Term translation method and system based on artificial intelligence and storage medium

Abstract

The invention relates to the field of terminology information processing and application, in particular to a terminology translation method and system based on artificial intelligence and a storage medium. The method comprises the steps of extracting a first term text from a first language, combining the first term text with a sentence where the first term text is located to form first input information, and adding a cue word text to the first input information to obtain second input information; processing the second input information through a large language model to obtain first field information; combining the first term text with the first domain information to form a first defined text through a large language model; vectorizing the first definition text to obtain a first semantic vector; and performing similarity comparison after repeating the steps. The model performs semantic analysis on the basis of the terms and the sentences where the terms are located, information parameters in a specific language environment are combined, candidate translations corresponding to the terms are searched firstly on the basis of a large language model in the technical field, the candidate translations are selected, and a translation result with the highest matching degree is output.

Description

Technical Field

[0001] The present invention relates to the field of professional term information processing and application, and particularly to a term translation method, system and storage medium based on artificial intelligence. Background Art

[0002] Currently, term translation and management mainly rely on term sets and dictionaries published in professional fields. However, they cannot correctly match polysemous words and require manual judgment and selection. Secondly, term sets are static and have a limited update frequency, making it difficult to keep up with the rapid development of professional fields and the emergence of new terms. Moreover, the high operation and maintenance costs of term databases limit the application and update speed. And when it comes to term expansion in different fields and different languages, the support ability is insufficient.

[0003] Therefore, it is necessary to provide a term translation method, system and storage medium based on artificial intelligence to break the above limitations. Summary of the Invention

[0004] The present invention provides a term translation method, system and storage medium based on artificial intelligence, including the following three aspects:

[0005] In the first aspect, a term translation method based on artificial intelligence includes the following steps:

[0006] S1. Select a first language, extract a first term text from the first language, combine the first term text with the sentence where the first term text is located to form a first input information, and add a prompt text to the first input information to obtain a second input information; process the second input information through a large language model to obtain a first domain information;

[0007] S2. Combine the first term text with the first domain information and form a first definition text through a large language model;

[0008] S3. Vectorize the first definition text to obtain a first semantic vector;

[0009] S4. Select a second language and repeat steps S1 to S3 to obtain a second semantic vector;

[0010] S5. After setting a similarity threshold, compare the similarity between the first semantic vector and the second semantic vector with the threshold as a reference and make a judgment.

[0011] Further, it further includes step S6, specifically, if the similarity is greater than the threshold, it is determined that they are the same term, and if the similarity value is less than or equal to the threshold, it is determined that they are non - same terms.

[0012] Further, the prompt word is used to ensure that the model is more accurate when processing term domain classification.

[0013] Further, the first set threshold is 90%.

[0014] Furthermore, in step S2, the target language terms are retrieved according to the classification in the existing term library according to the field, and are screened by the large language model for the field and context.

[0015] Further, before performing step S5, the system further includes an annotation correction parameter. If the sum of the annotation correction parameter and the similarity value is greater than the first set threshold, it is determined to be the same term.

[0016] Further, the system further includes a term evidence-based module that classifies the terms and forms a level parameter, and loads the level parameter when generating the second language text.

[0017] In a second aspect, a term translation system based on artificial intelligence includes a semantic vectorization algorithm model, a text classification model, and a vector comparison model. The semantic vectorization algorithm model is used to vectorize the meaning of the text content, the text classification model is used to obtain the classification to which a specific term belongs, and the vector comparison model is used to compare the similarity degree of different vectors.

[0018] In a third aspect, a readable storage medium stores execution instructions, and when the execution instructions are executed by a processor, they are used to implement the method described in the first aspect.

[0019] The present technical solution adopts advanced large language model semantic understanding technology. After the text is input into the model, the model will perform semantic analysis based on the "term" and the "sentence where it is located", and output the information parameters of the "term" in the specific language environment of the "sentence where it is located". Based on the technical field, the large language model will first search for candidate translations corresponding to the term, select the candidate translations, and output the translation result with the highest matching degree. Description of the Drawings

[0020] Figure 1 is a logic block diagram;

[0021] Figure 2 is a reference diagram for threshold setting. Detailed Embodiments

[0022] To enable those skilled in the art to better understand the technical solution of the present application, the present application will be described in detail below with reference to the drawings and specific embodiments. The embodiments of the present application will be further described in detail below with reference to the drawings and specific examples, but shall not be construed as a limitation to the present application.

[0023] The terms "first", "second" and similar words used in this application do not denote any order, quantity or importance, but are only used for distinction. Words such as "including" or "comprising" mean that the elements before this word cover the elements listed after this word, and do not exclude the possibility of also covering other elements. The execution order of each step in the method described in this application in combination with the drawings is not limited. As long as the logical relationship between each step is not affected, several steps can be integrated into a single step, a single step can be decomposed into multiple steps, and the execution order of each step can also be adjusted according to specific requirements.

[0024] It should also be understood that the term "and / or" in this application is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this application generally represents an "or" relationship between the front and back associated objects.

[0025] Embodiment 1

[0026] Reference Figure 1 , this embodiment discloses a method for term translation based on artificial intelligence, including the following steps:

[0027] S1. Select a first language, extract a first term text from the first language, combine the first term text with the sentence where the first term text is located to form a first input information, and add a prompt text to the first input information to obtain a second input information; process the second input information through a large language model to obtain a first domain information.

[0028] Hereinafter, term 1 is used to simplify and replace the first term text, language A is used to simplify and replace the first language, definition 1 is used to simplify and replace the first definition text, and sentence 1 is used to simplify and replace the sentence where the first term text is located.

[0029] The description is as follows: The first step is to extract the term in a sentence in language A to obtain term 1. Use "term 1" and "the sentence where term 1 is located" as the input information of the model. Under the prompt words designed by us, let the large language model judge what specific field the term "term 1" belongs to based on the input information. Since the meaning of a term may be different in different fields. For example, the word "apple" may refer to "Apple Inc." in the computer field. And in the plant field, it refers to "a kind of fruit". The purpose is to enable the fine-tuned large language model to judge the specific field of "term 1" in the sentence through these two pieces of information, namely "term 1" and "the sentence where term 1 is located".

[0030] S2. Combine the first term text with the first domain information and form a first definition text through a large language model;

[0031] It is described as follows that this step obtains the "term domain" and "term 1" from the previous step model. These two pieces of information are input into the large language model. Here, we use another set of prompt words to fine-tune the large language model, enabling the large language model to accurately generate the definition of "term 1" in this "term domain". Here, the characteristic that the large language model has a huge knowledge base is utilized to enable the model to generate the definition of the term. The reason for generating the definition is that the term itself is restricted by the number of characters, and usually, fewer characters are required to express a complex concept. At this time, if only the semantics of the term is vectorized, there will be too many words that are not similar in vectors between languages, and there will be a large deviation in the definition of the words. Therefore, we choose to first let the model understand the meaning of the term in the sentence to judge its domain, and then let the model expand the definition for this term. S3. Vectorize the first definition text to obtain the first semantic vector;

[0032] It is described as follows that this step uses a semantic vectorization model to vectorize the generated "definition of term 1", thus changing the "term definition" into a "semantic vector" that can be used by the computer for semantic comparison.

[0033] S4. Select a second language and repeat steps S1 to S3 to obtain a second semantic vector; Refer to Figure 1 , S5. After setting a similarity threshold, compare the similarity between the first semantic vector and the second semantic vector with the threshold as a reference and make a judgment.

[0034] It also includes step S6, specifically, if the similarity is greater than the threshold, it is determined that they are the same term, and if the similarity value is less than or equal to the threshold, it is determined that they are non - same terms.

[0035] The prompt words are used to ensure that the model is more accurate in processing term domain classification.

[0036] The first set threshold is 90%. Refer to Figure 2 , Calculate the similarity between the vectors in definition vector set 1 and the vectors in definition vector set 2 in turn. If their similarity is higher than the threshold, it is considered that the meanings of the two are the same, such as vector A and vector F.

[0037] In step S2, the target - language term is retrieved and classified according to the existing term library according to the domain, and the domain and context are screened by the large language model.

[0038] Before performing step S5, the system also includes an annotation correction parameter. If the sum of the annotation correction parameter and the similarity value is greater than the first set threshold, it is determined that they are the same term.

[0039] The system also includes a term evidence - based module, which classifies the terms and forms a rank parameter, and loads the rank parameter for the generated second - language text.

[0040] The application is described as follows

[0041] #Role: Terminology Management Expert

[0042] ##Profile

[0043] - Proficient Languages: Japanese, English

[0044] - Role Description: Terminology management experts are good at accurately identifying, researching, and maintaining professional terms in a multilingual environment to ensure the consistency and professionalism of translations.

[0045] - Knowledge Reserve: Medicine, Neurosurgery

[0046] ##background:

[0047] Source of Vocabulary: Medical Professional Works

[0048] Vocabulary Language: Japanese

[0049] ##input:

[0050] | <Term> | <Sentence where it appears> |

[0051] ##task: According to the term and the sentence where the term appears, sequentially complete the following tasks:

[0052] 1. Classify the vocabulary according to the classification criteria of Chinese professional disciplines.

[0053] ##Rules:

[0054] - Return NA if no result can be output.

[0055] ##Print:

[0056] | <Term> | <Field Information> |

[0057] ##Initialization

[0058] As a role <role>Have <profile>The described ability, refer to <background>Complete with the information provided <task>, strictly abide by <rules>, in order to <print>The required output result. Take a deep breath before drawing a conclusion.

[0059] #Role: Neurosurgery Terminology Expert

[0060] ##Profile

[0061] - Proficient Languages: Japanese, Chinese

[0062] - Role Description: A terminology management expert is good at accurately identifying, researching, and maintaining professional terms in a multilingual environment to ensure the consistency and professionalism of translations.

[0063] - Knowledge Reserve: Medicine, Neurosurgery

[0064] ##background:

[0065] - Lexical Source: Medical Professional Works

[0066] - Lexical Language: Japanese

[0067] ##Input:

[0068] | <Term> | <Domain Information> |

[0069] ##task: Complete the following tasks in sequence:

[0070] - Based on the input <Domain Information>, determine the definition of the term. The term definition should include the following key elements: the core concept of the term, the attributes and characteristics of the term, and the scope and limitations of the term.

[0071] ##Rules:

[0072] - If unable to output the result, return NA.

[0073] ##Print:

[0074] | <Term> | <Definition> |

[0075] ##Initialization

[0076] As a role <role>Have <profile>The described capabilities, refer to <background>The information provided, receive <input> the information, complete <task>, strictly abide by <rules>, with <print>The required output result. Please think step by step and take a deep breath before reaching a conclusion.

[0077] Example 2

[0078] Based on Example 1, this example elaborates on an artificial intelligence-based term translation system, including a semantic vectorization algorithm model, a text classification model, and a vector comparison model. The semantic vectorization algorithm model is used to vectorize the meaning of text content, the text classification model is used to obtain the classification to which a specific term belongs, and the vector comparison model is used to compare the similarity degrees of different vectors. This example elaborates on using this system to compare whether two words in different languages refer to the same concept by computer; specifically, it is used to analyze whether a certain term is consistent between texts in two different languages.

[0079] For example

[0080] Language A: Intracranial aneurysm is an abnormal dilation of the blood vessel wall that may have a serious impact on the health of patients. (Sentence 1)

[0081] Language B: Intracranial aneurysms are abnormal dilations of blood vessels that can have severe consequences for patients' health. (Sentence 2)

[0082] For the two highlighted words in Language A and Language B, non-professionals cannot confirm whether these two words refer to the same concept. Using this system can determine whether these two words refer to the same concept.

[0083] The specific steps are as follows:

[0084] 1. Input "intracranial aneurysm" (Term 1) and "Sentence 1" into the large language model (term classification model) fine-tuned with specific prompt words. The large language model will judge and output the professional field to which the term "intracranial aneurysm" belongs - "neurosurgery" according to the meaning of "intracranial aneurysm" in "Sentence 1".

[0085] 2. Input "Neurosurgery" and "Intracranial aneurysm" into the large language model (term definition model) fine-tuned with specific prompts. The large language model will generate the definition of "Intracranial aneurysm" in the field of "Neurosurgery". Thus, the "term definition" of "Intracranial aneurysm" in Language A is obtained - "Intracranial aneurysm is a vascular disease in the field of neurosurgery, referring to the local abnormal dilation or protrusion of cerebral arteries, usually occurring at the bifurcation of blood vessels. It can be congenital or caused by acquired factors such as arteriosclerosis, infection, etc. If an intracranial aneurysm ruptures, it may lead to subarachnoid hemorrhage, seriously endangering life. Therefore, the timely detection and treatment of intracranial aneurysms are crucial to prevent potential serious complications." This is Definition 1.

[0086] 3. Input "Intracranial aneurysms" (Term 2) and "Sentence 2" into the large language model (term classification model) fine-tuned with specific prompts. The large language model will judge and output the professional field where the term "Intracranial aneurysms" is located according to the meaning of "Intracranial aneurysms" in "Sentence 2" - "neurosurgery".

[0087] 4. Input "neurosurgery" and "Intracranial aneurysms" into the large language model (term definition model) fine-tuned with specific prompts. The large language model will generate the definition of "Intracranial aneurysms" in the field of "neurosurgery". Thus, the "term definition" of "Intracranial aneurysms" in Language B is obtained - "In neurosurgery, 'Intracranial aneurysms' refer to localized, abnormal expansions or bulges in the walls of the arteries within the brain, often occurring at bifurcations of blood vessels. These aneurysms can be congenital or caused by factors such as arteriosclerosis or infection. If ruptured, they can lead to subarachnoid hemorrhage, a severe medical condition requiring prompt diagnosis and treatment.” (Definition 2)

[0088] 5. Through the semantic vectorization algorithm, obtain the semantic vectors of "Definition 1" and "Definition 2". Our system first uses a multilingual text embedding model to obtain the embedding vectors of each defined term. This model is trained using a large amount of multilingual corpora in order to capture and retain the semantic information of the same concepts in different languages.

[0089] The core of this model uses the deep learning algorithm of the "neural network language model", whose goal is to learn a function that can map text content to a continuous vector space, such that in this space, words that are semantically similar are also close to each other in the vector space. In fact, this process is a kind of "semantic vectorization".

[0090] In specific operations, we will input the definitions of each term word in Language A and Language B into this multilingual word embedding model, and the model will output the corresponding embedding vectors. The "Definition 1 vector" and the "Definition 2 vector" are obtained in this way.

[0091] 6. Measure the similarity between the "Definition 1 vector" and the "Definition 2 vector". The specific measurement method used is to calculate their cosine similarity. The calculation formula is as follows:

[0092]

[0093] where A and B are the two vectors we want to compare, and '·' represents the dot product operation, ||A|| 2 and ||B|| 2 represent the two-norms (i.e., the lengths of the vectors) of vectors A and B respectively.

[0094] At this time, these two vectors are generated by our system, which uses a multilingual word embedding model to generate the "Definition 1 vector" and the "Definition 2 vector". In this model, the embedding vectors obtained for the same or similar concepts in different languages should be close. This is based on the distributional hypothesis theory, which holds that the meaning of a word can be defined by the context of the words around it. That is, if two words have similar roles in their respective contexts, then their semantics should also be similar. Therefore, in a cross-lingual environment, even for words in different languages, as long as they refer to the same or similar concepts, their embedding vectors should be close. In our system, if the cosine value between the "Definition 1 vector" and the "Definition 2 vector" is greater than or equal to 0.85, we will consider that the words represented by these two vectors are almost the same semantically, that is, they refer to the same concept. For example, if the cosine similarity of the vectors of the words "intracranial aneurysm" in language A and "Intracranial aneurysms" in language B exceeds 0.85, then we can consider that their meanings are the same.

[0095] Example 3

[0096] This example elaborates on the establishment of a term base for this technology based on Example 2. The difference from Example 2 is that this example shows the scenario in which this technology is used in "two texts with different languages but similar content". This technology can automatically find corresponding terms in the two texts and quickly establish a term base. When the user has two documents in different languages but with similar content. There may be multiple terms with the same concept but different languages.

[0097] For example:

[0098] When having both the "Clinical Management Guidelines for Unruptured Intracranial Aneurysms in China (2024 Edition)" (Document 1) and the "2023 Guideline for the Management of Patients With Aneurysmal Subarachnoid Hemorrhage: A Guideline From the American Heart Association / American Stroke Association" (Document 2). Since both of these documents are about the clinical treatment of aneurysms, there may be many corresponding Chinese and English terms. (It should be noted that these two documents are not in a translation relationship, but are formulated by experts in their respective countries). Users can quickly establish Chinese-English term pairs from these two documents through the patented technology of this patent, so as to establish a term library. The specific steps are as follows.

[0099] 1. For the texts in Document 1 and Document 2, use the end punctuation marks in punctuation to split the sentences (the end punctuation marks can be defined by yourself. The default is the period, exclamation mark, and question mark). Obtain the sentence set of Document 1 (Sentence Set 1) and the sentence set of Document 2 (Set 2).

[0100] 2. For all the sentences in Set 1 and Set 2, use the term extraction model to extract terms. Obtain the "terms" and "sentences where the terms are located" in all of Set 1, and the "terms" and "sentences where the terms are located" in all of Set 2.

[0101] Term extraction uses a large language model. This model has been fine-tuned with prompts. The prompts are used as follows.

[0102]

[0103]

[0104] 3. For each term, perform the steps (125 or 345) to obtain the "term definition vector" in "Example 1". At this time, obtain the definition vector set of all terms in Document 1 (Definition Vector Set 1), and the definition vector set of all terms in Document 2 (Definition Vector Set 2).

[0105] 4. Traverse "Definition Vector Set 1", and calculate the cosine similarity with the vectors in "Definition Vector Set 2" in turn. If the cosine value is greater than 0.85, it is considered that the two have the same meaning and form a term pair (refer to Example 1).

[0106] The pseudo-code example is as follows:

[0107]

[0108]

[0109] Example 4

[0110] The difference from Example 3 is that this example demonstrates the application of the technology in the translation scenario. First, the text to be translated in a certain language can be analyzed, and then retrieved in the term base, and the accurate translation of the term can be obtained quickly.

[0111] For example:

[0112] When preparing to translate "Clinical Management Guidelines for Unruptured Intracranial Aneurysms in China (2024 Edition)" (Document 1) into English. Assume that we already have a corresponding term base. The data structure of the term base needs to include either "the term string itself", "term examples", or "term definitions". Or the term base itself is a vector database.

[0113] The specific steps are as follows.

[0114] 1. For the text in Document 1, use a sentence splitting model to perform sentence splitting respectively. Obtain the sentence set of Document 1 (Sentence Set 1)

[0115] 2. For all sentences in Set 1, use a term extraction model to extract terms. Obtain the "terms" and "sentences where the terms are located" in the entire Set 1.

[0116] 3. For each term in Set 1, perform the steps (125 or 345) to obtain the "term definition vector" in Example 1. At this time, obtain the definition vectors of all terms in Document 1 (Definition Vector Set 1)

[0117] 4. Use the term definition vectors obtained in step 3 to complete term retrieval in the term base.

[0118] 5. If the terms extracted from Document 1 are stored in the term base, the corresponding terms in the term base will be obtained based on semantics.

[0119] Because this method refers to semantics, it can find the corresponding terms in the term base more accurately than traditional string retrieval.

[0120] Example 5

[0121] This example illustrates a readable storage medium in which execution instructions are stored, and when the execution instructions are executed by a processor, they are used to implement the method in Example 1.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.< / print> < / rules> < / task> < / background> < / profile> < / role> < / print> < / rules> < / task> < / background> < / profile> < / role>

Claims

1. A terminology translation method based on artificial intelligence, characterized in that: The steps include: S1, selecting a first language, extracting a first term text from the first language, combining the first term text with the sentence where the first term text is located to form first input information, adding a prompt word text to the first input information to obtain second input information; processing the second input information through a large language model to obtain first domain information; S2. Combining the first term text with the first domain information through a large language model to form a first definition text; S3, vectorizing the first definition text to obtain a first semantic vector; S4, select a second language, and repeat steps S1 to S3 to obtain a second semantic vector; S5. After setting a similarity threshold, the similarity between the first semantic vector and the second semantic vector is compared with the threshold as a reference and a judgment is made.

2. The term translation method based on artificial intelligence according to claim 1, characterized in that: The step S6 is also included, specifically, if the similarity is greater than the threshold, it is determined to be the same term, and if the similarity is less than or equal to the threshold, it is determined to be a non-identical term.

3. The domain term translation method based on artificial intelligence according to claim 1 is characterized in that: The hint words are used to ensure that the model is more accurate in processing term domain classification.

4. The term translation method based on artificial intelligence according to claim 2 is characterized in that: The first set threshold is 90%.

5. The term translation based on artificial intelligence according to any one of claims 1 to 4 The method is characterized in that Step S2 classifies and retrieves the target language terms from an existing term library according to the field, and performs field and context screening using a large language model.

6. The terminology translation method based on artificial intelligence according to claim 5 is characterized in that: Before performing step S5, the system further includes an annotation correction parameter. If the sum of the annotation correction parameter and the similarity value is greater than the first set threshold, it is determined to be the same term.

7. The terminology translation method based on artificial intelligence according to claim 5, characterized in that: The system also includes a terminology evidence-based module, which classifies terms into levels and forms level parameters, and loads the level parameters when generating the second language text.

8. A terminology translation system based on artificial intelligence, characterized in that: It includes a semantic vectorization algorithm model, a text classification model and a vector comparison model. The semantic vectorization algorithm model is used to vectorize the meaning of text content, the text classification model is used to obtain the category to which a specific term belongs, and the vector comparison model is used to compare the similarity between different vectors.

9. A readable storage medium, characterized in that: The readable storage medium stores execution instructions, which are used to implement the method according to any one of claims 1 to 6 when executed by a processor.

Images