Query optimization method based on large model thinking chain

Abstract

The invention relates to a query optimization method based on a large model thinking chain, which combines a multi-modal large model and an innovative thinking chain technology to plan and gradually optimize complex queries, so that a multi-modal retrieval system can more accurately understand user requirements and meet the retrieval intention of the user, thereby remarkably improving the retrieval effect.

Description

Technical Field

[0001] The present invention relates to a query optimization method based on the thought chain of a large model. Background Art

[0002] Users' queries are usually complex, containing fuzzy or verbose text descriptions, making it difficult for existing retrieval systems to accurately capture the true intentions of users. When directly using the original query for image retrieval, it is often impossible to fully explore the actual needs of users, resulting in insufficient accuracy and relevance of retrieval results.

[0003] Traditional query rewriters are restricted by the model scale and pre-training data, lacking the ability to deeply understand cross-modal semantic knowledge.

[0004] The complexity of the query optimization task is reflected in multiple aspects, such as entity optimization, concept optimization, layout relationship optimization, and entity interaction optimization, etc., all of which require personalized customized processing. When directly rewriting complex queries with large models, it is difficult to accurately grasp cross-modal complex intentions, thus reducing the optimization effect. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: In order to overcome the above technical problems, the present invention provides a query optimization method based on the thought chain of a large model.

[0006] The technical solution adopted by the present invention to solve its technical problems is: A query optimization method based on the thought chain of a large model, including the following steps:

[0007] a. Input the content to be queried into a multi-modal large model assistant. For complex tasks that require multi-step reasoning, use the thought chain technology to decompose the complex reasoning process in the query into several easy-to-process steps, and then analyze potential semantic defects, such as concept deviation, fuzzy spatial relationship, understanding obstacles for challenging entities, etc.;

[0008] b. Based on the semantic defect analysis results, plan a query optimization scheme, and clarify the optimization goals and strategies;

[0009] c. Rewrite the query according to the optimization scheme to generate an optimized query expression;

[0010] d. Use the optimized query to perform multi-modal tasks, such as multi-modal retrieval, to further improve the retrieval effect.

[0011] Further, in step a, the thought chain can be modeled as a chain decomposition of conditional probabilities: Where Q is the input question to be answered, A is the generated final answer, and S1, S2,..., Sn are implicitly or explicitly generated intermediate reasoning steps.

[0012] Furthermore, for large language models based on Transformer, the generation process can be represented as autoregressive sequence prediction: where yt is each step of generation, including intermediate steps and the final answer, and the chain of thought guides the model to focus on logical coherence by explicitly requiring the model to output the reasoning steps S.

[0013] Furthermore, during training or prompting, the model is encouraged to generate logical chains of reasoning. For few-shot learning, the objective can be expressed as: where θ are the model parameters and S are the reasoning steps in the human annotations or examples.

[0014] The beneficial effect of the present invention is that a query optimization method based on the chain of thought of a large model combines a multimodal large model with an innovative chain-of-thought technology to plan and gradually optimize complex queries, enabling the multimodal retrieval system to more accurately understand user needs and meet their retrieval intent, thereby significantly improving the retrieval effect. Detailed implementation

[0015] A query optimization method based on the chain of thought of a large model of the present invention includes the following steps:

[0016] a. Input the content to be queried into a multimodal large model assistant. For complex tasks that require multi-step reasoning, use the chain-of-thought technology to decompose the complex reasoning process in the query into several easily processed steps, and then analyze potential semantic defects, such as concept deviations, fuzzy spatial relationships, understanding obstacles for challenging entities, etc.;

[0017] b. Based on the semantic defect analysis results, plan a query optimization plan and clarify the optimization objectives and strategies;

[0018] c. Rewrite the query according to the optimization plan to generate an optimized query expression;

[0019] d. Use the optimized query to perform multimodal tasks, such as multimodal retrieval, to further improve the retrieval effect.

[0020] Furthermore, in step a, the chain of thought can be modeled as a chain decomposition of conditional probabilities: where Q is the input question to be answered, A is the generated final answer, and S1, S2, …, Sn are implicitly or explicitly generated intermediate reasoning steps.

[0021] Furthermore, for large language models based on Transformer, the generation process can be represented as autoregressive sequence prediction: Where yt is each step of the generation, including intermediate steps and the final answer. The thinking chain guides the model to focus on logical coherence by explicitly requiring the model to output the reasoning step S.

[0022] Furthermore, during training or prompting, the model is encouraged to generate logical reasoning chains. For few-shot learning, the goal can be expressed as: Here, θ is the model parameter and S is the inference step in the manual annotation or example.

[0023] Example 1:

[0024] For a math problem Q, the model's thought chain for generating a solution might be as follows:

[0025] Q: Xiao Ming had 5 apples. He ate 2 and bought 3 more. How many apples does he have now?

[0026] S1: Initial number of apples = 5;

[0027] S2: remaining after eating = 5-2 = 3;

[0028] S3: Total after purchase = 3 + 3 = 6;

[0029] A: 6.

[0030] Example 2:

[0031] For the language question Q, the model's thought chain for generating the answer might be as follows:

[0032] Q: There is a mountain on top of a mountain. Guess a word?

[0033] S1: There is another mountain character above the mountain character;

[0034] S2: Two mountain characters stacked up and down form the character "出";

[0035] A: 6.

[0036] Compared with the existing technology, the query optimization method based on the big model thinking chain of the present invention plans and gradually optimizes complex queries by combining the multimodal big model with the innovative thinking chain technology, so that the multimodal retrieval system can more accurately understand user needs and meet their retrieval intentions, thereby significantly improving the retrieval effect.

[0037] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A query optimization method based on the large model's chain of thought, characterized in that It includes the following steps: a. Input the content to be queried into the multi-modal large model assistant. For complex tasks that require multi-step reasoning, use the chain of thought technique to decompose the complex reasoning process in the query into several easy-to-handle steps, and then analyze potential semantic defects, such as concept deviation, fuzzy spatial relationships, understanding obstacles for challenging entities, etc.; b. Based on the results of semantic defect analysis, plan a query optimization plan, and clarify the optimization goals and strategies; c. Rewrite the query according to the optimization plan to generate an optimized query expression; d. Use the optimized query to execute multi-modal tasks, such as multi-modal retrieval, to further improve the retrieval effect.

2. The query optimization method based on the large model's chain of thought according to claim 1, wherein In step a, the chain of thought can be modeled as a chain decomposition of conditional probabilities: Q)·P(S2∣Q,S1)……P(A∣Q,S1,S2,…,S n )), where Q is the input question to be answered, A is the generated final answer, and S1, S2, …, Sn are implicitly or explicitly generated intermediate reasoning steps.

3. The query optimization method based on the large model thought chain according to claim 2, wherein For a Transformer-based large language model, the generation process can be represented as autoregressive sequence prediction: where yt is each step of the generation, including intermediate steps and the final answer, and the chain of thought guides the model to focus on logical coherence by explicitly asking the model to output the reasoning steps S.

4. The query optimization method based on the large model thought chain according to claim 2, characterized in that During training or prompting, the model is encouraged to generate logical reasoning chains. For few-shot learning, the objective can be expressed as: where θ are the model parameters and S are the reasoning steps in the human annotations or examples.

Images