A method for generating controllable answers to aerospace information based on large language models

Abstract

The present invention belongs to the technical field of big data and artificial intelligence, and discloses a method for generating controllable answers to information in the aerospace field based on a large language model, so that the large language model can generate content-controllable answers based on real-time data in the background of an aerospace business system. First, the slot big data model is trained by a sample set after slot conversion. Then, questions are asked to the slot big model to obtain answers containing slot characters. Then, various types of constrained words in the sentence where the slot characters are located are obtained through word segmentation and part-of-speech classification, and content queries are performed in the aerospace business database based on the constrained words. Finally, the slot characters are replaced with the specific content after the query to obtain the final controllable answer. The present invention solves the problem that the answer content generated by the existing big data model is arbitrarily fabricated and inconsistent with the real-time data by training the large language model with slot samples and performing a database search with joint context constraints on the slot characters in the generated answer.

Description

Technical Field

[0001] The present invention belongs to the technical field of big data and artificial intelligence, and relates to a method for generating controllable answers to information in the aerospace field based on a large language model. More specifically, it relates to a method for automatically generating answers to professional knowledge questions in the aerospace field through a large language model, and supports automatic retrieval, replacement and update of some time-varying information in the answers generated by the large language model through manual intervention. Background Art

[0002] With the gradual emergence of intelligent generative question-answering systems such as ChatGPT and Wenxinyiyan, large language models have been widely used in the field of mass entertainment. A large number of people can ask questions to large language models and get more intelligent answers than some previous question-answering systems. However, large language models still have defects such as content fabrication and knowledge forgetting. These defects are acceptable in some mass scenarios such as chat entertainment and document editing, but they are unacceptable for vertical professional fields such as aerospace and military that have high requirements for the rigor of generating answers.

[0003] At the same time, data in professional fields such as aerospace are subject to real-time changes, such as the real-time location of satellites, whether they can communicate, and the areas that can be photographed. However, large language models need to be trained offline. When the large language model is trained and put into use, the data and knowledge used in model training are already outdated data. Answers generated based on outdated data are completely uncontrollable and unacceptable in professional fields, and operations based on such answers may even cause accidents.

[0004] Therefore, it is urgent to study a method for controlling the generation of answers using a large language model suitable for the aerospace professional field, so that the large language model can generate answers with controllable content based on the real-time data of the aerospace business system background, avoiding the problem of arbitrary fabrication of answer content and inconsistency with real-time data. Summary of the invention

[0005] The purpose of the present invention is to overcome the problems that the large language model trained offline is not compatible enough with real-time data and the generated answer content is uncontrollable. A method for generating controllable answers for aerospace information based on a large language model is proposed. The large language model is trained with a slot sample set to have the ability to generate slot answers. Subsequently, the database approximate fields and field retrieval constraints are matched according to the slot context. According to the field retrieval constraints, a database query statement is generated and the slot content is queried and filled to obtain the final controllable answer generation result.

[0006] The technical solution adopted by the present invention is:

[0007] A method for generating controllable answers to aerospace information based on a large language model comprises the following steps:

[0008] (1) Slot-processing professional documents in the aerospace field, replacing time-varying parameters that change with the passage of time with fixed slot characters, and obtaining professional documents after slot conversion; the time-varying parameters include time, quantity, location, and task;

[0009] (2) The professional documents after slot conversion are used as a training sample set to train the large language model to obtain a slot large language model, and questions are asked to the slot large language model to obtain answers containing slot characters, where the content at the slot character position is content related to the time-varying parameter;

[0010] (3) The slot characters in the answer containing the slot characters are converted into specific content in the background database of the aerospace business system, thereby obtaining the final controllable answer.

[0011] Furthermore, step (3) specifically includes the following steps:

[0012] (301) Sequentially divide the paragraph containing each slot character into sentences according to the period, and segment the sentences containing the corresponding slot character and remove stop words to obtain a context vocabulary set of the corresponding slot;

[0013] (302) Classifying the context vocabulary set into parts of speech to obtain time constraint words, location constraint words, object constraint words, and content attribute words; wherein the time constraint words are used to constrain the time range of the content to be filled in the slot, the location constraint words are used to constrain the location range of the content to be filled in the slot, the object constraint words are used to constrain the object range corresponding to the content to be filled in the slot, and the content attribute words are used to constrain the content type corresponding to the content to be filled in the slot;

[0014] (303) Calculate the approximation between the content attribute words and all the fields in the background database of the aerospace business system to obtain the most relevant fields, and use the time constraint words, location constraint words and object constraint words as constraint conditions to query the most relevant fields, and replace the current slot characters with the query results.

[0015] Compared with the background technology, the present invention has the following advantages:

[0016] The present invention proposes a new method for generating controllable answers to aerospace information based on a large language model. By training the large language model with slot samples and performing database retrieval with joint context constraints on the slot characters in the generated answers, the problem that the answers generated by the existing large data models are arbitrarily fabricated and inconsistent with real-time data is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall process framework design diagram of the present invention.

[0018] Figure 2 It is a schematic diagram of the answer generated by the slot large language model of the present invention and the final controllable answer after the slot is replaced. DETAILED DESCRIPTION

[0019] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0020] Figure 1 It is a principle flow chart of a specific implementation of the method for generating controllable answers to aerospace information based on a large language model of the present invention.

[0021] In this embodiment, if Figure 1 The method for generating controllable answers to aerospace information based on a large language model includes the following steps:

[0022] (1) First, the professional documents in the aerospace field are slotted. The time-varying parameters such as time, quantity, location, and task that will change over time are replaced with fixed slot characters to obtain professional documents after slot conversion.

[0023] (2) The professional documents after slot conversion are then used as a training sample set to train a large language model (such as ChartGLM, GTP4ALL, MOSS) to obtain a slot large language model, and questions are asked to the slot large language model to obtain answers containing slot characters. The content at the slot character position is content related to the time-varying parameters.

[0024] (3) Then, the slot characters in the answer containing the slot characters are converted into specific contents in the background database of the aerospace business system to obtain the final controllable answer. The comparison diagram of the answer containing the slot characters output by the slot big data model and the final controllable answer after querying and replacing the answer through the background database of the aerospace business system is shown in the figure below. Figure 2 shown.

[0025] Wherein, step (3) comprises the following steps:

[0026] (301) First, the paragraph where the first slot character is located is divided into sentences according to periods, and the sentence containing the slot character is segmented and stop words are removed to obtain the context vocabulary set of the slot.

[0027] (302) Then, the context vocabulary set is classified into parts of speech to obtain time constraint words, location constraint words, object constraint words and content attribute words; the time constraint words are used to constrain the time range of the content that should be filled in the slot, such as this year, yesterday, May 30, etc.; the location constraint words are used to constrain the location range of the content that should be filled in the slot, such as Shijiazhuang, Beijing, etc.; the object constraint words are used to constrain the object range corresponding to the content that should be filled in the slot, such as high-resolution satellite, Kashgar ground station, etc.; the content attribute words are used to constrain the content type corresponding to the content that should be filled in the slot, such as: running time, running posture, remaining energy, etc.

[0028] (303) Then, the Bert model is used to calculate the approximation between the content attribute words and all the fields in the background database of the aerospace business system to obtain the most relevant field name; and the time constraint words, location constraint words and object constraint words are used as constraint conditions to generate SQL query statements to query the field in the database, and the query results are replaced with the position of the current slot character.

[0029] (304) All slot characters are processed in sequence according to steps (301)-(303), and all slot characters in the generated answer are converted into specific content in the background database of the aerospace business system.

[0030] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for generating controllable answers to aerospace information based on a large language model, characterized in that: The following steps are involved: (1) The professional documents in the aerospace field are slotted, and the time-varying parameters that change with the passage of time are replaced with fixed slot characters to obtain the professional documents after slot conversion; the time-varying parameters include time, quantity, location and task; (2) The professional documents after slot conversion are used as a training sample set to train the large language model to obtain a slot large language model, and questions are asked to the slot large language model to obtain answers containing slot characters. The content at the slot character position is the content related to the time-varying parameter; (3) Convert the slot characters in the answer containing the slot characters into specific content in the background database of the aerospace business system, thus obtaining the final controllable answer; Wherein, step (3) specifically includes the following steps: (301) Sequentially divide each paragraph containing the slot character into sentences according to the period, and segment the sentences containing the corresponding slot character and remove stop words to obtain a context vocabulary set of the corresponding slot; (302) Classifying the context vocabulary set into parts of speech to obtain time constraint words, location constraint words, object constraint words, and content attribute words; wherein the time constraint words are used to constrain the time range of the content to be filled in the slot, the location constraint words are used to constrain the location range of the content to be filled in the slot, the object constraint words are used to constrain the object range corresponding to the content to be filled in the slot, and the content attribute words are used to constrain the content type corresponding to the content to be filled in the slot; (303) Calculate the approximation between the content attribute words and all the fields in the background database of the aerospace business system to obtain the most relevant fields, and use the time constraint words, location constraint words and object constraint words as constraint conditions to query the most relevant fields and replace the current slot characters with the query results.

Images