Medical literature interpretation method and system based on large language model

Abstract

The invention relates to the technical field of language analysis, and discloses a medical literature interpretation method and system based on a language large model, and the method comprises the steps: constructing a relationship network between words, and obtaining a first relationship network; determining a first association relationship between each text word and other words in the text, and endowing the first association relationship with a weight according to the number of association times; constructing a relationship network for words in the text according to the text content to obtain a second relationship network; determining a second association relationship between each text word and other words in the text, and endowing the second association relationship with a weight; and according to the first relationship network and the weight endowed by the first relationship network, performing optimization of the association path, and obtaining an interpretation result through content interpretation on the path. The defects of a traditional calculation method in the aspects of processing efficiency and semantic understanding are overcome. The accuracy and efficiency of literature analysis are greatly improved, and the convenience of knowledge acquisition is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of language analysis, and in particular to a method and system for interpreting medical literature based on a large language model. Background Art

[0002] With the advent of the information age, the computing needs that need to be processed have also increased exponentially. Faced with massive amounts of data, with the rapid development of the biomedical field and the publication of a large number of scientific research results, the number of medical literature has increased exponentially, and researchers and medical workers are facing the challenge of information overload. How to efficiently obtain and understand the latest medical information has become a key issue in scientific research and clinical practice. Traditional literature reading and analysis methods are difficult to cope with massive data and are prone to information omissions.

[0003] In recent years, artificial intelligence technology, especially the progress in natural language processing (NLP) and deep learning, has provided a new way to solve this problem. Language models based on large-scale pre-training (such as BERT, GPT, etc.) have performed well in tasks such as semantic understanding, information extraction, and content generation, showing great potential in medical literature processing. These models can automatically identify medical terms, symptom descriptions, pathological relationships, and other information in the literature through semantic analysis and context understanding, and can extract core points from complex texts.

[0004] In addition, language models developed specifically for the medical field (such as BioBERT, ClinicalBERT, etc.) combined with pre-training of medical-specific corpora can more accurately identify the semantic relationship between medical terms. The application of these technologies not only helps to accelerate the dissemination of medical knowledge, but also provides researchers and doctors with more intelligent literature assistance tools, improving the efficiency of knowledge acquisition and the scientific nature of decision-making. Summary of the invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that the existing document interpretation methods have the problems of low efficiency, insufficient semantic understanding, and difficulty in automatically generating optimization problems of coherent interpretation content.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for interpreting medical literature based on a large language model, comprising:

[0008] Obtain the medical literature that needs to be interpreted and perform word segmentation;

[0009] For the medical word library, a relationship network between words is constructed to obtain a first relationship network;

[0010] Matching the words in the text with the words in the medical word library to obtain a matching result of the text words; determining a first association relationship between each text word and other words in the text according to the first relationship network, and assigning a weight to the first association relationship according to the number of associations;

[0011] According to the text content, a relationship network is constructed for the words in the text to obtain a second relationship network; a second association relationship between each text word and other words in the text is determined, and a weight is assigned to the second association relationship;

[0012] According to the first relationship network and the weight assigned by the first relationship network, the associated path is optimized, and the interpretation result is obtained by interpreting the content on the path.

[0013] As a preferred solution of the medical literature interpretation method based on the language large model described in the present invention, the word segmentation processing includes using the BioBERT model to segment the content in the text;

[0014] The medical word library includes standard expressions and synonyms of all medical-specific words, and the standard expression and synonyms of each word are assimilated as a matchable word.

[0015] As a preferred solution of the medical literature interpretation method based on the language large model described in the present invention, wherein: the first relationship network includes, while entering the medical word library, entering the relationship between each word;

[0016] The words in the medical word library are used as nodes, and the words with specific semantics or logical relationships are connected, and each connection line is given a weight of unit 1.

[0017] As a preferred solution of the medical literature interpretation method based on the language large model described in the present invention, wherein: matching the words in the text with the words in the medical word library includes converting the words in the text and the words in the medical word library into vector representations using a word embedding model;

[0018] Calculate the i-th word vector x in the text i and the jth word vector x in the medical word library j The similarity between

[0019] The similarity S ij It is expressed as:

[0020]

[0021] When S ij >sim thresholdWhen , the match is successful; each word in the text is matched one by one with all the words in the medical word library in turn; during the matching process, x is calculated one by one j The corresponding standard expressions and their synonyms are i The similarity of , take the minimum value as the similarity S ij The final result;

[0022] For the i-th word vector x i The set of all matched word vectors XL i ={x i1 ,x i2 ,...,x iL};

[0023] Among them, x iL represents the Lth successfully matched word vector in the medical word library;

[0024] The first association relationship includes, after completing the word matching, obtaining all word vectors of the matched set in the text that is not an empty set; taking any two word vectors x i and x v , for x i Each element x in the matched set ia , analyze x ia With x v The connection relationship between the elements in the matched set is weighted by accumulating the number of connections;

[0025]

[0026] Among them, XX iv Represents the i-th word vector x in the text i and the vth word vector x in the text v The association weight of a represents the i-th word vector x in the text. i The index of the element in the matched set; b represents the vth word vector x in the text v The index of the element in the matched set; A represents the i-th word vector x in the text i The number of elements in the matched set; B represents the vth word vector x in the text v The number of elements in the matched set; x ia Represents the i-th word vector x in the text i The ath element in the matched set; x vb Represents the vth word vector x in the text v The bth element in the matched set; f(x ia , x vb ) represents x iaand x vb Is there a discriminant function for association? In the first relationship network, if x ia and x vb If there is a connection between them, the record is 1; if x ia and x vb If there is no connection, the record is 0.

[0027] As a preferred solution of the method for interpreting medical literature based on a large language model described in the present invention, the second relationship network includes, for the input text T = {w1, w2, ..., w d}, use the pre-trained language model to encode the context of each word;

[0028] After encoding, we get the context embedding vector for each word:

[0029]

[0030] in, Represents word w i The context embedding of w is obtained through the output of the large model, which contains the context embedding vector of the semantic information of the word in a specific context; d represents the dth word, where d represents the number of words; [i] represents the specific location of the index model output;

[0031] After obtaining the context embedding vector of each word, a neural network module is used to calculate the semantic relevance weight between each pair of words;

[0032] The formula for calculating correlation is:

[0033]

[0034] Among them, g ij Represents a word pair (w i , w r )’s semantic relevance score; and represents a trainable weight matrix; and is the bias vector, σ is the activation function; express and The splicing, word w r The context embedding vector of g ir ∈(0, 1), 1 represents word w i and w r has the strongest correlation, and 0 indicates no correlation;

[0035] After completing the calculation of the semantic association of each pair of words,ir After multiplying by the adjustable coefficient E, the actual association weight between every two words in the text is obtained.

[0036] As a preferred embodiment of the method for interpreting medical literature based on a large language model of the present invention, the optimization of the association path includes setting the starting word to be the first word in the text whose similarity with any word in the medical word library is greater than sim threshold The word, set to w start ;

[0037] From start Analyze the weight sum of other words in the text, select the word with the largest weight sum as the next word in the path, and connect them in sequence until the end;

[0038] The weight sum is expressed as:

[0039] W(cz)=P(XX cz ,0)+E×g cz

[0040] Among them, W(cz) represents word w c and w z The weight and g cz Represents word w c and w z The semantic relevance score of P(XX cz , 0) represents the selection function, if w c and w z The similarity between them and any word in the medical word library is greater than sim threshold , the output is XX cz , XX cz Represents word w c and w z The first association weight of c and w z There is no similarity with any word in the medical word library greater than sim threshold , the output is 0.

[0041] As a preferred solution of the medical literature interpretation method based on the language large model described in the present invention, the content interpretation includes taking the word string in the path as input, performing integrity assessment on the word string through a trained bidirectional LSTM network, and outputting an integrity score T;

[0042] If T is greater than the set threshold value θ, it is judged that the input word string is sufficient to constitute a coherent content; otherwise, it is judged to be unrecognizable; when the output is an unrecognizable judgment result, E starts from the initial value and increases by 1 each time. After the increase is completed, the path optimization is re-performed and the optimization result is interpreted again until the content of the path can be interpreted;

[0043] Use the pre-trained Seq2Seq generative model to generate a coherent interpretation text by calling the generation function and inputting the word string into the model:

[0044] D = Seq2SeqModel(S formatted )

[0045] Where D is the generated interpretation content, Seq2SeqModel represents the BART model used to generate coherent interpretation content, and S formatted Represents the input word string.

[0046] A medical literature interpretation system based on a language large model using the method of the present invention, wherein:

[0047] The collection unit obtains the medical literature that needs to be interpreted and performs word segmentation processing;

[0048] The word library unit constructs a relationship network between words for the medical word library to obtain a first relationship network;

[0049] The first analysis unit matches the words in the text with the words in the medical word library to obtain a matching result of the text words; determines a first association relationship between each text word and other words in the text according to the first relationship network, and assigns a weight to the first association relationship according to the number of associations;

[0050] The second analysis unit constructs a relationship network for the words in the text according to the text content to obtain a second relationship network; determines a second association relationship between each text word and other words in the text, and assigns a weight to the second association relationship;

[0051] The interpretation unit optimizes the associated path according to the first relationship network and the weight assigned by the first relationship network, and obtains the interpretation result by interpreting the content on the path.

[0052] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein: the processor implements the steps of any one of the methods of the present invention when executing the computer program.

[0053] A computer-readable storage medium stores a computer program, wherein: when the computer program is executed by a processor, the steps of any one of the methods of the present invention are implemented.

[0054] Beneficial effects of the present invention: The medical literature interpretation method based on the language big model provided by the present invention realizes the accurate interpretation of medical literature by constructing a multi-layer relationship network of the medical vocabulary library and utilizing the context analysis capability of the language big model. By optimizing the path generation method, the present invention can effectively extract the core medical information in the literature and automatically generate coherent interpretation content, thereby overcoming the shortcomings of traditional computing methods in processing efficiency and semantic understanding. Compared with the prior art, the present invention greatly improves the accuracy and efficiency of literature analysis, provides intelligent auxiliary tools for medical researchers and clinicians, and greatly improves the convenience of knowledge acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0056] Figure 1 An overall flow chart of a method for interpreting medical literature based on a large language model provided for the first embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0058] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a method for interpreting medical literature based on a language large model, comprising:

[0059] S1: Obtain the medical literature that needs to be interpreted and perform word segmentation.

[0060] Word segmentation processing includes using the BioBERT model to segment the content in the text.

[0061] Furthermore, BioBERT is a language model pre-trained based on medical corpus, which can identify professional medical vocabulary and phrases in the literature (such as drug names, symptoms, treatment methods, etc.), thereby more accurately separating key information units in medical literature.

[0062] Word segmentation is the basis of semantic analysis. After dividing the content of the document into words, it can provide basic data for the subsequent construction of the semantic relationship network between words, so that the model can analyze the contextual semantic association between words. Word segmentation can remove irrelevant or auxiliary words in the document, retain meaningful medical terms and keywords, reduce the complexity of semantic analysis, and improve the accuracy of interpretation.

[0063] S2: For the medical word library, a relationship network between words is constructed to obtain a first relationship network.

[0064] The medical word library includes standard expressions and synonyms of all medical-specific words, and the standard expression and synonyms of each word are assimilated as a matchable word.

[0065] The first relationship network includes, while entering the medical word library, entering the relationship between each word, using the words in the medical word library as nodes, connecting words with specific semantic or logical relationships, and giving each connection line a weight of unit 1.

[0066] It is worth mentioning that by node-processing the words in the medical vocabulary and building semantic relationship connections between words, a semantic network can be formed. Each connecting line represents a specific relationship between two medical words (such as the therapeutic relationship between drugs and diseases, the manifestation relationship between diseases and symptoms, etc.), which helps to analyze the logical or functional associations between words. There are a large number of synonyms and different standard expressions in medical vocabulary. Through assimilation processing, they are mapped to a standardized word node, allowing the system to recognize these synonymous expressions and reduce misunderstandings, ensuring the accuracy of vocabulary relationships.

[0067] By assigning weights to the relationships between words (such as unit 1), a standardized semantic structure is formed. Subsequent analysis can use this weight information to calculate the word association path, facilitating path optimization and semantic depth analysis.

[0068] It is also important to know that in medical literature, many medical terms do not exist independently, but appear in pairs or groups. For example, some drugs have a therapeutic relationship with specific diseases, and the description of some diseases includes specific symptoms. The first relationship network can reflect these combined relationships and provide the necessary contextual support for the interpretation process. In medical literature, terms such as diseases, symptoms, drugs, and treatment plans are related to each other. The first relationship network provides a semantic network that allows the system to reason and select association paths based on the relationship between terms to explore deeper medical information. The first relationship network not only provides a direct mapping of lexical relationships, but also helps determine the strength of association between terms through weighting. With the weight information, paths can be prioritized according to the strength of association in subsequent path optimization, ensuring that the generated content is more semantically coherent and logically clear.

[0069] S3: Match the words in the text with the words in the medical word library to obtain the matching results of the text words; determine the first association relationship between each text word and other words in the text according to the first relationship network, and assign a weight to the first association relationship according to the number of associations.

[0070] Specifically, a word embedding model is used to convert words in the text and words in the medical word library into vector representations. Calculate the i-th word vector x in the text i and the jth word vector x in the medical word library j The similarity between .

[0071] The similarity S ij It is expressed as:

[0072]

[0073] When S ij >sim threshold When , the match is successful; each word in the text is matched one by one with all the words in the medical word library in turn; during the matching process, x is calculated one by one j The corresponding standard expressions and their synonyms are i The similarity of , take the minimum value as the similarity S ij The final result.

[0074] For the i-th word vector x i The set of all matched word vectors XL i ={x i1 ,x i2 ,...,x iL}.

[0075] Among them, x iL Represents the Lth successfully matched word vector in the medical word library.

[0076] It should be noted that words in a text may be expressed in different ways, especially in the medical field, where the same concept may have multiple synonyms or near synonyms. By matching each word with the standard expression and its synonyms in the medical word library one by one, it is ensured that the same term in different expressions can be identified, and the minimum similarity value is selected as the final matching standard to obtain higher matching accuracy. The expression of medical terms is often diverse, and different words may express similar or related medical concepts. The word embedding model is used to convert the words in the text and the words in the medical word library into vector representations, which can quantify the semantic similarity between words. Through similarity calculation, these semantically associated words can be effectively captured, helping the system to still recognize the same medical concept when the words are expressed in a variety of ways. For each word in the text, the set of all matched word vectors (where each element represents a successfully matched word vector) is used for subsequent association relationship analysis. By analyzing these successfully matched word vector sets, the relationship between words can be better captured, and potential medical topics and concept associations in the literature can be identified, thereby providing data support when constructing the first relationship network.

[0077] Assume that the word in the text is "heart disease", and the medical word library contains "heart disease" (standard expression) and its synonyms "coronary heart disease" and "cardiovascular disease". Through the word embedding model, the similarity between "heart disease" and each synonym is calculated. If a synonym has the lowest similarity with "heart disease" and reaches the matching threshold, then the synonym will be considered a successful match and enter the set. This ensures that different term expressions point to standard medical concepts and ensures the accuracy of the analysis.

[0078] The first association relationship includes, after completing the word matching, obtaining all word vectors of the matched set in the text that is not an empty set; taking any two word vectors x i and x v , for x i Each element x in the matched set ia , analyze x ia With x v The connection relationship between the elements in the matched set is weighted by accumulating the number of connections.

[0079]

[0080] Among them, XX iv Represents the i-th word vector x in the text i and the vth word vector x in the text v The association weight of a represents the i-th word vector x in the text. i The index of the element in the matched set; b represents the vth word vector x in the text vThe index of the element in the matched set; A represents the i-th word vector x in the text i The number of elements in the matched set; B represents the vth word vector x in the text v The number of elements in the matched set; x ia Represents the i-th word vector x in the text i The ath element in the matched set; x vb Represents the vth word vector x in the text v The bth element in the matched set; f(x ia , x vb ) represents x ia and x vb Is there a discriminant function for association? In the first relationship network, if x ia and x vb If there is a connection between them, the record is 1; if x ia and x vb If there is no connection, the record is 0.

[0081] It is important to mention that there is a close logical or semantic relationship between some medical terms in the text, such as the relationship between disease and symptoms, and between drugs and therapies. By analyzing whether there is a connection between the elements in each two word vector sets and accumulating them, the correlation of these terms can be accurately identified, thereby constructing a first correlation relationship that reflects the actual knowledge structure in medical literature. The first correlation relationship not only identifies whether there is a correlation between terms, but also distinguishes the strength of the correlation. The weight is assigned by accumulating the number of connections between elements in the matched set, which can reflect the strength of the correlation. The higher the weight, the higher the correlation between the two terms, which are mentioned together many times in the literature. This design helps the model to give priority to high-weight paths in subsequent analysis, making the content interpretation more accurate. In medical literature, some terms often appear in fixed combinations. Through this step, these co-occurrence patterns can be captured, thereby providing data support for analyzing the relationship between medical concepts. For example, cardiovascular disease and hypertension may co-occur many times. By accumulating the weights of co-occurrence relationships, these fixed combinations can be identified, laying the foundation for the concept of medical knowledge graph.

[0082] The weighted first association is crucial in the subsequent path optimization and content generation process. By giving priority to term paths with higher weights, the system can more accurately generate interpretation content based on real semantics, improving the coherence and logic of the interpretation content. This provides reliable support for the final output of the model, ensuring that the interpretation results can accurately reflect the medical information in the literature.

[0083] S4: constructing a relationship network for words in the text according to the text content to obtain a second relationship network; determining a second association relationship between each text word and other words in the text, and assigning a weight to the second association relationship.

[0084] Furthermore, the second relationship network includes, for the input text T = {w1, w2, ..., w d}, using a pre-trained language model to encode the context of each word.

[0085] After encoding, we get the context embedding vector for each word:

[0086]

[0087] in, Represents word w i The context embedding of w is obtained through the output of the large model, which contains the context embedding vector of the semantic information of the word in a specific context; d Represents the dth word, where d represents the number of words; [i] represents the specific location of the index model output.

[0088] After obtaining the context embedding vector of each word, a neural network module is used to calculate the semantic relevance weight between each pair of words. The formula for calculating relevance is:

[0089]

[0090] Among them, g ij Represents a word pair (w i , w r )’s semantic relevance score; and represents a trainable weight matrix; and is the bias vector, σ is the activation function (such as Sigmoid function); express and The splicing, word w r The context embedding vector of g ir ∈(0, 1), 1 represents word w i and w r has the strongest association, and 0 indicates no association.

[0091] After completing the calculation of the semantic association of each pair of words, ir After multiplying by the adjustable coefficient E, the actual association weight between every two words in the text is obtained.

[0092] What I want to say is that the concatenated vector contains w i , wr The contextual information of the two words is therefore more helpful in capturing their semantic relevance. This concatenation method allows the model to focus on the semantic relationship between the two words in context while learning. Through concatenation, the model can perform nonlinear operations based on the combined information of the two words, such as weighting of the weight matrix and transformation of the activation function, which allows the model to perform complex interactive calculations on the context vectors of the two words and extract more advanced semantic association features. After inputting the concatenated context vector, the neural network can learn more complex relationships through multiple levels of weights, not just the features of a single word. This makes the model more discriminative when faced with diverse contexts and complex semantic relationships. Using the same concatenated input format for all word pairs helps keep the neural network calculations simple and consistent, and avoids inconsistencies or information loss when calculating word vectors one by one.

[0093] It is also important to know that the second relationship network is a relationship network generated based on the context of the text, which is different from the static relationship of the first relationship network based on the medical vocabulary. Through the second relationship network, the system can build a vocabulary relationship network that is closer to the actual context of the text, reflecting the real association of words in the context of the document. This dynamic relationship helps to capture the subtle semantic associations in the document, so as to more accurately understand the connection between medical concepts. The semantic relevance weight in the second relationship network is calculated by neural network. Using the context embedding vectors of the words, by splicing and inputting them into the neural network module, the semantic relevance score of each pair of words can be obtained. The relevance score ranges from 0 to 1, indicating the degree from no association to the strongest association. In this way, the system can quantify the relationship between words in detail, making the generated relationship network more accurate and expressive. The dynamic semantic relationship of the second relationship network provides richer contextual basis for content generation. A high-weighted semantic relevance score means that the contextual association of the two words in the document is strong, which is very important for content generation. By giving priority to high-weight paths, the system can generate more coherent interpretation content that conforms to actual semantic logic, thereby ensuring the accuracy and natural fluency of the output content. Suppose a medical document contains two words, "heart disease" and "blood pressure". Through context encoding and neural network relevance calculation, their context embedding vectors are obtained, and a high semantic relevance score (such as 0.8) is calculated. This score indicates that in the specific context of the current document, "heart disease" and "blood pressure" have a strong semantic relevance. This high relevance can prompt the system to prioritize connecting these two concepts when interpreting the content, thereby generating explanatory content that contains the relationship between the two, such as "heart disease patients often have blood pressure problems."

[0094] S5: optimizing the associated path according to the first relationship network and the weight assigned by the first relationship network, and obtaining an interpretation result by interpreting the content on the path.

[0095] Set the starting word to be the first word in the text that has a similarity greater than sim threshold The word, set to w start . From w start Analyze the weight sum of other words in the text, select the word with the largest weight sum as the next word in the path, and connect them in sequence until the end.

[0096] The weight sum is expressed as:

[0097] W(cz)=P(XX cz ,0)+E×g cz

[0098] Among them, W(cz) represents word w c and w z The weight and g cz Represents word w c and w z The semantic relevance score of P(XX cz , 0) represents the selection function, if w c and w z The similarity between them and any word in the medical word library is greater than sim threshold , the output is XX cz , XX cz Represents word w c and w z The first association weight of c and w z There is no similarity with any word in the medical word library greater than sim threshold , the output is 0.

[0099] In the literature, the relationship network between medical terms and concepts may be complex. Through path optimization, the most semantically related paths can be identified. Starting from the first word with a high similarity to the medical vocabulary, by selecting the next word with the largest weight and the largest sum, the system ensures that it can move along the path with the closest semantic association when interpreting. This allows the interpretation content to focus on the core semantics and avoid interference from irrelevant information. During the path optimization process, the most relevant word order is selected by accumulating weights and the sum, ensuring the coherence of the interpretation content. Each step selects the best next word based on the weight and the sum, so that the system can generate interpretation content that conforms to the logical order and avoids breaks or semantic jumps in the interpretation results.

[0100] By introducing the semantic relevance score to calculate the weight and sum, path optimization not only considers the direct relationship between the two words when selecting the next word, but also refers to the strength of their semantic relationship. The selection function is used to determine the relevance of each pair of words and determine their weights, ensuring that the path optimization process gives priority to word combinations with high similarity to the terms in the medical word library, thereby improving the accuracy and semantic consistency of interpretation.

[0101] The word string in the path is taken as input, and the integrity of the word string is evaluated through the trained bidirectional LSTM network, and an integrity score T is output.

[0102] If T is greater than the set threshold θ, it is judged that the input word string is sufficient to constitute a coherent content; otherwise, it is judged to be unrecognizable; when the output is an unrecognizable judgment result, E starts from the initial value and increases by 1 each time. After the increase is completed, the path optimization is re-performed and the optimization result is interpreted again until the content of the path can be interpreted.

[0103] It should be noted that when the score T does not reach the set coherence threshold, it indicates that the interpretation content may lack sufficient information. By gradually increasing the weight coefficient E, we can pay more attention to the core semantic relationship in the path optimization process, increase the proportion of semantics, and thus strengthen the semantic relevance of the interpretation content. The increasing semantic weight makes the system more inclined to choose highly related vocabulary combinations when optimizing the path, so as to ensure the accuracy and depth of the final interpretation content.

[0104] When the score T does not meet the threshold requirement, the gradual increase of the weight coefficient E will guide the system to re-optimize the path and add more keywords. By introducing more related words, the system can capture richer semantic information and make the interpretation content more hierarchical and coherent. This design ensures that the interpretation content contains enough key information while making the content more complete and thorough. When the interpretation content is not coherent enough, direct output will affect the quality of the interpretation, so this step avoids the generation of meaningless or incomplete interpretations through completeness evaluation. When the content is judged as "unrecognizable", the system will trigger the semantic enhancement mechanism and introduce more related information in the optimization process to ensure that the output content meets the set coherence standard, thereby improving the quality of the interpretation results. By repeatedly increasing the semantic weight and re-optimizing the path, an adaptive interpretation content optimization mechanism is formed. This design enables the system to adaptively adjust the degree of semantic attention according to the interpretation score, and continuously optimize the path selection until the content that meets the coherence requirements is generated. The iterative mechanism ensures that the system can adapt to text content of different complexities, making the interpretation results more accurate and reliable.

[0105] Suppose the word string "coronary heart disease, myocardial ischemia, surgery" obtained by path optimization is scored by the bidirectional LSTM network, and the result TTT does not reach the threshold. This indicates that the path may lack a certain degree of semantic coherence. The system will start the process of increasing the weight coefficient EEE, re-optimize the path, and may introduce related words such as "coronary artery" or "interventional treatment" to further enhance the medical semantic information in the interpretation content. This iteration is repeated until the interpretation content with sufficient semantic information and clear logic is generated, such as "coronary heart disease may cause myocardial ischemia, which usually requires interventional treatment or surgery."

[0106] Use the pre-trained Seq2Seq generative model to generate a coherent interpretation text by calling the generation function and inputting the word string into the model:

[0107] D = Seq2SeqModel(S formatted )

[0108] Where D is the generated interpretation content, Seq2SeqModel represents the BART model used to generate coherent interpretation content, and S formatted Represents the input word string.

[0109] It should be noted that the BART model, as a type of Seq2Seq architecture, has been pre-trained on a large-scale corpus and has strong natural language understanding and generation capabilities. It can generate coherent and fluent interpretation texts based on the context after receiving the input key word string. This generation capability is particularly suitable for inferring and completing complete medical interpretation content from incomplete term strings, making the output more readable and logical. Seq2Seq models such as BART have the characteristics of processing incomplete inputs, and can understand the potential semantics and generate coherent text from the incomplete information in the form of word strings. Compared with traditional rule matching or template filling, BART can accurately infer reasonable interpretation content in the case of incomplete or fragmented input through the understanding of context association, thereby filling in the semantic gaps in the word string. The bidirectional encoder part of the BART model can capture the contextual semantics of the input word string, while its decoder part can generate coherent text with long-range dependencies. For the interpretation of medical literature, the contextual dependency and logical relationship between words are crucial. The bidirectional encoding capability of BART ensures that the input word string is reasonably connected at the semantic level, thereby outputting interpretation content that conforms to medical logic. Seq2Seq models such as BART can combine pre-trained knowledge during the generation process to bring more details and extensions to word strings. For example, when the input contains keywords such as "heart disease", "treatment" and "drug", BART can combine the medical knowledge learned during its pre-training to automatically generate explanations about treatment plans or drug selection for heart disease. This design greatly enriches the level and amount of information in the interpretation content, making the output more professional and practical.

[0110] Assuming the input word string is "coronary heart disease", "interventional treatment", and "drug", BART can generate a complete interpretation based on this input, such as "Coronary heart disease is a common cardiovascular disease, which is usually treated with interventional treatment and combined with drug control." This interpretation is not just a simple combination of word strings, but provides a complete medical background, treatment methods, and measures. This is the value of the BART generative model design.

[0111] On the other hand, this embodiment also provides a medical literature interpretation system based on a language large model, which includes:

[0112] The collection unit obtains the medical literature that needs to be interpreted and performs word segmentation processing.

[0113] The vocabulary unit constructs a relationship network between words for the medical vocabulary to obtain a first relationship network.

[0114] The first analysis unit matches the words in the text with the words in the medical word library to obtain a matching result of the text words; determines the first association relationship between each text word and other words in the text according to the first relationship network, and assigns a weight to the first association relationship according to the number of associations.

[0115] The second analysis unit constructs a relationship network for words in the text according to the text content to obtain a second relationship network; determines a second association relationship between each text word and other words in the text, and assigns a weight to the second association relationship.

[0116] The interpretation unit optimizes the associated path according to the first relationship network and the weight assigned by the first relationship network, and obtains the interpretation result by interpreting the content on the path.

[0117] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0118] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0119] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0120] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0121] Example 2 is an embodiment of the present invention, which provides a method for interpreting medical literature based on a large language model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0122] This experiment uses "intelligent interpretation system A" (based on the present invention) and "standard interpretation system B" (control system) to evaluate the performance, and examines the system's accuracy, time consumption, stability of multiple recognitions, content coherence, and user satisfaction under different experimental conditions such as the amount of literature, text complexity, and number of keywords, as shown in Table 1.

[0123] The experimental environment includes three test document sets, each containing 50, 100 and 150 documents. The documents cover complex medical topics such as cardiovascular disease, cancer, diabetes, etc., with complex content structure and rich terminology. The number of key terms in each document ranges from 20 to 100, involving different diseases, treatment methods, drugs, etc. During the experiment, both System A and System B identified key terms from each set of documents, made semantic associations, and interpreted content generation.

[0124] The process of intelligent interpretation system A: System A uses the BioBERT model to perform word segmentation, identify medical terms, and match them with the medical vocabulary. By constructing the first relationship network, the standardized expression of the term and its synonyms are mapped to the same node, and the initial weights are assigned according to the semantic relationship and logical association between the terms. Next, system A uses the BART generation model to construct the second relationship network, generates the semantic relevance weights of word pairs through context embedding, and uses the path optimization algorithm to select the path with the highest relevance in the literature to generate coherent content. Finally, system A uses a bidirectional LSTM network to score the coherence of the generated path, and increases the weight to adjust the optimization path when the score is insufficient.

[0125] Standard interpretation process of system B: System B uses the traditional keyword matching method, without relationship network and path optimization, and only identifies keywords in the text and outputs them in simple combination. System B lacks semantic association when generating content, and only generates content based on surface matching, lacks coherence and association optimization, and has limited interpretation effect.

[0126] Table 1 Experimental data table

[0127]

[0128]

[0129] From the above table data, it can be seen that the term recognition accuracy of system A is always above 95% in different numbers of documents, while the accuracy of system B is between 80% and 85%. The high accuracy of system A comes from the medical term recognition ability of the BioBERT model, as well as the standardization and synonym mapping processing through the medical word library and the first relationship network. This enables system A to accurately identify key terms in medical documents and reduce misidentification. In contrast, system B has limited recognition ability for synonyms and medical terms and low accuracy due to the lack of vocabulary standardization, which cannot meet the needs of high-precision medical document interpretation.

[0130] The time consumption of system A is slightly higher than that of system B. As the number of documents increases, the processing time of system A increases more significantly. System A is time-consuming because it involves relationship network construction, context encoding, and path optimization. However, these steps significantly improve the quality and logic of the interpretation content. In contrast, system B is less time-consuming, but the interpretation content is fragmented and cannot generate a complete semantic structure. Although system A is more time-consuming, the increased time investment is reasonable and necessary for medical literature analysis scenarios that require high-quality interpretation.

[0131] System A performed well in terms of multiple recognition stability, with stability maintained between 92% and 94%, significantly higher than System B's 80% to 85%. System A ensures that relatively consistent interpretation content can be generated during each interpretation process through the construction of a multi-layer relationship network and a dynamic weight adjustment mechanism. System B lacks a semantic network and path optimization mechanism, and the results are highly volatile, and the stability of multiple recognitions cannot be guaranteed. For the interpretation of medical literature, stability is a key factor in ensuring content coherence and accurate information. The advantages of System A in this regard further prove its reliability in practical applications. (The stability of multiple recognition results refers to the degree of consistency of the recognized content during the test. In this experiment, the minimum value of the content consistency is taken as the "multiple recognition result")

[0132] System A always maintains a content coherence score of more than 9.2, while System B's score is between 6.8 and 7.0. The path optimization algorithm used by System A can give priority to paths with higher semantic weights, forming semantically coherent interpretation content when generating content. However, System B only extracts keywords, and the generated interpretation content lacks logic and contextual association, so the content coherence is poor. The design of System A is particularly suitable for the coherent interpretation of complex content in medical literature, and can provide more logical content output.

[0133] System A has a higher user satisfaction score than System B, with a stable score between 9.1 and 9.3, while System B's score is only between 6.3 and 6.8. System A performs well in terms of accuracy, coherence, and stability of multiple recognitions, making users more satisfied with its interpretation results. Especially in the application of medical literature interpretation, the semantic association and optimization mechanism of System A make it more suitable for professional needs. System B's interpretation content is fragmented and difficult to meet user needs, and its satisfaction score is low.

[0134] It can be seen from the experimental data that system A is significantly better than system B in terms of accuracy, multiple recognition stability and content coherence. Although system A is slightly more time-consuming, it has demonstrated outstanding innovation and reliability in overall performance by optimizing the interpretation path, improving semantic associations and ensuring content coherence. The design of system A not only improves the quality of interpretation, but also meets the high standards of medical literature interpretation for accurate information, logical coherence and consistent content. These experimental data fully demonstrate the advantages of the present invention, especially in the interpretation of complex medical content, which has extremely high application value.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for interpreting medical literature based on a large language model, characterized in that: include: Obtain the medical literature that needs to be interpreted and perform word segmentation; For the medical word library, a relationship network between words is constructed to obtain a first relationship network; Matching the words in the text with the words in the medical word library to obtain a matching result of the text words; determining a first association relationship between each text word and other words in the text according to the first relationship network, and assigning a weight to the first association relationship according to the number of associations; According to the text content, a relationship network is constructed for the words in the text to obtain a second relationship network; a second association relationship between each text word and other words in the text is determined, and a weight is assigned to the second association relationship; According to the first relationship network and the weight assigned by the first relationship network, the associated path is optimized, and the interpretation result is obtained by interpreting the content on the path.

2. The method for interpreting medical literature based on a language large model as claimed in claim 1, characterized in that: The word segmentation processing includes using the BioBERT model to segment the content in the text; The medical word library includes standard expressions and synonyms of all medical-specific words, and the standard expression and synonyms of each word are assimilated as a matchable word.

3. The method for interpreting medical literature based on a language large model as claimed in claim 2, characterized in that: The first relationship network includes, while entering the medical word library, entering the relationship between each word; The words in the medical word library are used as nodes, and the words with specific semantics or logical relationships are connected, and each connection line is given a weight of unit 1.

4. The method for interpreting medical literature based on a language large model as claimed in claim 3, characterized in that: Matching the words in the text with the words in the medical word library includes converting the words in the text and the words in the medical word library into vector representations using a word embedding model; Calculate the i-th word vector x in the text i and the jth word vector x in the medical word library j The similarity between The similarity S ij It is expressed as: When S ij >sim threshold When , the match is successful; each word in the text is matched one by one with all the words in the medical word library in turn; during the matching process, x is calculated one by one j The corresponding standard expressions and their synonyms are i The similarity of , take the minimum value as the similarity S ij The final result; For the i-th word vector x i The set of all matched word vectors XL i ={x i1 ,x i2 ,...,x iL }; Among them, x iL represents the Lth successfully matched word vector in the medical word library; The first association relationship includes, after completing the word matching, obtaining all word vectors of the matched set in the text that is not an empty set; taking any two word vectors x i and x v , for x i Each element x in the matched set ia , analyze x ia With x v The connection relationship between the elements in the matched set is weighted by accumulating the number of connections; Among them, XX iv Represents the i-th word vector x in the text i and the vth word vector x in the text v The association weight of a represents the i-th word vector x in the text. i The index of the element in the matched set; b represents the vth word vector x in the text v The index of the element in the matched set; A represents the i-th word vector x in the text i The number of elements in the matched set; B represents the vth word vector x in the text v The number of elements in the matched set; x ia Represents the i-th word vector x in the text i The ath element in the matched set; x vb Represents the vth word vector x in the text v The bth element in the matched set; f(x ia , x vb ) represents x ia and x vb Is there a discriminant function for association? In the first relationship network, if x ia and x vb If there is a connection between ia and x vb If there is no connection, the record is 0.

5. The method for interpreting medical literature based on a language large model as claimed in claim 4, characterized in that: The second relationship network includes, for the input text T={w1,w2,…,w d }, use the pre-trained language model to encode the context of each word; After encoding, we get the context embedding vector for each word: in, Represents word w i The context embedding of w is obtained through the output of the large model, which contains the context embedding vector of the semantic information of the word in a specific context; d represents the dth word, where d represents the number of words; [i] represents the specific location of the index model output; After obtaining the context embedding vector of each word, a neural network module is used to calculate the semantic relevance weight between each pair of words; The formula for calculating correlation is: Among them, g ij Represents a word pair (w i , w r )’s semantic relevance score; and represents a trainable weight matrix; and is the bias vector, σ is the activation function; express and The splicing, word w r The context embedding vector of g ir ∈(0, 1), 1 represents word w i and w r has the strongest correlation, and 0 indicates no correlation; After completing the calculation of the semantic association of each pair of words, ir After multiplying by the adjustable coefficient E, the actual association weight between every two words in the text is obtained.

6. The method for interpreting medical literature based on a language large model as claimed in claim 5, characterized in that: The optimization of the association path includes setting the starting word to be the first word in the text whose similarity with any word in the medical word library is greater than sim threshold The word, set to w start ; From start Analyze the weight sum of other words in the text, select the word with the largest weight sum as the next word in the path, and connect them in sequence until the end; The weight sum is expressed as: W(cz)=P(XX cz ,0)+E×g cz Among them, W(cz) represents word w c and w z The weight and g cz Represents word w c and w z The semantic relevance score of P(XX cz , 0) represents the selection function, if w c and w z The similarity between them and any word in the medical word library is greater than sim threshold , the output is XX cz , XX cz Represents word w c and w z The first association weight of c and w z There is no similarity with any word in the medical word library greater than sim threshold , the output is 0.

7. The method for interpreting medical literature based on a language large model as claimed in claim 6, characterized in that: The content interpretation includes taking the word string in the path as input, performing integrity assessment on the word string through a trained bidirectional LSTM network, and outputting an integrity score T; If T is greater than the set threshold value θ, it is judged that the input word string is sufficient to constitute a coherent content; otherwise, it is judged to be unrecognizable; when the output is an unrecognizable judgment result, E starts from the initial value and increases by 1 each time. After the increase is completed, the path optimization is re-performed and the optimization result is interpreted again until the content of the path can be interpreted; Use the pre-trained Seq2Seq generative model to generate a coherent interpretation text by calling the generation function and inputting the word string into the model: D=Seq2SeqModel(S formatted ) Where D is the generated interpretation content, Seq2SeqModel represents the BART model used to generate coherent interpretation content, and S formatted Represents the input word string.

8. A medical literature interpretation system based on a language large model using the method according to any one of claims 1 to 7, characterized in that: The collection unit obtains the medical literature that needs to be interpreted and performs word segmentation processing; The word library unit constructs a relationship network between words for the medical word library to obtain a first relationship network; The first analysis unit matches the words in the text with the words in the medical word library to obtain a matching result of the text words; determines a first association relationship between each text word and other words in the text according to the first relationship network, and assigns a weight to the first association relationship according to the number of associations; The second analysis unit constructs a relationship network for the words in the text according to the text content to obtain a second relationship network; Determine a second association relationship between each text word and other words in the text, and assign a weight to the second association relationship; The interpretation unit optimizes the associated path according to the first relationship network and the weight assigned by the first relationship network, and obtains the interpretation result by interpreting the content on the path.

9. A computer device comprising: A memory and a processor; the memory stores a computer program, wherein the processor implements the steps of any method as claimed in claim 1 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images