Retrieval enhancement planning method and device for data quality management

Abstract

The invention discloses a retrieval enhancement planning method and device for data quality management, and the method comprises the steps: S1, carrying out the semantic similarity search of user query data in a preset database, and generating a retrieval quotation according to a query result if there is a similar query; s2, if no similar query exists, extracting a main semantic component of user query data, and checking whether the user query data is consistent with robot action matching dependency or not according to the main semantic component, if the robot action matching dependency is consistent, continuing the next step, and if the robot action matching dependency is not consistent, discarding the data; s3, inputting the user query data, the retrieval quotation, the fixed example and the main semantic component into a preset LLM planner to generate a logic answer; and S4, inputting the logic answer into an actuator for execution. According to the method, the overall performance of retrieving an enhanced large language model (RALMs) is improved by improving retrieval precision, enhancing data quality and improving semantic comprehension, and then the task planning capacity of an overall intelligent system is provided.

Description

Technical Field

[0001] The present invention relates to the technical field of data quality management, and more specifically, to a retrieval enhancement planning method and device for data quality management. Background Art

[0002] With the emergence of large language models (LLMs), researchers have started a new wave of research on integrating LLMs as few-shot planners with embodied AI. These models can convert natural language into formal representations, such as planning domain definition language (PDDL), first-order logic expressions, XML, etc., showing great potential. At the same time, retrieval-augmented generation (RAG) is a quite promising advanced method that uses vector databases to efficiently store and retrieve vectorized knowledge representations, enabling the knowledge base of the LLM to be dynamically adjusted without changing the underlying language model. Therefore, if RAG is made compatible with the LLM planner, we can timely adjust the memory stored by the robot.

[0003] However, RAG encounters some limitations of its own when applied to embodied intelligence. Existing embedding generation defects may lead to misjudgment of sentences with similar contexts but opposite semantics. For example, the two sentences "Can you turn on the air conditioner?" and "Can you turn off the air conditioner?" Although the high-dimensional vectors represented in the vector space are very close, the action meanings they convey are completely opposite. These limitations result in poor data quality in the robot's knowledge base. When these semantically inconsistent sentences are input into the language model as background knowledge, it may disrupt the answer generation of the model, which is particularly harmful to the embodied intelligence system and may cause the robot's actions to violate the user's intended commands. This deviation may have a serious impact on the performance and reliability of the system.

[0004] Regarding the above problems, the relevant research work is as follows:

[0005] LLMs Planner: Researchers have used LLMs as planners in robot task planning, including fine-tuning small-scale LLMs through a large number of annotations (Lykov and Tsetserukou, 2023; Lykov et al., 2023) and prompt engineering through few-shot examples to stimulate the best performance of the LLM. Specifically, for prompt engineering based on few-shot examples, previous studies either directly presented a fixed and typical set of examples (Ahn et al., 2022; Huang et al., 2022; Liu et al., 2023a; Chen et al., 2024), or retrieved examples closely related to the query in the repository (Song et al., 2022). However, these methods not only have high costs in knowledge storage and retrieval due to simple vector search, but also the examples become outdated because the knowledge base does not provide real-time updates.

[0006] In terms of data quality management: Functional dependencies (FDs) [Codd, 1971] were initially introduced in the 1970s to represent data integrity constraints and relationships. Based on FDs, conditional functional dependencies (CFDs) [Fan et al., 2008] have been proposed for data cleaning purposes. They use conditions to specify subsets of tuples for which the dependencies hold. Subsequently, matching dependencies (MDs) [Fan et al., 2011] were proposed to identify records representing the same real-world entity. Approximate functional dependencies (AFDs) [Karegar et al., 2021] were proposed to tolerate partially violating tuples to better handle noisy datasets. In addition, association rules (ARs), initially used to capture item relationships in transaction data, have also been widely studied for data repair and association analysis of relational data. Meanwhile, the mining of data dependencies can refer to the following research [Song and Chen, 2009; Schirmer et al., 2020; Fan et al., 2010; Santhya et al., 2014]. Similar rules have been applied to graphs [Galárraga et al., 2013; Cao et al., 2023; Fan et al., 2022] to analyze social networks [Erlandsson et al., 2016; Cagliero and Fiori, 2013] by extracting relationships. Graph association rules (GARs) [Fan et al., 2015, 2016, 2020] have directly defined association rules on graphs for graph data analysis [Fang et al., 2016; Song et al., 2016] and knowledge graph search [Namaki et al., 2017]. However, all these data quality rules are designed for relations or graphs, and it is difficult for them to support vector data quality management for RALMs.

[0007] Therefore, it is indeed necessary to develop a retrieval-enhanced planning method and device for data quality management to optimize the data quality of the robot knowledge base by managing vector data through the QMRA framework, which is an important complement and extension to existing research. This method has not been fully explored and applied in the prior art. Through the QMRA framework, it is possible to directly optimize data quality problems, providing a brand-new solution to improve the overall performance of the embodied intelligent planning system. Summary of the Invention

[0008] The purpose of the present invention is to provide a retrieval-enhanced planning method and device for data quality management to overcome the defects existing in the prior art.

[0009] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] A retrieval enhancement planning method for data quality management, comprising the following steps:

[0011] S1. Perform semantic similarity search on the user query data in a preset database. If there are similar queries, generate retrieval citations according to the query results;

[0012] S2. If there are no similar queries, extract the main semantic components of the user query data, and check whether the user query data is consistent with the robot action matching dependency according to the main semantic components. If the robot action matching dependency is consistent, continue to the next step; if not, discard the data;

[0013] S3. Input the user query data, retrieval citations, fixed examples, and main semantic components into a preset LLM planner to generate a logical answer;

[0014] S4. Input the logical answer into an executor for execution.

[0015] Further, the step S2 further includes: pruning the vector index of the user query data with inconsistent robot action matching dependencies, so that the database can update its vector search index.

[0016] Further, the step of extracting the main semantic components of the user query data in the step S2 includes: extracting the basic semantic components of the user query data, and the basic semantic components include predicates and objects.

[0017] Further, the step S3 further includes performing conflict checking on the actions of the logical answer, and the formula for the conflict checking is:

[0018]

[0019] In the formula, indicates that the vectors corresponding to two user query data s1 and s2 are contradictory, (s1[o i = i s2[o i ) indicates that for each i, the object o i in sentence s1 is the same as that in sentence s2; then indicates that the action p i in sentence s1 is in conflict with that in sentence s2.

[0020] Further, the pruning vector index specifically includes: processing each user query data, retrieving a set of sentences similar to the user query data, and comparing the conflict between the user query data and the retrieved set of sentences. When a retrieved set of sentences conflicts with a set number of user query data, the similarity links of the set number of user query data in the vector database are cut off.

[0021] Further, after step S5, there is also a step of evaluating the execution result, and the evaluation is environmental perception or human evaluation.

[0022] The present invention also provides a device for implementing the above-mentioned retrieval enhancement planning method for data quality management, including:

[0023] A semantic similarity search module, configured to perform semantic similarity search on user query data in a preset database, and generate retrieval citations according to the query results if there are similar queries;

[0024] An extraction module, configured to extract the main semantic components of the user query data if there are no similar queries, and check whether the user query data is consistent with the robot action matching dependency according to the main semantic components. If the robot action matching dependency is consistent, proceed to the next step; if not, discard the data;

[0025] A generation module, configured to input the user query data, retrieval citations, fixed examples, and main semantic components into a preset LLM planner to generate a logical answer;

[0026] An input module, configured to input the logical answer into an executor for execution.

[0027] Compared with the prior art, the advantages of the present invention are as follows: By improving retrieval accuracy, enhancing data quality, and improving semantic understanding, the present invention improves the overall performance of retrieval-enhanced large language models (RALMs), thereby providing the task planning ability of the overall embodied intelligent system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0029] Figure 1 is a flowchart of the retrieval enhancement planning method for data quality management of the present invention.

[0030] Figure 2It is the flowchart of conflict checking in the present invention.

[0031] Figure 3 It is the schematic diagram of the witness theorem in the present invention.

[0032] Figure 4 It is the schematic diagram of the HNSW index pruning example in the present invention.

[0033] Figure 5 It is the specific schematic diagram of the present invention.

[0034] Figure 6 It is the framework diagram of the retrieval enhancement planning device for data quality management in the present invention. Detailed implementation manners

[0035] The preferred embodiments of the present invention will be elaborated in detail below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making the protection scope of the present invention more clearly defined.

[0036] The concept of the present invention is centered around the Quality-Managed Retrieval-Augmented LLM Planner (QMRA-LLM-planner) framework for data quality management. By introducing Robot Action Matching Dependencies (RAMDs), using large models to identify and prune information inconsistent with the query context in the retrieval results, improving the quality of the vector database to enhance the accuracy of Retrieval-Augmented Large Language Models (RALMs), and thus enhancing the planning ability of the overall embodied intelligent system.

[0037] In this embodiment, the retrieval-augmented large language model combines information retrieval and large language models, generates more accurate and rich answers by retrieving a large amount of relevant information, and thus improves the overall performance and application scope of the model. Generally speaking, the retrieval enhancement technology represents data in vector form, constructs an index and uses an approximate query method to retrieve answers related to the input question, and then inputs the answers into the large language model to supplement background knowledge for the large language model, thereby improving the overall performance of the model.

[0038] Refer to Figure 1 , this embodiment discloses a retrieval enhancement planning method for data quality management, including the following steps:

[0039] Step S1: Perform semantic similarity search on the user query data in a preset database. If there are similar queries, generate retrieval citations according to the query results.

[0040] Step S2: If there is no similar query, the QMRA framework will use a large language model to extract the main semantic components of the user query data, and check whether the user query data is consistent with the robot action matching dependencies according to the main semantic components. If the robot action matching dependencies are consistent, proceed to the next step; if not, discard the data.

[0041] Step S3: Input the user query data, retrieved citations, fixed examples, and main semantic components into a preset LLM planner to generate a logical answer.

[0042] Step S4: Input the logical answer into an executor for execution, and evaluate the execution result. The evaluation is either environmental perception or human evaluation.

[0043] In this embodiment, the step S3 also includes a conflict check for action contradictions in the logical answer.

[0044] For action contradictions, two natural language instructions s1 and s2 and their corresponding embedding vectors V s1 and V s2 The conflict relationship between them is defined. If s1 and s2 are similar in context but opposite in semantics, there is a conflict between them, denoted as This means that although the two sentences may be similar in text, the actions indicated by the two are impossible to be executed simultaneously in the actual scenario.

[0045] For conflict checking, the inconsistency between robot actions is defined through logical rules to ensure the consistency and accuracy of information in the process of converting from natural language instructions to robot actions. The formula for conflict checking is:

[0046]

[0047] In the formula, indicates that the vectors corresponding to the two user query data s1 and s2 are contradictory, (s1[o i = i s2[o i ) means that for each i, the object o i in sentence s1 is the same as that in sentence s2; then indicates that the action p i in sentence s1 is in conflict with that in sentence s2 and cannot coexist.

[0048] It can be described in natural language as: If two instructions require opposite actions for the same object, there is a conflict. This detection can be achieved by comparing the action tuples (object and action pair) in the sentences.

[0049] In this embodiment, for the detected conflicts, the process of eliminating conflicts is achieved by pruning the edges corresponding to the connection vectors in the index graph. The aim is to identify natural language sentences that are syntactically and contextually similar but semantically opposite. Due to their high similarity, their embedding vectors may be close in the vector space, resulting in being wrongly linked together when constructing the vector graph index. Once a conflict is confirmed, the corresponding edge is cut to correct the wrong index.

[0050] In this embodiment, the pruning of the vector index for user query data with inconsistent robot action matching dependencies enables the database to update its vector search index.

[0051] As Figure 2 shown, in step S2, the extraction of the main semantic components mainly involves the LLM decomposing the sentence into main components and serving as a comparator to check the consistency between the retrieved context and the query.

[0052] This embodiment designs a prompt containing rules and instructions. In the rule part of the prompt, the meaning of RAMDs is explained to the LLM in natural language; in the instruction part, the LLM is required to extract and compare sentence components according to RAMDs and determine whether the data is consistent.

[0053] The LLM will first extract the basic semantic components of sentences s1 and s2, including the predicate (pre) and the object (obj). After comparison, the LLM will find that "turn on" and "turn off" are semantically conflicting, which is represented as Meanwhile, s1[o] = i s2[o]. Therefore, the LLM will return that there is a conflict.

[0054] Regarding the pruning of the vector index, this embodiment first proposes an index pruning algorithm based on the witness theorem, which can identify contexts in the vector database that are wrongly regarded as similar.

[0055] Witness Theorem:

[0056] The index pruning algorithm specifically includes: If the conflict judgment between query q and sentence s1 is different from the conflict judgment between q and s2, then q witnesses the contradiction between s1 and s2. When a sufficient number of witnesses are collected, it can be determined that s1 and s2 are actually also in conflict. In this case, the vector index can be modified by cutting the similarity link between s1 and s2. The pseudocode can be expressed as follows:

[0057]

[0058] The index pruning algorithm first processes each query, retrieves the set of sentences similar to it, and compares the conflict between these sentences and the query. When a query serves as a "witness" for a sufficient number of sentence pairs, proving that they are actually inconsistent, the algorithm cuts off the similarity links of these sentences in the vector database. This process is repeated until all queries are processed, thereby updating and optimizing the index structure of the vector database to ensure the accuracy of retrieval results and data quality.

[0059] Taking the example of the HNSW index, Figure 4 this process can be better understood. A and B are considered similar and connected together. However, when the LLM determines that q1 and q2 are similar to A but not to B, while q3 is similar to B but not to A, there is already sufficient evidence to show that A and B should be separated. Therefore, QMRA will cut off the edge between A and B, and the vector search index is pruned.

[0060] Example: In a coffee shop environment, as Figure 5 shown, where a humanoid robot serves as a waiter. This environment is digitally constructed using the MO-vln (Multi-Task benchmark for Open-Set Zero-Shot Vision-and-Language Navigation) simulator. The robot waiter needs to perform visual segmentation and detection to obtain information about objects from different angles.

[0061] Customer instruction: In this scenario, the customer says to the robot waiter, "Turn on the hall light." This instruction is passed as input q to the system.

[0062] Normal RAG LLM planner (Regular Retrieval-Augmented LLM Planner)

[0063] Retrieve relevant context: The system retrieves the context related to query q, such as:

[0064] "Turn off the hall light."

[0065] "Can you turn on the hall light, please?"

[0066] Error message leads to incorrect target state: Due to the existence of the error message, the LLM generates an incorrect target state Closed_HallLight, resulting in the action generator producing an incorrect action sequence. Eventually, the instruction does not receive the correct response.

[0067] QMRA-LLM-planner (Retrieval-Augmented LLM Planner for Data Quality Management)

[0068] Pruned Index Structure: By pruning the index structure, the system retrieves contexts that match the query q, such as:

[0069] "Can you turn on the hall light, please?"

[0070] Correct Goal State and Action Sequence: The LLM generates the correct goal state Active_HallLight, and the action generator constructs a behavior tree to guide the robot to turn on the light. Finally, the robot correctly turns on the light and responds to the instruction.

[0071] Table 1 below shows the comparison of the experimental results of the QMRA framework under different language models and different index structures.

[0072] Table 1

[0073]

[0074] See Figure 6 As shown, the present invention also provides an apparatus for implementing the above-mentioned retrieval-augmented planning method for data quality management, including: a semantic similarity search module 1 for performing semantic similarity search on user query data in a preset database, and generating retrieval citations according to the query results if there are similar queries; an extraction module 2 for extracting the main semantic components of the user query data if there are no similar queries, and verifying whether the user query data is consistent with the robot action matching dependency according to the main semantic components. If the robot action matching dependency is consistent, proceed to the next step; if not, discard the data; a generation module 3 for inputting the user query data, retrieval citations, fixed examples, and main semantic components into a preset LLM planner to generate a logical answer; and an input module 4 for inputting the logical answer into an executor for execution.

[0075] By directly addressing the data quality issues in the prior art, the QMRA framework provides a novel solution. By improving retrieval accuracy, enhancing data quality, and improving semantic understanding, the overall performance of retrieval-augmented large language models (RALMs) is improved, thereby providing the task planning ability of the overall embodied intelligent system. Specifically, it has the following advantages:

[0076] 1. Enhance semantic understanding. By introducing Robot Action Matching Dependencies (RAMDs), the LLM can handle relatively complex semantic relationships, including identifying and processing semantically opposite expressions. This ability, lacking in the prior art, makes the QMRA framework more accurate in processing natural language, enabling it to better understand and respond to user query requirements. By extracting and comparing the main semantic components of sentences, the QMRA framework can deeply understand the semantic content of sentences and match it with the query context to determine its relevance and conflicts.

[0077] 2. Improve data quality. The prior art often neglects data quality issues in the knowledge base, resulting in retrieval results that may contain information irrelevant or contradictory to the query, thus reducing the quality of generated answers. The QMRA framework actively improves the data quality of the knowledge base through pruning algorithms, rather than merely relying on existing data passively. This proactive data quality management strategy makes the QMRA framework more reliable in processing retrieval results, reducing the spread of incorrect information and thus improving the overall quality of the data.

[0078] 3. Increase memory capacity. Due to the deficiencies in combining retrieval enhancement with large language models in the prior art, it is difficult to integrate this technology into embodied intelligent systems. The present invention manages the data quality in the vector database through the Robot Action Matching Dependencies (RAMDs) rule, effectively combining the knowledge base with the large language model, enabling the system to quickly and accurately update the robot's memory through the update of the knowledge base, further enhancing the memory capacity and task planning ability of the embodied intelligent system.

[0079] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, the patent owner may make various modifications or alterations within the scope of the appended claims. As long as it does not exceed the protection scope described by the claims of the present invention, it shall fall within the protection scope of the present invention.

Claims

1. A retrieval enhancement planning method for data quality management, characterized in that: The following steps are involved: S1. Search the user query data for semantic similarity in a preset database. If there are similar queries, generate search citations based on the query results. S2. If there is no similar query, extract the main semantic components of the user query data, and check whether the user query data is consistent with the robot action matching dependency based on the main semantic components. If the robot action matching dependency is consistent, proceed to the next step; if not, discard the data; S3, inputting user query data, retrieved citations, fixed examples and main semantic components into a preset LLM planner to generate a generative logical answer; S4. Input the logic answer into the executor for execution.

2. The retrieval enhancement planning method for data quality management according to claim 1, characterized in that: The step S2 also includes: pruning vector indexes of user query data with inconsistent robot action matching dependencies, so that the database can update its vector search index.

3. The retrieval enhancement planning method for data quality management according to claim 1, characterized in that: The step of extracting the main semantic components of the user query data in step S2 includes: extracting basic semantic components of the user query data, and the basic semantic components include a predicate and an object.

4. The retrieval enhancement planning method for data quality management according to claim 1, characterized in that: The step S3 also includes performing a conflict check on the actions of the logical answer, and the formula for the conflict check is: In the formula, Indicates that the vectors corresponding to the two user query data s1 and s2 are contradictory, (s1[o i ]= i s2[o i ]) means that for each i, the force object o in sentence s1 i It is the same as in sentence s2; Then it means the action p in sentence s1 i This conflicts with what is in sentence s2.

5. The retrieval enhancement planning method for data quality management according to claim 2, characterized in that: The pruning vector index specifically includes: processing each user query data, retrieving a sentence set similar to the user query data, and comparing the conflict between the user query data and the queried sentence set. When a queried sentence set conflicts with a set number of user query data, the similarity links of the set number of user query data in the vector database are cut off.

6. The retrieval enhancement planning method for data quality management according to claim 1, characterized in that: The step S5 also includes a step of evaluating the execution result, which is environmental perception or human evaluation.

7. A device for implementing the retrieval enhancement planning method for data quality management according to any one of claims 1 to 6, characterized in that: include: The semantic similarity search module is used to search for semantic similarity of user query data in a preset database, and generate search citations based on the query results if there are similar queries; An extraction module is used to extract the main semantic components of the user query data when there is no similar query, and to check whether the user query data is consistent with the robot action matching dependency according to the main semantic components, and to proceed to the next step if the robot action matching dependency is consistent, and to discard the data if it is inconsistent; A generation module, for inputting user query data, retrieved citations, fixed examples and main semantic components into a preset LLM planner to generate a generation logic answer; The input module is used to input the logic answer into the executor for execution.

Images