LLM term understanding method, system and device and storage medium

Abstract

The invention discloses an LLM term understanding method, system and device and a storage medium, and the method comprises the steps: obtaining and recognizing professional terms in user input, carrying out the LLM input enhancement through constructing scene description for each professional term, sending the enhanced user input into a pre-trained LLM to complete semantic analysis, and carrying out the semantic analysis. And performing multi-dimensional reordering on a semantic analysis result in combination with LLM initial confidence, context alignment and user preference, taking the paraphrase with the highest score as the optimal semantics input this time, performing ambiguity processing on the optimal semantics, completing a semantic analysis process, and realizing iterative optimization based on user feedback. According to the method, the accuracy of preliminary understanding is ensured, the interpretability and the preintervention of the whole term understanding process are improved, the method can adapt to habits and requirements of different users, and continuous iterative optimization can be performed to keep pace with the times.

Description

Technical Field

[0001] The present invention relates to a method for term understanding, and particularly to an LLM term understanding method, system, device, and storage medium. Background Art

[0002] Large Language Models (LLMs) have made significant progress in general natural language understanding, but still face challenges when dealing with "typical terms" commonly used in specific professional fields or specific user groups. These terms often have polysemy, that is, the same term may have very different meanings in general context and professional context. For example, when a user inputs "frequency modulation", its meaning in everyday terms is "adjusting the sound frequency, radio, etc.", while in the power professional field, it more often means "adjusting the power generation to stabilize the grid frequency". Existing LLMs, due to the extensiveness of their training data, often lack an in-depth understanding of the precise meaning, nuances, and unique referentiality of specific domain-specific terms in a specific context, and cannot make a correct initial interpretation in professional scenarios, thus affecting subsequent intention judgment.

[0003] Currently, the accuracy of intention recognition mainly depends on the initial semantic segmentation and entity recognition of the input text. Once the model fails to accurately grasp the meaning of the core professional terms in the initial stage, subsequent intention judgment is likely to deviate from the right track. Even if subsequent clarification is carried out, it will increase the interaction cost.

[0004] In addition, it is difficult for existing technologies to effectively quantify and utilize users' personalized usage habits, understanding preferences for specific professional terms, as well as the general importance and recognized interpretations of a certain term in a specific domain. Even with user portraits or knowledge bases, this information often intervenes in the later stage of semantic parsing, resulting in a very limited early intention guiding role and making it more difficult to quickly adapt to the evolution of domain knowledge and the changes in users' expression habits. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide an LLM term understanding method and system, which improve the model's perception ability of domain knowledge and users' implicit intentions, and dynamically and efficiently adjust the importance of specific terms for specific users or the entire domain and their association strength with specific intentions according to users' real-time feedback, so as to achieve the system's rapid adaptation to and continuous maintenance of specific domains. Another object of the present invention is to provide a device and a computer-readable storage medium for implementing the above LLM term understanding method.

[0006] Technical Solution: The LLM term understanding method of the present invention includes:

[0007] Preprocess the user input and perform professional term recognition to form a set of professional terms;

[0008] LLM input enhancement is achieved by constructing scenario-based guided descriptions for each term in the terminology set.

[0009] Perform LLM semantic parsing with input enhancement;

[0010] The LLM semantic parsing results are reordered in multiple dimensions, including: initial LLM confidence, context alignment, and user preference, to obtain the optimal semantic representation of the current input parsing.

[0011] Ambiguity processing is performed on the optimal semantic representation to complete the semantic parsing process and iterative optimization based on user feedback.

[0012] Optionally, the preprocessing of user input may employ NLP techniques, such as word segmentation and part-of-speech tagging.

[0013] Optionally, the terminology recognition includes: comparing the preprocessed user input with the terminology list of the dynamic knowledge base DKB using one or more of string matching, fuzzy matching, and word vector-based similarity calculation methods, to identify all the terminology in the user input, and denoting them as the set PTidentified = {pt1, pt2, ..., pt...} n}

[0014] Optionally, the construction of a contextualized guided description for LLM input enhancement for each term in the terminology set includes:

[0015] Find the current professional terminology pt from the Dynamic Knowledge Base (DKB). i All related definitions;

[0016] Calculate the relevance score between the usage scenario of each definition and the user's current input, and select the definition with the highest relevance score as the current term. i The best guiding interpretation;

[0017] According to the best guiding interpretation, the current professional term pt is... i Construct its bootstrap description;

[0018] All identified technical terms are combined into a general guiding context, which serves as the contextualized guiding description for the user input.

[0019] Optionally, the LLM semantic parsing after input enhancement includes: combining user input and its contextualized guided description to form enhanced LLM input and inputting it into a pre-trained LLM; the LLM performs semantic parsing for each identified technical term (pt). i Output K possible semantic understanding results (SSR) jj = 1, ..., K and their initial confidence scores in LLM ConfLLM(SSR) j ).

[0020] Optionally, the multi-dimensional reordering includes:

[0021] Obtain the identified technical terms (pt) i In SSR j The best interpretation in the context; among them, SSR j j = 1, ..., K represents the input augmentation LLM for each identified terminology pt i Output K possible semantic understanding results;

[0022] SSR is performed on the user input based on the best interpretation. j Context alignment within the scene, and calculate the context alignment degree of the current scene;

[0023] Calculate SSR based on global preferences and user preferences j DKB preference score in the scenario;

[0024] Based on semantic understanding results SSR j Calculate SSR using LLM initial confidence, context alignment, and DKB preference score. j The scores are reordered, and the semantic understanding result with the highest score is taken as the optimal semantic representation result for parsing the user input.

[0025] Furthermore, the SSR j The context alignment in a scene is calculated using the following formula:

[0026]

[0027] For each technical term pt i S align (SSR j ,pt i The calculation formula is as follows:

[0028] S align (SSR j ,pt i )=Similarity(Interpi,best_match.Contextual_Keywords,Q user \PT identified )

[0029] Where Interpi,best_match is an SSR j The best interpretation in the context, Contextual_Keywords is the description corresponding to the best interpretation, Quser \PT identified For non-technical parts of user input.

[0030] Furthermore, the calculation of SSR j DKB Preference Score in Scenario pref_total (SSR j The following formula applies:

[0031]

[0032] S pref (SSR j ,pt i )=w1·Interpi,best_match.GPS+w2·Interpi,best_match.UPS

[0033] Where w1 and w2 are the weight coefficients of global preference and user preference, respectively; Interpi is the technical term pt. i Corresponding definitions (may include one or more); GPS stands for Global Preference Score, i.e., Interp. i A score representing the prevalence or correctness within the domain; UPS is a user preference score, i.e., a user's historical preference for Interp. i User-specific scores based on preferences.

[0034] Furthermore, the calculation of SSR j The reordering score is calculated using the following formula:

[0035] S final (SSR j )=λ1·normalize(Conf LLM (SSR j ))+λ2·normalize(S align_total (SSR j ))+λ3·normalize(S p ref_total(SSR j ))

[0036] Where λ1, λ2, and λ3 are hyperparameters used to balance the initial confidence level of the LLM. LLM (SSR j ), Context Alignment S align_total (SSR j ) and DKB preference score S pref_total (SSR j ), where normalize is the normalization function.

[0037] Optionally, the ambiguity handling includes:

[0038] Compare the second-ranked score in the multi-dimensional re-ranking with the optimal semantic representation;

[0039] If the difference between the two scores is less than a given threshold, or the score of the optimal semantic representation is less than a given threshold, then the user's current input is considered ambiguous. The ambiguous result or the optimal semantic representation below the given threshold must be submitted to the user for confirmation. Based on user feedback, UPS records are updated, or new definitions for technical terms are added to the review database. These are periodically reviewed by industry experts and added to the dynamic knowledge base (DKB) after passing the review. Based on user feedback, an adjusted optimal semantic representation is generated, and subsequent business operations are executed based on this result.

[0040] If the given threshold is met, subsequent business operations are performed based on the optimal semantic representation, and the semantic parsing process is completed.

[0041] The LLM terminology understanding system includes:

[0042] The terminology recognition module is used to preprocess user input and identify all technical terms to form a terminology set.

[0043] The input enhancement module is used to construct guided descriptions for each technical term. It combines the guided descriptions of all identified technical terms into a general guided context, which serves as the contextualized guided description for the user input. This contextualized ...

[0044] The semantic parsing module is used to perform multi-dimensional reordering of the LLM semantic parsing results after input augmentation to obtain the optimal semantic representation of the current input parsing;

[0045] The ambiguity handling and iterative optimization module is used to handle ambiguity in the optimal semantic representation, complete the semantic parsing process, and perform iterative optimization based on user feedback.

[0046] The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement some or all of the steps in the LLM terminology understanding method described above.

[0047] The computer-readable storage medium has a computer program stored thereon, which, when executed by a processor, implements some or all of the steps in the LLM terminology understanding method described above.

[0048] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0049] 1. By assigning dynamic personalized weights (for specific users) and global weights (for the entire professional field) to professional terms, and introducing these weight information to intervene or guide the large language model when it performs preliminary semantic understanding and intent assumptions, the likelihood of the model moving toward the correct professional intent in the early stages of parsing is significantly increased, reducing subsequent chain errors caused by initial comprehension biases.

[0050] 2. By continuously learning from user feedback, we establish unique understanding preferences (personalized weights) for each user regarding professional terms, thereby meeting the technical needs of "personalized" semantic parsing services and better aligning with individual users' work habits and expression styles.

[0051] 3. The global weighting system can dynamically reflect the general importance of professional terms in a specific field, their mainstream interpretations, and their relevance to core business intentions by summarizing feedback from group users and continuous maintenance by domain experts, thus keeping knowledge "fresh" and keeping it up-to-date.

[0052] 4. The adoption of DKB and reordering algorithms makes the model's understanding of terms easier to analyze and adjust. When unexpected term interpretations occur, intervention can be made by adjusting DKB or weights, demonstrating good interpretability and operability. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0055] like Figure 1 As shown, an LLM terminology comprehension method includes the following steps:

[0056] (1) Receive user input Quser, identify professional terms in it, and form a set of professional terms.

[0057] (11) Before performing technical term recognition, NLP and other technologies are used to preprocess the user input Quser, such as word segmentation and part-of-speech tagging.

[0058] (12) Compare the output of step (11) with the PT list in the dynamic knowledge base DKB. The comparison method includes, but is not limited to, string matching, fuzzy matching, or word vector-based similarity calculation, to identify all professional terms in the input. The set of identified terms is denoted as PTidentified = {pt1, pt2, ..., pt...} n}

[0059] Taking the power system as an example, a user inputs the following message: "Urgent! The No. 2 main transformer at the 500kV Mingguang substation just tripped. Check the protection logic in effect at that time and the relevant tripping history. Also, see what impact this event has on the power flow of the line?"

[0060] The construction and maintenance of the dynamic knowledge base (DKB) are shown in Table 1 below.

[0061] Table 1

[0062]

[0063] The identified set of technical terms is PTidentified = {"main transformer", "trip", "protection logic", "trip history", "power flow"}.

[0064] (2) For each pt in PTidentified i Build scenario-based guided descriptions for LLM input enhancement.

[0065] (21) As shown in Table 1 above, since each term has multiple definitions, the corresponding professional term pt is found from the Dynamic Knowledge Base (DKB). i All related definitions are denoted as InterpSet i ={Interp i,1 Interp i,2 ,...}. Among them, Interp i,1 Interp i,2 ... represents the i-th technical term pt in PTidentified. i All definitions of .

[0066] Regarding the term "main transformer": the user input clearly points to an "emergency" "tripping" event, which is highly related to the "locating specific equipment during fault analysis" applicable scenario in serial number 1. Therefore, this term corresponds to the interpretation of serial number 1.

[0067] Similarly:

[0068] The term "trip" corresponds to the definition in number 3;

[0069] The term "protection logic" corresponds to the definition in number 5;

[0070] The term "tripping history" corresponds to the definition in number 7;

[0071] The term "trend" corresponds to the definition in number 8.

[0072] (22) Calculate the relevance of the usage scenario of each definition to the user's current input, and denote the set of relevance scores as Score. interpi={score interpi,1 score interpi,2 ,...,score interpi,t} indicates the current technical term pt i The relevance scores of the t definitions to the user's current input are used. The definition with the highest score is taken as the current term pt. i The best guiding definition, and based on the best definition, the current professional term pt i Constructing Guided Descriptions (GD) of Terminology i .

[0073] The template for guided descriptions is similar: The user mentions the term "${PT}" in their input. Based on the user's usage habits and the context, the most likely meaning of this term is "${Interp}".

[0074] Therefore, for the term "main variable", the final guiding description is as follows:

[0075] The user mentioned the term "main transformer" in their input. Based on the user's usage habits and the context, the most likely meaning of this term is: "the main transformer in the power system".

[0076] (23) Guided descriptions of all identified technical terms GD i Combined into a general guiding context GC = {GD1, GD2, ..., GD...} n} serves as a contextualized guidance description for the user input.

[0077] For example, in one embodiment, the final GC is as follows:

[0078] The user mentioned the term "main transformer" in their input. Based on the user's usage habits and scenario, this term most likely means "the main transformer in the power system"; the user mentioned the term "trip" in their input. Based on the user's usage habits and scenario, this term most likely means "the circuit breaker automatically trips due to a fault or protection action"; the user mentioned the term "protection logic" in their input. Based on the user's usage habits and scenario, this term most likely means "the set of rules used by relay protection devices to determine faults and issue trip commands"; the user mentioned the term "trip history" in their input. Based on the user's usage habits and scenario, this term most likely means "the record of past trip events of the equipment"; the user mentioned the term "power flow" in their input. Based on the user's usage habits and scenario, this term most likely means "the flow and distribution of electrical energy in the power system".

[0079] Emphasizing terminology and its definitions in GC helps large models extract correct SSRs using their contextual learning capabilities. This GC does not include terminology.

[0080] (3) Perform LLM semantic parsing after input enhancement.

[0081] (31) The original user input Quser is combined with the guiding context GC to form the enhanced LLM input Qaugmented. The construction of Qaugmented includes, but is not limited to, combinations of user input (i.e., the original problem) Quser followed by GC.

[0082] (32) Input Qaugmented into the pre-trained LLM and require the LLM to be tailored to each identified pt. i Output K (K>=0) possible semantic understanding results (SSR candidates). Each SSR j It should be clearly stated that it is related to pt i A more detailed understanding is needed. The final output is CSSR = {(SSR1, ConfLLM(SSR1)), ..., (SSR...}. K ,ConfLLM(SSR K ))}, where ConfLLM(SSR j ) represents the confidence level of the LLM in the output.

[0083] Below are two examples of generated SSR candidates, simplified in JSON format.

[0084] SSR_A (Our expected optimal solution, initial LLM confidence level: 0.78):

[0085]

[0086] SSR_B (High-confidence LLM solution, but with slight misunderstandings; initial LLM confidence: 0.85):

[0087]

[0088] (4) Based on context awareness and user preferences, the LLM semantic parsing results are reordered in multiple dimensions.

[0089] (41) Get each pt in PTidentified i In the current SSR j The best interpretation of Interpi, best_match in the scenario. Methods for obtaining this include, but are not limited to, using SSR. j Each pt i The description is compared with all the definitions in DKB, and the one with the highest similarity is Interpi,best_match.

[0090] The above SSR A The terminology mapping in this scenario is as follows:

[0091] Interp1, best_match(SSR A ) = Interp 1,1

[0092] Interp2, best_match(SSR A ) = Interp 2,1

[0093] Interp3, best_match(SSR A ) = Interp 3,1

[0094] Interp4, best_match(SSR A ) = Interp 4,1

[0095] Interp5, best_match(SSR A ) = Interp 5,1

[0096] The above SSR B The terminology mapping in this scenario is as follows:

[0097] Interp1, best_match(SSR B ) = Interp 1,1

[0098] Interp2, best_match(SSR B ) = Interp 2,1

[0099] Interp3, best_match(SSR B ) = Interp 3,2

[0100] Interp4, best_match(SSR B ) = Interp 4,1

[0101] Interp5, best_match(SSR B ) = Interp 5,2

[0102] Among them, Interp i,1 Interp i,kFor the technical term pt i Corresponding definitions: Since a term may have multiple definitions, they are numbered k in order. Depending on the context, for the technical term "pt",... i The best interpretation may also differ.

[0103] (42) Calculate SSR j The context alignment score for the scene is shown in the following formula:

[0104]

[0105] For each technical term pt i S align (SSR j ,pt i The calculation formula is as follows:

[0106] S align (SSR j ,pt i )=Similarity(Interpi,best_match.Contextual_Keywords,Q user \PT identified )

[0107] Where Interpi,best_match is an SSR j The best interpretation in this context, Q user \PT identified For non-technical parts of the user input, Contextual_Keywords provides the best-in-class description.

[0108] SSR A The context alignment score in the scene is:

[0109] S align (SSR A ,pt1)=0.5*0.6+0.5*0.9=0.75

[0110] S align (SSR A ,pt2)=0.5*0.8+0.5*0.95=0.875

[0111] S align (SSR A ,pt3)=0.5*0.95+0.5*0.95=0.95

[0112] S align (SSR A ,pt4)=0.5*0.7+0.5*0.8=0.75

[0113] S align (SSR A ,pt5)=0.5*0.9+0.5*0.95=0.925

[0114] Salign_total(SSR A )=Avg(0.75,0.875,0.95,0.75,0.925)=0.85

[0115] SSR B The context alignment score in the scene is:

[0116] S align (SSR B ,pt1)=0.75

[0117] S align (SSR B ,pt2)=0.875

[0118] S align (SSR B ,pt3)=0.5*0.3+0.5*0.4=0.35

[0119] S align (SSR B ,pt4)=0.75

[0120] S align (SSR B ,pt5)=0.5*0.3+0.5*0.4=0.35

[0121] S align_total (SSR B )=Avg(0.75,0.875,0.35,0.75,0.35)=0.615

[0122] (43) Calculate SSR j The DKB preference score in this scenario is shown in the following formula:

[0123]

[0124] S pref (SSR j ,pt i )=w1·Interpi,best_match.GPS+w2·Interpi,best_match.UPS

[0125] Here, w1 and w2 are the weight coefficients of global preference and user preference, respectively.

[0126] For example, when w1 = 0.5 and w2 = 0.5, then SSR A The DKB preference score in this scenario is:

[0127] S pref (SSR A ,pt1)=0.5*0.9+0.5*0.95=0.925

[0128] S pref (SSR A ,pt2)=0.5*0.95+0.5*0.98=0.965

[0129] S pref (SSR A ,pt3)=0.5*0.85+0.5*0.9=0.875

[0130] S pref (SSR A ,pt4)=0.5*0.9+0.5*0.92=0.91

[0131] S pref (SSR A ,pt5)=0.5*0.88+0.5*0.92=0.90

[0132] S pref_total (SSR A )=Avg(0.925,0.965,0.875,0.91,0.90)=0.915;

[0133] SSR B The DKB preference score in this scenario is:

[0134] S pref (SSR B ,pt1)=0.925

[0135] S pref (SSR B ,pt2)=0.965

[0136] S pref (SSR B ,pt3)=0.5*0.7+0.5*0.6=0.65

[0137] S pref (SSR B ,pt4)=0.91

[0138] S pref (SSR B,pt5)=0.5*0.8+0.5*0.75=0.775

[0139] S pref_total (SSR B )=Avg(0.925,0.965,0.65,0.91,0.775)=0.845

[0140] (44) Calculate SSR j The reordering score is shown in the following formula:

[0141] S final (SSR j )=λ1·normalize(ConfLLM(SSR j ))+λ2·normalize(S align_total (SSR j ))+λ3·normalize(S pref_total (SSR j ))

[0142] Where λ1, λ2, and λ3 are hyperparameters used to balance the initial confidence of LLM, context alignment, and DKB preference score, and normalize is the normalization function.

[0143] When λ1=0.3, λ2=0.4, λ3=0.3, SSR A The final reordering score is:

[0144] S final (SSR A = 0.3·0.78 + 0.4·0.85 + 0.3·0.915 = 0.234 + 0.34 + 0.2745 = 0.8485;

[0145] SSR B The final reordering score is:

[0146] S final (SSR B =0.3·0.85+0.4·0.615+0.3·0.845=0.255+0.246+0.2535=0.7545.

[0147] (45) According to S final The score for element SSR j Sort the data, and denote the sorted result as C. SSR_ranked In C SSR_ranked In the middle, select S such that final The highest-scoring SSR j SSR is the optimal semantic representation for this input parsing. optimal .

[0148] Therefore, take SSR A The result of the final semantic parsing will be used as the input for subsequent business processing.

[0149] (5) Perform ambiguity processing on the optimal semantic representation to complete the semantic parsing process and iterative optimization based on user feedback.

[0150] (51) If SSR optimal With S final The score difference between the second-place finishers is less than a specified threshold (e.g., 0.05); or SSR optimal If the score is less than a specified threshold (e.g., 0.8), it is considered that there is ambiguity in the understanding of the user's current input, and the process jumps to step (52); otherwise, based on the parsing result SSR optimal The semantic parsing process is complete once subsequent business operations are performed.

[0151] (52) Ambiguous results or SSRs below a specified threshold will be returned. optimal The results are then given to the user for confirmation. If the user corrects the definition of a term, or selects option C... SSR_ranked Other candidate semantic understanding results (SSR) k Based on the user's selection, update the corresponding UPS record (e.g., increase the weight of the user's definition of the term); if the user provides a new definition for a term, add the new definition to the review pool, and add it to the dynamic knowledge base DKB after periodic review by industry experts.

[0152] (53) Select the SSR by the user j As a new SSR optimal And perform subsequent business operations based on this result.

[0153] This invention also provides an LLM terminology understanding system, comprising:

[0154] The terminology recognition module is used to preprocess user input and identify all technical terms to form a terminology set.

[0155] The input enhancement module is used to construct guided descriptions for each technical term. It combines the guided descriptions of all identified technical terms into a general guided context, which serves as the contextualized guided description for the user input. This contextualized ...

[0156] The semantic parsing module is used to perform multi-dimensional reordering of the LLM semantic parsing results after input augmentation to obtain the optimal semantic representation of the current input parsing;

[0157] The ambiguity handling and iterative optimization module is used to handle ambiguity in the optimal semantic representation, complete the semantic parsing process, and perform iterative optimization based on user feedback.

[0158] To implement the above-mentioned method for understanding LLM terminology, the present invention provides an electronic device and a computer-readable storage medium.

[0159] The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement some or all of the steps in the LLM terminology understanding method described above.

[0160] The computer-readable storage medium has a computer program stored thereon, which, when executed by a processor, implements some or all of the steps in the LLM terminology understanding method described above.

Claims

1. A method for understanding LLM terminology, characterized in that, The method includes: User input is preprocessed and technical terms are identified to form a set of technical terms; LLM input enhancement is achieved by constructing scenario-based guided descriptions for each term in the terminology set. Perform LLM semantic parsing with input enhancement; The LLM semantic parsing results are reordered in multiple dimensions to obtain the optimal semantic representation of the current input parsing. The dimensions include: initial LLM confidence, context alignment, and user preference. Ambiguity processing is performed on the optimal semantic representation to complete the semantic parsing process and iterative optimization based on user feedback.

2. The LLM terminology understanding method according to claim 1, characterized in that, The construction of contextualized guided descriptions for LLM input enhancement for each term in the terminology set includes: Find the current professional terminology pt from the Dynamic Knowledge Base (DKB). i All related definitions; Calculate the relevance score between the usage scenario of each definition and the user's current input, and select the definition with the highest relevance score as the current term. i The best guiding interpretation; According to the best guiding interpretation, the current professional term pt is... i Construct its bootstrap description; All identified technical terms are combined into a general guiding context, which serves as the contextualized guiding description for the user input.

3. The LLM terminology understanding method according to claim 1, characterized in that, The multi-dimensional reordering includes: Obtain the identified technical terms (pt) i In SSR j The best interpretation in the context; among them, SSR j j = 1, ..., K represents the input augmentation LLM for each identified terminology pt i Output K possible semantic understanding results; SSR is performed on the user input based on the best interpretation. j Context alignment within the scene, and calculate the context alignment degree of the current scene; Calculate SSR based on global preferences and user preferences j DKB preference score in the scenario; Based on semantic understanding results SSR j Calculate SSR using LLM initial confidence, context alignment, and DKB preference score. j The scores are reordered, and the semantic understanding result with the highest score is taken as the optimal semantic representation result for parsing the user input.

4. The LLM terminology understanding method according to claim 3, characterized in that, The SSR j The context alignment in a scene is calculated using the following formula: For each technical term pt i S align (SSR j ,pt i The calculation formula is as follows: S align (SSR j ,pt i )=Similarity(Interp i,best_match .Contextual_Keywords,Q user \PT identified ) Among them, Interp i,best_match For SSR j The best interpretation in the context, Contextual_Keywords is the description corresponding to the best interpretation, Q user \PT identified For non-technical parts of user input.

5. The LLM terminology comprehension method according to claim 3, characterized in that, The calculation of SSR j DKB Preference Score in Scenario pref_total (SSR j The following formula is used: S pref (SSR j ,pt i )=w1·Interp i,best_match .GPS+w2·Interp i,best_match .UPS Where w1 and w2 are the weight coefficients of global preference and user preference, respectively; Interp i For the technical term pt i The corresponding definitions are: GPS stands for Global Preference Score; UPS stands for User Preference Score.

6. The LLM terminology comprehension method according to claim 3, characterized in that, The calculation of SSR j The reordering score is calculated using the following formula: S final (SSR j )=λ1·normalize(Conf LLM (SSR j ))+λ2·normalize(S align_total (SSR j ))+λ3·normalize(S pref_total (SSR j )) Where λ1, λ2, and λ3 are hyperparameters used to balance the initial confidence level of the LLM. LLM (SSR j ), Context Alignment S align_total (SSR j ) and DKB preference score S pref_total (SSR j ), where normalize is the normalization function.

7. The LLM terminology understanding method according to claim 1, characterized in that, The ambiguity handling includes: Compare the second-ranked score in the multi-dimensional re-ranking with the optimal semantic representation; If the difference between the two scores is less than a given threshold, or the score of the optimal semantic representation is less than a given threshold, then the user's current input is considered ambiguous. The ambiguous result or the optimal semantic representation below the given threshold must be submitted to the user for confirmation. Based on user feedback, UPS records are updated, or new definitions for technical terms are added to the review database. These are periodically reviewed by industry experts and added to the dynamic knowledge base (DKB) after passing the review. Based on user feedback, an adjusted optimal semantic representation is generated, and subsequent business operations are executed based on this result. If the given threshold is met, subsequent business operations are performed based on the optimal semantic representation, and the semantic parsing process is completed.

8. An LLM terminology comprehension system, characterized in that, The system includes: The terminology recognition module is used to preprocess user input and identify all technical terms to form a terminology set. The input enhancement module is used to construct guided descriptions for each technical term. It combines the guided descriptions of all identified technical terms into a general guided context, which serves as the contextualized guided description for the user input. This contextualized ... The semantic parsing module is used to perform multi-dimensional reordering of the LLM semantic parsing results after input augmentation to obtain the optimal semantic representation of the current input parsing; The ambiguity handling and iterative optimization module is used to handle ambiguity in the optimal semantic representation, complete the semantic parsing process, and perform iterative optimization based on user feedback.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the LLM terminology understanding method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the LLM terminology understanding method as described in any one of claims 1 to 8.

Images