Intention recognition method and device in environmental part and related equipment under observable conditions

Abstract

The application discloses an intention recognition method and device under an environment partially observable condition and related equipment, and relates to the technical field of intention recognition. The method specifically comprises the following steps: obtaining an original observation sequence, wherein the original observation sequence comprises a plurality of observation values arranged according to timestamps; in the case that there is data missing in the original observation sequence, performing prediction based on the original observation sequence to obtain a predicted value; complementing the original observation sequence based on the predicted value to obtain a target observation sequence; taking the target observation sequence as input data and taking a plan library as prior knowledge, inputting the target observation sequence and the plan library into an identifier to perform intention recognition, and obtaining a first plan sequence; in the case that the first plan sequence does not satisfy a preset condition, inputting the target observation sequence and the plan library into a pre-trained large language model to perform intention recognition, and obtaining a second plan sequence. The method can improve the accuracy of the intention recognition result.

Description

Technical Field

[0001] This invention belongs to the field of intent recognition technology, specifically relating to an intent recognition method, apparatus, and related equipment under partially observable environmental conditions. Background Technology

[0002] The core of cognitive intent recognition lies in inferring the plans and purposes of the identified object based on acquired observation data. However, in real-world environments, observations are often missing or difficult to obtain for various reasons. Most existing methods assume that the observation data is fully observable, or handle missing observations by adding "missing" branches to decision trees. These methods are time- and space-intensive.

[0003] The advantage of the SBR algorithm is that it can provide a complete plan sequence in explicit recognition scenarios. By labeling observations, propagating labels, and backtracking labels, it can identify paths when needed. However, in unobservable dynamic environments, the SBR algorithm can only pass labels by assuming labels. If the observation sequence is too long, it will lead to label explosion, thus consuming a lot of space. Therefore, it can be seen that the accuracy of the current technology in the case of missing observations is poor. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the accuracy of the above-mentioned methods commonly used in the prior art is poor when the observation values ​​are missing. In order to solve the above problem, the present invention provides an intention recognition method, device and related equipment under partially observable environmental conditions.

[0005] The content of this invention includes:

[0006] In a first aspect, embodiments of the present invention provide an intent recognition method under partially observable environmental conditions, comprising:

[0007] Obtain the original observation sequence, which includes multiple observations arranged by timestamps, and the observations are used to characterize the actions performed by the observed object;

[0008] In the case of missing data in the original observation sequence, a prediction is made based on the original observation sequence to obtain a predicted value, which is used to characterize the action performed by the predicted observed object.

[0009] The original observation sequence is completed based on the predicted values ​​to obtain the target observation sequence;

[0010] The target observation sequence is used as input data, and the plan library is used as prior knowledge. The input is used to identify the intent by the recognizer to obtain the first plan sequence. The recognizer is used to identify the intent based on the SBR algorithm. The plan library includes multiple pre-stored plan sequences.

[0011] If the first plan sequence does not meet the preset conditions, the target observation sequence and the plan library are input into a pre-trained large language model for intent recognition to obtain a second plan sequence. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected by child nodes. The root node is used to represent the purpose, and the leaf node is used to represent the action.

[0012] Optionally, the step of predicting based on the original observation sequence to obtain the predicted value includes:

[0013] Based on the observed values, a vector retrieval operation is performed in a pre-built knowledge base to obtain candidate predicted values. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequences.

[0014] Based on the candidate predicted values ​​and the observed values, a vector retrieval operation is performed in the knowledge base, and the predicted value is determined from the candidate predicted values ​​based on the retrieval results.

[0015] Optionally, the step of performing a vector retrieval operation based on the observed values ​​in a pre-built knowledge base to obtain candidate predicted values ​​includes:

[0016] Based on each observation, a first matching sequence is retrieved from the knowledge base, wherein the first matching sequence is a pre-stored plan sequence with the observation as the leaf node;

[0017] The root node of the first matching sequence is determined as the prefix anchor point;

[0018] Based on each of the prefix anchor points, a second matching sequence is retrieved from the knowledge base. The second matching sequence is a pre-stored plan sequence with the prefix anchor point as the root node.

[0019] The leaf nodes of the second matching sequence are determined as the candidate predicted values.

[0020] Optionally, the step of performing a vector retrieval operation in the knowledge base based on the candidate predicted values ​​and the observed values, and determining the predicted value from the candidate predicted values ​​based on the retrieval results, includes:

[0021] Based on each of the candidate predicted values, a third matching sequence is retrieved from the knowledge base, the third matching sequence being a pre-stored plan sequence with the leaf nodes being the candidate predicted values; and based on each of the observed values, a fourth matching sequence is retrieved from the knowledge base, the fourth matching sequence being a pre-stored plan sequence with the leaf nodes being the observed values.

[0022] The third matching sequence, the fourth matching sequence, and the original observation sequence are input into a pre-trained large language model for prediction processing to obtain the predicted value.

[0023] Optionally, the step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes:

[0024] A first prompt message is constructed based on the target observation sequence and the plan library. The first prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the plan library.

[0025] The first prompt message is input into a large language model for processing to obtain the second plan sequence.

[0026] Optionally, the step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes:

[0027] Based on each target observation, a fifth matching sequence is retrieved from the knowledge base. The target observation includes the observed value and the predicted value. The fifth matching sequence is a pre-stored plan sequence with the target observation as the leaf node. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequence.

[0028] A second prompt message is constructed based on the fifth matching sequence, the target observation sequence, and the plan library. The second prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the fifth matching sequence and the plan library.

[0029] The second prompt message is input into the large language model for processing to obtain the second plan sequence.

[0030] Optionally, the step of predicting based on the original observation sequence to obtain the predicted value includes:

[0031] The original observation sequence is used as input data, and the plan library is used as prior knowledge. The data is then input into a large language model for prediction to obtain the predicted value.

[0032] Secondly, embodiments of the present invention provide an intent recognition device under partially observable environmental conditions, comprising:

[0033] The acquisition module is used to acquire the original observation sequence, which includes multiple observation values ​​arranged according to timestamps, and the observation values ​​are used to characterize the actions performed by the observed object;

[0034] The prediction module is used to make predictions based on the original observation sequence when there are missing data in the original observation sequence, and to obtain predicted values. The predicted values ​​are used to characterize the actions performed by the observed objects as predicted.

[0035] The completion module is used to complete the original observation sequence based on the predicted value to obtain the target observation sequence;

[0036] A recognizer is used to perform intent recognition by taking the target observation sequence as input data and the plan library as prior knowledge to obtain a first plan sequence. The recognizer is used to perform intent recognition based on the SBR algorithm. The plan library includes multiple pre-stored plan sequences.

[0037] A pre-trained large language model is used to perform intent recognition based on the target observation sequence and the plan library when the first plan sequence does not meet the preset conditions, and obtain a second plan sequence. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected by child nodes. The root node is used to represent the purpose and the leaf node is used to represent the action.

[0038] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement the steps in the intention recognition method under partially observable environmental conditions as described in the first aspect.

[0039] Fourthly, embodiments of the present invention provide a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the intent recognition method under partially observable environmental conditions as described in the first aspect.

[0040] The beneficial effects of this invention are as follows: In the embodiments of this invention, an original observation sequence is obtained, which includes multiple observations arranged according to timestamps; when data is missing in the original observation sequence, prediction is performed based on the original observation sequence to obtain predicted values; the original observation sequence is then completed based on the predicted values ​​to obtain a target observation sequence. The target observation sequence is used as input data, and a plan library is used as prior knowledge, which is input into a recognizer for intent recognition to obtain a first plan sequence; if the first plan sequence does not meet preset conditions, the target observation sequence and the plan library are input into a pre-trained large language model for intent recognition to obtain a second plan sequence. On the one hand, predicting missing data to obtain a complete target observation sequence when data is missing in the original observation sequence ensures the integrity of the input observation sequence and improves the effectiveness of the input data. On the other hand, using a large language model for intent recognition when the recognizer's output does not meet preset requirements ensures that an output result can be obtained regardless of the situation, improving the effectiveness and accuracy of intent recognition. Attached Figure Description

[0041] Figure 1 A flowchart of an intent recognition method under partially observable environmental conditions provided in an embodiment of the present invention;

[0042] Figure 2 This is one of the example diagrams for the planning library;

[0043] Figure 3 A schematic diagram of the symbolic plan prediction and recognition model based on LLM enhancement provided in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the process for obtaining predicted values ​​based on the original observation sequence, provided in an embodiment of the present invention.

[0045] Figure 5 Example diagram of the path sequence containing the retrieved existing observations;

[0046] Figure 6 An example diagram showing the path sequence of the same destination obtained through retrieval;

[0047] Figure 7 A schematic diagram of an LSBR model based on RAG technology provided in an embodiment of the present invention;

[0048] Figure 8 This is the second example diagram of the plan library;

[0049] Figure 9 A schematic diagram of an intent recognition device under partially observable environmental conditions provided in an embodiment of the present invention;

[0050] Figure 10This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0051] In the embodiments of this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first," "second," etc., in the specification of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are usually of the same class, without limiting the number of objects. For example, the first object can be one or multiple.

[0052] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0053] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0054] This application provides an intention recognition method, apparatus, and related equipment under partially observable environmental conditions, aiming to improve the accuracy of intention recognition under partially observable environmental conditions.

[0055] Please see Figure 1 , Figure 1 This is a flowchart illustrating the intent recognition method under partially observable environmental conditions provided in an embodiment of the present invention. The method specifically includes the following steps:

[0056] Step 101: Obtain the original observation sequence, which includes multiple observations arranged according to timestamps, and the observations are used to characterize the actions performed by the observed object.

[0057] Step 102: In the case of missing data in the original observation sequence, a prediction is made based on the original observation sequence to obtain a predicted value, which is used to characterize the action performed by the observed object as predicted.

[0058] Step 103: Complete the original observation sequence based on the predicted value to obtain the target observation sequence.

[0059] Step 104: The target observation sequence is used as input data, the plan library is used as prior knowledge, and the input is used to the recognizer for intent recognition to obtain the first plan sequence. The recognizer is used to perform intent recognition based on the Symbolic Plan Recognition (SBR) algorithm. The plan library includes multiple pre-stored plan sequences.

[0060] Step 105: If the first plan sequence does not meet the preset conditions, input the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected through child nodes. The root node is used to represent the purpose, and the leaf node is used to represent the action.

[0061] An observed object refers to an entity that is monitored or tracked in a specific environment, and its behavior can be characterized through a series of observations. These observations typically originate from sensors, log records, or other data acquisition devices, and are recorded in the form of timestamps to form the original observation sequence. For example, in autonomous driving scenarios, the observed object could be a vehicle, pedestrian, or traffic light; in human-machine interaction systems, the observed object could be a user or an operating device.

[0062] The object of observation is an entity capable of performing observation, including vehicles, autonomous robotic arms, or robots. The object of observation varies across different application domains, and no specific limitation is made here. For example, in a smart customer service scenario, the object of observation is the user using a mobile phone; in an industrial monitoring scenario, the object of observation is the worker operating the equipment or an autonomous robotic arm; and in the field of agent behavior analysis, the object of observation is the agent.

[0063] An action refers to a specific behavior or state change performed by an observed object at a particular moment, and it is reflected through observation values. For example, in autonomous driving scenarios, actions could be vehicle acceleration, deceleration, steering, etc.; in human-computer interaction systems, actions could be clicking, swiping, voice input, etc. Action recognition relies on modeling and analyzing the state changes of the observed object.

[0064] An action is a basic behavioral unit actually performed by an observed object and captured by the system. It has a clear timestamp and semantic meaning. The timestamp is used to characterize the time when the observed object performs the action to determine the order of different actions. These actions are explicit and recordable operations. In different application domains, the specific type of action varies depending on the observed object. For example, in some embodiments, this method is applied to agent behavior analysis, where the observed object is an agent, and the action is the agent's behavior or state, such as location movement (moving to a certain area), tool use (picking up or putting down an item), and communication behavior.

[0065] It should be understood that, in this embodiment, the actions included in the original observation sequence are called observations, representing the raw data obtained through actual observation. Each observation corresponds to a timestamp, and arranging multiple observations in order of timestamps constitutes the original observation sequence. The original observation sequence reflects the behavioral trajectory of the observed object, but due to network interruptions, sensor failures, communication interruptions, occlusion, system communication failures, etc., data may be missing in the original observation sequence, resulting in an incomplete behavioral chain. In specific implementations, one or more observations may be missing in the original observation sequence, and the missing observations may be located at any position in the original observation sequence.

[0066] To compensate for this incompleteness, when data gaps exist in the original observation sequence, predicted values ​​are obtained based on the original observation sequence, and the original observation sequence is completed based on the predicted values, thereby generating the target observation sequence—that is, a logically coherent and complete action flow after completion. In this embodiment, the predicted actions are referred to as predicted values, thus distinguishing them from the actual observed values.

[0067] The following is an example using a specific implementation. For instance, when the observed object is a user, the original observation sequence is: ["Search for product A", "Click on product A", "Add to cart", "Redirect to homepage", data missing] etc. In the original observation sequence, the action after "Redirect to homepage" is not observed, therefore there is missing data. Based on the original observation sequence, predictions are made, resulting in the predicted values ​​"Enter checkout page", "Select shipping address", and "Submit order". The predicted values ​​are then padded to the original observation sequence according to timestamps, resulting in the target observation sequence as follows: ["Search for product A", "Click on product A", "Add to cart", "Redirect to homepage", "Enter checkout page", "Select shipping address", "Submit order"].

[0068] A plan library, also known as a domain representation, refers to the topological structure of a trusted plan library. It contains a collection of all known, standard plan sequences; essentially, it's a domain knowledge graph or behavioral template library that defines the reasonable and typical action sequences that can be followed to achieve a certain goal within that domain. In this embodiment, the plan library includes multiple pre-stored plan sequences. Each pre-stored plan sequence describes the complete execution path corresponding to a goal (such as "completing an online purchase"). This path is organized in a tree or linear structure, decomposing layer by layer from the root node representing the final goal to the leaf nodes representing specific actions. Nodes between the root node and the leaf nodes are called child nodes.

[0069] For a concrete example, please see Figure 2 The plan library consists of a root node representing the purpose, leaf nodes representing the actions, and multiple child nodes. In addition, it includes the topology of the plan library, including the depth of the plan library, the number of child nodes in the decomposition of compound actions (called the And branch factor), the decomposition method of compound actions (called the OR branch factor), and the order constraints between actions (i.e., the parameters of the actions, such as action A occurring before action B). Figure 2 The plan library shown represents the actions and plan structure of the described agent, where the root node represents the goal, such as B9, B6, B15, B12, and B3 in the figure. The set of all root nodes in the plan library is the goal set.

[0070] It should be understood that the planning sequence is a complete sequence containing the root node corresponding to the goal, that is, it includes the root node representing the goal, the leaf nodes representing the actions, and the child nodes connecting the root node and the leaf nodes. The observation sequence, however, only contains actions and does not include anything representing the goal. In this embodiment, the first planning sequence, the second planning sequence, and the pre-stored planning sequence are all planning sequences; "first," "second," etc., are only used to distinguish similar objects. The original observation sequence and the target observation sequence are both observation sequences.

[0071] In this embodiment, the observed object is first observed, and the actions performed by the observed object at different times are recorded. The observed values ​​are then arranged in the order of timestamps to obtain the original observation sequence. After obtaining the original observation sequence, it is determined whether there is any missing data in the original observation sequence.

[0072] If there are missing data in the original observation sequence, prediction is needed based on the original observation sequence to fill in the missing data and obtain the target observation sequence. If there are no missing data in the original observation sequence, the original observation sequence can be directly used as the target observation sequence and input into the recognizer for intent recognition to obtain the first plan sequence. The specifics are not limited here.

[0073] It should be understood that the specific method for obtaining the predicted value based on the original observation sequence is not limited here. Optionally, in some embodiments, the step of predicting based on the original observation sequence to obtain the predicted value includes:

[0074] The original observation sequence is used as input data, and the plan library is used as prior knowledge. The data is then input into a large language model for prediction to obtain the predicted value.

[0075] In this embodiment, the original observation sequence is input into a pre-trained Large Language Model (LLM). The LLM uses the plan library as contextual information and combines it with the plan library to perform sequence modeling, complete any missing or incomplete parts that may exist in the original observation sequence, and obtain the predicted value.

[0076] First, prompts are constructed based on the original observation sequence and structured prior knowledge in the plan library, and then input into the large language model. Based on the received prompts, the large language model performs the following processing steps: First, it parses the semantics and temporal order of observations in the original observation sequence; second, it identifies one or more candidate plan sequences that match the observations by combining the path structure of each pre-stored plan sequence in the plan library; third, it infers the observation most likely to appear at the missing time position based on the logical dependencies, execution order constraints, and path consistency constraints between actions in the candidate plan sequences; finally, it generates one or more predicted values ​​that conform to the context and belong to the leaf nodes of the plan library.

[0077] As a specific implementation, the prompt restricts the output format of the large language model to JSON: {{"timestamp": "observation"}}, and the predicted value must be a leaf node in the plan library—that is, a node that has no child nodes in the structure; if a node is followed by a more indented child node, it is not a leaf node and cannot be used as a predicted value.

[0078] In other embodiments, the prompt also provides path consistency constraints, a first critical constraint, a second critical constraint, and a depth constraint. The path consistency constraint ensures that all observations (including missing ones) represent a complete path from the root node to a leaf node. The first critical constraint requires that all first-level child nodes (direct children of the root) of all paths must be identical. The second critical constraint requires that the predicted value and its adjacent second- and third-level child nodes must be identical, as they belong to the same plan. The depth constraint requires that the plan sequence must be equal to the depth; the length of all plan sequences must equal the depth, and arbitrary combinations of nodes are not allowed.

[0079] In some embodiments, the predicted values ​​are output as structured data, containing the action content and its corresponding timestamp location information. Subsequently, the predicted values ​​are inserted into the corresponding missing positions in the original observation sequence according to their timestamps, thereby obtaining the completed target observation sequence for subsequent intent recognition processing.

[0080] With the above settings, the entire prediction process is completed by the large language model under the symbolic constraints provided by the plan library. This not only utilizes its powerful context modeling and generation capabilities, but also ensures the rationality and interpretability of the prediction results in the task logic.

[0081] It should be understood that the SBR algorithm identifies the output plan sequence through labeling and strong order constraints. In some embodiments, the target observation sequence may not meet the strong constraints, causing the recognizer to fail to output the first plan sequence, which is reflected in the algorithm as an empty output. In some embodiments, the preset condition is that the first plan sequence is empty.

[0082] If the target observation sequence experiences congestion or fails to meet the constraints during propagation in the SBR algorithm (i.e., the first plan sequence does not satisfy the preset conditions), LLM generates the final recognition result (i.e., the second plan sequence) based on prior domain knowledge and the guidance of the target observation sequence, and uses the second plan sequence as the result obtained from the final intent recognition. Although the second plan sequence cannot completely satisfy the strong order constraint, it can be completely output.

[0083] In this embodiment, the missing action is predicted using LLM to obtain the predicted value, and the original observation sequence is then completed to obtain the target observation sequence. This ensures that the observation sequence always maintains its integrity, allowing for rapid identification using SBR to obtain the first planning sequence. Simultaneously, if the first planning sequence output by SBR fails to meet preset conditions, a second planning sequence is output using LLM, ensuring that there is always an identification result output to assist in the final decision-making action.

[0084] Optionally, in some embodiments, the step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes:

[0085] A first prompt message is constructed based on the target observation sequence and the plan library. The first prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the plan library.

[0086] The first prompt message is input into a large language model for processing to obtain the second plan sequence.

[0087] It should be understood that, in this embodiment, the large language model for intent recognition and the aforementioned large language model for action prediction can be the same model or different models. In some embodiments, the specific structural parameters of the large language model can be adjusted according to different needs, or the large language model can be guided to perform different functions by setting prompt messages.

[0088] Specifically, a structured prompt message (i.e., the first prompt message) is first constructed and sent to a pre-trained large language model. The first prompt message contains key information such as task description, output format requirements, execution steps, plan library structure, and existing observation data, which is used to guide the large language model to complete the mapping reasoning from the target observation sequence to the second plan sequence based on the plan library.

[0089] After receiving the first prompt message, the large language model parses the task description and clarifies that its core task is to generate a complete path sequence for each target observation (including the observed value and the predicted value) in the target observation sequence, starting from the root node, passing through the intermediate nodes and ending at the corresponding leaf node (which is the target observation value), and each path must contain timestamp information.

[0090] Then, the large language model retrieves the tree-like organizational structure provided in the plan library, traverses the entire plan library, and identifies all leaf nodes (i.e., nodes without child nodes) as the basic candidate set for subsequent matching. Through this step, the large language model ensures that subsequent path generation only uses valid leaf nodes, avoiding the generation of non-standard or fictitious action paths.

[0091] Then, the large language model reads the target observations from the target observation sequence and analyzes the leaf nodes corresponding to each target observation one by one. For each target observation, the large language model searches for the leaf node with the same semantics in the plan library and traces its complete path from the root node to form a hierarchical sequence from the root node to that leaf node.

[0092] Then, based on the structural relationships of the plan library, the large language model constructs a "plan path" that conforms to the format requirements for each target observation. The format is the string: "root node1 node2 ... leaf node timestamp". The node names in the path are arranged in hierarchical order from the root node to the leaf node, and the corresponding timestamp value is appended at the end to ensure the consistency between the path and the time series.

[0093] Then, the large language model sorts all the generated "planned paths" in ascending order by timestamp to ensure that the time order of the output results is consistent with the target observation sequence.

[0094] Finally, the large language model outputs all "planned paths" directly as a list of strings, without adding any additional explanations or descriptions, according to the output format requirements. The final output is a list of strings with a length equal to the number of target observations, where each path follows a structured format of "root → child node → leaf node → timestamp".

[0095] As a specific example, such as Figure 3 As shown, this embodiment of the invention provides an LLM-augmented Symbolic Plan Prediction and Recognition Model (LLM-augmented SBR, LSBR) to implement the intent recognition method under partially observable environmental conditions provided in this embodiment. In this embodiment, the domain representation T (i.e., the plan library) and the target set G jointly define the semantic space of LLM in a specific application context. These domain priors are input to the recognizer, which is responsible for structured parsing of the input target observation sequence O and recognizing and outputting a first plan sequence. This process reflects the recognizer's ability to extract patterns and decode semantics from the observation sequence. If the target observation sequence experiences congestion or does not conform to the constraint rules during propagation in the SBR algorithm, LLM generates the final recognition result (i.e., a second plan sequence) through prior domain knowledge and guidance from the target observation sequence. In this embodiment, the above framework ensures that there is always a recognition result output to assist in the final decision-making action.

[0096] While the LLM-based SBR prediction and recognition model can predict missing sequences and ensure completeness, thus working in conjunction with SBR to output recognition results, the increasing amount of domain knowledge and the growing number of leaf nodes and root-to-leaf paths in the plan library can lead to LLM memory confusion, resulting in incorrect inference and recognition results.

[0097] In some embodiments, Retrieval-Augmented Generation (RAG) technology is used to identify intent under partially observable environmental conditions.

[0098] Optionally, in some embodiments, the step of predicting based on the original observation sequence to obtain a predicted value includes:

[0099] Based on the observed values, a vector retrieval operation is performed in a pre-built knowledge base to obtain candidate predicted values. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequences.

[0100] Based on the candidate predicted values ​​and the observed values, a vector retrieval operation is performed in the knowledge base, and the predicted value is determined from the candidate predicted values ​​based on the retrieval results.

[0101] The knowledge base is the contextual knowledge referenced by the large language model in the generated results. After the knowledge base is built, the RAG technology can be used to reason based on the most similar knowledge, greatly reducing the scale of domain knowledge. Thanks to the parsing capability of the SBR algorithm in the process of initializing domain knowledge, the SBR knowledge base (i.e., the knowledge base) can be built automatically. Specifically, since SBR initialization requires parsing all plan library paths, root nodes, and leaf nodes, they can be automatically stored in the format of the knowledge base without consuming additional memory and time.

[0102] In some embodiments, a plan library derives a knowledge base, which is a vector storage. Specifically, the plan library is a tree structure, and the knowledge base is the parsing of the tree structure, resulting in paths from the root node to the leaf node; the knowledge base itself is in the form of a vector database.

[0103] Optionally, in some embodiments, before making predictions based on the original observation sequence to obtain predicted values, the method further includes:

[0104] The knowledge base is constructed based on the plan library.

[0105] Specifically, in this embodiment, a self-construction algorithm is used to encode symbolic domain knowledge and behavioral paths composed of action nodes into a searchable semantic knowledge base, thereby realizing the conversion from a plan base to a knowledge base.

[0106] The algorithm's input consists of formally defined domain knowledge D and the set of terminal actions generated by the SBR initialization algorithm. and behavioral path set Each path Represented as a sequence of actions .

[0107] First, node documents are constructed to obtain the action document set. Specifically, for each terminal action The algorithm generates the corresponding document representation, as follows:

[0108] .

[0109] At the same time, construct metadata records:

[0110] .

[0111] This process ensures that all executable basic actions provide semantic knowledge for subsequent planning reasoning and labels their type attributes to support type-based query filtering.

[0112] Secondly, path document construction is performed to obtain the path document set. Specifically, for each behavioral path Decompose it into ordered action steps And generate a structured text description, as follows:

[0113] .

[0114] Similar to the construction of node documents, the metadata for constructing path documents is as follows:

[0115] .

[0116] This process explicitly encodes the topological features of the path (start and end nodes and length), making it easier to capture semantic similarity and functional equivalence between paths in a high-dimensional space.

[0117] Next, knowledge base integration is performed. Specifically, the action document set generated above is integrated. With path document set Combined into a unified document collection:

[0118] .

[0119] Simultaneously, the metadata of the node documents and path documents are merged to obtain the element data set. and assign a unique identifier sequence Finally, ChromaDB—a lightweight, persistent embedded knowledge base—is used to create the specified database. Documents, metadata, and IDs are injected through the corresponding interface to complete the vectorized storage of knowledge.

[0120] In this embodiment, an effective mapping from a discrete symbolic system to a continuous semantic space is achieved through a knowledge base self-construction algorithm based on SBR. When used later, structured prompts can guide the embedded model to retain action semantics and path logic.

[0121] It should be understood that, in practical implementation, different plan libraries can be obtained based on different domains, and thus different domain-specific knowledge bases can be constructed based on these plan libraries. In practical implementation, the corresponding knowledge base is determined based on the domain corresponding to the original observation sequence, and subsequent retrieval is performed within that domain-specific knowledge base.

[0122] In this embodiment, firstly, a vector retrieval operation is performed in a pre-built knowledge base based on the observed values ​​to obtain candidate predicted values. Specifically, in some embodiments, performing a vector retrieval operation in a pre-built knowledge base based on the observed values ​​to obtain candidate predicted values ​​includes:

[0123] Based on each observation, a first matching sequence is retrieved from the knowledge base, wherein the first matching sequence is a pre-stored plan sequence with the observation as the leaf node;

[0124] The root node of the first matching sequence is determined as the prefix anchor point;

[0125] Based on each of the prefix anchor points, a second matching sequence is retrieved from the knowledge base. The second matching sequence is a pre-stored plan sequence with the prefix anchor point as the root node.

[0126] The leaf nodes of the second matching sequence are determined as the candidate predicted values. For each observation in the original observation sequence, a corresponding first query statement is constructed, and the path with the leaf node as the observation is retrieved from the knowledge base based on the second query statement to obtain the first matching sequence. Then, for each first matching sequence, the root node corresponding to the first matching sequence is determined as the prefix anchor point, and a second query statement is constructed again based on each prefix anchor point. The path with the root node as the prefix anchor point is retrieved from the knowledge base based on the second query statement to obtain the second matching sequence. The leaf nodes of all second matching sequences are collected as candidate predicted values, and after deduplication of all candidate predicted values, the final set of candidate predicted values ​​is obtained.

[0127] Optionally, in some embodiments, performing a vector retrieval operation in the knowledge base based on the candidate predicted values ​​and the observed values, and determining the predicted value from the candidate predicted values ​​based on the retrieval results, includes:

[0128] Based on each of the candidate predicted values, a third matching sequence is retrieved from the knowledge base, the third matching sequence being a pre-stored plan sequence with the leaf nodes being the candidate predicted values; and based on each of the observed values, a fourth matching sequence is retrieved from the knowledge base, the fourth matching sequence being a pre-stored plan sequence with the leaf nodes being the observed values.

[0129] The third matching sequence, the fourth matching sequence, and the original observation sequence are input into a pre-trained large language model for prediction processing to obtain the predicted value.

[0130] After obtaining candidate predicted values, a two-stage retrieval query is constructed (a third query statement and a fourth query statement). The third query statement is used to query paths in the knowledge base with candidate predicted values ​​as leaf nodes, and the fourth query statement is used to query paths in the knowledge base with observed values ​​as leaf nodes. Based on the third and fourth query statements, two searches are performed in the knowledge base to obtain the third and fourth matching sequences. Based on the third and fourth matching sequences, the original observation sequence, and the timestamps to be predicted, corresponding prompt words are constructed, and a large language model is called to perform predictions to obtain the predicted values.

[0131] In some embodiments, corresponding prompt words are constructed based on the third matching sequence, the fourth matching sequence, the original observation sequence, and the timestamp to be predicted, and a large language model is called to make predictions to obtain initial prediction results. The initial prediction results are then de-labeled, have escaped quotation marks, and are formatted and normalized, parsed into a JSON dictionary, and the final prediction value is obtained.

[0132] For ease of understanding, a specific embodiment will be used as an example below. Please refer to... Figure 4 , Figure 5 and Figure 6 First, based on the observations in the original observation sequence (e.g. Figure 4 In The path matching retrieval is performed in the knowledge base to obtain multiple existing observation behavior sequences (ObservationPath), which is the first matching sequence. Each sequence starts with the root node and extends along the behavior sequence to the leaf node, where the leaf node represents the actual observable behavior or state value (i.e., the observation value). For example, as shown in Figure 5, multiple complete paths containing observation values ​​(such as A81 and A92) are matched. Each path is labeled with semantic information such as node ID, preconditions (such as B77 precondition

[100] >B78 (ID: 102)) and path length.

[0133] To further narrow down the candidate space and enhance the semantic consistency of predictions, the goal node (Goal Node) to which the corresponding observation belongs in each ObservationPath is extracted, i.e., the root node in the first matching sequence, as a prefix anchor. Based on the principle that two consecutive observations belong to the same goal, multiple observation sequences (GoalPaths) with the same goal are further retrieved based on the prefix anchor, i.e., the second matching sequence. As shown in Figure 6, each GoalPath has root node as root node, and its first-level child nodes are the extracted goal nodes (such as B84, B28, B56, etc.), and subsequent nodes continue to the final observation.

[0134] In some embodiments, after retrieving the first matching sequence and the second matching sequence, these first matching sequence and the second matching sequence are passed as context input to the large language model. The large language model evaluates the structural similarity, semantic coherence and purpose consistency between the sequences, thereby inferring the most likely missing observation and obtaining the predicted value.

[0135] In some embodiments, since a knowledge base has been constructed, further retrieval can be performed based on the knowledge base during the process of using a large language model to identify intent and obtain a second plan sequence, thereby improving the accuracy of the second plan sequence.

[0136] Optionally, in some embodiments, the step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes:

[0137] Based on each target observation, a fifth matching sequence is retrieved from the knowledge base. The target observation includes the observed value and the predicted value. The fifth matching sequence is a pre-stored plan sequence with the target observation as the leaf node. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequence.

[0138] A second prompt message is constructed based on the fifth matching sequence, the target observation sequence, and the plan library. The second prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the fifth matching sequence and the plan library.

[0139] The second prompt message is input into the large language model for processing to obtain the second plan sequence.

[0140] For each target observation (including observed value and predicted value) in the target observation sequence, construct the corresponding third query statement, and retrieve the path of the leaf node as the observation value from the knowledge base based on the third query statement to obtain the fifth matching sequence.

[0141] A second prompt message is constructed based on the fifth matching sequence, the target observation sequence, and the plan library. This second prompt message requires the large language model to extract the complete path from the fifth matching sequence, and specifies the output format and method for each path. By calling the large language model in this way, the second plan sequence can be obtained.

[0142] As a specific example, such as Figure 7 As shown, this embodiment of the invention also provides an LSBR model (LLM-RAG-augmented SBR, LRSBR) based on Retrieval-Augmented Generation (RAG) technology. The core idea of ​​the LRSBR algorithm is to predict missing or fuzzy observation sequences through a collaborative mechanism of structured analysis, semantic vectorized embedding, RAG retrieval, and LLM inference completion. The overall framework adopts a multi-module collaborative mechanism, integrating key technologies such as domain knowledge representation, sequence recognition, semantic embedding, vector indexing, and context enhancement to construct a prediction model that supports self-construction of the knowledge base and LLM retrieval enhancement.

[0143] Specifically, after acquiring the original observation sequence, if the original observation sequence is complete, the Sequence Recognizer (SBR) performs structured processing on the input sequence, combining the plan library T and the target set G to generate the first plan sequence P. If the observation sequence contains missing or incomplete semantic information, it is marked as "missing". At the same time, the original observation sequence is transformed into a high-dimensional semantic vector through an embedding model and stored in the knowledge base to support subsequent fast semantic retrieval.

[0144] As a specific embodiment, the intention recognition method under partially observable environmental conditions specifically includes the following steps:

[0145] Obtain the original observation sequence, which includes multiple observations arranged by timestamps, and the observations are used to characterize the actions performed by the observed object;

[0146] In the event of missing data in the original observation sequence, a first matching sequence is retrieved from the knowledge base based on each observation value. The first matching sequence is a pre-stored plan sequence with the observation value as the leaf node.

[0147] The root node of the first matching sequence is determined as the prefix anchor point;

[0148] Based on each of the prefix anchor points, a second matching sequence is retrieved from the knowledge base. The second matching sequence is a pre-stored plan sequence with the prefix anchor point as the root node.

[0149] The leaf nodes of the second matching sequence are determined as the candidate predicted values;

[0150] Based on each of the candidate predicted values, a third matching sequence is retrieved from the knowledge base, the third matching sequence being a pre-stored plan sequence with the leaf nodes being the candidate predicted values; and based on each of the observed values, a fourth matching sequence is retrieved from the knowledge base, the fourth matching sequence being a pre-stored plan sequence with the leaf nodes being the observed values.

[0151] The third matching sequence, the fourth matching sequence, and the original observation sequence are input into a pre-trained large language model for prediction processing to obtain the predicted value.

[0152] The original observation sequence is completed based on the predicted values ​​to obtain the target observation sequence;

[0153] The target observation sequence is used as input data, and the plan library is used as prior knowledge. The input is used to identify the intent by the recognizer to obtain the first plan sequence. The recognizer is used to identify the intent based on the SBR algorithm. The plan library includes multiple pre-stored plan sequences.

[0154] If the first plan sequence does not meet the preset conditions, a fifth matching sequence is retrieved from the knowledge base based on each target observation value. The target observation value includes the observed value and the predicted value. The fifth matching sequence is a pre-stored plan sequence with the target observation value as the leaf node. The knowledge base includes multiple plan sequence vectors, and the plan sequence vectors are used to represent the pre-stored plan sequence.

[0155] A second prompt message is constructed based on the fifth matching sequence, the target observation sequence, and the plan library. The second prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the fifth matching sequence and the plan library.

[0156] The second prompt message is input into a large language model for processing to obtain the second plan sequence. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected through child nodes. The root node is used to represent the purpose, and the leaf node is used to represent the action.

[0157] It should be understood that, in some embodiments, retrieval in the knowledge base can be achieved by invoking a retrieval algorithm, specifically, the retrieval algorithm includes the following steps:

[0158] Use an embedding model to convert query statements into semantic vectors;

[0159] Based on semantic vectors, approximate nearest neighbor search is used to perform vector similarity retrieval in the knowledge base to obtain retrieval results;

[0160] The search results are sorted by similarity and then filtered again using metadata filtering conditions.

[0161] Returns the final set of matching plan sequences.

[0162] It should be understood that the query statements used when the above retrieval algorithm is invoked at different stages will be different, and the final set of matching plan sequences returned will also be different. For example, the query statement can be the first query statement, the second query statement, the third query statement, the fourth query statement mentioned above, or a query statement constructed based on other retrieval requirements, which is not limited here.

[0163] It should be understood that in some embodiments, a knowledge base is created based on the name of the plan library. For example, each type of plan library is named 1-5-2-3-2-full-100_baseline, with 100 instances; therefore, the knowledge bases are 1-5-2-3-2-full-100_baseline_1 to 1-5-2-3-2-full-100_baseline_100. During prediction, based on the currently used knowledge base, its corresponding vector database instance is retrieved using its name. If the vector database does not exist, the domain-specific vector database is constructed and persisted using the structured parsing data generated during the SBR algorithm initialization phase.

[0164] In this embodiment, during the query processing phase, an embedding model is used to map the query statement to a vector space, and similarity matching is performed using a vector storage index to retrieve relevant candidate data from the database. The knowledge base used for retrieval is built once during the SBR algorithm initialization process and is used permanently. This process achieves semantic alignment from unstructured text to structured database content.

[0165] To verify the effectiveness of the proposed LR-SBR and L-SBR compared to existing techniques (taking SBR as an example), the Mirsky dataset is used for testing. The Mirsky dataset describes a plan library structure with different depths, decomposition methods, and implementations, allowing for comparison of the differences in time and space consumption between different algorithms.

[0166] First, in the domain of a plan library, a plan library consists of a goal (i.e., the root node used to represent the purpose), action nodes (i.e., leaf nodes), and non-terminal nodes (i.e., child nodes). In addition, it also includes the topology of the plan library, including the depth of the plan library, the number of sub-actions in the decomposition of a compound action (called the And branch factor), the decomposition method of the compound action (called the OR branch factor), and the order constraints between actions.

[0167] Based on the topology and basic attribute structure of the plan library, the naming convention for the plan library is defined as full-AlphabetSize-Plan-Goal-Depth-AND-OR, where:

[0168] Full: indicates complete order;

[0169] AlphabetSize: Represents the number of actions, i.e. the total number of actions that the identified person can perform;

[0170] Plan: This means that a plan is executed one at a time, that is, the identified subject executes a plan for one objective at a time, without considering interruptions to the plan;

[0171] Goal: Represents the target number, i.e., the number of target nodes contained in the plan library;

[0172] Depth: Indicates depth, that is, how many layers of complex actions need to be broken down into simple, executable actions;

[0173] AND: indicates the AND branching factor, that is, how many sub-actions a complex action is broken down into;

[0174] OR: Represents the OR branching factor, that is, how many ways a complex action can be decomposed;

[0175] For example, `full-20-1-5-2-2-3` indicates that the plan library is fully ordered, has 20 action nodes, executes one plan at a time, has 5 objectives, a depth of 2, an AND branch factor of 2, and an OR branch factor of 3. Visualize the plan library topology under different structural parameters: Figure 2 The full-20-1-5-2-2-3 diagram shows a shallow two-layer structure with an AND branch factor of 2 and an OR branch factor of 3, representing a many-to-many combination relationship between the target and the action. Figure 8 (full-20-1-5-2-5-1) increases the AND branching factor to 5 at the same depth, making each target associated with more child nodes and the structure more "flat"; the plan library full-20-1-5-6-2-1 deepens to 6 layers, with an AND branching factor of 2 and an OR branching factor of 1, presenting a deeper and more linear hierarchical structure.

[0176] The dataset selected in this embodiment is shown in Table 1. The three plan libraries under each feature are not entirely identical, and 100 independent plan library instances are generated under each configuration (defined by the Depth, AND, and OR parameters). The "Average Number of Nodes" in the table is the arithmetic mean of the total number of nodes in these 100 instances, hence it appears as a small value. The number of nodes when OR is 3 is less than the number of nodes when OR is 2 because there are 100 action spaces when OR is 2, while there are only 20 when OR is 3.

[0177]

[0178] The observation sequences of the data under the three selected features are of fixed length, representing the length of the sequence data in the planning library, as shown in Table 2. For example, Depth-6 indicates that the sequence length in the planning library is 27, and the OR-3 attribute indicates that the sequence length in the planning library is 2.

[0179]

[0180] First, Mirsky proposed that the experimental features to be compared in plan identification experiments are identification time and the number of nodes generated during identification. Therefore, this embodiment designs a dynamic observation missing identification experiment for three feature plan libraries: Depth2, Depth4, Depth6; AND2, AND3, AND5; and OR2, OR3, OR4, comparing identification time and the number of generated nodes. Specifically, the experiment constructs scenarios with 20%, 30%, and 50% observation missing values. Each plan library contains 100 different plan library examples. These 100 examples differ only in the plan and implementation nodes; their Depth, AND, and OR features are completely identical.

[0181] For evaluation indicators, three indicators were selected: Plan Completion Rate (PCR), Plan Goal Accuracy Rate (PGA), and Plan Sub-Goal Accuracy Rate (PSGA).

[0182]

[0183] As shown in Table 3, LRSBR has a comprehensive advantage in balancing completion rate and accuracy, and is especially suitable for complex plan libraries and high missing rate environments; while LSBR is still competitive in simple tasks with fewer missing items.

[0184] By conducting identification tests on datasets with different Depth, AND, and OR feature behavior libraries under uncertain environments, the effectiveness of the LSBR and LRSBR algorithms provided in this embodiment of the invention in terms of recognition accuracy and efficiency was verified. Experimental results show that, compared with the traditional SBR algorithm, LRSBR can improve the robustness of recognition by sacrificing some time without consuming additional space.

[0185] Please see Figure 9 This invention also provides an intent recognition device 900 under partially observable environmental conditions, comprising:

[0186] The acquisition module 901 is used to acquire the original observation sequence, which includes multiple observation values ​​arranged according to timestamps, and the observation values ​​are used to characterize the actions performed by the observed object.

[0187] Prediction module 902 is used to make a prediction based on the original observation sequence when there is data missing in the original observation sequence, and obtain a prediction value, wherein the prediction value is used to characterize the action performed by the observed object as predicted.

[0188] The completion module 903 is used to complete the original observation sequence based on the predicted value to obtain the target observation sequence;

[0189] Recognizer 904 is used to take the target observation sequence as input data and the plan library as prior knowledge to perform intent recognition and obtain a first plan sequence. The recognizer is used to perform intent recognition based on the SBR algorithm. The plan library includes multiple pre-stored plan sequences.

[0190] A pre-trained large language model 905 is used to perform intent recognition based on the target observation sequence and the plan library when the first plan sequence does not meet the preset conditions, and obtain a second plan sequence. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected by child nodes. The root node is used to represent the purpose and the leaf node is used to represent the action.

[0191] Optionally, the step of predicting based on the original observation sequence to obtain the predicted value includes:

[0192] Based on the observed values, a vector retrieval operation is performed in a pre-built knowledge base to obtain candidate predicted values. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequences.

[0193] Based on the candidate predicted values ​​and the observed values, a vector retrieval operation is performed in the knowledge base, and the predicted value is determined from the candidate predicted values ​​based on the retrieval results.

[0194] Optionally, the step of performing a vector retrieval operation based on the observed values ​​in a pre-built knowledge base to obtain candidate predicted values ​​includes:

[0195] Based on each observation, a first matching sequence is retrieved from the knowledge base, wherein the first matching sequence is a pre-stored plan sequence with the observation as the leaf node;

[0196] The root node of the first matching sequence is determined as the prefix anchor point;

[0197] Based on each of the prefix anchor points, a second matching sequence is retrieved from the knowledge base. The second matching sequence is a pre-stored plan sequence with the prefix anchor point as the root node.

[0198] The leaf nodes of the second matching sequence are determined as the candidate predicted values.

[0199] Optionally, the step of performing a vector retrieval operation in the knowledge base based on the candidate predicted values ​​and the observed values, and determining the predicted value from the candidate predicted values ​​based on the retrieval results, includes:

[0200] Based on each of the candidate predicted values, a third matching sequence is retrieved from the knowledge base, the third matching sequence being a pre-stored plan sequence with the leaf nodes being the candidate predicted values; and based on each of the observed values, a fourth matching sequence is retrieved from the knowledge base, the fourth matching sequence being a pre-stored plan sequence with the leaf nodes being the observed values.

[0201] The third matching sequence, the fourth matching sequence, and the original observation sequence are input into a pre-trained large language model for prediction processing to obtain the predicted value.

[0202] Optionally, the step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes:

[0203] A first prompt message is constructed based on the target observation sequence and the plan library. The first prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the plan library.

[0204] The first prompt message is input into a large language model for processing to obtain the second plan sequence.

[0205] Optionally, the step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes:

[0206] Based on each target observation, a fifth matching sequence is retrieved from the knowledge base. The target observation includes the observed value and the predicted value. The fifth matching sequence is a pre-stored plan sequence with the target observation as the leaf node. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequence.

[0207] A second prompt message is constructed based on the fifth matching sequence, the target observation sequence, and the plan library. The second prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the fifth matching sequence and the plan library.

[0208] The second prompt message is input into the large language model for processing to obtain the second plan sequence.

[0209] Optionally, the step of predicting based on the original observation sequence to obtain the predicted value includes:

[0210] The original observation sequence is used as input data, and the plan library is used as prior knowledge. The data is then input into a large language model for prediction to obtain the predicted value.

[0211] The intent recognition device 900 under partially observable environmental conditions provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0212] It should be noted that the division of units in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0213] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0214] like Figure 10 As shown, this application provides an electronic device 1000, including: a memory 1002, a processor 1001, and a program stored in the memory 1002 and executable on the processor 1001; the processor 1001 is used to read the program in the memory 1002 to implement the steps in the intention recognition method under partially observable environmental conditions as described above.

[0215] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the various processes of the above-described embodiment of the intent recognition method under observable environmental conditions, and achieves the same technical effect. To avoid repetition, it will not be described again here. The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as compact disks (CDs), digital video discs (DVDs), Blu-ray discs (BDs), high-definition versatile discs (HVDs), etc.), and semiconductor storage (such as read-only memory (ROMs), erasable programmable read-only memory (EPROMs), electrically erasable programmable read-only memory (EEPROMs), non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0216] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0217] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0218] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this application without departing from its spirit, and all of these modifications are within the scope of protection of this application.

Claims

1. An intent recognition method under partially observable environmental conditions, characterized in that, include: Obtain the original observation sequence, which includes multiple observations arranged by timestamps. The observations characterize the actions performed by the observed object. The observed object is an entity capable of performing the observation, including vehicles, autonomous robotic arms, or robots. The action refers to the specific behavior or state change performed by the observed object at a certain moment. In the case of the observed object being a vehicle, the action includes vehicle acceleration, deceleration, and steering. In the case of the observed object being an autonomous robotic arm or robot, the action includes positional movement. The observations are obtained from sensors, log records, or other data acquisition devices. In the event of missing data in the original observation sequence, a prediction is made based on the original observation sequence to obtain a predicted value. The predicted value is used to characterize the action performed by the observed object. The missing data is caused by network interruption, sensor failure, communication interruption, occlusion, or system communication failure. The original observation sequence is completed based on the predicted values ​​to obtain the target observation sequence; The target observation sequence is used as input data, and the plan library is used as prior knowledge. The input is used to identify the intent by the recognizer to obtain the first plan sequence. The recognizer is used to identify the intent based on the Symbolic Plan Recognition (SBR) algorithm. The plan library includes multiple pre-stored plan sequences. If the first plan sequence does not meet the preset conditions, the target observation sequence and the plan library are input into a pre-trained large language model for intent recognition to obtain a second plan sequence. The second plan sequence is used as the intent recognition result to assist decision-making. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected by child nodes. The root node is used to represent the purpose, and the leaf node is used to represent the action.

2. The intention recognition method under partially observable environmental conditions according to claim 1, characterized in that, The prediction based on the original observation sequence to obtain the predicted value includes: Based on the observed values, a vector retrieval operation is performed in a pre-built knowledge base to obtain candidate predicted values. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequences. Based on the candidate predicted values ​​and the observed values, a vector retrieval operation is performed in the knowledge base, and the predicted value is determined from the candidate predicted values ​​based on the retrieval results.

3. The intention recognition method under partially observable environmental conditions according to claim 2, characterized in that, The step of performing vector retrieval operations based on observations in a pre-built knowledge base to obtain candidate predicted values ​​includes: Based on each observation, a first matching sequence is retrieved from the knowledge base, wherein the first matching sequence is a pre-stored plan sequence with the observation as the leaf node; The root node of the first matching sequence is determined as the prefix anchor point; Based on each of the prefix anchor points, a second matching sequence is retrieved from the knowledge base. The second matching sequence is a pre-stored plan sequence with the prefix anchor point as the root node. The leaf nodes of the second matching sequence are determined as the candidate predicted values.

4. The intention recognition method under partially observable environmental conditions according to claim 2, characterized in that, The step of performing a vector retrieval operation in the knowledge base based on the candidate predicted values ​​and the observed values, and determining the predicted value from the candidate predicted values ​​based on the retrieval results, includes: Based on each of the candidate predicted values, a third matching sequence is retrieved from the knowledge base, the third matching sequence being a pre-stored plan sequence with the leaf nodes being the candidate predicted values; and based on each of the observed values, a fourth matching sequence is retrieved from the knowledge base, the fourth matching sequence being a pre-stored plan sequence with the leaf nodes being the observed values. The third matching sequence, the fourth matching sequence, and the original observation sequence are input into a pre-trained large language model for prediction processing to obtain the predicted value.

5. The intention recognition method under partially observable environmental conditions according to claim 1, characterized in that, The step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes: A first prompt message is constructed based on the target observation sequence and the plan library. The first prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the plan library. The first prompt message is input into a large language model for processing to obtain the second plan sequence.

6. The intention recognition method under partially observable environmental conditions according to claim 1, characterized in that, The step of inputting the target observation sequence and the plan library into a pre-trained large language model for intent recognition to obtain a second plan sequence includes: Based on each target observation, a fifth matching sequence is retrieved from the knowledge base. The target observation includes the observed value and the predicted value. The fifth matching sequence is a pre-stored plan sequence with the target observation as the leaf node. The knowledge base includes multiple plan sequence vectors, which are used to represent the pre-stored plan sequence. A second prompt message is constructed based on the fifth matching sequence, the target observation sequence, and the plan library. The second prompt message is used to instruct the large language model to generate a corresponding plan sequence for each timestamp in the target observation sequence based on the fifth matching sequence and the plan library. The second prompt message is input into the large language model for processing to obtain the second plan sequence.

7. The intention recognition method under partially observable environmental conditions according to claim 1, characterized in that, The prediction based on the original observation sequence to obtain the predicted value includes: The original observation sequence is used as input data, and the plan library is used as prior knowledge. The data is then input into a large language model for prediction to obtain the predicted value.

8. An intent recognition device under partially observable environmental conditions, characterized in that, include: The acquisition module is used to acquire the original observation sequence, which includes multiple observation values ​​arranged according to timestamps. The observation values ​​are used to characterize the actions performed by the observed object. The observed object is an entity capable of performing the observation, including a vehicle, an autonomous robotic arm, or a robot. The action refers to the specific behavior or state change performed by the observed object at a certain moment. In the case of the observed object being a vehicle, the action includes the vehicle's acceleration, deceleration, and steering. In the case of the observed object being an autonomous robotic arm or robot, the action includes positional movement. The observation values ​​are obtained from sensors, log records, or other data acquisition devices. The prediction module is used to make predictions based on the original observation sequence when there is data missing in the original observation sequence, and to obtain a predicted value. The predicted value is used to characterize the action performed by the observed object. The data missing is caused by network interruption, sensor failure, communication interruption, occlusion or system communication failure. The completion module is used to complete the original observation sequence based on the predicted value to obtain the target observation sequence; A recognizer is used to perform intent recognition by taking the target observation sequence as input data and the plan library as prior knowledge to obtain a first plan sequence. The recognizer is used to perform intent recognition based on the SBR algorithm. The plan library includes multiple pre-stored plan sequences. A pre-trained large language model is used to perform intent recognition based on the target observation sequence and the plan library when the first plan sequence does not meet the preset conditions, and obtain a second plan sequence. The second plan sequence is used as the intent recognition result to assist decision-making. The plan sequence is a complete path from the root node to the leaf node. The root node and the leaf node are connected by child nodes. The root node is used to represent the purpose and the leaf node is used to represent the action.

9. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; characterized in that the processor is configured to read the program from the memory to implement the steps in the intention recognition method under partially observable environmental conditions as claimed in any one of claims 1 to 7.

10. A readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps in the intention recognition method under partially observable environmental conditions as described in any one of claims 1 to 7.

Images