A time sequence consistent evidence organization method for aging-oriented question answering

Abstract

The application provides a time sequence consistent evidence organization method for time limit question answering, comprising: obtaining time limit question answering text; obtaining a candidate evidence set; obtaining priori scores corresponding to each candidate evidence; constructing action scoring references corresponding to the candidate evidence based on the priori scores, and modeling an evidence selection process as a sequential decision process; in each round of candidate evidence selection, constructing a current state according to a current selected evidence set, an unselected candidate set, a time slot coverage state, a semantic abstract state and a remaining budget, calculating action probabilities corresponding to candidate selection actions and a stop action through a strategy network, and selecting a target candidate evidence to join the selected evidence set based on the action probabilities, or selecting the stop action based on the action probabilities; when the strategy network selects the stop action, or the current remaining budget is less than a preset budget threshold, or the current selection round number reaches a preset maximum selection round number, outputting the selected evidence set and generating a corresponding question answering result.

Description

Technical Field

[0001] This application belongs to the field of natural language processing technology, and in particular relates to a method for organizing time-consistent evidence for time-sensitive question answering. Background Technology

[0002] With the development of large language models and retrieval enhancement generation technology, timely question answering, which involves time constraints, event evolution and historical state queries, has gradually become an important research direction in the field of intelligent question answering. Therefore, how to obtain effective evidence that meets the time conditions has become a key issue in timely question answering.

[0003] In existing technologies, candidate evidence is typically retrieved from a knowledge base or document collection based on the user-input question text. Then, a semantic relevance model is used to rank the candidate evidence, and the ranking result is input into a question-answering generation model to generate the final answer. Some solutions also introduce time filtering, timestamp matching, or time encoding to process the time information in the candidate evidence, thereby improving time matching capabilities in time-sensitive question-answering scenarios. Furthermore, there are also solutions that progressively filter candidate evidence through multi-stage ranking or dynamic evidence selection.

[0004] However, most existing technologies focus on matching the local relevance between candidate evidence and the question text, and do not give enough consideration to the temporal consistency between candidate evidence and the overall time coverage. This can easily lead to problems such as time conflicts, incomplete time coverage and evidence redundancy, thus affecting the accuracy of the question-and-answer results. Summary of the Invention

[0005] Therefore, it is necessary to provide a time-consistent evidence organization method for time-sensitive question answering to address the aforementioned technical issues.

[0006] According to a first aspect of the present invention, a method for organizing time-consistent evidence for time-sensitive question answering is provided, comprising: Obtain the timely Q&A text and parse the time constraints in the timely Q&A text; A set of candidate evidence is obtained based on the time-sensitive question and answer text; Semantic scores and temporal scores corresponding to candidate evidence are calculated based on semantic channels and temporal channels respectively, and the semantic scores and temporal scores are dynamically fused based on the temporal constraints to obtain the prior scores corresponding to each candidate evidence. Based on the prior scores, action scoring references corresponding to candidate evidence are constructed, and the evidence selection process is modeled as a sequential decision-making process. In each round of candidate evidence selection, the current state is constructed based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget. The action probabilities corresponding to the candidate selection action and the stop action are calculated through the policy network. The target candidate evidence is selected to be added to the set of selected evidence based on the action probabilities, or the stop action is selected based on the action probabilities. When the policy network chooses to stop the action, or the current remaining budget is less than the preset budget threshold, or the current selection rounds reach the preset maximum selection rounds, the selected evidence set is output and the corresponding question and answer results are generated.

[0007] In some embodiments, parsing the time constraints in the time-sensitive question-and-answer text includes: Extract the main content constraints and time constraints from the time-sensitive Q&A text; The time constraints are normalized to construct time cards, which include time relationship types, time anchors, and time granularity. The time relationship types are used to characterize the relationship between the time constraints and the time anchors.

[0008] In some embodiments, semantic scores and temporal scores corresponding to candidate evidence are calculated based on semantic channels and temporal channels, respectively, and the semantic scores and temporal scores are dynamically fused based on the temporal constraints, including: Construct semantic feature vectors and temporal feature vectors corresponding to candidate evidence; Calculate the time interaction features based on the time constraints and the time intervals of the candidate evidence; A semantic score is calculated based on the semantic feature vector, and a time score is calculated based on the time feature vector and the time interaction feature. Based on the degree of explicitness of the time constraints in the time-sensitive question-and-answer text, a dynamic fusion weight is determined to dynamically fuse the semantic score and the time score.

[0009] In some embodiments, it also includes: A loss function is constructed based on the prior score, and the loss function includes ranking loss and time loss; The ranking loss is constructed based on the matching labels of candidate evidence and standard evidence, and is used to optimize the ranking effect of candidate evidence; the time loss is constructed based on the time matching labels of candidate evidence and time constraints, and is used to optimize the accuracy of time scoring. The semantic channel and the temporal channel are optimized based on the ranking loss and the time loss.

[0010] In some embodiments, the step of constructing the current state in each round of candidate evidence selection based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget includes: The current information digest is calculated based on the currently selected set of evidence to represent the information content that has been acquired. The time slot coverage status is calculated based on the time constraints to characterize the coverage of the time constraints by the currently selected evidence set.

[0011] In some embodiments, the step of calculating the action probabilities corresponding to candidate selection actions and stop actions through a policy network, and selecting target candidate evidence to add to the selected evidence set based on the action probabilities, or selecting a stop action based on the action probabilities, includes: For each unselected candidate evidence, a candidate action representation is constructed, which includes candidate evidence encoding features, the prior score, the semantic summary state, the time slot coverage state, and the remaining budget information; Construct a current action space, which includes stop actions and candidate selection actions corresponding to each unselected candidate piece of evidence; The action scores of the stop action and each candidate selection action are calculated based on the policy network, and illegal actions corresponding to the selected candidates, candidates exceeding the budget, or candidates with index abnormalities are masked. The action scores after masking are normalized to obtain action probabilities, and candidate actions or stop actions are selected based on the action probabilities. When a candidate action is selected, the corresponding candidate evidence is added to the set of selected evidence and the current state set is updated. When a stop action is selected, the current set of selected evidence is output.

[0012] In some embodiments, before calculating the action probabilities corresponding to candidate selection actions and stop actions through the policy network, the following steps are included: By using an inter-candidate interaction encoder, the candidate action representations corresponding to the unselected candidate evidence are interactively encoded to obtain interactively enhanced candidate action representations. The action score of the corresponding candidate action is calculated based on the enhanced candidate action representation.

[0013] In some embodiments, calculating the time slot coverage state based on the time constraints includes: Construct a corresponding set of time slots based on the aforementioned time constraints; Calculate the time slot coverage vector based on the coverage of each time slot according to the candidate evidence; The overall time coverage rate is calculated based on the time slot coverage vector to obtain the time slot coverage status.

[0014] In some embodiments, calculating the current information digest based on the currently selected set of evidence includes: The selected evidence set is aggregated using the mean pooling method to obtain the first semantic summary; The selected evidence set is aggregated using attention pooling to obtain a second semantic summary.

[0015] In some embodiments, it also includes: Construct a composite reward function, which includes evidence recall reward, question answering accuracy reward, time consistency reward, time slot coverage reward, and evidence quantity penalty item; The reward value corresponding to the evidence selection trajectory is calculated based on the composite reward function, and the policy network used to select target candidate evidence round by round is updated by reinforcement learning based on the reward value.

[0016] In some embodiments, it also includes: Based on the matching labels between candidate evidence and standard evidence, and the time matching labels between candidate evidence and time constraints, the prior scoring model used to generate the prior score is trained under supervision. The parameters of the prior scoring model are fixed, and the policy network used for selecting target candidate evidence round by round is trained by reinforcement learning based on the composite reward function; The prior scoring model and the policy network are jointly fine-tuned based on a preset learning rate, which is less than the learning rate in the supervised pre-training stage.

[0017] According to a second aspect of the present invention, a temporally consistent evidence organization system oriented towards time-sensitive question answering is provided, comprising: The acquisition module is used to acquire timely question and answer texts and acquire a set of candidate evidence based on the timely question and answer texts. The processing module is used to parse the time constraints in the time-sensitive question-and-answer text; calculate the semantic score and time score corresponding to the candidate evidence based on the semantic channel and the time channel respectively, and dynamically fuse the semantic score and time score based on the time constraints to obtain the prior score corresponding to each candidate evidence; construct the action scoring reference corresponding to the candidate evidence based on the prior score, and model the evidence selection process as a sequential decision-making process; in each round of candidate evidence selection, construct the current state based on the current set of selected evidence, the set of unselected candidates, the time slot coverage state, the semantic summary state, and the remaining budget, calculate the action probabilities corresponding to the candidate selection action and the stop action through the policy network, and select the target candidate evidence to add to the set of selected evidence based on the action probability, or select the stop action based on the action probability; The judgment module is used to determine whether the policy network chooses to stop, or whether the current remaining budget is less than a preset budget threshold, or whether the current selection rounds have reached a preset maximum selection rounds, and generates corresponding question and answer results; The output module is used to output the selected evidence set and generate corresponding question and answer results.

[0018] According to a third aspect of the present invention, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above-described time-consistent evidence organization method for time-sensitive question answering.

[0019] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the time-consistent evidence organization method for time-sensitive question-and-answer purposes as described above.

[0020] According to the present invention, candidate evidence is scored and fused through semantic and temporal channels, so that the candidate evidence screening process considers both semantic relevance and temporal consistency. Target candidate evidence is selected round by round based on the time slot coverage status and semantic summary status, enabling the evidence organization process to be dynamically adjusted in conjunction with the overall status of the currently selected evidence. This reduces the occurrence of time conflicts and invalid redundancy among candidate evidence, thereby improving the rationality of evidence organization in time-sensitive question-and-answer scenarios and enhancing the temporal consistency and accuracy of question-and-answer results.

[0021] Furthermore, by constructing time slot coverage status and calculating overall time coverage, the system can characterize the time coverage of the currently selected evidence set. By generating semantic summaries based on mean pooling and attention pooling, and updating the parameters of the evidence selection strategy in conjunction with a composite reward function, the candidate evidence selection process can take into account time coverage, semantic relevance, and evidence quantity control, thereby improving the overall coordination in the evidence selection process, enhancing the ability to organize complex time-sensitive information, and realizing dynamic evidence optimization in time-sensitive question-and-answer scenarios.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 A schematic flowchart of a time-consistent evidence organization method for time-sensitive question answering according to an embodiment of the present invention is shown; Figure 2 It shows Figure 1 A schematic flowchart of the method for step S300 shown; Figure 3 It shows Figure 1 A schematic flowchart of the method for step S500 shown; Figure 4 A structural block diagram of a time-consistent evidence organization system oriented towards time-sensitive question answering according to an embodiment of the present invention is shown; Figure 5 A schematic structural diagram of a computer device according to an embodiment of the present invention is shown. Detailed Implementation

[0025] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, it should be noted that, for ease of description, only the parts relevant to this application are shown in the accompanying drawings, not the entire structure. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0026] The terms “comprising” and “having”, and any variations thereof, used in this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] It should be noted that the purpose of this application is to provide an adaptive evidence organization scheme for time-sensitive question answering that balances temporal constraints and evidence accuracy. Specifically, in business scenarios with strong time constraints, such as time-sensitive knowledge question answering, enhanced retrieval generation, dynamic information retrieval, fact-checking, and time-sensitive compliance auditing, this application no longer relies solely on semantic similarity ranking or a fixed number of searches for evidence selection. Instead, it comprehensively considers semantic relevance, temporal consistency, completeness of time slot coverage, evidence redundancy, and the accuracy of the final question answer for overall decision optimization. For example, in the time-sensitive fact-finding question answering scenario, based on the explicit time constraints extracted from the user's question, candidate evidence is scored using both semantic and temporal channels. Through sequential selection and stopping decisions, a set of evidence that is temporally consistent, content-rich, and concise in quantity is automatically selected, ultimately improving the accuracy and temporal reliability of the question answering results.

[0029] In one exemplary embodiment, such as Figure 1 As shown, a time-consistent evidence organization method for time-sensitive question answering is provided, which includes the following steps S100 to S600: Step S100: Obtain the time-sensitive Q&A text and parse the time constraints in the time-sensitive Q&A text.

[0030] In this embodiment, time constraints are represented by explicit time points, time intervals, relative times, or implicit time limits contained in the user's question. These constraints are used to locate and filter evidence that conforms to time logic and to limit the time range for evidence selection.

[0031] Specifically, the time constraint conditions are, for example, 2023, 2019 to 2023, the last three years, the latest version, or currently valid, and are used to filter out evidence that matches the time range from the candidate evidence.

[0032] It should be noted that the time constraints are obtained by conducting logical structure analysis on the time-sensitive question and answer text. This analysis process involves breaking down and extracting the core content, keywords, and time-limited words from the text.

[0033] In some embodiments, the time constraints in parsing timely question-and-answer text include: Extract the main content constraints and time constraints from the timely Q&A text.

[0034] Time constraints are normalized to construct time cards, which include time relationship types, time anchors, and time granularity. The time relationship type is used to characterize the relationship between time constraints and time anchors.

[0035] In this embodiment, by separating the main content constraints and time constraints from the time-sensitive question-and-answer text, and normalizing the time constraints into a unified time card representation, different forms of time expression can be mapped to a standardized time semantic structure. Based on the time cards, information such as time relationships, time ranges, and time granularity can be further extracted and used as input for subsequent time interaction feature construction, time score calculation, and time slot coverage analysis. This enhances the ability to constrain time consistency during evidence retrieval and evidence organization, and improves the matching degree between the final question-and-answer results and time requirements.

[0036] Step S200: Obtain a set of candidate evidence based on the timely question and answer text.

[0037] In this embodiment, the candidate evidence set is labeled to obtain evidence related to the timely question and answer text. The labeling process includes sorting labels and time labels.

[0038] In this embodiment, during the training phase, ranking labels and time labels can be constructed for the candidate evidence set to train the prior scoring module. During the inference phase, the candidate evidence set can carry or be parsed to obtain information such as candidate evidence encoding features, basic semantic scores, candidate time intervals, and time missing markers, and the prior scoring module calculates the prior score. The ranking labels are used to determine whether the content of the candidate evidence is valid standard evidence compared to the core text content. If so, it is marked as valid; otherwise, it is marked as invalid, thus distinguishing the purpose of the evidence and assisting in the ranking of evidence. The time labels are used to verify whether the time range of the evidence matches the timeliness requirements of the question and answer. They are determined by combining two types of constraint scenarios: a single time point and a continuous time interval. If the evidence time range contains a specified time point, is earlier or later than that time point overall, or overlaps with a specified interval, it is determined to meet the timeliness requirements and a label is constructed. These two types of labels guide the model to distinguish valid evidence and learn time features, ensuring that the selected evidence matches the question and answer content and time sequence rules.

[0039] Step S300: Calculate the semantic score and time score corresponding to the candidate evidence based on the semantic channel and the time channel respectively, and dynamically fuse the semantic score and time score based on the time constraint to obtain the prior score corresponding to each candidate evidence.

[0040] In this embodiment, the prior score is a comprehensive consideration of the semantic relevance and temporal matching degree of each candidate evidence. This step can be used as a priority basis for evidence selection and to initially screen evidence with a high degree of matching.

[0041] Figure 2 It shows Figure 1 A schematic flowchart of the method for step S300 is shown. Figure 2 As shown, steps S301 to S304 are included: Step S301: Construct the semantic feature vector and temporal feature vector corresponding to the candidate evidence.

[0042] In this embodiment, the semantic feature vector is used to characterize the semantic relationship between candidate evidence and time-sensitive question and answer text, and the temporal feature vector is used to characterize the temporal matching relationship between candidate evidence and time constraints.

[0043] The semantic feature vector includes a local semantic encoding vector generated based on the local context information of the candidate evidence, a global semantic encoding vector generated based on the global context information, and semantic matching features between the question text and the candidate evidence, which are used to characterize the semantic content of the candidate evidence and its relevance to the question and answer text; the temporal feature vector includes the temporal consistency features between the candidate evidence and the time constraints, the time constraint feature vector corresponding to the time-sensitive question and answer text, and the time missing marker features corresponding to the candidate evidence, which are used to characterize the matching between the temporal information of the candidate evidence and the time constraints.

[0044] It should be noted that the time consistency feature is used to reflect the correspondence between the time interval of candidate evidence and the time constraint conditions, the time constraint feature vector is used to characterize the time relationship type, time anchor point and time granularity information in the timeliness question and answer text, and the time missing marker feature is used to characterize whether there is a time information missing in the candidate evidence.

[0045] Step S302: Calculate the time interaction features based on the time constraints and the time intervals of the candidate evidence.

[0046] In this embodiment, the time interaction feature is represented as a specific time relationship value between candidate evidence and time constraints, which is used to calculate a specific time feature vector. The time interaction feature includes interval overlap rate and time distance.

[0047] Step S303: Calculate the semantic score based on the semantic feature vector, and calculate the time score based on the time feature vector and time interaction features.

[0048] In this embodiment, the semantic score is used to characterize the degree of semantic relevance between candidate evidence and time-sensitive question-and-answer text, and the time score is used to characterize the degree of temporal consistency between candidate evidence and time constraints.

[0049] This application adopts a dual-channel scoring mechanism with independent semantic and temporal channels. The semantic channel score is used to reflect the matching of candidate evidence with the time-sensitive question and answer text in terms of event content, entity relationship and semantic expression. The temporal channel score is used to reflect the temporal matching relationship between the time information corresponding to the candidate evidence and the time constraints.

[0050] Since there are situations in time-sensitive question-and-answer scenarios where "semantic relevance is not consistent with time", this application models semantic relevance and time consistency independently, so that the scoring process of candidate evidence can take into account both semantic content matching and time constraint matching. This avoids the problem of time conflict or time offset when ranking based solely on semantic relevance, and provides a basis for the dynamic fusion and sequential selection of candidate evidence.

[0051] Step S304: Determine the dynamic fusion weight based on the clarity of the time constraints in the time-sensitive question-and-answer text, so as to dynamically fuse the semantic score and the time score.

[0052] Clarity refers to whether the time range corresponding to the time constraint is clear, whether the time boundary is clear, and whether the time reference is unique. The higher the clarity of the time constraint, the more specific the time range is defined in the time-sensitive question-and-answer text, and the stronger the constraint effect of time information in the candidate evidence screening process. The lower the clarity of the time constraint, the more ambiguous or uncertain the target time range is, and the more important the role of semantic content in the candidate evidence screening process.

[0053] Specifically, when the time constraint is a vague time expression such as "recently" or "recently," the degree of clarity of the time constraint is low because the corresponding time range is not fixed. Therefore, the fusion weight corresponding to the semantic score is relatively increased during the dynamic fusion process. When the time constraint is a clear time expression such as "last year," "2021," or "May 2023," the degree of clarity of the time constraint is high because the corresponding time range is relatively certain. Therefore, the fusion weight corresponding to the time score is relatively increased during the dynamic fusion process, so that the candidate evidence screening process is more in line with the time consistency requirements in the time-sensitive question-and-answer scenario.

[0054] In one specific embodiment, dynamic fusion is calculated using the following formula: ; Where, α i Here, g represents the gating weights, σ ​​represents the normalization process, and g represents the gating weights. θ (x i (This is) a multilayer perceptron. ∈(0,1) is the prior score, u i sem Score the semantic channel, ui temp Scoring is done based on the time channel. When the time conditions of the problem are clearly defined, 1-α i It tends to increase (i.e., α) i (reduce), thereby giving time channel u i temp Higher weighting; α when time conditions are ambiguous i The tendency is towards increasing, biased towards semantic channel u i sem .

[0055] It should be noted that the above formula is only an exemplary implementation method for illustrating the dynamic fusion process. Its purpose is to illustrate the dynamic weight adjustment relationship between the semantic channel and the time channel. The fusion form can be adjusted in different implementation scenarios.

[0056] In some possible embodiments, it also includes: A loss function is constructed based on prior scores, and the loss function includes ranking loss and time loss. Among them, the ranking loss is constructed based on the matching labels of candidate evidence and standard evidence, and is used to optimize the ranking effect of candidate evidence; the time loss is constructed based on the time matching labels of candidate evidence and time constraints, and is used to optimize the accuracy of time scoring. The parameters of the semantic channel and the temporal channel are optimized based on the ranking loss and the temporal loss.

[0057] In a specific implementation, the sorting loss and time loss are calculated using the following formula: ; ; p i =σ(α i u i sem + (1-α i )u i temp ); p i temp = σ(u i temp ); Among them, W + W - W + t W - t y represents the weights of positive and negative samples. i The ranking label used to characterize whether candidate evidence is standard evidence, p i t represents the prior probability or ranking probability after fusion. iThe time label used to characterize whether candidate evidence meets the time constraints, p i temp The probability of time matching for candidate evidence.

[0058] Step S400: Construct action scoring references corresponding to candidate evidence based on prior scores, and model the evidence selection process as a sequential decision-making process.

[0059] In this embodiment, the round-by-round evidence selection process is modeled as a sequential decision-making process. The state s in round t... t Including the selected evidence set S t The remaining unselected candidate set R t Semantic summary status g t Time slot coverage status m t and the remaining budget b t The policy network is based on state s t Output motion space A t The probability of each action is given, and the action space includes the stop action and the candidate action corresponding to each unselected candidate evidence.

[0060] This step differs from traditional one-time batch retrieval and evidence selection methods. Instead of directly determining the final evidence set based solely on a single round of ranking results, it utilizes prior scoring to provide action scoring references for subsequent evidence selection processes, modeling the evidence selection process as a sequential decision-making process. Within this sequential decision-making framework, the system divides the evidence selection task into multiple consecutive decision rounds. Each round dynamically decides on subsequent selection actions based on the current evidence organization status. This allows the evidence selection process to comprehensively consider factors such as the information content carried by the selected evidence, time coverage, and remaining selection budget, gradually constructing an evidence set that satisfies temporal consistency and content integrity, thus supporting the subsequent generation of accurate and timely question-and-answer results.

[0061] In step S500, in each round of candidate evidence selection, the current state is constructed based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget. The action probabilities corresponding to the candidate selection action and the stop action are calculated through the policy network. The target candidate evidence is selected to be added to the set of selected evidence based on the action probability, or the stop action is selected based on the action probability.

[0062] In this embodiment, the evidence selection process is modeled as a continuous sequential decision-making process. In each round of decision-making, the system constructs a current state based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget, and inputs the current state into the policy network. The policy network calculates the action probabilities corresponding to candidate selection actions and stop actions based on the current state to determine whether to continue supplementing new evidence or terminate the evidence selection process. Compared with traditional schemes that rely on a fixed number of evidence or preset stop rules, this embodiment incorporates the stop action as a decision action with equal status to the candidate evidence selection action into a unified decision-making framework. This allows the system to autonomously determine whether to continue selecting new evidence based on the current evidence organization effect, thereby reducing the possibility of introducing redundant evidence while ensuring time consistency and information integrity, and improving the overall quality of the evidence organization results.

[0063] Figure 3 It shows Figure 1 A schematic flowchart of the method for step S500 is shown. Figure 3 As shown, the steps include S501 to S507 as follows: Step S501: Calculate the current information digest based on the currently selected evidence set to represent the currently acquired information content.

[0064] In some possible embodiments, the current message digest is calculated based on the currently selected set of evidence, including: The selected evidence set is aggregated using the mean pooling method to obtain the first semantic summary.

[0065] The selected evidence set is aggregated using attention pooling to obtain a second semantic summary.

[0066] The semantic summary utilizes pooling operations to compress and integrate the vector features of selected evidence, forming a concise representation that feeds back the acquired information to the policy network. Furthermore, the first semantic summary is obtained by aggregating features using mean pooling, a form that is easy to implement and meets the initial needs of the solution; the second semantic summary is obtained by aggregating features using attention pooling, which further optimizes the effect and improves overall performance. Both types of semantic summaries compress the vector representation of evidence through pooling operations, forming a concise information representation that feeds back the acquired information to the policy network, supporting subsequent evidence selection.

[0067] Step S502: Calculate the time slot coverage status based on the time constraint conditions to characterize the coverage of the time constraint by the currently selected evidence set.

[0068] In some possible embodiments, calculating the time slot coverage state based on time constraints includes: Construct a set of corresponding time slots based on time constraints.

[0069] The time slot coverage vector is calculated based on the coverage of each time slot using candidate evidence.

[0070] The overall time coverage rate is calculated based on the time slot coverage vector to obtain the time slot coverage status.

[0071] Among them, the time slot coverage vector is used to provide detailed information on each time slot and the overall coverage, thereby identifying the time intervals in which evidence has not yet been supplemented.

[0072] Step S503: Construct a candidate action representation for each unselected candidate evidence. The candidate action representation includes candidate evidence coding features, prior score, semantic summary status, time slot coverage status, and remaining budget information.

[0073] In this embodiment, the candidate action representation is used to characterize the feature information of the corresponding actions of unselected candidate evidence in the current decision round, providing input for subsequent action score calculation. Unlike the method of scoring based solely on the content of the candidate evidence itself, this embodiment considers not only the coding features and prior scores of the candidate evidence when constructing the candidate action representation, but also introduces the current semantic summary state, time slot coverage state, and remaining budget information, so that the candidate action representation can simultaneously reflect the value of the candidate evidence itself and the current evidence organization state.

[0074] Among them, candidate evidence encoding features are used to characterize the semantic content of candidate evidence; prior scores are used to characterize the initial matching degree between candidate evidence and the current time-sensitive question-and-answer requirements; semantic summary status is used to characterize the information content contained in the currently selected evidence set; time slot coverage status is used to characterize the coverage of the target time range by the currently selected evidence set; and remaining budget information is used to characterize the evidence space that can be selected subsequently. By integrating the above information to construct candidate action representations, the subsequent decision-making process can not only focus on the relevance of the candidate evidence itself, but also comprehensively consider its information supplementation role and time coverage improvement role in the existing evidence set, thereby improving the completeness and temporal consistency of the evidence organization results.

[0075] In some embodiments, multiple candidate action representations can be jointly modeled through an inter-candidate interactive coding mechanism to uncover the correlation and complementarity between different candidate evidence, thereby obtaining enhanced candidate action representations and providing richer contextual information for subsequent action score calculation.

[0076] Step S504: Construct the current action space, which includes the stop action and the candidate selection action corresponding to each unselected candidate evidence.

[0077] In this embodiment, the action space defines all candidate actions that the policy network can execute in the current decision round, providing a decision range for subsequent action score calculation and action selection. Each unselected candidate piece of evidence corresponds to a candidate selection action; executing the corresponding candidate selection action indicates that the corresponding candidate evidence is added to the set of selected evidence. The stop action indicates that the current evidence selection process is terminated and the current set of selected evidence is output.

[0078] By incorporating the stopping action and the candidate selection action into a unified action space, the policy network can not only decide which candidate evidence to select in the next round, but also autonomously determine whether the currently selected evidence set already meets the question-answering requirements. Compared to stopping with a fixed number of pieces of evidence, a fixed number of selection rounds, or preset rules, this embodiment uses the stopping behavior as a learnable decision action in the decision-making process. This allows the evidence organization process to dynamically adjust the selection strategy based on the information completeness, time coverage, and remaining budget status of the current evidence set, improving the flexibility and accuracy of the evidence organization results.

[0079] In some embodiments, assuming the current set of unselected candidates contains N candidate pieces of evidence, the current action space includes N candidate selection actions and one stop action. As the evidence selection process progresses, selected candidate evidence is removed from the set of unselected candidates, and the corresponding candidate selection actions no longer participate in subsequent decisions, thus enabling the action space to be dynamically updated with the decision-making process. By constructing a dynamically changing action space, the policy network can always make decisions based on the current range of available evidence, improving the rationality and computational efficiency of the evidence selection process.

[0080] Step S505: Calculate the action scores of the stop action and each candidate selection action based on the policy network, and mask the illegal actions corresponding to the selected candidates, candidates exceeding the budget, or candidates with index abnormalities.

[0081] In this embodiment, the policy network is used to evaluate the decision value of each action in the current action space and to provide a basis for calculating the probability of subsequent actions. The action score characterizes the execution value of the corresponding action in the current decision round; a higher action score indicates that executing the action is more conducive to improving the completeness, temporal consistency, and question-answering accuracy of the final evidence organization result.

[0082] Specifically, the policy network receives current state information and candidate action representations as input, and evaluates the value of the stop action and each candidate selection action to obtain the corresponding action score. The action score for a candidate selection action represents the potential benefit of adding the corresponding candidate evidence to the selected evidence set; the action score for a stop action represents whether the current selected evidence set already has sufficient information completeness and time coverage, thus determining whether the subsequent evidence selection process should be terminated.

[0083] To prevent invalid decisions from affecting the selection of subsequent actions, this embodiment further masks illegal actions. Specifically, actions corresponding to already selected candidates indicate duplicate selection of evidence already added to the evidence set; actions corresponding to candidates exceeding the budget indicate that execution would violate the current budget constraint; and actions corresponding to candidates with index anomalies indicate that the candidate evidence is non-existent, invalid, or inaccessible. For these illegal actions, the corresponding action scores are masked through a masking mechanism, preventing them from participating in the probability calculation of subsequent actions.

[0084] By introducing an illegal action masking mechanism, the actual decision space of the policy network can be effectively reduced, avoiding the waste of computational resources on evaluating invalid actions. At the same time, it ensures that the subsequent action selection results meet the budget constraints and evidence selection rules, thereby improving the stability of the evidence organization process and the efficiency of decision-making.

[0085] Compared to the traditional approach that only ranks candidate evidence, this embodiment integrates the stopping action and the candidate selection action into the value assessment process, enabling the policy network to simultaneously learn which piece of evidence to select and when to stop selecting, thereby achieving adaptive control of the evidence organization process.

[0086] Step S506: Normalize the action scores after masking to obtain action probabilities, and select candidate actions or stop actions based on the action probabilities.

[0087] In this embodiment, after calculating the action score for each action and processing the illegal action mask, the action scores for the remaining valid actions are normalized to obtain the action probability for each action. The action probability characterizes the priority of executing the corresponding action in the current decision state; a higher action probability indicates that the corresponding action better meets the current evidence organization requirements.

[0088] After obtaining the action probabilities, the system performs action selection decisions based on the action probabilities corresponding to each action. When a candidate action is selected, the corresponding candidate evidence is added to the selected evidence set to further supplement the information content or time coverage of the current evidence set; when a stop action is selected, it means that the current selected evidence set can meet the requirements of the current question-answering task for content completeness and time consistency, thereby terminating the subsequent evidence selection process.

[0089] Compared with the traditional method of directly selecting a predetermined number of evidence based on a fixed ranking result, this embodiment uses action probability to uniformly model different decision actions, enabling the evidence selection process to dynamically adjust subsequent decision behaviors according to the current state. This takes into account multiple factors such as evidence relevance, time consistency, and evidence quantity control, thereby improving the rationality and accuracy of the final evidence organization result.

[0090] In some embodiments, the action probability is determined by the policy distribution output by the policy network, and the selection result corresponding to each action will affect the construction of the decision state in the next round. When a candidate action is selected, the selected evidence set, semantic summary state, time slot coverage state, and remaining budget are all updated accordingly; when the stop action is selected, the current decision trajectory ends, and the current selected evidence set is output as the evidence basis for the subsequent question-answering generation stage. Through the above mechanism, the evidence organization process forms a continuously iterative sequential decision-making process, realizing dynamic optimization of the evidence selection process.

[0091] Step S507: When a candidate action is selected, the corresponding candidate evidence is added to the set of selected evidence and the current state set is updated. When a stop action is selected, the current set of selected evidence is output.

[0092] In this embodiment, when a candidate selection action is chosen, the corresponding candidate evidence is added to the selected evidence set, and the current state set is reconstructed based on the updated selected evidence set for use in the next round of evidence selection. Specifically, if a candidate piece of evidence is selected, the evidence is added to the selected set, and the semantic summary state, time slot coverage state, and remaining budget information in the current state set are updated simultaneously, so that subsequent decisions can be dynamically adjusted based on the latest evidence organization state.

[0093] When the "stop" action is selected, it indicates that the currently selected evidence set has met the requirements of information completeness and time consistency for the current question-answering task. At this point, the subsequent evidence selection process is terminated, and the currently selected evidence set is output as the final evidence organization result. By unifying the execution results of candidate selection actions and stop actions into the state transition process, continuous iterative updates and adaptive termination control of the evidence selection process can be achieved, thereby improving the organization quality of the final evidence set and decision-making efficiency.

[0094] In some embodiments, before calculating the action probabilities corresponding to candidate selection actions and stop actions through the policy network, the following steps are included: By interactively encoding the candidate action representations corresponding to the unselected candidate evidence through an inter-candidate interaction encoder, the interactively enhanced candidate action representations are obtained.

[0095] The action score of the corresponding candidate action is calculated based on the candidate action representation after interaction enhancement.

[0096] In this embodiment, before calculating the action score corresponding to the candidate selection action, the candidate action representation corresponding to multiple unselected candidate evidence is first jointly modeled using an inter-candidate interaction encoder to mine the correlation, complementarity and redundancy relationships between candidate evidence, thereby obtaining the interactively enhanced candidate action representation.

[0097] Since candidate evidence is usually not independent of each other, different candidate evidence may have overlapping content, overlapping time, supplementary information, or logical connections. It is difficult to accurately reflect its actual role in the current set of unselected candidates by relying solely on the characteristics of a single candidate evidence. Therefore, an inter-candidate interactive encoder is used to enable information exchange among multiple candidate action representations, so that each candidate action representation can retain its own characteristic information while also incorporating contextual information provided by other candidate evidence.

[0098] After obtaining the enhanced candidate action representation, the action score of the corresponding candidate action is calculated based on the enhanced candidate action representation. The action score not only reflects the degree of matching between the candidate evidence itself and the current question-and-answer requirement, but also reflects its information supplement value and relative contribution to the existing candidate set, thereby improving the accuracy of subsequent action decisions and the overall quality of evidence organization results.

[0099] In step S600, the strategy network selects to stop the action, or when the current remaining budget is less than the preset budget threshold, or when the current selection rounds reach the preset maximum selection rounds, and outputs the selected evidence set and generates the corresponding question and answer results.

[0100] In this embodiment, when the policy network selects the stop action, it indicates that the currently selected evidence set has met the requirements of information integrity and time consistency for the current question-answering task, and the subsequent evidence selection process is terminated. In addition, invalid actions such as duplicate selection, exceeding the budget, and index anomalies are masked to prevent them from participating in action probability normalization. When the policy network selects the stop action, the evidence selection process is terminated and the currently selected evidence set is output. The stop action is forcibly executed when the remaining budget is 0 or the number of selection rounds reaches the preset maximum number of rounds. When stopping evidence selection, the final selected evidence set is directly output. Invalid actions can be determined through probability distribution calculation. The action score of the stop action is determined by the stop action network MLP.stop The action score of the candidate selection action is calculated based on the semantic summary state, time slot coverage state, and remaining budget; the action score of the candidate selection action is determined by the candidate selection network MLP. sel Calculated based on candidate action representations.

[0101] In some possible embodiments, it also includes: Construct a composite reward function, which includes evidence recall reward, question answering accuracy reward, time consistency reward, time slot coverage reward, and evidence quantity penalty.

[0102] The reward value corresponding to the evidence selection trajectory is calculated based on the composite reward function, and the policy network used to select target candidate evidence round by round is updated by reinforcement learning based on the reward value.

[0103] In this embodiment, the evidence recall reward incentivizes the algorithm model to select as much relevant and valid evidence as possible; the question-answering accuracy reward is used to benchmark the final answer effect and improve the overall question-answering quality; the temporal consistency reward ensures that the selected evidence is logically consistent in time sequence; the time slot coverage reward encourages the model to complete the information for each time period and improves temporal integrity; and the evidence quantity penalty is used to constrain the behavior of selecting too much redundant evidence, avoiding complex content and wasted resources. The entire composite reward system takes into account the validity of evidence, the accuracy of the answer, the temporal compliance, and the conciseness of the content. After completing an evidence selection trajectory, the corresponding reward value is calculated based on the composite reward function to comprehensively evaluate the quality of the evidence organization result. Based on the reward value, the strategy network used to select target candidate evidence round by round is updated through reinforcement learning, so that the strategy network gradually learns better evidence selection strategies, improving the decision quality and question-answering effect in the subsequent evidence organization process.

[0104] Specifically, the composite reward function is calculated using the following formula: ; Among them, R term For the compound reward function, R ret As a reward for evidence recall, R qa As a reward for question-answering accuracy, R time For time consistency rewards, R cov The reward is for time slot coverage, and the penalty is for the amount of evidence. λ, 1-λ, η, ξ, and γ are all weighting coefficients. This invention does not impose specific numerical limits on the weighting coefficients, and their values ​​can be flexibly adjusted according to task requirements.

[0105] Optionally, when the composite reward function is sparse, stepwise reward shaping can be introduced for optimization. The stepwise reward function can be calculated using the following formula: ; Among them, F(S)t Let ) be the potential function corresponding to set S, which can be obtained by a weighted combination of evidence recall revenue and time slot coverage revenue, and cost(e) at Let represent the cost value corresponding to the single piece of evidence selected in round t, and β be the penalty coefficient. The single-step reward calculated in each round can be applied to the temporal difference iterative optimization process.

[0106] In some embodiments, it also includes: Supervised training is performed on the prior scoring model used to generate prior scores based on the matching labels between candidate evidence and standard evidence, as well as the time matching labels between candidate evidence and time constraints. The parameters of the prior scoring model are fixed, and the policy network used for selecting target candidate evidence round by round is trained by reinforcement learning based on the composite reward function; The prior scoring model and policy network are jointly fine-tuned based on a preset learning rate, which is smaller than the learning rate in the supervised pre-training stage.

[0107] In this embodiment, a phased training approach is used to optimize the prior scoring model and the policy network. In the first phase, the prior scoring model is trained under supervision using matching labels between candidate evidence and standard evidence, as well as time matching labels between candidate evidence and time constraints. This allows the prior scoring model to learn the semantic relevance and time matching relationships of candidate evidence, thereby generating accurate prior scores. In the second phase, the parameters of the prior scoring model are fixed, and the policy network is trained using a composite reward function. This allows the policy network to learn the impact of different evidence selection behaviors on the final question-answering performance, gradually obtaining a better evidence organization strategy. In the third phase, the prior scoring model and the policy network are jointly fine-tuned, enabling them to further synergistically optimize, thus balancing the accuracy of prior scores and the ability to organize evidence, improving the overall timeliness of question answering performance.

[0108] The joint fine-tuning phase uses a learning rate lower than that of the supervised training phase, which helps to achieve stable optimization of model parameters while retaining the knowledge learned in the pre-training phase, thus avoiding large fluctuations in the training process.

[0109] According to the present invention, candidate evidence is scored and fused through semantic and temporal channels, enabling the candidate evidence selection process to simultaneously consider semantic relevance and temporal consistency. Target candidate evidence is selected round by round based on the time slot coverage status and semantic summary status, allowing the evidence organization process to dynamically adjust in conjunction with the overall status of the currently selected evidence. This reduces temporal conflicts and invalid redundancy among candidate evidence, thereby improving the rationality of evidence organization in time-sensitive question-and-answer scenarios and enhancing the temporal consistency and accuracy of the question-and-answer results. Furthermore, by constructing the time slot coverage status and calculating the overall time coverage rate, the system can characterize the time coverage of the currently selected evidence set. By generating semantic summaries based on mean pooling and attention pooling, and updating the parameters of the evidence selection strategy using a composite reward function, the candidate evidence selection process can balance time coverage, semantic relevance, and evidence quantity control, thereby improving the overall coordination of the evidence selection process and enhancing the ability to organize complex time-sensitive information, achieving dynamic evidence optimization in time-sensitive question-and-answer scenarios.

[0110] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0111] Step S701: Obtain the time-sensitive question and answer text: "As of May 6, 2021, who is the Prime Minister of the United Kingdom?". Perform semantic parsing and time entity extraction on this text to obtain the corresponding time constraint condition of May 6, 2021, which requires that the selected evidence must cover this time point.

[0112] Step S702: Based on the above question-and-answer text, four candidate pieces of evidence are retrieved from the knowledge base. The specific information is as follows: Evidence 1: "Boris's term of office is from July 2019 to September 2022", corresponding to the time range of [July 2019, September 2022]. Evidence 2: "Sunak took office in October 2022", corresponding to the time interval [October 2022, +∞); Evidence 3: "Boris in 2021", corresponding to the time range [2021-01, 2021-12]; Evidence 4: Other statements that are irrelevant to the semantics of the question and lack a clearly defined valid time interval.

[0113] In step S703, evidence 1 and 3 receive higher prior scores because they are semantically related and their time covers the query node; evidence 2 receives lower prior scores because its time is not matched and evidence 4 is semantically irrelevant.

[0114] Step S704: Based on the prior scores, construct a candidate evidence sequence [1, 3, 2, 4] from high to low, and then conduct round-by-round evidence selection: Round 1: The initial set of selected evidence is empty, the time slot coverage status is uncovered, and the semantic summary status is empty. Calculate the scores for each candidate action based on the current status, and select evidence 1 with the highest prior score to add to the selected set; simultaneously update the semantic summary status and the time slot coverage status, at which point the time slot coverage rate is 1.

[0115] Round 2: The semantic summary status is updated based on the selected evidence 1, and the time slot coverage status shows the covered query time nodes. However, the policy network makes a comprehensive judgment based on the evidence recall reward, question answering accuracy reward, and evidence quantity penalty. It determines that continuing to select evidence 3 can improve evidence coverage and answer confidence, and the gain is higher than the cost of adding new evidence. Therefore, evidence 3, which is complementary to the content of evidence 1, is selected and added to the selected set.

[0116] Round 3: Generate the final semantic summary based on the selected evidence 1 and 3. Both the time slot coverage state and the evidence coverage state meet the requirements. If the reward gain for continuing to select candidate evidence is lower than the cost penalty, the policy network chooses to stop and outputs the set of selected evidence.

[0117] In step S705, the policy network makes a comprehensive judgment based on the time slot coverage status, evidence coverage status, and evidence quantity penalty, and then selects a stop action. The final set of selected evidence is output as {1, 3}, and the question-answering result is generated based on this set: "As of May 6, 2021, the Prime Minister of the United Kingdom is Boris Johnson".

[0118] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0119] Based on the same inventive concept, embodiments of this application also provide a temporally consistent evidence organization system for implementing the above-described time-sensitive question-and-answer approach. For example... Figure 4 As shown, the system includes an acquisition module S601, a processing module S602, a judgment module S603, and an output module S604. The acquisition module S601 is used to acquire timely question and answer texts and acquire a set of candidate evidence based on the timely question and answer texts. The processing module S602 is used to parse the time constraints in the timely question-and-answer text; calculate the semantic score and time score corresponding to the candidate evidence based on the semantic channel and the time channel respectively, and dynamically fuse the semantic score and time score based on the time constraints to obtain the prior score corresponding to each candidate evidence; construct the action scoring reference corresponding to the candidate evidence based on the prior score, and model the evidence selection process as a sequential decision-making process; in each round of candidate evidence selection, construct the current state based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget, calculate the action probability corresponding to the candidate selection action and the stop action through the policy network, and select the target candidate evidence to be added to the set of selected evidence based on the action probability, or select the stop action based on the action probability; The judgment module S603 is used to determine whether the policy network has selected the stop action, or whether the current remaining budget is less than the preset budget threshold, or whether the current selection round has reached the preset maximum selection round, and generates the corresponding question and answer results; The output module S604 is used to output the selected evidence set and generate the corresponding question and answer results.

[0120] Other features of this system are consistent with the aforementioned time-consistent evidence organization method for time-oriented question answering, and will not be repeated here.

[0121] The modules in the aforementioned regional decomposition system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0122] In one embodiment, Figure 5 A schematic structural diagram of a computer device according to an embodiment of the present invention is shown, such as... Figure 5 As shown, the computer device can be a terminal, and its internal structure diagram can be as follows: Figure 5As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a time-consistent evidence organization method for timely question answering. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0123] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0124] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described time-consistent evidence organization method for time-sensitive question answering.

[0125] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described time-consistent evidence organization method for time-sensitive question answering.

[0126] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0127] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0128] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0129] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for organizing time-consistent evidence for time-sensitive question-and-answer purposes, characterized in that, include: Obtain the timely Q&A text and parse the time constraints in the timely Q&A text; A set of candidate evidence is obtained based on the time-sensitive question and answer text; Semantic scores and temporal scores corresponding to candidate evidence are calculated based on semantic channels and temporal channels respectively, and the semantic scores and temporal scores are dynamically fused based on the temporal constraints to obtain the prior scores corresponding to each candidate evidence. Based on the prior scores, action scoring references corresponding to candidate evidence are constructed, and the evidence selection process is modeled as a sequential decision-making process. In each round of candidate evidence selection, the current state is constructed based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget. The action probabilities corresponding to the candidate selection action and the stop action are calculated through the policy network. The target candidate evidence is selected to be added to the set of selected evidence based on the action probabilities, or the stop action is selected based on the action probabilities. When the policy network chooses to stop the action, or the current remaining budget is less than the preset budget threshold, or the current selection rounds reach the preset maximum selection rounds, the selected evidence set is output and the corresponding question and answer results are generated.

2. The method according to claim 1, characterized in that, The process of parsing the time constraints in the time-sensitive question-and-answer text includes: Extract the main content constraints and time constraints from the time-sensitive Q&A text; The time constraints are normalized to construct time cards, which include time relationship types, time anchors, and time granularity. The time relationship types are used to characterize the relationship between the time constraints and the time anchors.

3. The method according to claim 1, characterized in that, Semantic and temporal scores for candidate evidence are calculated based on semantic and temporal channels, respectively, and the semantic and temporal scores are dynamically fused based on the temporal constraints, including: Construct semantic feature vectors and temporal feature vectors corresponding to candidate evidence; Calculate the time interaction features based on the time constraints and the time intervals of the candidate evidence; A semantic score is calculated based on the semantic feature vector, and a time score is calculated based on the time feature vector and the time interaction feature. Based on the degree of explicitness of the time constraints in the time-sensitive question-and-answer text, a dynamic fusion weight is determined to dynamically fuse the semantic score and the time score.

4. The method according to claim 3, characterized in that, Also includes: A loss function is constructed based on the prior score, and the loss function includes ranking loss and time loss; The ranking loss is constructed based on the matching labels of candidate evidence and standard evidence, and is used to optimize the ranking effect of candidate evidence; the time loss is constructed based on the time matching labels of candidate evidence and time constraints, and is used to optimize the accuracy of time scoring. The semantic channel and the temporal channel are optimized based on the ranking loss and the time loss.

5. The method according to claim 1, characterized in that, In each round of candidate evidence selection, the current state is constructed based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget, including: The current information digest is calculated based on the currently selected set of evidence to represent the information content that has been acquired. The time slot coverage status is calculated based on the time constraints to characterize the coverage of the time constraints by the currently selected evidence set.

6. The method according to claim 5, characterized in that, The step of calculating the action probabilities corresponding to candidate selection actions and stop actions through a policy network, and selecting target candidate evidence to add to the selected evidence set based on the action probabilities, or selecting a stop action based on the action probabilities, includes: For each unselected candidate evidence, a candidate action representation is constructed, which includes candidate evidence encoding features, the prior score, the semantic summary state, the time slot coverage state, and the remaining budget information; Construct a current action space, which includes stop actions and candidate selection actions corresponding to each unselected candidate piece of evidence; The action scores of the stop action and each candidate selection action are calculated based on the policy network, and illegal actions corresponding to the selected candidates, candidates exceeding the budget, or candidates with index abnormalities are masked. The action scores after masking are normalized to obtain action probabilities, and candidate actions or stop actions are selected based on the action probabilities. When a candidate action is selected, the corresponding candidate evidence is added to the set of selected evidence and the current state set is updated. When a stop action is selected, the current set of selected evidence is output.

7. The method according to claim 6, characterized in that, Before calculating the action probabilities corresponding to candidate selection actions and stop actions through the policy network, the following steps are included: By using an inter-candidate interaction encoder, the candidate action representations corresponding to the unselected candidate evidence are interactively encoded to obtain interactively enhanced candidate action representations. The action score of the corresponding candidate action is calculated based on the enhanced candidate action representation.

8. The method according to claim 5, characterized in that, Calculating the time slot coverage state based on the aforementioned time constraints includes: Construct a corresponding set of time slots based on the aforementioned time constraints; Calculate the time slot coverage vector based on the coverage of each time slot according to the candidate evidence; The overall time coverage rate is calculated based on the time slot coverage vector to obtain the time slot coverage status.

9. The method according to claim 5, characterized in that, The current information digest is calculated based on the currently selected set of evidence, including: The selected evidence set is aggregated using the mean pooling method to obtain the first semantic summary; The selected evidence set is aggregated using attention pooling to obtain a second semantic summary.

10. The method according to claim 1, characterized in that, Also includes: Construct a composite reward function, which includes evidence recall reward, question answering accuracy reward, time consistency reward, time slot coverage reward, and evidence quantity penalty item; The reward value corresponding to the evidence selection trajectory is calculated based on the composite reward function, and the policy network used to select target candidate evidence round by round is updated by reinforcement learning based on the reward value.

11. The method according to any one of claims 1-10, characterized in that, Also includes: Based on the matching labels between candidate evidence and standard evidence, and the time matching labels between candidate evidence and time constraints, the prior scoring model used to generate the prior score is trained under supervision. The parameters of the prior scoring model are fixed, and the policy network used for selecting target candidate evidence round by round is trained by reinforcement learning based on the composite reward function; The prior scoring model and the policy network are jointly fine-tuned based on a preset learning rate, which is less than the learning rate in the supervised pre-training stage.

12. A temporally consistent evidence organization system oriented towards time-sensitive question answering, characterized in that, include: The acquisition module is used to acquire timely Q&A text; A set of candidate evidence is obtained based on the time-sensitive question and answer text; The processing module is used to parse the time constraints in the time-sensitive question and answer text; Semantic scores and temporal scores corresponding to candidate evidence are calculated based on semantic channels and temporal channels respectively, and the semantic scores and temporal scores are dynamically fused based on the temporal constraints to obtain the prior scores corresponding to each candidate evidence. Based on the prior scores, action scoring references corresponding to candidate evidence are constructed, and the evidence selection process is modeled as a sequential decision-making process. In each round of candidate evidence selection, the current state is constructed based on the current set of selected evidence, the set of unselected candidates, the time slot coverage status, the semantic summary status, and the remaining budget. The action probabilities corresponding to the candidate selection action and the stop action are calculated through the policy network. The target candidate evidence is selected to be added to the set of selected evidence based on the action probabilities, or the stop action is selected based on the action probabilities. The judgment module is used to determine whether the policy network chooses to stop, or whether the current remaining budget is less than a preset budget threshold, or whether the current selection rounds have reached a preset maximum selection rounds, and generates corresponding question and answer results; The output module is used to output the selected evidence set and generate corresponding question and answer results.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the time-consistent evidence organization method for time-sensitive question answering as described in any one of claims 1-11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the time-consistent evidence organization method for time-based question answering as described in any one of claims 1-11.

Images