules.

[0045] In another exemplary embodiment of this application, step 402 is replaced by steps 501 to 505.

[0046] Step 501: When the English place name is a common place name, the English-Chinese place name translation supervision dataset is divided into a common place name translation training set, a common place name translation verification set, and a common place name translation test set in a ratio of 8:1:1.

[0047] Step 502: When the English place name is a derived place name, select an equal amount of derived place name data and ordinary place name data from the English-Chinese place name translation supervision dataset, and divide them into a derived place name translation training set, a derived place name translation verification set, and a derived place name translation test set according to a ratio of 7:1.5:1.5.

[0048] Step 503: Combine the ordinary place name translation training set and the derived place name translation training set into a place name translation training dataset.

[0049] Step 504: Combine the ordinary place name translation test set and the derived place name translation test set into a place name translation test dataset.

[0050] Step 505: Combine the ordinary place name translation verification set and the derived place name translation verification set into a place name translation verification dataset.

[0051] Specifically, in this embodiment, the ordinary place name dataset consists of 20,000 entries, divided into a training set (16,000 entries), a validation set (2,000 entries), and a test set (2,000 entries) in a ratio of 8:1:1; the derived place name dataset consists of 10,069 entries containing derivation relationship data (positive examples) and an equal amount of ordinary place name data (negative examples), divided into a training set (7,048 entries), a validation set (1,510 entries), and a test set (1,511 entries) in a ratio of 7:1.5:1.5.

[0052] In another exemplary embodiment of this application, the prompt word template in step 303 above includes: a place name transliteration prompt word template, a common place name prompt word template, and a derived place name prompt word template.

[0053] The place name transliteration prompt template is a prompt template constructed according to the place name transliteration rules, and the place name transliteration prompt template includes proper name transliteration instantiation prompts.

[0054] As an alternative implementation, the proper name transliteration instantiation prompt word is obtained according to the following formula.

[0055] .

[0056] in, It is the instantiation prompt word of the proper name; It is a fixed template: "Englishplace name words and their letter sequences is [content]"; It is the i-th word of the place name. It represents the corresponding letter sequence, and N is the total number of words.

[0057] The common place name prompt template is a prompt template constructed based on the translation characteristics of common place names. The common place name prompt template includes instantiated prompts for proper noun transliteration and instantiated prompts for common noun semantic translation.

[0058] The generic translation instantiation prompt is obtained using the following formula.

[0059] .

[0060] Wherein, C represents the place name category information; It is a generic translation instantiation prompt; It is a fixed template "The category information of geographic name is [content]", stored in the hasGeographicCategory property.

[0061] The derived place name prompt template is a prompt template constructed based on the translation characteristics of derived place names. The derived place name prompt template includes instantiated prompts for proper name transliteration, instantiated prompts for common name translation, instantiated prompts for derived place name discrimination, instantiated prompts for derived parts, and instantiated prompts for original place name categories.

[0062] As an optional implementation, the derived place name discrimination instantiation prompt word is obtained according to the following formula.

[0063] .

[0064] in, It is a derivation place name discrimination instantiation prompt; It uses the fixed template "Is this derived toponym? The answer is [content]"; it is determined by the hasGeographicCategory attribute, and if the geographic category is "derived toponym", then it is correct.

[0065] The instantiation prompt for the derived part is obtained using the following formula.

[0066] .

[0067] in, Instantiate prompt words for derived parts; It is a fixed template: "The derived part is [dp] The translation result of the derived part is [dpt]"; dp is the source language derived part; dpt is the target language derived part.

[0068] The original place name category instantiation prompt words are obtained according to the following formula.

[0069] .

[0070] in, It is a prompt word for the instantiation of the original place name category; It is a fixed template "The category of original place names is [content]"; oc is the original place name category. The original place name is found through the derivation relationship, and then the hasGeographicCategory attribute of the original place name is obtained.

[0071] In another exemplary embodiment of this application, step 304 is replaced by steps 601 to 604.

[0072] Step 601: Load the pre-trained general domain large language model, configure the input and output dimensions, and set the Lora parameters.

[0073] Step 602: For any training sample in the place name translation training dataset, construct an instantiated prompt word based on the English place name and attribute information of the training sample, combined with the prompt word template.

[0074] Step 603: Using the instantiated prompt word as input and the Chinese reference translation of the training sample as the target output, construct the input-output training pair corresponding to the training sample.

[0075] Step 604: Based on several input-output training pairs of the place name translation training dataset, train the general domain large language model and fine-tune the LoRa parameters to obtain the place name translation large language model. During training, calculate the error between the model's predicted translation and the Chinese reference translation using the cross-entropy loss function, update the LoRa parameters through an optimizer, and evaluate the performance on the place name translation validation dataset after each training round. Stop training and save the optimal model parameters when the performance on the place name translation validation dataset shows no improvement for a preset number of rounds. The LoRa parameters include the target module, the rank of the low-rank matrix, the scaling factor, and the learning rate. Weight decay, rounds, and batch size.

[0076] Specifically, in this embodiment, the input dimension is a maximum sequence length of 300, and the output dimension is the Chinese vocabulary dimension; the optimizer is the Adam optimizer; the performance evaluation metric on the validation set is Bleu-2; the preset number of rounds is 3, which optimizes the model's understanding and translation ability of place name prompts while reducing computational cost. The target module is "query_key_value" (the attention mechanism query-key-value module to be fine-tuned); the rank of the low-rank matrix is ​​R=128; the scaling factor Lora_alpha=2048, and the learning rate is... Weight decay is (To prevent overfitting), the number of epochs is 15; the batch size is 4.

[0077] In another exemplary embodiment of this application, before step 202 described above, steps 701 to 704 are also included.

[0078] Step 701: Extract data from the preprocessed English-Chinese place name translation supervision dataset and public geographic databases to populate the knowledge base. Specifically, in this embodiment, the English-Chinese place name translation supervision dataset containing English place names and standard Chinese translations is obtained from the Chinese Academy of Surveying and Mapping; the English place names are 30,069 records of Australian and American English place names (including place name text, geographic category, and other information) crawled from the Geoname website.

[0079] Step 702: Obtain the entities and attributes of English place names in the knowledge base using named entity recognition technology; the attributes include "hasTargetLanguageName" and "hasGeographicCategory", "hasTargetLanguageName" includes the Chinese translation result corresponding to the entity; "hasGeographicCategory" includes the geographical category corresponding to the entity, and the geographical category is a common place name, a derived place name, a original place name, a proper name derived part, a generic name derived part, and a fully derived part.

[0080] Step 703: Based on rule matching and semantic analysis, analyze the semantic relationships between entities and obtain the relationships between English place names; the relationships include: derived place names are completely derived from the original place names; derived place names are derived from the proper nouns of the original place names; derived place names are derived from the common nouns of the original place names; the derived part belongs to a certain derived place name; and the original place name is a subset of common place names.

[0081] Step 704: The extracted "entity-attribute-relationship" data is stored in RDF format to obtain the place name translation knowledge graph and deployed through a graph database. Specifically, in this embodiment, the graph database is Neo4j.

[0082] In another exemplary embodiment of this application, step 203 above, in which the corresponding prompt word template is selected based on the attribute information corresponding to the English place name, is replaced by steps 801 to 805.

[0083] Step 801: Determine whether the attribute information includes the derived part and the original geographic category to obtain the second determination result.

[0084] Step 802: If the second judgment result is yes, then select the derived place name prompt word template.

[0085] Step 803: If the second judgment result is negative, then determine whether the attribute information includes a geographical category to obtain the third judgment result.

[0086] Step 804: If the third judgment result is yes, then select the ordinary place name prompt word template.

[0087] Step 805: If the third judgment result is negative, then select the place name transliteration prompt template.

[0088] Regarding translation accuracy, this application effectively avoids the error propagation problem of traditional techniques, significantly improving the accuracy of English-Chinese place name translation. Experimental data shows that the optimal solution achieves a Bleu-2 value of 81.76% in the translation of ordinary place names, which is 21.26% higher than the traditional pipeline method; the average performance of derived place name translation is improved by 27.70%, with the optimal solution achieving a Bleu-2 value of 72.60%, and it can completely maintain the semantic relationship between the derived place names and the original place names. For example, "Lake Youngs" and "Lake Youngs Trail" are translated as "Yangsi Lake" and "Yangsi Lake Road" respectively, and the translation of the derived part is completely consistent. This advantage stems from the synergistic effect of steps 301-304 and 701-704: Traditional pipeline methods break down place name translation into multiple independent sub-tasks, making it easy for errors from earlier steps to propagate to later steps. However, this application uses an integrated template of "property name transliteration prompts + generic name semantic prompts" to allow the large language model to complete the joint translation of "property name + generic name" in a single inference, fundamentally eliminating error propagation. The dedicated prompt words designed for derived place names can also avoid semantic "disconnection". At the same time, the contextual information of "place name-category-related place names" provided by the knowledge graph can help the model solve problems such as "polysemy" and "geographical category confusion", such as clarifying that "Course" in "Golf Course" should be translated as "site" rather than "route", further ensuring the accuracy of the translation.

[0089] In terms of model training and translation inference efficiency, this application balances performance and computational cost, demonstrating a significant efficiency advantage. During model training, the LoRa fine-tuning technique updates only the parameters of key modules in the attention layer of the ChatGLM model, representing less than 0.1% of the total model parameters. On a V100 32G graphics card, 15 rounds of training require only 8 hours, representing a 500% improvement in efficiency compared to full fine-tuning. During translation inference, the response time for a single place name is ≤0.5 seconds, supporting batch translations of ≥200 entries per second, meeting the demands of massive data processing such as million-level bilingual map place name annotations. This advantage stems from the technical design of step 604: traditional full fine-tuning requires updating all parameters of a large language model, consuming significant computational resources and time-consuming processes. LoRa fine-tuning, however, approximates the parameter update amount using a low-rank matrix, selectively optimizing only the core modules affecting translation performance, thus reducing computational costs and accurately improving the model's place name translation capabilities. Simultaneously, the data standardization preprocessing in step 302 and the graph database index retrieval in step 704 reduce additional data cleaning and information query time, ensuring efficient workflow integration.

[0090] In terms of scenario adaptability, this application can directly serve professional scenarios such as bilingual map production, significantly reducing the application threshold. In bilingual map production, the place names translated by this method conform to the "property name transliteration standard" and the "accuracy of generic name translation," and the translation of derived place names within the same area maintains a high degree of consistency. For example, "Lake Youngs," "Lake Youngs Trail," and "Southeast Lake Youngs Road" are translated as "Yangsi Lake," "Yangsi Lake Road," and "Southeast Yangsi Lake Road," respectively, without the need for secondary manual proofreading. The efficiency of place name labeling using this application is 80 times higher than that of manual translation, and the translation error rate is reduced from 3.5% to below 0.2%. This advantage stems from the scenario-based design of steps 201-204 and the experimental data collection and preprocessing of steps 401-402: Steps 401-402 select Geoname official website place name data that matches the bilingual map scenario and introduce a supervised dataset that conforms to national translation standards to ensure that the model learns standardized translation results; the closed loop of "knowledge graph retrieval → prompt word instantiation → model reasoning" in steps 201-204 can transform the "category accuracy" and "derivation consistency" requirements required for map production into model prompt information, and the translation results can be directly used for map annotation, reducing manual intervention.

[0091] In terms of generalization ability, this method can be extended to the translation of place names in other alphabetic languages ​​such as French, German, and Russian, without the need to reconstruct the overall framework. When migrating the application, it is only necessary to replace the "corresponding letter sequence" in step 504 with the letter sequence of the target language, and to expand the target language attributes of the knowledge graph in step 202. For example, the sequence of the French place name "Mont Blanc" is "Mo nt B lanc". By adding the French category "mountain" for "Mont" to the knowledge graph, accurate translation can be achieved. Experiments have verified that when migrating to the translation of French place names, the BLEu-2 value still remains above 78%, and the generalization ability far exceeds that of traditional single-language translation methods. This advantage stems from the modular and scalable design of steps 303 and 202: the prompt word template in step 303 adopts a "fixed structure + variable placeholders" approach, with "letter sequence prompts" relying on the commonalities of alphabetic languages ​​and "category prompts" based on universal geographical categories across languages, requiring only the replacement of placeholder content; the knowledge graph ontology in step 202 is a three-layer abstract structure of "concept-relationship-attribute", with core concepts and relationships being universal geographical semantic logic across languages, requiring only the expansion of target language information at the attribute layer to quickly adapt to new language translation needs.

[0092] Based on the same inventive concept, this application also provides a place name intelligent translation system based on LoRa fine-tuning and knowledge graphs for implementing the aforementioned system. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more place name intelligent translation system embodiments based on LoRa fine-tuning and knowledge graphs provided below can be found in the limitations of the place name intelligent translation method based on LoRa fine-tuning and knowledge graphs described above, and will not be repeated here.

[0093] In one exemplary embodiment, such as Figure 3 As shown, a place name intelligent translation system based on LoRa fine-tuning and knowledge graph is provided.

[0094] The English place name input module is used to receive English place names input by the user.

[0095] The knowledge graph retrieval module is used to send the English place names to the place name translation knowledge graph to retrieve the attribute information and relationship information corresponding to the English place names.

[0096] The prompt word generation module is used to select the corresponding prompt word template according to the attribute information corresponding to the English place name, and generate an instantiated prompt word based on the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template; the prompt word template is an integrated template of proper noun transliteration and common noun translation.

[0097] The English place name translation module is used to input the instantiated prompt words into the place name translation language model, and output the Chinese place name translation result through context understanding and semantic mapping. The place name translation language model is a large language model trained on a general domain based on a place name translation training dataset, and the Lora parameters are fine-tuned. The Chinese place name translation result is a structured intermediate translation result, which includes at least the final translated name field, and further includes at least one of the proper name transliteration field and the generic name semantic translation field.

[0098] The consistency verification module is used to determine whether the Chinese place name translation result meets the consistency verification standard and obtain a first judgment result; when the first judgment result is yes, the structured intermediate translation result is output as the Chinese place name translation result; when the first judgment result is no, the prompt word generation module is called.

[0099] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 4 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores input English place names. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it can implement the intelligent place name translation method based on LoRa fine-tuning and knowledge graphs provided in the previous embodiment.

[0100] Those skilled in the art will understand that Figure 4 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0101] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0102] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0103] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0104] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0105] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for intelligent translation of place names based on LoRa fine-tuning and knowledge graphs, characterized in that, include: Receive English place names input by the user; The English place name is sent to the place name translation knowledge graph to retrieve the attribute information and relationship information corresponding to the English place name; Select the corresponding prompt word template based on the attribute information corresponding to the English place name, and generate an instantiated prompt word based on the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template. The prompt word template is an integrated template that combines proper noun transliteration and common noun translation; The instantiated prompt words are input into the place name translation language model, and the Chinese place name translation result is output through context understanding and semantic mapping; the place name translation language model is a language model obtained by training a general domain language model based on the place name translation training dataset and fine-tuning the Lora parameters; the Chinese place name translation result is a structured intermediate translation result, which includes at least the final translated name field, and further includes at least one of the proper name transliteration field and the common name semantic translation field; Determine whether the Chinese place name translation result meets the consistency check standard to obtain the first judgment result; if the first judgment result is yes, output the structured intermediate translation result as the Chinese place name translation result; if the first judgment result is no, jump to the step "select the corresponding prompt word template according to the attribute information corresponding to the English place name, and generate an instantiated prompt word according to the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template".

2. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 1, characterized in that, Before inputting the instantiated prompt words into the place name translation language model and outputting the Chinese place name translation result through contextual understanding and semantic mapping, the process also includes: obtaining the place name translation training dataset and training the place name translation language model, specifically including: Obtain a supervised dataset for English-Chinese place name translation; the supervised dataset includes English place names and Chinese reference translations; the English place names are common place names or derived place names; The English-Chinese place name translation supervision dataset is preprocessed and divided to obtain a place name translation training dataset, a place name translation test dataset, and a place name translation verification dataset. The place name translation training dataset includes several training samples. Each training sample includes an English place name, its corresponding attribute information, and a Chinese reference translation. When the English place name is a common place name, the attribute information of the English place name includes the geographical category. When the English place name is a derived place name, the attribute information of the English place name includes the geographical category, the derived part, and the original geographical category. Based on the translation characteristics of the common place names and the derived place names, prompt word templates are constructed respectively, and structured output specifications are configured for the prompt word templates, so that the place name translation big language model generates structured intermediate translation results according to preset fields; the structured output specifications define at least the final translation field, and further define at least one of the proper name transliteration field, common name semantic translation field, and derived element translation field; Based on the place name translation training dataset, combined with the prompt word template and the structured output specification, a general domain large language model is trained, and the Lora parameters are fine-tuned to obtain the place name translation large language model.

3. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 2, characterized in that, The English-Chinese place name translation supervision dataset is preprocessed and divided to obtain a place name translation training dataset, a place name translation test dataset, and a place name translation verification dataset, specifically including: Invalid data was removed using regular expressions, and the accuracy of place name category labeling was manually verified. All data is converted into a structure of "English place name - attribute information - Chinese reference translation", and divided into place name translation training dataset, place name translation test dataset, and place name translation verification dataset according to preset partitioning rules, specifically including: When the English place name is a common place name, the English-Chinese place name translation supervision dataset is divided into a common place name translation training set, a common place name translation verification set, and a common place name translation test set in a ratio of 8:1:

1. When the English place name is a derived place name, an equal amount of derived place name data and ordinary place name data are selected from the English-Chinese place name translation supervision dataset, and divided into a derived place name translation training set, a derived place name translation verification set, and a derived place name translation test set in a ratio of 7:1.5:1.

5. The training sets for general place name translation and derived place name translation are combined to form a place name translation training dataset. The test set for translating ordinary place names and the test set for translating derived place names are combined to form a test dataset for translating place names. The ordinary place name translation verification set and the derived place name translation verification set are combined into a place name translation verification dataset.

4. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 2, characterized in that, The prompt word templates include: place name transliteration prompt word templates, common place name prompt word templates, and derived place name prompt word templates; the place name transliteration prompt word templates are constructed according to place name transliteration rules, and include proper name transliteration instantiation prompt words; the common place name prompt word templates are constructed according to the translation characteristics of common place names, and include proper name transliteration instantiation prompt words and common name translation instantiation prompt words; the derived place name prompt word templates are constructed according to the translation characteristics of derived place names, and include proper name transliteration instantiation prompt words, common name translation instantiation prompt words, derived place name discrimination instantiation prompt words, derived part instantiation prompt words, and original place name category instantiation prompt words.

5. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 4, characterized in that, The instantiated prompt word for the proper name transliteration is obtained according to the following formula: ； in, It is the instantiation prompt word of the proper name; It is a fixed template: "English placename words and their letter sequences is [content]"; It is the i-th word of the place name. It represents the corresponding letter sequence, where N is the total number of words; The generic translation instantiation prompt word is obtained according to the following formula: ； Wherein, C represents the place name category information; It is a generic translation instantiation prompt; It is a fixed template "The category information of geographic name is [content]", stored in the hasGeographicCategory property.

6. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 4, characterized in that, The derived place name discrimination instantiation prompt word is obtained according to the following formula: ； in, It is a derivation place name discrimination instantiation prompt; It uses the fixed template "Is this aderived toponym? The answer is [content]"; it is determined by the hasGeographicCategory attribute, and if the geographic category is "derived toponym", then it is "yes". The derived part instantiation prompt word is obtained according to the following formula: ； in, Instantiate prompt words for derived parts; It is a fixed template: "The derived part is [dp] The translation result of the derived part is [dpt]"; dp is the source language derived part; dpt is the target language derived part. The instantiated prompt words for the original place name category are obtained according to the following formula: ； in, It is a prompt word for the instantiation of the original place name category; It is a fixed template "The category of original place names is [content]"; oc is the original place name category. The original place name is found through the derivation relationship, and then the hasGeographicCategory property of the original place name is obtained.

7. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 2, characterized in that, Based on the aforementioned place name translation training dataset and the aforementioned prompt word template, a general-domain large language model is trained, and the Lora parameters are fine-tuned to obtain the aforementioned place name translation large language model, specifically including: Load the pre-trained general-domain large language model, configure the input and output dimensions, and set the Lora parameter; For any training sample in the place name translation training dataset, based on the English place name and attribute information of the training sample, and combined with the prompt word template, an instantiated prompt word is constructed; Using the instantiated prompt words as input and the Chinese reference translation of the training samples as the target output, an input-output training pair corresponding to the training samples is constructed. Based on several input-output training pairs of the place name translation training dataset, a general-domain large language model is trained, and the LoRa parameters are fine-tuned to obtain the place name translation large language model. During the training of the general-domain large language model, the error between the model's predicted translation and the Chinese reference translation is calculated using the cross-entropy loss function, and the LoRa parameters are updated through an optimizer. After each round of training, the performance is evaluated on the place name translation validation dataset. When the performance on the place name translation validation dataset does not improve for a consecutive preset number of rounds, training is stopped and the optimal model parameters are saved. The LoRa parameters include the target module, the rank of the low-rank matrix, the scaling factor, the learning rate, the weight decay, the round number, and the batch size.

8. The method for intelligent English-Chinese place name translation based on LoRa fine-tuning and knowledge graph as described in claim 1, characterized in that, Before sending the English place name to the place name translation knowledge graph to retrieve the corresponding attribute information and relationship information, the method further includes: constructing the place name translation knowledge graph, specifically including: Data was extracted from the preprocessed English-Chinese place name translation supervision dataset and public geographic databases to populate the knowledge base; The entity and attributes of English place names in the knowledge base are obtained through named entity recognition technology; the attributes include "hasTargetLanguageName" and "hasGeographicCategory", the "hasTargetLanguageName" includes the Chinese translation result corresponding to the entity; the "hasGeographicCategory" includes the geographical category corresponding to the entity, the geographical category is a common place name, a derived place name, a original place name, a proper noun derived part, a generic noun derived part, and a fully derived part; Based on rule matching and semantic analysis, the semantic relationships between entities are analyzed to obtain the relationships between English place names; the relationships include: derived place names are completely derived from the original place names; derived place names are derived from the proper nouns of the original place names; derived place names are derived from the common nouns of the original place names; the derived part belongs to a certain derived place name; and the original place name is a subset of common place names. The extracted "entity-attribute-relationship" data is stored in RDF format to obtain a place name translation knowledge graph, which is then deployed through a graph database.

9. The intelligent place name translation method based on LoRa fine-tuning and knowledge graph as described in claim 1, characterized in that, Based on the attribute information corresponding to the English place name, select the corresponding prompt word template, specifically including: Determine whether the attribute information includes derived parts and the original geographic category to obtain a second determination result; If the second judgment result is yes, then select the derived place name prompt word template; If the second judgment result is negative, then the attribute information is judged to include geographical category, and the third judgment result is obtained. If the third judgment result is yes, then select the ordinary place name prompt template; If the third judgment result is negative, then select the place name transliteration prompt template.

10. A place name intelligent translation system based on LoRa fine-tuning and knowledge graph, characterized in that, include: The English place name receiving module is used to receive English place names input by the user; The knowledge graph retrieval module is used to send the English place name to the place name translation knowledge graph and retrieve the attribute information and relationship information corresponding to the English place name. The prompt word generation module is used to select the corresponding prompt word template according to the attribute information corresponding to the English place name, and generate instantiated prompt words according to the attribute information and relationship information corresponding to the English place name, combined with the corresponding prompt word template. The prompt word template is an integrated template that combines proper noun transliteration and common noun translation; The English place name translation module is used to input the instantiated prompt words into the place name translation language model, and output the Chinese place name translation result through context understanding and semantic mapping; the place name translation language model is a language model obtained by training a general domain language model based on the place name translation training dataset and fine-tuning the Lora parameters; the Chinese place name translation result is a structured intermediate translation result, which includes at least the final translated name field, and further includes at least one of the proper name transliteration field and the generic name semantic translation field; The consistency verification module is used to determine whether the Chinese place name translation result meets the consistency verification standard and obtain a first judgment result; when the first judgment result is yes, the structured intermediate translation result is output as the Chinese place name translation result; when the first judgment result is no, the prompt word generation module is called.

Images