Method and device for processing semantic text and execution action based on limited hand gesture action repository

Abstract

The present application relates to the field of natural language resource scheduling processing, and more particularly to a semantic text and execution action processing method and device based on a limited sign language action resource library, the method comprising: obtaining an original text, a text structure result with an available state, a structured rewriting candidate, and an action capability set; extracting a protected semantic slot with an original text character position; retaining an original text branch and forming a candidate branch; directly selecting the original text branch or a conservative branch when the analysis state does not belong to an allowed state; performing semantic constraint judgment; performing resource constraint judgment, resolving an accurate action identifier in an action alias order and verifying that all actions are executable; jointly determining according to the results of the semantic constraint judgment and the resource constraint judgment, selecting the candidate branch or reverting to the original text branch or the conservative branch; outputting an action plan unit bound with an original text span, a resource state, and an acceptance track when accepting, and outputting a conservative result and a rejection reason when rejecting, thereby reducing structured risk and realizing verifiable and traceable rewriting.

Description

Technical Field

[0001] This invention relates to the field of natural language resource scheduling and processing, and in particular to a method and apparatus for processing semantic text and execution actions based on a limited sign language action resource library. Background Technology

[0002] Limited sign language action resource libraries typically cannot cover all surface lexical forms of the input text. Processing solely by string matching or general synonym replacement may result in the silent loss of key semantic information in the original text when candidates cover numbers, times, entities, negation, or modal information. Furthermore, the same surface word can represent different concepts in different parts of speech or contexts, potentially leading to the acceptance of incorrect actions. Candidates may be semantically valid, but their target actions may be absent from the current action capability set, ambiguous, or isolated. When a candidate requests partially executable actions while others are not, existing technologies often output the executable portion, resulting in incomplete results. In actual processing, when the structural analysis component is unavailable, the system continues to rewrite using residual lexical units or old caches, treating "cannot be determined" errors as "rewrite allowed." Moreover, rewriting before verification overwrites the original text, making it impossible to definitively recover the original branch after a failure. The output only contains the action name without the original text span, resource status, or decision reason, making it difficult to verify why a particular request changed the output. Candidates that lack sufficient independent evidence, are not in a production licensed state, or pose a risk of being blocked are adopted, causing candidates that have not yet met the conditions for use to change their action output.

[0003] Therefore, a processing method is needed that binds the text structure state, protected semantics, candidate applicable conditions, and current capabilities of action resources in the same request, so that every change in output is traceable and that the original branch or a predefined conservative result is stably maintained when any necessary condition is not met. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a semantic text and execution action processing method and apparatus based on a limited sign language action resource library, in order to overcome structural error acceptance caused by problems such as loss of key semantic information, misuse of context, unexecutable actions, and inability to track failures in existing technologies.

[0005] To achieve the above objectives, this invention provides a semantic text and execution action processing method based on a limited sign language action resource library, comprising: Retrieve the original text, the text structure result with available states, one or more structured rewrite candidates, and the set of action capabilities requested for the current request; Based on the text structure results, at least one protected semantic slot with the original character position is extracted to determine the set of protected semantic slots; Retain the original text branch corresponding to the original text, and form candidate branches based on the structured rewrite candidates; When the analysis state of the text structure result is not in an allowed state, record the analysis unavailable state, do not call the candidate action parsing, and directly select the original text branch or output the predefined conservative branch. The candidate branch is subjected to semantic constraint judgment, which includes: verifying whether the candidate branch is in a predetermined production license state and there is no risk of blocking; verifying the independence of the number, source or content identity of evidence according to preset evidence rules; verifying boundary conditions and part-of-speech conditions according to the original text span covered by the candidate branch; verifying context conditions; and verifying that the candidate branch does not cause the semantic loss of the protected semantic slot. Based on the set of action capabilities, resource constraint judgment is performed on the action alias sequence requested by the candidate branch. The resource constraint judgment includes: parsing the precise action identifier according to the order of the action alias sequence and according to the preset deterministic rules, and verifying that all actions requested by the action alias sequence have a parsing result that is allowed to be executed; when any action alias cannot determine the target action, or the parsing result does not belong to the set of allowed execution states, the candidate branch is rejected. A joint determination is made based on the results of the semantic constraint judgment and the resource constraint judgment to determine the branch selection result. The candidate branch is selected only when both the semantic constraint judgment and the resource constraint judgment pass; otherwise, the original text branch or the predefined conservative branch is selected. Based on the branch selection result, when the candidate branch is accepted, an action plan unit containing the precise action identifier, original text span, resource status, and acceptance trajectory is output; when the candidate branch is rejected, the action plan unit of the candidate branch is not output, but a structured conservative result, rejection reason, and non-execution status are output.

[0006] Preferably, the text structure result includes at least an analysis state and lexical units with character positions; the analysis state includes a successful analysis state, a partially usable state, and an unusable state. When the analysis state of the text structure result is a partially usable state, the required structural fields of the candidate branch are determined according to the applicable conditions of the candidate branch, and if any of the required structural fields is missing or unreliable, the candidate branch is prevented from changing its output in this request. The required structural fields are determined by the candidate rules: all candidates require lexical units, character start and end positions, and sentence boundaries; candidates with limited parts of speech must cover the parts of speech of the candidate span; candidates involving entities, numbers, or time must have corresponding entity, numerical, or time spans and reliability indicators; candidates involving negation or modality must have negation or modality markers and their spans; candidates requiring predicates, arguments, or dependency contexts must have corresponding predicates, semantic roles, or dependency fields; candidates with necessary context or blocking context must have bounded text windows that do not cross sentence boundaries and window reliability.

[0007] Preferably, extracting the protected semantic slots includes: extracting at least one of entity, number, time, negation, and modality information from the text structure result as protected semantic slots with original character positions. Each protected semantic slot includes at least slot type, surface text, original character span, and recognition basis. For overlapping recognition results, deduplication is performed according to a preset deterministic priority rule to form a set of slots without repetition; the deterministic priority rule includes: when the same time expression is simultaneously recognized as an entity and time, it is recorded first by the time slot; numerical entities are recorded first by the number slot. Verifying that the candidate branch does not cause semantic loss of the protected semantic slots includes: when the candidate branch intersects with a hard protected slot and does not retain the surface information of the hard protected slot, it is allowed to continue only when there is a predetermined equivalence preservation relationship; the hard protected slots include at least one of entity slots, number slots, time slots, negation slots, and modality slots; wherein, entities, negations, and modality must not be rewritten solely based on candidate self-declaration, and numbers and time can be declared preserved when there is independent evidence and a clear equivalence relationship.

[0008] Preferably, verifying the context conditions includes: within a bounded context constrained by sentence boundaries, verifying the necessary context conditions and blocking context conditions of the candidate branch. Ordinary context is determined within a bounded window after removing the target surface word; when the blocking item is a longer compound word containing the target surface word, the blocking context is determined within a window containing the target surface word, and the bounded window does not cross sentence boundaries.

[0009] Preferably, the structured rewriting candidate is an information set containing the following fields: original surface terms and original text span, standard expression, applicable parts of speech, necessary context, blocking context, action alias sequence, concept identifier, anonymous evidence summary, status and risk marker, and slot retention declaration. There is no mandatory processing order for the fields; only the internal order of the action alias sequence must be strictly maintained. In the resource constraint judgment, when any action alias in the action alias sequence cannot determine the target action, or when the current resource state of any action among all actions does not belong to the set of allowed execution states, the candidate branch is rejected and the output of some actions of the candidate branch is prohibited.

[0010] Preferably, the resource constraint judgment and the assembly of the action planning unit use the same action capability set snapshot identifier.

[0011] Preferably, when multiple candidate branches cover the same original text span, the highest candidate and its difference from the second highest candidate are determined based on preset semantically relevant ranking factors. These ranking factors include, but are not limited to: prioritizing non-overlapping spans, prioritizing longer spans, production license status, lower risk, higher evidence level, and stable record number. When the highest candidate is not unique or the difference does not reach a preset threshold, all candidate branches corresponding to that original text span are rejected, and the original text branch or the predefined conservative branch is selected; wherein, the stable record number is only used for fixed trajectory arrangement and should not be used alone as a basis for semantic disambiguation.

[0012] Preferably, the predefined conservative branch includes at least one of the following: maintaining the original text and marking it as unparsed, outputting a pre-configured word-by-word plan or finger spelling plan, returning to an unexecutable state, and maintaining the previously confirmed output, wherein the finger spelling plan is based on Chinese finger letters; the acceptance trajectory includes at least the selected branch, the reason for acceptance or rejection, the precise action identifier, the original text span, and the current resource status; and the acceptance trajectory records the same request binding relationship of the following information: the original text span covered by the candidate branch, the snapshot identifier of the action capability set on which the resource constraint judgment is based, the results of the semantic constraint judgment and the resource constraint judgment, and the reason for finally selecting or rejecting the candidate branch.

[0013] Preferably, local non-manual prompts or sentence-level non-manual prompts corresponding to the original text span are generated. The non-manual prompts referred to in this application are non-manual feature prompt information used to convey non-hand movement information such as question type and degree information to the target execution end, and may include types such as facial expressions and torso posture; this application does not limit its specific type set. The local non-manual prompt includes type, intensity, original text span, and target action index, and is bound to the corresponding action planning unit; the sentence-level non-manual prompt is not bound to a single action. The types of the non-manual prompts include degree, emphasis, comparison, specific question, or non-question, and the intensity is a discrete level; degree information is bound to the action it modifies, and sentence-level question information acts on the entire sentence; a explicitly unsupported state is output for parallel capabilities not supported by the target execution end, the parallel capabilities including spatial location, role consistency, and class labeling channels.

[0014] Preferably, the action capability set is a view of existing action resources that can be queried in this request. Each action includes a precise action identifier and the current resource status, which includes, but is not limited to, executable, isolated, missing, conflicting, or unresolved. The action capability set describes the availability status of the action resource, rather than the movement capability of the execution end joint. The set of allowed execution states is predetermined by the deployment configuration, and only states that have been verified as playable are allowed to pass; missing, multiple precise action conflicts, isolation, service unavailable, and unverified are considered disallowed by default.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves verifiable and traceable structured text rewriting decisions through a dual-branch structure of original text and candidates, extraction and preservation verification of protected semantic slots with original text positions, ordered parsing of action aliases and atomic resource verification, and joint determination of semantic constraint judgment and resource constraint judgment, which significantly reduces the risk of accepting structural errors under limited sign language action resources. Furthermore, this invention reduces the risk of silent loss of entity, number, time, negation, or modal information during candidate rewriting by using protected semantic slots with original text positions and slot-preserving relationships; Furthermore, the present invention, through the original text / candidate dual branch and analysis state gate, can deterministically maintain the original result even when the analysis component is unavailable or the candidate fails; Furthermore, this invention uses resource constraint judgment to prevent missing, conflicting, isolated, unverified, or unexecutable actions from entering the output as partial plans; Furthermore, by binding the original text span, resource status, and acceptance trajectory of the same request, the present invention makes the rewriting decision replayable and verifiable. Attached Figure Description

[0016] Figure 1This is a flowchart of a semantic text and execution action processing method based on a limited sign language action resource library, according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0019] In this embodiment, semantic constraint judgments are performed on the text structure results and structured candidates; resource constraint judgments are performed based on the action aliases requested by the structured candidates and the snapshot of the action capability set requested in this instance. The original text branch, the semantic constraint judgment result, and the resource constraint judgment result all enter the branch selector. The candidate branch is adopted only when both the semantic constraint judgment and the resource constraint judgment pass. The branch selection result, along with the action identifiers and resource states already verified in the same capability snapshot, enters the plan assembly to avoid using contradictory resource states between various judgments and the plan output. Non-manual prompts are optional plan enrichment information, extracted from the original text or normalized text and mapped back to the original text span, and are not necessary conditions for semantic constraint judgments and resource constraint judgments.

[0020] Please see Figure 1 The diagram shows a flowchart of a semantic text and execution action processing method based on a limited sign language action resource library according to an embodiment of the present invention. This embodiment of the semantic text and execution action processing method based on a limited sign language action resource library includes: Step S1: Obtain the original text, text structure results, one or more structured rewriting candidates, and the current action capability set. Specifically, the action capability set should reflect the current request point in time, rather than just using the old state when the candidates were created. In this embodiment, the text structure results include at least the analysis state and lexical units with character positions. The analysis state includes successful analysis, partially available, and unavailable states. The action capability set is a view of existing action resources that can be queried for this request. Each action contains a precise action identifier and the current resource state, which includes executable, isolated, missing, conflicted, or unresolved.

[0021] Step S2: Extract protected semantic slots. Extract entity, number, time, negation, and modal slots with original text spans from the structural results. In this embodiment, protected semantic slots are original text information units that cannot be silently deleted or altered by the candidate. Each slot includes at least slot type, surface text, original text character span, and identification criteria. Overlapping identification results are deduplicated; for example, when the same time expression is simultaneously identified as both an entity and a time, it is preferentially recorded as a time slot; numerical entities are preferentially recorded as number slots.

[0022] Step S3: Establish a dual branch consisting of the original text and candidate branches. Save the original text and its structural trajectory as the original text branch. For each matching position of a candidate surface word, calculate the candidate span and form an independent candidate branch. The original text branch is not overwritten until any candidate has completed semantic constraint and resource constraint judgments.

[0023] Step S4: Execute the analysis state gate. If the analysis state is not an allowed state, record that the analysis is unavailable, do not call the candidate action parsing, and directly maintain the original branch or output a predefined conservative state. If the analysis state is partially available, in the preferred embodiment, further calculate the set of required fields for this candidate. Only when all required fields exist and are marked as reliable is it allowed to continue; otherwise, it is rejected by default and subsequent processing is stopped (i.e., fail-closed).

[0024] Step S5 involves performing boundary, part-of-speech, and context checks. This involves determining whether the candidate span is a complete word or phrase boundary, whether the part-of-speech of the covered word belongs to the allowed candidate set, and whether the necessary and blocking contexts are satisfied. For ordinary context checks, it is preferable to perform the checks within a bounded window after removing the target surface word to avoid candidates satisfying the conditions on their own. When the blocking term is a longer compound word containing the target surface word, the checks are performed within a window containing the target word to prevent the target word from being incorrectly rewritten as a compound word fragment. The window should preferably not cross sentence boundaries.

[0025] Step S6: Perform evidence and protected slot determination. Verify that the candidate is in a production license state, the number of anonymous evidences reaches the threshold, the evidence identity meets the independence requirements, and there is no risk of blocking. In one implementation, each evidence record stores an anonymous source identifier and a content summary; records with the same anonymous source identifier or duplicate content summaries are counted as only one piece of evidence. The threshold for the number of independent evidences can be provided by configuration, for example, at least two. Obtain the protected slots that intersect with the candidate span. If the slot surface text still appears in the specification expression, it is considered directly retained; if it does not appear directly, there must be a pre-approved equivalence retention statement. Entities, negations, and modalities must not be rewritten solely based on the candidate's self-declaration in a preferred implementation; numbers and times can be declared retained when there is independent evidence and a clear equivalence relationship.

[0026] Step S7: Parse the precise actions and verify the resource status. Parse the precise action identifiers one by one in the original order of the candidate action aliases. If any action alias cannot determine the target action according to the preset deterministic rules, the target action is missing, or the parsing result does not belong to the set of allowed execution states, it is considered that the action parsing is incomplete. The number of requested actions in the candidates must be equal to the number of successfully parsed actions; outputting only part of the actions is prohibited.

[0027] Step S8: Select a branch and perform a deterministic backtracking. A candidate branch is used only when the analysis status, boundary, part-of-speech tag, context, evidence, protected slot, and all action resources pass simultaneously. If any necessary condition fails, the original branch or a predefined conservative branch is selected, and a stability reason code is recorded. When multiple candidates exist simultaneously, a preferred implementation can sort them sequentially by non-overlapping span, longer span priority, production license status, lower risk, higher evidence level, and stability record number; if the highest candidate cannot be determined for the same span, that span is rejected. Semantic constraint judgment and resource constraint judgment work together on the branch selector; if either judgment fails, a backtracking to the original branch or a conservative branch occurs.

[0028] Step S9: Output a planned or conservative result. When a candidate is accepted, an executable action plan is output. Each plan unit includes at least a sequence number, action alias, precise action identifier, corresponding original text span, and current resource status, along with an acceptance trajectory. In one implementation, an action variant identifier, normalized span, resource identifier set, and local non-manual prompts are also output. When a candidate is rejected, one of the following conservative results is output based on the deployment configuration: retain the original text and mark it as unparsed, use a pre-configured word-by-word or finger spelling plan, return to an unexecutable state, or retain the previously confirmed output, wherein the finger spelling plan is based on the Chinese finger alphabet scheme specified in the Ministry of Education's "Chinese Finger Alphabet Scheme" (GF 0021-2019). This invention does not require that original text without resources necessarily be converted into executable actions; its essential technical effect is to prevent candidates from erroneously changing the output when they are rejected.

[0029] In this embodiment, the text structure results involved include at least the analysis status and lexical units with character positions. The analysis status includes successful analysis, partially available, and unavailable states. The action capability set is a view of existing action resources that can be queried for this request. Each action contains a precise action identifier and the current resource status, which includes executable, isolated, missing, conflicting, or unresolved. The action capability set should reflect the current request point in time, rather than just using the old status when the candidate was created. In this embodiment, the text structure results may also include parts of speech, sentence boundaries, entities, time, negation, modality, numbers, and semantic roles. Character positions use a zero-starting, left-closed, right-open interval [start, end].

[0030] In this embodiment, a precise action identifier can consist of a basic action identifier and optional meaning identifiers, variant identifiers, direction identifiers, regional identifiers, or version identifiers. Different meanings of the same basic action or variants from different dictionary sources can be distinguished by sub-identifiers. Sub-identifiers participate in the uniqueness determination of action parsing, and this method does not limit the specific representation of the sub-identifiers. For example, when the same basic action 'time' has different meanings in different dictionary sources, it can be recorded as action ability items with different sub-identifiers.

[0031] In this embodiment, the snapshot of the action capability set can be an entity copy, version reference, transaction view, or query token bound to this request; the resource status includes, but is not limited to, executable, isolated, missing, conflicted, or unresolved, and the specific status set and reason code are predetermined by the deployment configuration.

[0032] In this embodiment, the structured rewrite candidates, action capability items and their status, and evidence summary can be provided by a data source that has been established before the request; the data source can be updated, and subsequent requests read new capability views. This method does not limit the way the data source is updated.

[0033] In this embodiment, if the analysis state is not an allowed state, the analysis is recorded as unavailable, the candidate action is not invoked for parsing, and the original branch or a predefined conservative state is directly maintained. If the analysis state is partially available, in the preferred embodiment, the required field set of this candidate is further calculated. Only when all required fields exist and are marked as reliable is further allowed to continue; otherwise, it is also rejected by default and subsequent processing is stopped.

[0034] In one implementation, the set of action capabilities can be provided separately for different target execution ends, and this method does not limit the number, type or capability configuration of the target execution ends.

[0035] In this embodiment, the rules for determining the required fields of the candidates are as follows: all candidates must have lexical units, character start and end positions, and sentence boundaries; candidates with limited parts of speech must cover the parts of speech of the candidate span; candidates involving entities, numbers, or time must have the corresponding entity, numerical value, or time span and reliability flags; candidates involving negation or modality must have negation or modality flags and their spans; candidates requiring predicates, arguments, or dependency contexts must have corresponding predicates, semantic roles, or dependency fields; candidates with necessary context or blocking contexts must have bounded text windows that do not cross sentence boundaries and window reliability. For example, if a candidate only requires lexical units, parts of speech, and time span, but the analysis results lack dependency tasks unrelated to the candidate, the missing tasks can be recorded in the trajectory and the process can continue; if the time identification field is missing, the candidate cannot be accepted on the grounds of "no time slot".

[0036] In this embodiment, entity, number, time, negation, and modal slots with original text spans are extracted from the structural results. In this embodiment, protected semantic slots are original text information units that cannot be deleted or altered by candidate silent methods. Each slot includes at least a slot type, surface text, original text character span, and identification criteria. Overlapping identification results are deduplicated; for example, when the same time expression is simultaneously identified as both an entity and a time, it is preferentially recorded as a time slot; numerical entities are preferentially recorded as number slots. In one implementation, necessary predicates can also be recorded as tracking slots.

[0037] In this embodiment, the acceptance trajectory includes at least the selected branch, the reason for acceptance or rejection, the precise action identifier, the original text span, and the current resource status. The acceptance trajectory records the following information regarding the binding relationship of the same request: the original text span covered by the candidate branch, the snapshot identifier of the action capability set on which the resource constraint judgment is based, the results of the semantic constraint judgment and the resource constraint judgment, and the reason for finally selecting or rejecting the candidate branch. The acceptance trajectory also records the circumstances under which the candidate branch is rejected and the reasons therefor.

[0038] In this embodiment, the structured rewriting candidate is an information set containing the following fields: original surface terms and original text span, standard expression, applicable part of speech, necessary context, blocking context, action alias sequence, concept identifier, anonymous evidence summary, status and risk marker, slot preservation statement. There is no mandatory processing order for each field, only the internal order of the action alias sequence must be strictly maintained.

[0039] In this embodiment, it is determined whether the candidate span is a complete word or phrase boundary, whether the part of speech of the covered word belongs to the allowed candidate set, and whether the necessary context and blocking context are satisfied. Ordinary context is preferably determined within a bounded window after removing the target surface word to avoid candidates satisfying the conditions on their own. When the blocking item is a longer compound word containing the target surface word, it is determined within a window containing the target word to prevent the target word from being incorrectly rewritten as a compound word fragment. The window preferably does not cross sentence boundaries.

[0040] In this embodiment, the resource constraint judgment and action planning unit assembly use the same action capability set snapshot identifier. The action capability set snapshot includes at least a snapshot identifier, generation time or version, precise action identifier, number of alias resolution results, executable flag, resource status, resource reference or summary, and status source. When a snapshot version change is detected before assembly, the resource constraint judgment is re-executed based on the new action capability set snapshot, or a predefined conservative branch is selected. The set of allowed execution states for resource constraint judgment is predetermined by the deployment configuration; for example, only states verified as playable are allowed. Missing states, multiple conflicting precise actions, isolation, service unavailability, and unverified states are considered disallowed by default. In one implementation, candidate resolution filters resolution results according to allowed execution states and selects the target action according to preset deterministic rules; unverified or disallowed resources cannot enter the secure release result through the final release gate.

[0041] In this embodiment, when multiple candidate conflict handling occurs, and candidate spans overlap or multiple candidates exist within the same span, the following deterministic rules are preferably adopted: First, perform the checks for semantic constraint judgment and resource constraint judgment separately; then, form overlapping groups according to the starting point from smallest to largest; within each group, prioritize longer spans; for the same span, compare the production license status level, risk number, independent evidence level, and pre-registered business priority. The conflict is only adopted when the highest candidate is unique and the confidence or rule difference with the second-highest candidate reaches a preset minimum value; otherwise, reject the entire conflict group and maintain the original text or conservative branch. Stable record numbers are only used for fixed trajectory arrangement and should not be used to forcibly select semantically indistinguishable candidates as production results.

[0042] This embodiment also includes an optional plan enrichment step: generating local non-manually controlled prompts or sentence-level non-manually controlled prompts corresponding to the original text span. Local non-manually controlled prompts include type, intensity, original text span, and target action index, and are bound to the corresponding action plan unit; sentence-level non-manually controlled prompts are not bound to a single action. The types of non-manually controlled prompts include degree, emphasis, comparison, specific question, or non-question, and the intensity is a discrete level; degree information is bound to the action it modifies, and sentence-level question information applies to the entire sentence; for parallel capabilities not supported by the target execution end, an explicit lack of support status is output, and parallel capabilities include spatial location, role consistency, and class tagging channels.

[0043] This embodiment includes an acquisition module, which is used to acquire the original text, the text structure result with available status, one or more structured rewriting candidates, and the set of action capabilities currently requested.

[0044] In this embodiment, a protected semantic slot extraction module is included, which is used to extract at least one protected semantic slot with the original character position based on the text structure result, so as to determine the protected semantic slot set.

[0045] In this embodiment, a branching module is included, which is used to retain the original text branches corresponding to the original text and to form candidate branches based on the structured rewrite candidates. The branching module saves the original text and its structural trajectory as the original text branches, calculates the candidate span for each matching position of the candidate surface words and forms an independent candidate branch, and does not cover the original text branches before any candidate completes various judgments such as semantic constraint judgment and resource constraint judgment.

[0046] In this embodiment, a state gate module is included, which is used to record the analysis unavailable state when the analysis state of the text structure result is not an allowed state, without calling the candidate action parsing, and directly selecting the original text branch or outputting a predefined conservative branch.

[0047] This embodiment includes a semantic constraint judgment module, which performs semantic constraint judgment on candidate branches. The semantic constraint judgment module verifies whether the candidate branch is in a predetermined production license state and does not pose a blocking risk; verifies the independence of the quantity, source, or content identity of evidence according to preset evidence rules; verifies boundary conditions and part-of-speech conditions based on the original text span covered by the candidate branch; verifies context conditions; and verifies that the candidate branch does not cause semantic loss of the protected semantic slot. The semantic constraint judgment module verifies necessary context conditions and blocking context conditions within a bounded context constrained by sentence boundaries. Ordinary context is judged within a bounded window after removing the target surface word; when the blocking item is a longer compound word containing the target surface word, the blocking context is judged within the window containing the target surface word.

[0048] This embodiment includes a resource constraint judgment module, which performs resource constraint judgment on the action alias sequence requested by the candidate branch based on the action capability set. The resource constraint judgment module parses the precise action identifier according to the order of the action alias sequence and preset deterministic rules, and verifies that all actions requested by the action alias sequence have a parsing result that allows execution. When any action alias in the action alias sequence cannot determine the target action, or when the current resource state of any action in all actions does not belong to the set of allowed execution states, the resource constraint judgment module rejects the candidate branch and prohibits the output of some actions of the candidate branch.

[0049] This embodiment includes a branch selection module, which performs a joint judgment based on the results of semantic constraint judgment and resource constraint judgment to determine the branch selection result. The branch selection module selects a candidate branch only when both semantic constraint judgment and resource constraint judgment pass; otherwise, it selects the original text branch or a predefined conservative branch.

[0050] In this embodiment, an output module is included, which is used to output an action plan unit that is bound to the precise action identifier, the original text span, the resource status and the acceptance trajectory when the candidate branch is accepted, based on the branch selection result; when the candidate branch is rejected, the action plan unit of the candidate branch is not output, but a structured conservative result, the rejection reason and the non-execution status are output.

[0051] In this embodiment, the motion planning unit is an upstream execution description that binds the original text span, motion resource identifier, and resource status. It is not equivalent to joint rotation, posture trajectory, or motion data. The motion planning unit can be converted into motion data by subsequent motion resource parsing, playback, or combination modules. This application does not require skeletal motion redirection, inverse kinematics solving, or target bone configuration as necessary steps.

[0052] In one specific embodiment, the input text "What time is it now?" is used as an example for illustration: The input text is "What time is it now?", the analysis status is successful, and the structure result is as follows: Original text "What time is it now?"; Lexical unit 1 "Now", span [0,2), part of speech NT; Lexical unit 2 "What time", span [2,4), part of speech NT; Time span 1 "Now", span [0,2); Time span 2 "What time", span [2,4); Entity span is empty; Warning is empty. The system obtains two time slots: "Now / [0,2)" and "What time / [2,4)". The candidate only covers [2,4), so "What time" enters the slot of this candidate for judgment, and "Now" is kept in its original position. The original branch stores "what time is it now", and the anonymous candidate record A contains: the original surface word "what time", the candidate span [2,4), the standard expression "how much time", the applicable part of speech NT, the action alias "time, how much", the concept "hour inquiry", the status is the production license status, the anonymous evidence summary is two registered and independent evidence records, and the slot is equivalent to the time slot. Semantic constraint judgment stage: The analysis status is qualified; the candidate span is a complete word unit; the part of speech belongs to the allowed set; there are no blocking conditions in the context; the number and independence of anonymous evidence reach the threshold; the candidate is in the production license status, the candidate covers "several" time slots, and has the reviewed time slot equivalence preservation relationship, so there is no reason for time slot loss; Resource constraint judgment stage: The request action alias "Time" corresponds to the precise action identifier "Time (timing, opportunity, moment)", and its current state is playable and executable; the request action alias "How much" corresponds to the precise action identifier "Number (quantity, amount, how much)", and its current state is playable and executable. Both are determined according to preset deterministic rules and are currently executable. The number of parsed actions and the number of request actions are both 2. Among them, the variant marker is only an example and does not limit the specific identifier system. The system accepts the candidate branch and outputs two sequentially arranged planning units, each bound with the candidate original text span [2,4) and the current resource state; the decision reason is acceptance, the word "now" in the original text remains in its original position, and its subsequent action planning is not included in this candidate action output.

[0053] In the plan enrichment stage, the character "ji" in the original text "what time is it" is located in the original text span [2,3), which belongs to a specific interrogative marker. The system independently forms a sentence-level non-manual prompt from the original text: the type is interrogative, the standard value is specific interrogation, the intensity is medium, the surface text is "ji", the original text span is [2,3), and the target action index is empty. This prompt will not be lost even if the candidate "what time" is rewritten into the action sequence of "time, how many". For the spatial position, role consistency and class marker channels that are not supported by the target execution end, they are explicitly recorded as "not supported" in the planning unit; the above examples are only used to illustrate the processing flow and field relationship, and are not used to evaluate the semantic correctness or naturalness of specific sign language expressions.

[0054] Specifically, the rejection mechanism of this embodiment includes: Scenario 1: Analysis component unavailable: set the analysis state to unavailable, while deliberately retaining the old lexeme and old time span to simulate cache residue after component abnormality, the analysis state gate still records the analysis unavailability first; the number of action parsing calls is 0; the candidate is rejected and the original text or conservative branch is maintained, this scenario illustrates that old fields cannot bypass the analysis state gate; Scenario 2: Protected time slot lost: keep the analysis and action resources normal, but make the candidate neither contain the surface text of the original time slot nor declare that the time slot remains equivalent, the system records that the protected slot is lost: time, even if both actions can be parsed, the entire candidate is rejected; Scenario 3: Action resource isolation: keep all semantic conditions passed, but set the current state of the precise action corresponding to "how many" to isolated, the capability query only obtains one executable action, the number of parses is less than the requested number, the system records that the action is not parsed, clears the candidate actions and prohibits only executing the "time" action; Scenario 4: Compound word and context blocking: when the target surface word appears in a longer compound word, or hits a context that conflicts with the candidate concept, the candidate is rejected in the boundary condition judgment or context condition judgment, and the system will not apply the action just because the surface substrings are the same; Scenario 5: Multiple candidate conflicts: when there are multiple candidates in the same span that have passed all basic judgments, but the stable sorting still cannot determine the highest candidate, the system rejects this span and returns the original text or conservative branch, and must not randomly adopt a candidate, nor rely on a non-deterministic container traversal order; Scenario 6, Some Available Required Fields Missing: Keep the overall analysis status as partially available, while marking the time span field required by the candidate as unavailable. Even if the terms and parts of speech still exist, record the missing required field: time and reject the candidate. If only dependency fields unrelated to the candidate are missing, and the terms, parts of speech, time span, and context window are all reliable, then execution can continue and the missing task can be written into the track. Scenario 7, Capability snapshot changes: When the semantic conditions remain unchanged, snapshot A marks both target actions as playable, and the candidate can enter the plan. Subsequently, the resource status is updated to form snapshot B, in which one action becomes isolated. New requests using snapshot B must reject the entire candidate. If the same request detects that the snapshot has changed from A to B before the plan is assembled, the resource constraint judgment is re-executed or the conservative branch is returned. The plan containing only the first action must not continue to be output. Scenario 8, Candidate Conflict with Same Span: Two candidates cover the same span of the original text, both judged by separate semantic and resource constraints, but with the same production license status, risk, and evidence level, making it impossible to form a unique highest candidate with the minimum score difference. The system only arranges the two diagnostic records by stable record number, does not use stable record number for semantic disambiguation, returns the candidate conflict, and preserves the original text or conservative branch.

[0055] In this embodiment, the above steps can be implemented based on computer program modules. The system may include a text normalization and span mapping module, a structure result acquisition module, a protected slot extraction module, a candidate branch construction module, a semantic constraint judgment module, an action capability parsing module, a branch selection module, a non-manual prompt generation module, and a plan output module. The modules can be deployed on the same electronic device, or they can be implemented through a combination of local interfaces or network services.

[0056] In one implementation, the above steps process the request in the following order: perform deterministic normalization on the input text while preserving the mapping from normalized characters to original characters; obtain analysis results with state, lexical, part-of-speech, entity, time, and other structural fields; extract protected semantic slots to generate structured candidates; call the capability parser on candidates judged by semantic constraints, and the capability parser returns actions that meet the allowed conditions of the capability view for this request according to preset deterministic rules; segment the accepted candidates into actions, and retain the original text or conservative branches for rejected candidates; perform filtering and conservative sequence planning on the action units, while saving the action index mapping before and after processing; generate local or sentence-level non-manual prompts from the text and action units, and map the normalized span back to the original text span; combine each final action with the action capability state, local non-manual prompts, and original text span that have been verified by resource constraint judgment to form a planning unit; output the action sequence, planning unit, sentence-level non-manual prompts, rejection reason, uncertainty, and acceptance trajectory.

[0057] In this embodiment, the fields of the candidate records can be maintained in any order, and there is no mandatory processing order among the fields; what this method requires is the execution relationship between semantic constraint judgment, resource constraint judgment and action alias resolution, rather than the storage or maintenance order of the candidate fields.

[0058] The aforementioned capability resolver has explicit interface constraints: the input consists of sequentially arranged action aliases and a snapshot of the capability set requested in this request; the output returns actions that meet the allowed conditions of the capability view requested in this request according to predefined deterministic rules; if any action alias cannot determine the target action or the resolution result does not belong to the set of allowed execution states, a partial success result must not be returned.

[0059] In one embodiment, the electronic device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the semantic text and action processing method based on a limited sign language action resource library as described in any of the above embodiments.

[0060] In one embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the above-described method. The computer-readable storage medium may be a non-transitory storage medium capable of storing program code.

[0061] Those skilled in the art will understand that the module division and step sequence described in the above specific embodiments are merely preferred examples, and the present invention is not limited thereto. In alternative embodiments, the structural results can be provided by a local analyzer, a remote service, or a combination of multiple analysis components; character spans can be represented by other equivalent positions, as long as they can stably correspond to the original text; action capability sets can come from databases, memory indexes, file indexes, or capability query services; evidence identity can be represented by anonymous numbers, digests, or signatures, and evidence independence verification does not depend on the original source name; context can be represented by character windows, word windows, or intra-sentence dependency ranges, but boundaries should be limited and deterministic; conservative branches can be the original text unparsed state, a word-by-word plan, a finger-spelling plan, or maintaining existing outputs; decision reasons can be represented by enumeration codes, bit tags, or structured objects; the device can be executed centrally by a single processor or by multiple service cooperators. The execution is synchronized; the plan can be returned synchronously or persisted first and then read by the playback component; other stable keys can be used for candidate sorting, but random selection or silent overwriting of the original text is not allowed in case of conflict; non-manual prompts can be represented by local action segments, sentence-level ranges, or other explicit jurisdictions, as long as the triggering text and scope of action can be traced back; degree information can be retained as hand actions, converted to non-manual prompts, or enter a conservative state, but must not be silently deleted without a trajectory; compatible output can contain only action sequences, and detailed output can further include plan units, non-manual prompts, resource capabilities, and received trajectories; spaces, role consistency, or class-marked channels not supported by the target execution end can be explicitly marked as not supported, so as to maintain the stability of the output contract when the execution end is upgraded in the future.

[0062] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A semantic text and execution action processing method based on a limited sign language action resource library, characterized in that, include: Retrieve the original text, the text structure result with available states, one or more structured rewrite candidates, and the set of action capabilities requested for the current request; Based on the text structure results, at least one protected semantic slot with the original character position is extracted to determine the set of protected semantic slots; Retain the original text branch corresponding to the original text, and form candidate branches based on the structured rewrite candidates; When the analysis state of the text structure result is not in an allowed state, record the analysis unavailable state, do not call the candidate action parsing, and directly select the original text branch or output the predefined conservative branch. The candidate branch is subjected to semantic constraint judgment, which includes: verifying whether the candidate branch is in a predetermined production license state and there is no risk of blocking; verifying the independence of the number, source or content identity of evidence according to preset evidence rules; verifying boundary conditions and part-of-speech conditions according to the original text span covered by the candidate branch; verifying context conditions; and verifying that the candidate branch does not cause the semantic loss of the protected semantic slot. Based on the set of action capabilities, resource constraint judgment is performed on the action alias sequence requested by the candidate branch. The resource constraint judgment includes: parsing the precise action identifier according to the order of the action alias sequence and according to the preset deterministic rules, and verifying that all actions requested by the action alias sequence have a parsing result that is allowed to be executed; when any action alias cannot determine the target action, or the parsing result does not belong to the set of allowed execution states, the candidate branch is rejected. A joint determination is made based on the results of the semantic constraint judgment and the resource constraint judgment to determine the branch selection result. The candidate branch is selected only when both the semantic constraint judgment and the resource constraint judgment pass; otherwise, the original text branch or the predefined conservative branch is selected. Based on the branch selection result, when the candidate branch is accepted, an action plan unit containing the precise action identifier, original text span, resource status, and acceptance trajectory is output; when the candidate branch is rejected, the action plan unit of the candidate branch is not output, but a structured conservative result, rejection reason, and non-execution status are output.

2. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, The text structure result includes at least the analysis status and the word units with character positions; the analysis status includes analysis success, partially available, and unavailable status. When the analysis status of the text structure result is in a partially available state, the required structure fields of the candidate branch are determined according to the applicable conditions of the candidate branch, and when any of the required structure fields is missing or unreliable, the candidate branch is prevented from changing its output in this request. The required structural fields are determined by candidate rules, which include: all candidates must have lexical units, character start and end positions, and sentence boundaries; candidates with limited parts of speech must cover the parts of speech of the candidate span; candidates involving entities, numbers, or time must have corresponding entity, numerical, or time spans and reliability indicators; candidates involving negation or modality must have negation or modality markers and their spans; candidates requiring predicates, arguments, or dependency contexts must have corresponding predicates, semantic roles, or dependency fields; and candidates with necessary context or blocking context must have bounded text windows that do not cross sentence boundaries and window reliability.

3. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, Extracting the protected semantic slots includes: extracting at least one of entity, number, time, negation and modality information from the text structure result as protected semantic slots with original character positions; each of the protected semantic slots includes at least slot type, surface text, original character span and identification criteria; For overlapping identification results, deduplication is performed according to a preset deterministic priority rule to form a set of slots without repetition; the deterministic priority rule includes: when the same time expression is identified as both entity and time, it is recorded first by time slot; numerical entities are recorded first by number slot. Verifying that the candidate branch does not cause semantic loss of the protected semantic slot includes: when the candidate branch intersects with the hard protected slot and does not retain the surface information of the hard protected slot, it is allowed to continue only when there is a predetermined equivalence preservation relationship; the hard protected slot includes at least one of entity slots, number slots, time slots, negation slots, and modal slots; wherein, entities, negations, and modals must not be rewritten solely based on candidate self-declaration, and numbers and times can be declared preserved when there is independent evidence and a clear equivalence relationship.

4. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, The verification of context conditions includes: within a bounded context constrained by sentence boundaries, verifying the necessary context conditions and blocking context conditions of the candidate branch; wherein, the ordinary context is determined within a bounded window after removing the target surface word, and when the blocking item is a longer compound word containing the target surface word, the blocking context is determined within a window containing the target surface word, and the bounded window does not cross sentence boundaries.

5. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, The structured rewriting candidate is an information set containing the following fields: original surface terms and original text span, canonical expression, applicable part of speech, necessary context, blocking context, action alias sequence, concept identifier, anonymous evidence summary, state and risk marker, slot retention statement; There is no mandatory processing order for the fields; only the internal order of the action alias sequence must be strictly maintained. In the resource constraint judgment, if any action alias in the action alias sequence cannot determine the target action, or if the current resource state of any action in all actions does not belong to the set of allowed execution states, the candidate branch is rejected and the output of some actions of the candidate branch is prohibited.

6. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, The resource constraint judgment and the assembly of the action planning unit use the same action capability set snapshot identifier.

7. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 5, characterized in that, When multiple candidate branches cover the same original text span, the highest candidate and its difference from the second highest candidate are determined according to preset semantically related ranking factors; the ranking factors include, but are not limited to: priority based on non-overlapping spans, priority based on longer spans, production license status, lower risk, higher evidence level, and stable record sequence number; When the highest candidate is not unique or the difference does not reach the preset threshold, all candidate branches corresponding to the original text span are rejected, and the original text branch or the predefined conservative branch is selected; wherein, the stable record number is only used for fixed trajectory arrangement and shall not be used as a basis for semantic disambiguation alone.

8. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 6, characterized in that, The predefined conservative branch includes at least one of the following: keeping the original text and marking it as unparsed, outputting a pre-configured word-by-word plan or finger spelling plan, returning to an unexecutable state, or keeping the previously confirmed output, wherein the finger spelling plan is based on Chinese finger letters; The acceptance trajectory includes at least the selected branch, the reason for acceptance or rejection, the precise action identifier, the original text span, and the current resource status; and the acceptance trajectory records the same request binding relationship of the following information: the original text span covered by the candidate branch, the snapshot identifier of the action capability set on which the resource constraint judgment is based, the results of the semantic constraint judgment and the resource constraint judgment, and the reason for finally selecting or rejecting the candidate branch.

9. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, Generate local non-manual prompts or sentence-level non-manual prompts corresponding to the span of the original text; the local non-manual prompts include type, intensity, original text span, and target action index, and are bound to the corresponding action planning unit; the sentence-level non-manual prompts are not bound to a single action; The types of non-manual prompts include degree, emphasis, comparison, specific question, or non-question, and the intensity is a discrete level; degree information is bound to the action it modifies, and sentence-level question information applies to the entire sentence; The system outputs an explicit non-support status for parallel capabilities that are not supported by the target execution end, including spatial location, role consistency, and class tagging channels.

10. The semantic text and execution action processing method based on a limited sign language action resource library according to claim 1, characterized in that, The action capability set is a view of existing action resources that can be queried in this request. Each action includes a precise action identifier and the current resource status. The resource status includes, but is not limited to, executable, isolated, missing, conflicting, or unresolved. The set of allowed execution statuses is predetermined by the deployment configuration and only states that have been verified as playable are allowed to pass. Missing, multiple precise action conflicts, isolation, service unavailable, and unverified are considered as not allowed by default.

11. A semantic text and execution action processing device based on a limited sign language action resource library, characterized in that, include: The acquisition module is used to acquire the original text, the text structure result with available status, one or more structured rewriting candidates, and the set of action capabilities requested for the current request; The protected semantic slot extraction module is used to extract at least one protected semantic slot with the original character position based on the text structure result, so as to determine the protected semantic slot set; The branching module is used to retain the original text branches corresponding to the original text and to form candidate branches based on the structured rewriting candidates; The state gate module is used to record the analysis unavailable state when the analysis state of the text structure result is not an allowed state, without calling the candidate action parsing, and directly selecting the original text branch or outputting a predefined conservative branch; The semantic constraint judgment module is used to perform semantic constraint judgment on the candidate branch. The semantic constraint judgment includes: verifying whether the candidate branch is in a predetermined production license state and there is no risk of blocking; verifying the independence of the number, source or content identity of evidence according to preset evidence rules; verifying boundary conditions and part-of-speech conditions according to the original text span covered by the candidate branch; verifying context conditions; and verifying that the candidate branch does not cause the semantic loss of the protected semantic slot. The resource constraint judgment module is used to perform resource constraint judgment on the action alias sequence requested by the candidate branch based on the action capability set. The resource constraint judgment includes: parsing the precise action identifier according to the order of the action alias sequence and a preset deterministic rule, and verifying that all actions requested by the action alias sequence have a parsing result that is allowed to be executed; when any action alias cannot determine the target action, or the parsing result does not belong to the set of allowed execution states, the candidate branch is rejected. The branch selection module is used to make a joint judgment based on the results of the semantic constraint judgment and the resource constraint judgment to determine the branch selection result. The candidate branch is selected only when both the semantic constraint judgment and the resource constraint judgment pass; otherwise, the original text branch or the predefined conservative branch is selected. The output module is used to output an action plan unit that includes precise action identifier, original text span, resource status and acceptance trajectory when the candidate branch is accepted, based on the branch selection result; and to output a structured conservative result, rejection reason and non-execution status when the candidate branch is rejected, without outputting the action plan unit of the candidate branch.

12. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the semantic text and execution action processing method based on a limited sign language action resource library as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, implements the semantic text and action processing method based on a finite sign language action resource library as described in any one of claims 1 to 10.

Images