Language model illusion suppression dynamic decoding method based on multi-dimensional contrast signal

Abstract

The invention relates to the technical field of illusion suppression, in particular to a language model illusion suppression dynamic decoding method based on a multi-dimensional contrast signal, which comprises the following steps of: acquiring a language model input condition, extracting a word group connection relationship and promoting word item output, identifying a grammar sequence and an attachment mode, extracting an entity attribute and comparing reference time. And judging a semantic sequence, analyzing an action connection relationship, screening paths with the same structural form, expanding to next statement action construction, and introducing guidance to obtain a dynamic control signal set. According to the method, a word order and grammar dependency are sorted by constructing a lexical item path, a dynamic contrast mechanism is established in combination with entity reference time and knowledge node updating time, an action sequence is promoted by relying on verb phrases and conjunctions, inter-sentence repeated main sentences and structural convergence fragments are recognized, and a language skeleton with stable features is screened; time and action signals are introduced to guide the generation process, and semantic organization coherence, inter-sentence structure collaboration and content time sequence consistency are enhanced.

Description

Technical Field

[0001] This invention relates to the field of hallucination suppression technology, and in particular to a dynamic decoding method for hallucination suppression based on a language model of multidimensional contrast signals. Background Technology

[0002] The field of illusion suppression technology primarily involves the judgment and intervention control of the authenticity of output content in natural language generation tasks. Core aspects include identifying and avoiding problems such as factual conflicts, logical errors, and semantic isolation in the content generated by language models. This is achieved by constructing mechanisms such as factual consistency assessment, logical consistency verification, and linguistic consistency judgment to improve the credibility and stability of model-generated text. This technology systematically covers multiple key stages, including knowledge enhancement, training data screening, language logic modeling, and post-processing of output results, and is widely used in application scenarios with extremely high requirements for factual accuracy, such as legal interpretation, industry Q&A, and document review. Traditional language model illusion suppression methods refer to post-processing the generated text after the large language model has completed text generation, using external knowledge verification mechanisms or training data intervention mechanisms to identify and filter illusory content. Specifically, this includes enhancing the model's factual learning ability during the training phase through data cleaning and knowledge graph injection; improving the model's sensitivity to real-world corpora through model fine-tuning and comparative learning; or introducing a structured knowledge base after generation to compare and verify key entities and events in the text.

[0003] Existing technologies mainly rely on static text comparison and factual consistency judgment to process output content. They lack the ability to construct the path of language structure progression before the semantics are fully organized, making it difficult to identify the dependent positions and expression order between terms. When processing content with nested events or sequential actions, gaps in connection and logical misalignment are likely to occur. Fact node updates fail to respond in a timely manner during the text generation stage, causing delays in information expression and structural interruptions. Especially in text generation tasks that involve multi-turn action connections or cross-sentence structure tracking, semantic drift and logical fragmentation are likely to occur, affecting the integrity of the content and the controllability of the structure. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals, comprising the following steps: To achieve the above objectives, the present invention adopts the following technical solution: a dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals, comprising the following steps: S1: Obtain the input conditions of the language model, extract the connection relationship between the beginning and end of the phrase, advance the output of the word items according to the path, identify the word order relationship and dependency mode of verbs, nouns and connecting elements, and obtain the candidate path word item organization record; S2: Based on the candidate path terms, organize records, extract entity attribute terms, retrieve entity node update times from external knowledge bases, compare path reference times with entity times, and obtain path fact content time comparison results. S3: Based on the time comparison results of the factual content of the path, extract verb phrases and conjunctions, determine the semantic order, analyze the action connection relationship and semantic breakpoints in the upper and lower sentences, and obtain the internal logical coherence information table of the path. S4: Based on the internal logical coherence information table of the path, extract the main sentence segment, compare the segment position distribution according to word order and difference path, analyze the combination of repeated and identical segments in the text, filter the path with the same structural form, and obtain the main language path list. S5: Based on the main language path list covering semantic and action information, the information is extended to the next statement action construction as a guide source. The guide is introduced into the autoregressive decoding to obtain the dynamically generated control result.

[0005] As a further aspect of the present invention, the candidate path term organization record includes word order relationship, grammatical dependency mode, language structure position, and path term arrangement mode; the path factual content time comparison result includes entity category label, entity node, update timestamp, and reference time comparison information; the path internal logical coherence information table includes verb phrase order, sequential connection structure, semantic advancement direction, context connection information, and grammatical breakpoints; the main language path list includes semantic main clause, path text differences, word order repetition information, segment position relationship, and semantic consistency content; and the dynamic generation control result includes time sequence lag items, logical connection breakpoints, time-advanced entities, action information guidance, and dynamic control signal set.

[0006] As a further aspect of the present invention, the dependency mode refers to the subordinate and grammatical dependency relationship between verbs, nouns and connecting elements in the syntactic structure, and analyzes the structural position and word order relationship of terms in the path; The entity attribute phrases refer to noun phrases with entity features and corresponding category tags, which can be matched with entity nodes in external knowledge bases.

[0007] As a further aspect of the present invention, the path with the same structural form is a set of paths in which the segment order, word order arrangement and semantic expression are consistent in the differentiated candidate paths; The term "covered semantics" refers to a semantic segment within a language path that can fully express the original semantic core without logical gaps or conflicts.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Obtain the input conditions of the current task of the language model, extract the continuous word group structure in the input content, identify the cross-pointing relationship between the connectors and guide words between adjacent words in the sentence, extract the connection state vector between the first and last positions of the word group according to the connection rules, and obtain the word group connection state set. S102: Based on the set of word group connection states, the word items are expanded backward according to the connection relationship between word groups, the word item combination sequence in the progressive path is obtained, the position indexes of verbs, nouns and connecting elements in the path are identified, the order and syntactic dependency relationship logic are mapped, and the word order dependency structure matrix is ​​obtained. S103: Based on the word order dependency structure matrix, aggregate verbs, nouns and conjunctions within the hierarchical path, and divide them into corresponding word order position blocks in the path according to their subordinate relationship in the language structure, to obtain candidate path term organization records.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the path data in the candidate path term organization record, extract noun phrases with attributes of person name, place name, and organization name in the sentence, determine the entity attributes, and assign corresponding category labels to each type of entity phrase to obtain the set of entity attribute phrases in the corpus. S202: Based on the set of entity attribute phrases in the corpus, match the corresponding entity nodes in the external knowledge base, obtain the update timestamp of the entity nodes, synchronously extract the reference time of entity content in the corpus, and compare the time sequence to obtain the reference time sequence of entity nodes and paths. S203: Based on the entity node and path reference time series, locate the path segment whose reference time is earlier than the entity update time, compare and analyze the path structure with the corresponding position in the original corpus, and obtain the path fact content time comparison result.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Based on the time comparison results of the factual content of the path, extract the verb phrases and sequential conjunctions in the path, and analyze the grammatical dependency and positional relationship between the action words and conjunctions according to the order in which the words appear in the sentence to obtain the action connection sequence within the sentence; S302: Based on the intra-sentence action connection sequence, according to the word order structure of the sentences above and below the original text, compare the connection order and semantic extension direction between action pairs, segment the action segments with directional jumps and semantic interruptions, and obtain semantic connection abnormal path segments. S303: Based on the semantic connection anomaly path fragment, search for adjacent action phrases and connecting components in the path structure, extract and splice the path content of the semantically coherent region, and arrange the extended action chain in sentence order to obtain the path's internal logical coherence information table.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the internal logical coherence information table of the path, extract the main clause segment in the path, break down the noun phrases and prepositional conjunctions that are dependent on the verb in the sentence, define the sentence segment boundary according to the grammatical position, and filter the sentence parts with structural expression features to obtain the main clause segment extraction result; S402: Based on the main clause segment extraction results, analyze the start and end order of the segment content in the differential path, identify the segment parts with consistent structural order according to the word order connection method, exclude the segment position offset path, and obtain the segment order matching result; S403: Call the segment sequence matching result, locate the path content containing the corresponding segment from the candidate paths, identify the paths where the content covered by the segments has a positional correspondence, filter the paths with consistent content expression, and obtain the main language path list.

[0012] As a further aspect of the present invention, the process of extracting the main clause segment in the path specifically involves: traversing the paths listed in the internal logical coherence information table of the path, identifying the descriptive sentence segments with subject-predicate structure in the path, and extracting the noun phrases and prepositional conjunctions attached to each verb during the identification process; The process of defining the boundaries of a sentence segment according to its grammatical position is as follows: based on the position of the verb in the sentence, locate the related subject and relational limiting components forward, identify the object or modifying components that are grammatically connected with the verb backward, and locate the beginning and end positions of the sentence segment through syntactic dependency relations. The process of selecting sentence segments with structural expression features is as follows: based on the completeness of the subject-verb-object components and the semantic orientation features of the connecting components, the sentence segments with grammatical continuity and coherence are grammatically judged, and the sentence segments with structural expression features are extracted into the main clause segment extraction results.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Based on the path of the main language path list, retrieve the time tag of the action semantic segment, and compare the order with the time encoding sequence of the preceding and following action items. When the current item's encoding value is after the previous item's encoding value, locate the corresponding action item as a time sequence lag item, and obtain the action time sequence lag index set. S502: Based on the action time sequence lag index set, retrieve the adjacent action items of the corresponding path, perform state judgment on the logical connection structure of the action items, and when the connection state is a broken value, regard the corresponding path as an abnormal logical structure path. Then, according to the action node time tag order, exclude action entities whose starting time tag is after the logical preceding time tag to obtain a semantic structure continuous path set. S503: Based on the continuous path set of the semantic structure, extract data segments covering semantic tags and action codes, and compare the semantic type field of the action code with the verb category field of the action node to filter out the encoding structure that can be extended to the action construction of the next statement. Use the data as dynamic control input parameters and synchronize with the state transition matrix to obtain the dynamic generation control result.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, word path is constructed to sort out word order and grammatical dependencies, and a dynamic comparison mechanism is established by combining entity reference time and knowledge node update time. The action sequence is advanced by relying on verb phrases and conjunctions, repeating main clauses and structurally similar segments between sentences are identified, language skeletons with stable features are selected, and time and action signals are introduced to guide the generation process, thereby enhancing the semantic organization coherence, inter-sentence structural synergy and content temporal consistency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals, comprising the following steps: S1: Obtain the input conditions of the current task of the language model, extract the beginning and end connection relationship of the phrases in the input content, and gradually output the terms through the path. Identify the word order relationship and grammatical dependency between phrases in the path, distinguish the language structure position of verbs, nouns and conjunctions, and obtain the candidate path term organization record. S2: Based on the candidate path term organization record, extract the phrases with entity attributes and assign entity category labels. At the same time, extract the corresponding entity nodes and the current update timestamp from the external knowledge base. Compare the reference time of the entity content in the path with the entity update time. Associate the paths with time differences with the original path corpus to obtain the path fact content time comparison results. S3: Based on the time comparison results of the factual content of the path, extract verb phrases and sequential conjunctions in the path, determine the semantic progression direction according to the adjacent order of the words in the sentence, compare the connection mode between actions according to the original context order, segment the paragraph grammar with semantic interruption and causal jump, and obtain the internal logical coherence information table of the path. S4: Based on the internal logical coherence information table of the path, extract the main sentence segments with semantic expression function from the path, compare the segment texts in the differentiated path, analyze the repetition and positional similarity in word order arrangement, filter the path content with consistent semantic expression, and obtain the main language path list. S5: Based on the paths in the main language path list, alternately locate the time sequence lag items and logical connection breakpoints in the paths, exclude paths including entities with advanced time and interrupted action sequence in order, extend the part of the path structure that covers semantic and action information as the source of guidance to the next statement to construct the action, introduce guidance as a set of dynamic control signals in the autoregressive decoding process, and obtain the dynamically generated control result.

[0023] The candidate path term organization record includes word order relationship, grammatical dependency mode, language structure position, and path term arrangement. The path factual content time comparison results include entity category labels, entity nodes, update timestamps, and reference time comparison information. The path internal logical coherence information table includes verb phrase order, sequential connection structure, semantic progression direction, contextual connection information, and grammatical breakpoints. The main language path list includes semantic main clauses, path text differences, word order repetition information, paragraph positional relationships, and semantic consistency content. The dynamic generation control results include time sequence lag items, logical connection breakpoints, time-advanced entities, action information guidance, and dynamic control signal set.

[0024] Please see Figure 2 The specific steps of S1 are as follows: S101: Obtain the input conditions of the current task of the language model, extract the continuous word group structure in the input content, identify the cross-pointing relationship between the connectors and guide words between adjacent words in the sentence, extract the connection state vector between the first and last positions of the word group according to the connection rules, and obtain the word group connection state set. First, a set of consecutively occurring terms is extracted from the language model input, and the presence of conjunctions or introductory words is identified. Adjacent term pairs are traversed, and their connectivity is compared one by one to determine the presence of restrictive introductory words or progressive connectors, such as "and," "or," "with," and "make." During processing, the connection symbols between each pair of terms are sequentially marked as connection relationship mapping records, and their semantic correlation is further analyzed. For example, if a term pair contains a bidirectional connecting word, it is considered a strong connection; otherwise, it is a weak connection. Then, based on connection rules, it is determined whether there is a potential logical relationship of mutual dependence or semantic progression between the beginning and end of each term group. The judgment logic is mainly based on part-of-speech collocation. Common features of regularity and dependency sequences include, for example, the priority of guiding direction when a verb is paired with a noun is that the verb comes first and bears the semantic core. If a pairing such as "activate-neuron" appears, then based on syntactic experience, "activate" is judged to be the main controlling term, and the weight distribution of the connection state before and after is deduced in reverse. In addition, if there is a progressive pairing such as "through-signal-transmission", it can be split into two connection processes, namely "through-signal" and "signal-transmission", and judged according to the connection rules to obtain its connection direction and priority. By comparing all input terms in the above way, the connection state between each pair of terms is converted into an encoded value, and the connection feature of the term is uniformly preserved in the form of a state vector to obtain the set of phrase connection states.

[0025] S102: Based on the set of phrase connection states, the word items are expanded backward according to the connection relationship between phrases, the word item combination sequence in the progressive path is obtained, the position index of verbs, nouns and connecting elements in the path is identified, the order and syntactic dependency relationship logic are mapped, and the word order dependency structure matrix is ​​obtained. First, the connection state vectors in the set are traversed. Starting from the connection state of each pair of words, the position of the word in the original sentence is determined. It is then determined whether there is a valid backward connection relationship for the current word. If so, its successor word is included in the semantic progression path and the process continues until the path terminates. During the path progression, the word order in the current path is recorded at each step forward, and the part-of-speech identifier of each word is extracted sequentially to construct a part-of-speech tag sequence. For example, in "schedule", "request", and "resource", "schedule" is a verb (V), "request" is a verb (V), and "resource" is a noun (N). Then, the position index of these words in the sentence is recorded. For example, in the sentence "user requests to schedule resources", "user" is position 1, "request" is position 2, "schedule" is position 3, and "resource" is position 4. Next, it is determined whether there is a dependency relationship between adjacent words. The dependency relationship refers to the grammatical subordinate relationship between words. For example, in "schedule resources", "resource" is dependent on the verb "schedule". The relationship between "scheduling → resource" is recorded as a verb-object relationship. This annotation can be achieved through dependency analysis. After constructing multiple path progression combinations, a word order dependency structure matrix can be constructed based on three dimensions: term index, part-of-speech tag, and dependency relationship. The first dimension is the term order index, the second dimension is the part-of-speech tag sequence, and the third dimension is the dependency direction and subordinate tag. Each ternary combination represents the grammatical structural unit of a term node in the path. For example, "request → resource" can be represented as: index 2 (request), part-of-speech tag V, dependency N (resource). In practical applications, this matrix can be used for structural control in language generation tasks. Taking the example sentence "user requests scheduling resources for allocation" as an example, the constructed path combination is "user—request—scheduling—resource—allocation", with the corresponding part-of-speech tag sequence being NVVNV, the term index sequence being 1-2-3-4-5, and the dependency relationship being annotated as "request → user" as a subject-verb relationship, "scheduling → resource" as a verb-object relationship, and "allocation → resource" as an agent relationship, ultimately resulting in a word order dependency structure matrix.

[0026] S103: Based on the word order dependency structure matrix, aggregate verbs, nouns and conjunctions within the hierarchical path, and divide them into corresponding word order position blocks in the path according to their subordinate relationship in the language structure to obtain candidate path word item organization records; First, obtain the verbs, nouns, and connectives corresponding to the path nodes in the matrix. Extract the index position, part-of-speech tag, and semantic dependency of each term under each path. Expand the path sequentially according to the semantic progression order. During the expansion, identify the verb nodes belonging to the main action and record the noun nodes they depend on. Combine the position of the connective to determine whether the associated verb or noun combination forms a complete sentence expression. If the verb node is located at the beginning of the path, it indicates the starting term of the action, and its dependent object or connective should be processed first. Conversely, if the verb node appears at the end of the path, it is necessary to check whether its preceding dependent node intersects with the action at the beginning of the path to confirm whether there is an action chain state in the current path. In the example, for the path "user initiates request and waits for response", "initiates" should be identified as the main action, "request" should be positioned as a noun node, and "and" should be the connective. Then, "wait" is identified as another action term, and "response" is identified as a noun attached to a verb. The term groups are located according to the "initiate-request" and "wait-response" methods. When reorganizing the paths according to the order between the groups, the preceding action path and the following semantic unit are respectively assigned to different path position blocks. Throughout the process, the semantic association judgment is made based on the directionality and hierarchical identifiers of the terms in the word order dependency structure matrix to ensure that each verb and noun has a valid connection relationship. If there is a path segment with unclear connecting components or lacking a dependent target, the path segment is not included in the sorting scope. After completing the above processing in all paths, the nouns and connecting components associated with the verbs are aggregated according to the verb combination as the core. Finally, the expression units with clear language composition positions and tight term combinations in each path are extracted as output to obtain the candidate path term organization record.

[0027] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the path data in the candidate path term organization record, extract noun phrases with attributes of person name, place name, and organization name in the sentence, determine the entity attributes, and assign corresponding category labels to each type of entity phrase to obtain the set of entity attribute phrases in the corpus. First, each noun phrase within each path is traversed to check for phrases containing proper nouns such as personal names, place names, or institutional names. Specifically, this involves analyzing whether the noun phrases contain typical personal name identifiers like "Mr," "Professor," or "Li Mou," or place name identifiers like "street," "province," or "lake," or institutional features like "company," "research institute," or "neighborhood committee." The contextual modification structure, adjacent constituent words, and their collocation patterns in the word order of these phrase combinations are all included in the judgment criteria to confirm their entity affiliation. In practice, if the path data contains the statement "XX teaches at Guangzhou University," then "XX" can be identified as a personal name phrase, and "Guangzhou University" as an institutional phrase. The start and end index positions in the path are then marked accordingly. The term number in the original path is extracted and recorded as the noun recognition result. Then, in the entity attribute judgment stage, the overlap between the term and the existing entries in the selected entity lexicon needs to be further compared. If the overlap ratio exceeds the set matching threshold, it can be regarded as a valid match. The matching threshold is set to 0.8, which means that more than 80% of the term length must be consistent with a certain entity phrase in the entity lexicon in order to be judged as a valid recognition result. After confirming that the recognition is valid, the noun phrase is classified into one of the three sets of person name, place name or organization name, and a corresponding category label is assigned. Person name is labeled as PER, place name is labeled as LOC, and organization name is labeled as ORG. Finally, after traversing all paths, all term phrases with entity semantic features and completed annotation are extracted to obtain the entity attribute phrase set in the corpus.

[0028] S202: Based on the set of entity attribute phrases in the corpus, match the corresponding entity nodes in the external knowledge base, obtain the update timestamp of the entity nodes, synchronously extract the reference time of entity content in the corpus, and compare the time sequence to obtain the reference time sequence of entity nodes and paths. First, the original text information of each phrase is obtained, and its category tag is retrieved, such as PER for personal names, LOC for place names, and ORG for organizations. After obtaining the tag, the entity phrase is used as the search condition to access an external knowledge base and locate the entity node text information that is identical to or has a similarity greater than a set matching threshold. The matching threshold is set to 0.85. If the phrase "Guangzhou University" has a complete match in the knowledge base, the "update timestamp" field of the node's attribute information is directly extracted. This field is pure numerical time series data without specific formatting and is temporarily stored as the entity node's time item. During the matching process, if the entity phrase is "XX" and there are extended entity names such as "Professor XX" or "Doctor XX" in the knowledge base, priority matching is performed based on parameters such as character matching degree and term coverage ratio, and the node with the highest coverage ratio is selected. As the corresponding target, the position of the entity phrase in each path sentence in the corpus is then retrieved synchronously. The path is traced back to the nearest verb or time expression component, and the corresponding reference time item is extracted. The reference time item is set as the core time word content in the sentence that indicates the occurrence of the event. If there is text such as "XX once worked at Guangzhou University" in the path, the time expression corresponding to "once" in "once worked" is extracted and defined as "path reference time". It is compared with the update timestamp in the entity node corresponding to "XX". If the update time is 500103 and the reference time is 500050, the time interval between the two is 53, which means that the reference time is earlier than the node update time. This difference is temporarily stored as a time sequence reference item. Finally, the corresponding entries of each entity node time and path reference time are sorted according to the path number to obtain the entity node and path reference time sequence.

[0029] S203: Based on the time series of entity nodes and path references, locate the path segments whose reference time is earlier than the entity update time, compare and analyze the path structure with the corresponding position in the original corpus, and obtain the time comparison results of the path fact content. First, compare the positions according to the entity update time and the path reference time in each sequence of data. Record the path numbers with all path reference time values less than the corresponding entity update time values into the abnormal path set. Set the time difference threshold to 10. If the reference time of "XX" in the path "P32" is 234 and the entity update time is 256, then record "P32" into the abnormal path set and mark this path reference as premature. Subsequently, call the original path text corresponding to each path number in the abnormal path set, locate the specific position where the entity phrase appears in the path, and at the same time intercept the consecutive sentence segments before and after it as the comparison corpus. Perform repositioning operations on the verb position, entity arrangement, and syntactic features in the context of the main sentence segment in the path. For example, analyze the path "XX is currently the president of Guangzhou University". Its reference time is 234 and the entity update time is 256. Then this path is suspected of having a time-lagged reference. When comparing its sentence pattern and structure with the context original text "XX was recently appointed as the president of Guangzhou University", extract the predicate verb "was appointed" in the sentence as the time-indicating verb, and judge the logical order between the predicate position where "is currently" is located and "was appointed". If the logical relationship is broken and the time-indicating actions are inconsistent before and after, then it is determined that there is a semantic conflict in "is currently" in the path. Further compare the position index between the verb "is currently" and the preposed time component "recently" in the path to identify whether the usage position of this verb in the original corpus is significantly different from the actual update cycle of the entity node. If the difference value exceeds the acceptable time sequence tolerance range in the context syntax, then further lock this path as a fact mismatch path. At the same time, extract the adjacent sentence segments above and below at this place in the path hierarchy, and analyze whether there is a break in the logical direction with other action descriptions in this path paragraph. If so, jointly judge it as a structurally abnormal path segment. Finally, based on the time sequence mismatch label and syntactic logic abnormality label marked in the path, obtain the time comparison result of the path fact content.

[0030] Please refer to Figure 4 , the specific steps of S3 are as follows: S301: Based on the time comparison result of the path fact content, extract the verb phrases and sequential conjunctions in the path. According to the order in which the lexical items appear in the sentence, analyze the grammatical attachment method and position association relationship between the action words and the conjunctions to obtain the intra-sentence action connection sequence; First, the main sentence content of each path segment is retrieved, and verb phrases centered around verbs and conjunctions indicating sequential usage are extracted sequentially. For example, in the path "He got up, ate breakfast, and then went to work," the verb phrases "got up," "ate breakfast," and "went to work," as well as the conjunctions "after" and "then," are extracted. The extracted results are rearranged according to the left-to-right order of the terms in the original path text, and a list of the original index positions of the terms in the sentence is created for subsequent dependency tracking. Each pair of verb phrases and conjunctions is then combined into term pairs. By locating the grammatical attribute nodes of the verbs in the sentence, it is determined whether they are the core dependency points of the conjunctions. For example, it is determined whether "ate breakfast" is logically driven by "after." If the verb phrase appears after the conjunction and there are no other independent predicate phrases inserted between them, then the verb phrase is considered a core dependency point of the conjunction. The verb phrase is the semantic sequence target of the conjunction. Then, grammatical dependency verification is performed on each pair of verb phrases and conjunctions. The verb subordination marker and conjunction progression level marker provided in the syntactic tree structure are compared. If the two are separated by a level of less than or equal to 3 in the syntactic dependency path and the dependency direction is from the conjunction to the verb phrase, it is determined to be a valid dependency relationship. Its original position index information in the sentence is recorded. For example, the dependency path level between "get up" and "after" is 2, and the direction is from "after" to "get up". The connection is then considered valid. Subsequently, all verb phrase and conjunction pairs that are identified as valid dependencies in the entire path are serialized and output to construct a list of terms sorted by word order index. The dependency level and direction relationship between each pair of terms are marked in the list. Finally, the action connection sequence in the sentence is obtained.

[0031] S302: Based on the action connection sequence within a sentence, according to the word order structure of the sentences above and below the original text, the connection order and semantic extension direction between action pairs are compared, and action segments with directional jumps and semantic interruptions are segmented to obtain semantically connected abnormal path segments. First, the dependency order between each group of action words and conjunctions is traversed and read, extracting their position index in the original path text. Then, the sentence connection boundaries are extracted according to the context level of the paragraph in which the path is located. For example, if there is an action group "open the door - walk into the house - do homework" in the sequence, the sentence structure "he opened the door, then walked into the house, and finally did homework" in the path is extracted, and its start index and end index in the overall path are recorded. Then, the content of the sentences above and below the corresponding index in the original path is called to analyze whether the connection direction shows a progressive relationship. If the conjunctions between adjacent actions are sequential conjunctions such as "then", "next", "following", etc., and the action positions move sequentially, then the segment is marked as a normal sequence segment; otherwise, if the conjunction is "but", If there are issues such as subject shifts or predicate omissions, the "However" clause is marked as a semantic direction shift segment. Further semantic breakage analysis is performed on the marked abnormal action segments. First, action verbs are extracted sequentially from the action group. Then, common hyponyms and hypernyms are compared against the context in the language resource database. For example, if there is a lack of semantic extension between "closing the window" and "doing homework," it is determined that a connecting action is missing. Finally, the sentence indexes, conjunction order, and logical direction between verb meanings of all action segments are compared item by item. If a clear dependency path or logical coherence cannot be established between any group of actions, it is marked as an abnormal action connection segment and included in the output results, ultimately yielding semantically connected abnormal path fragments.

[0032] S303: Based on the semantic connection abnormal path fragments, find the action phrases and connecting components that are adjacent to the preceding and following parts in the path structure, extract and splice the path content of the semantically coherent region, and arrange the extended action chain according to the sentence order to obtain the path's internal logical coherence information table. First, the starting and ending indices of the abnormal action connection region in the path are located. Then, verb phrases and conjunctions between the end of the preceding sentence and the beginning of the following sentence are searched sequentially. If a connection gap is detected between action phrases such as "enter the house" and "start doing homework," action candidates such as "open the door" and "return to the room" are searched backwards and inserted under semantic consistency conditions. Subsequently, the verb phrases and conjunctions within the spliced ​​region are rearranged, and an intra-sentence action list is generated based on the order of appearance in the corpus. Each verb-conjunction pair in this list is arranged according to the actual word order in the sentence, and its attached subject phrase, subject position index, and verb action tense are marked. Further detection is performed to check for action jumps before and after splicing. If the subject changes, and there is no jump change, the combination is retained and included in the path expansion. If there is a semantic change, the expanded content is excluded. Then, the retained combination of actions is bidirectionally connected with the actions in the sentences before and after it. The connecting words are replaced with context. For example, if the original path uses "then", but the expanded path is more suitable for connecting with "next" or "then", the replaced connecting word is retained. The confirmation process is carried out based on the semantic closeness calculation value of the sentences before and after. All confirmed path segments are rearranged by sentence. The action chains between the segments are encoded, and the action sequence number and connection relationship in each path segment are recorded. Finally, the path after splicing each action chain is stored in the path database and mapped to the initial path index position to obtain the path's internal logical coherence information table.

[0033] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the path's inherent logical coherence information table, extract the main clause segment from the path, break down the noun phrases and prepositional conjunctions that are dependent on the verb in the sentence, define the sentence segment boundaries according to grammatical position, and filter the sentence parts with structural expression features to obtain the main clause segment extraction results; First, call the sentence order arrangement and action chain position index recorded therein, and sequentially retrieve the component distribution from the beginning to the end of each sentence in the path to find the main clause area with an obvious subject-predicate structure and a complete syntactic composition. In the main clause, preferentially identify the components that have an attachment relationship with the verb, and use the noun phrases and prepositional conjunctions in these components as the participants in subsequent judgments. During the disassembling process, use the verb as the core and retrieve the continuously associated noun components within its forward and backward distances. For example, in the sentence "The student completes the homework", the "student" before the verb "completes" and the "homework" after it will be extracted, and their position indexes in the sentence will be marked. Then, divide the section containing the verb and the noun components on both sides, and use the position of the verb as the core anchor point of the sentence segment. According to whether there are grammatical positioning words such as "of", "in", "at", etc. (prepositional connecting components) before and after this anchor point, judge the boundary range of the sentence segment to which the verb phrase belongs, and further check whether there is a nested structure within this range. For example, in the sentence "Complete the report during the experiment", at this time, "during the experiment" should be regarded as a part of this sentence segment and its attached position should be recorded. Then, screen whether this sentence segment meets the structural expression characteristics. If the subject, predicate, and object are present in this section and there is a complete structural relationship semantically, then retain this sentence segment and include it in the result set. If the core components such as the subject or predicate are missing, or it is only a descriptive phrase for attributive modification, such as "due to weather reasons", then it is not included in the set of main clause segments. Finally, store the retained segments corresponding to their starting positions in the original sentence in order according to the path to obtain the result of main clause segment extraction.

[0034] S402: Based on the result of main clause segment extraction, correspondingly analyze the start and end order of the segment content in the differential path, identify the segment parts with consistent structural order according to the word order connection method, and exclude the path with offset segment positions to obtain the result of segment order matching; First, the start and end indexes of each main clause are extracted. Then, the main clause segments in each path are matched one-to-one with their positions in the original corpus according to the path number, providing a positioning reference for subsequent comparisons. Next, segments semantically similar to the main clause segments are extracted sequentially from the differentiated path set, using their sentence order within the path as the core comparison reference. First, it is determined whether the segment in the differentiated path is in the same index interval as the corresponding segment in the main path. If its start and end positions are within two sentences of the corresponding segment in the main clause, it is considered an alignable path; otherwise, it is directly marked as a positional offset path. Based on this, word order comparison is performed within the segments in the aligned paths, comparing the order of verbs and subject-object components in each sentence. If the word order matches the main path, then... If the verb and object are in the same order in the sentence and there are connecting elements on both sides that match the main path, then the sentence is considered to have a consistent word order connection. For such sentences, the path index and the corresponding content in the original corpus are retained for subsequent retrieval. For sentences that do not meet the above conditions, such as "object moved forward" or "verb omitted", it is determined whether the position of its keyword phrase is misaligned with the components in the main clause. If the number of misaligned sentences exceeds one, it is determined to be a sentence position offset path and the path is removed. Throughout the process, different sentences need to be processed uniformly in conjunction with the path number to prevent sentence confusion. At the same time, the start and end position marker values ​​during the comparison are retained as index data for path sequence analysis. Finally, the content whose sentence position and word order are consistent with the main path is selected to obtain the sentence order matching result.

[0035] S403: Call the segment sequence matching results, locate the path content containing the corresponding segment from the candidate paths, identify the paths where the content covered by the segments has a positional correspondence, filter the paths with consistent content expression, and obtain the list of main language paths; First, obtain the path number and sentence start / end index of each segment in the original path from the matching results. Then, locate each candidate path to see if it contains the aforementioned segment. If the segment content can be completely matched within the sentence order range of the path, mark the path as a covered path. Simultaneously, read the sentence text content corresponding to the segment in the path, comparing the positions of elements such as subject, verb, object, and prepositional phrase. Check whether the expression in the sentence is semantically identical to the segment in the main path, i.e., the verb-object combination does not undergo semantic expansion, the prepositional phrase does not change its subject, and the subject's referent remains unchanged. If any of these three criteria are incorrect... If the three criteria are the same, the path is considered to have a change in expression and is excluded from further processing. If all three criteria are consistent, the path is considered to have a consistent expression and continues to be processed. Then, a segment-level connectivity check is performed on the set of paths that have passed the checks. That is, the order of all segments in the path is compared with the order of the corresponding segments in the main path. If the order of two segments is swapped in the candidate path, the path is removed. Only the paths with unchanged segment order and consistent semantic expression are retained and enter into the final list. These paths are uniformly numbered and a main language path identifier is generated as the basis for subsequent calls, and finally the main language path list is obtained.

[0036] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the path of the backbone language path list, retrieve the time tag of the action semantic segment and compare the order with the time encoding sequence of the preceding and following action items. When the current item's encoding value is after the previous item's encoding value, locate the corresponding action item as a time lag item and obtain the action time lag index set. First, the time tag information carried in each action semantic segment needs to be extracted. This time tag can be derived from the text position of the action phrase in the path, explicit time expression in the context, or implicit time derivation results. For example, in "user clicks link and downloads file", the action "click" can be parsed as the starting action, and "download" is the subsequent action. The time tag of the former can be determined by the order of associated positions as t1, and the latter as t2. Second, the action item time tag is converted into a time code value. This value needs to be set in combination with its relative order in the main path and semantic trigger priority. For example, "click" as the main trigger action is assigned the code number 1001, and "download" is assigned 1002. Then, the code values ​​of the current item and the previous item are compared numerically. During the comparison process... The process involves calling the time-coded sequence of all action items in the current path, setting the code values ​​sequentially according to the order in which the actions appear in the semantic path, and establishing a mapping relationship between preceding and following items. The code values ​​of adjacent actions are directly subtracted. If the value of the current action code value minus the previous action code value is positive and greater than the preset minimum interval benchmark value of 20, it can be determined that the current item appears after the previous item, indicating that there is a time lag in its semantic progression. For example, if the code value of the previous item is 1040 and the current item is 1070, the difference is 30, which is greater than the minimum benchmark value, so the current item is included in the time sequence lag judgment range. The above code extraction and comparison process is repeated throughout the entire path traversal, and all action items that meet the lag condition are numbered and recorded, finally obtaining the action time sequence lag index set.

[0037] S502: Based on the action time sequence lag index set, retrieve the adjacent action items of the corresponding path, judge the state of the logical connection structure of the action items, and when the connection state is a broken value, regard the corresponding path as an abnormal logical structure path. Then, according to the action node time tag order, exclude action entities whose start time tag is after the logical preceding time tag to obtain a semantic structure continuous path set. First, each delayed action needs to be located in the semantic path according to the path identifier, determining the path node numbers of adjacent actions and their order of appearance in the text. Then, the words or identifier phrases connecting the current action to its adjacent actions need to be extracted one by one as the basis for judging the logical connection structure. For example, in the statement "After the user fills in the information and submits the form," if there is a connecting phrase "after" between "fill in" and "submit," it is marked as a normal connection. If the connecting phrase is missing in "After the user fills in the information and submits the form," the connection structure is marked as broken. When judging the connection state, a connection break threshold should be set as the judgment standard. This threshold can be set as the statistical value of the number of missing logical connection identifiers. When there is no logical connection identifier between two consecutive actions and the coding distance between the two actions exceeds the preset average action distance value of 40, it is marked as a broken state. For example, if "Operation A" is encoded as 1030 and "Operation B" is encoded as 1080, with a spacing of 50 and a missing connector, it is determined to be a broken connection, and the connection status is a broken value. When this status is met, the path is recorded as a logically abnormal path. Then, the time tags of each action node in the abnormal path need to be compared sequentially. If the time tag of the current action node is after the time tag of the logical predecessor, the action is determined to be an action entity with a delayed start time. This judgment process needs to traverse all action nodes in the logically abnormal path and extract their time tags one by one. For example, if the current action time code is 2100 and the previous action time code is 2050, then the current action time code is greater than the previous time code, which meets the lag judgment condition. The action entity is excluded from the continuous path set, while action entities that do not show time inversion are retained, and finally, a semantically continuous path set is obtained.

[0038] S503: Based on the continuous path set of semantic structure, extract data segments covering semantic tags and action codes, and compare the semantic type field of the action code with the verb category field of the action node to filter out the encoding structure that can be extended to the action construction of the next statement. Use the data as dynamic control input parameters and synchronize with the state transition matrix to obtain the dynamic generation control result. First, the action nodes in the path need to be located, and their corresponding semantic tags and action codes need to be extracted. The semantic tags can be determined by the event type mapped in the context of the verb phrase. For example, the action node "download" in "download report" can be mapped to the semantic tag "acquisition behavior", and its code is set to 3021. All action nodes and semantic tags in the path form a one-to-one code pair data fragment. Then, the semantic type field corresponding to the action code in the data fragment is called. The semantic type field indicates the behavior category to which the code belongs. Next, the verb category field of the corresponding action node is extracted. The verb category field is obtained by processing the original verb phrase through verb part-of-speech classification. For example, "send" in "send email" is classified as "transmission action", and the field value is set to T1. Then, the semantic type field and the verb category field need to be compared item by item. The comparison operation is performed in the order of the path. The corresponding values ​​of the two fields are looked up in the action number mapping table. If two identifier codes are the same, they are considered a match. For example, if the semantic type field value is A3 and the verb category field value is A3, and the corresponding identifier code is "C-12", then it is a matching structure. All encoded structures that meet this condition are filtered out for subsequent processing. During the filtering process, encoded data that do not meet the field consistency requirement must be excluded. For example, if the semantic type is B4 and the verb category is C2, the identifier codes are inconsistent, so this item is removed. For all the filtered encoded structure sets, a structure encoding column is constructed according to the order of actions in the path. This column is used as the dynamic control input parameter group input into the state transition rule table. The state transition rule table consists of the mapping relationship between the preset action state vector and the control input. For example, if the current state vector is S5 and the input control code is 3021, then the transition is to state S6. All input parameters are synchronously deduced with the current state to deduce the state change path, and finally the dynamically generated control result is obtained.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals, characterized in that, Includes the following steps: S1: Obtain the input conditions of the language model, extract the connection relationship between the beginning and end of the phrase, advance the output of the word items according to the path, identify the word order relationship and dependency mode of verbs, nouns and connecting elements, and obtain the candidate path word item organization record; S2: Based on the candidate path terms, organize records, extract entity attribute terms, retrieve entity node update times from external knowledge bases, compare path reference times with entity times, and obtain path fact content time comparison results. S3: Based on the time comparison results of the factual content of the path, extract verb phrases and conjunctions, determine the semantic order, analyze the action connection relationship and semantic breakpoints in the upper and lower sentences, and obtain the internal logical coherence information table of the path. S4: Based on the internal logical coherence information table of the path, extract the main sentence segment, compare the segment position distribution according to word order and difference path, analyze the combination of repeated and identical segments in the text, filter the path with the same structural form, and obtain the main language path list. S5: Based on the main language path list covering semantic and action information, the information is extended to the next statement action construction as a guide source. The guide is introduced into the autoregressive decoding to obtain the dynamically generated control result.

2. The hallucination suppression dynamic decoding method based on a multidimensional contrast signal language model according to claim 1, characterized in that, The candidate path term organization record includes word order relationship, grammatical dependency mode, language structure position, and path term arrangement. The path factual content time comparison result includes entity category label, entity node, update timestamp, and reference time comparison information. The path internal logical coherence information table includes verb phrase order, sequential connection structure, semantic progression direction, contextual connection information, and grammatical breakpoints. The main language path list includes semantic main clause, path text differences, word order repetition information, paragraph position relationship, and semantic consistency content. The dynamic generation control result includes time sequence lag items, logical connection breakpoints, time-advanced entities, action information guidance, and dynamic control signal set.

3. The hallucination suppression dynamic decoding method based on a multidimensional contrast signal language model according to claim 1, characterized in that, The dependency mode refers to the subordinate and grammatical dependency relationship between verbs, nouns and conjunctions in the syntactic structure, and analyzes the structural position and word order relationship of word items in the path; The entity attribute phrases refer to noun phrases with entity features and corresponding category tags, which can be matched with entity nodes in external knowledge bases.

4. The dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals according to claim 1, characterized in that, The path set with the same structural form is the path set in which the paragraph order, word order arrangement and semantic expression are consistent among the differentiated candidate paths; The term "covered semantics" refers to a semantic segment within a language path that can fully express the original semantic core without logical gaps or conflicts.

5. The dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Obtain the input conditions of the current task of the language model, extract the continuous word group structure in the input content, identify the cross-pointing relationship between the connectors and guide words between adjacent words in the sentence, extract the connection state vector between the first and last positions of the word group according to the connection rules, and obtain the word group connection state set. S102: Based on the set of word group connection states, the word items are expanded backward according to the connection relationship between word groups, the word item combination sequence in the progressive path is obtained, the position indexes of verbs, nouns and connecting elements in the path are identified, the order and syntactic dependency relationship logic are mapped, and the word order dependency structure matrix is ​​obtained. S103: Based on the word order dependency structure matrix, aggregate verbs, nouns and conjunctions within the hierarchical path, and divide them into corresponding word order position blocks in the path according to their subordinate relationship in the language structure, to obtain candidate path term organization records.

6. The dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Based on the path data in the candidate path term organization record, extract noun phrases with attributes of person name, place name, and organization name in the sentence, determine the entity attributes, and assign corresponding category labels to each type of entity phrase to obtain the set of entity attribute phrases in the corpus. S202: Based on the set of entity attribute phrases in the corpus, match the corresponding entity nodes in the external knowledge base, obtain the update timestamp of the entity nodes, synchronously extract the reference time of entity content in the corpus, and compare the time sequence to obtain the reference time sequence of entity nodes and paths. S203: Based on the entity node and path reference time series, locate the path segment whose reference time is earlier than the entity update time, compare and analyze the path structure with the corresponding position in the original corpus, and obtain the path fact content time comparison result.

7. The dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Based on the time comparison results of the factual content of the path, extract the verb phrases and sequential conjunctions in the path, and analyze the grammatical dependency and positional relationship between the action words and conjunctions according to the order in which the words appear in the sentence to obtain the action connection sequence within the sentence; S302: Based on the intra-sentence action connection sequence, according to the word order structure of the sentences above and below the original text, compare the connection order and semantic extension direction between action pairs, segment the action segments with directional jumps and semantic interruptions, and obtain semantic connection abnormal path segments. S303: Based on the semantic connection anomaly path fragment, search for adjacent action phrases and connecting components in the path structure, extract and splice the path content of the semantically coherent region, and arrange the extended action chain in sentence order to obtain the path's internal logical coherence information table.

8. The dynamic decoding method for hallucination suppression based on a multidimensional contrast signal language model according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the internal logical coherence information table of the path, extract the main clause segment in the path, break down the noun phrases and prepositional conjunctions that are dependent on the verb in the sentence, define the sentence segment boundary according to the grammatical position, and filter the sentence parts with structural expression features to obtain the main clause segment extraction result; S402: Based on the main clause segment extraction results, analyze the start and end order of the segment content in the differential path, identify the segment parts with consistent structural order according to the word order connection method, exclude the segment position offset path, and obtain the segment order matching result; S403: Call the segment sequence matching result, locate the path content containing the corresponding segment from the candidate paths, identify the paths where the content covered by the segments has a positional correspondence, filter the paths with consistent content expression, and obtain the main language path list.

9. The dynamic decoding method for hallucination suppression based on a language model using multidimensional contrast signals according to claim 8, characterized in that, The process of extracting the main clause segment from the path is as follows: traverse the paths listed in the internal logical coherence information table, identify the expression sentence segments with subject-predicate structure in the path, and extract the noun phrases and prepositional conjunctions attached to each verb during the identification process; The process of defining the boundaries of a sentence segment according to its grammatical position is as follows: based on the position of the verb in the sentence, locate the related subject and relational limiting components forward, identify the object or modifying components that are grammatically connected with the verb backward, and locate the beginning and end positions of the sentence segment through syntactic dependency relations. The process of selecting sentence segments with structural expression features is as follows: based on the completeness of the subject-verb-object components and the semantic orientation features of the connecting components, the sentence segments with grammatical continuity and coherence are grammatically judged, and the sentence segments with structural expression features are extracted into the main clause segment extraction results.

10. The dynamic decoding method for hallucination suppression based on a multidimensional contrast signal language model according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Based on the path of the main language path list, retrieve the time tag of the action semantic segment, and compare the order with the time encoding sequence of the preceding and following action items. When the current item's encoding value is after the previous item's encoding value, locate the corresponding action item as a time sequence lag item, and obtain the action time sequence lag index set. S502: Based on the action time sequence lag index set, retrieve the adjacent action items of the corresponding path, perform state judgment on the logical connection structure of the action items, and when the connection state is a broken value, regard the corresponding path as an abnormal logical structure path. Then, according to the action node time tag order, exclude action entities whose starting time tag is after the logical preceding time tag to obtain a semantic structure continuous path set. S503: Based on the continuous path set of the semantic structure, extract data segments covering semantic tags and action codes, and compare the semantic type field of the action code with the verb category field of the action node to filter out the encoding structure that can be extended to the action construction of the next statement. Use the data as dynamic control input parameters and synchronize with the state transition matrix to obtain the dynamic generation control result.

Images