Enterprise technical demand automatic identification method based on large model

Abstract

The application discloses an enterprise technical demand automatic identification method based on a large model, which comprises the following steps: step one, demand text standardization; step two, large model generates triggering evidence and inhibiting evidence; step three, generating semantic enhancement text and semantic intervention text; step four, constructing a directed influence matrix and generating a Classifier Trellis structure; step five, training a node classifier to output a node conditional probability; step six, fusing forward node confidence and reverse node confidence to obtain marginal confidence; step seven, based on the marginal confidence single-point flip and local recalculation, updating the label state and outputting the identification result. The application improves the consistency and stability of the multi-label identification of enterprise technical demands.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for automatically identifying enterprise technology needs based on a large model. Background Technology

[0002] Automatic identification of enterprise technical requirements falls under the category of text understanding and multi-label classification. Existing solutions typically target the requirement description text submitted by enterprises, identifying the requirement theme, technical field, application scenario, and constraints, and mapping them to a pre-set label set. Common implementations include matching methods based on keyword dictionaries and rule templates, text classification methods based on traditional supervised learning, fine-tuning classification methods based on pre-trained language models, and prompt-based classification methods based on large models. Enterprise technical requirement texts often come from complex sources, including colloquial expressions, cross-domain terminology, inconsistent formats, and incomplete content. This results in noise and structural gaps in the input text at the character, sentence, and field levels. Existing methods lack a unified and reusable processing link for sentence segmentation, noise reduction, and field standardization, leading to unstable semantics in the classification input and significant fluctuations in label output under different text expression methods.

[0003] In the multi-label recognition stage, existing technologies mostly use independent binary classifiers, chain classifiers, generative modeling from sequence to set, or graph-structured label modeling to characterize the dependencies between labels. Independent binary classifiers ignore label dependencies, chain classifiers are sensitive to label order and errors are easily propagated, graph-structured label modeling relies on prior graphs or additional annotation costs, and generative outputs are difficult to stably satisfy label set constraints. For practical applications of intent recognition and multi-label classification, as the size of the label set increases, there are directional influence relationships and hierarchical association relationships between labels. Existing solutions often lack an explicit quantification mechanism for the "strength of the directional influence of labels" and lack the process of using the influence strength to construct a decodable structured reasoning skeleton. This makes it difficult for the model output to simultaneously address label co-occurrence conflicts, differentiation of similar labels, and retention of weak evidence labels, which easily leads to problems such as insufficient consistency of label sets, missed detection of associated labels, and coexistence of conflicting labels.

[0004] While large-model-based prompting solutions can reduce manual feature engineering, they still suffer from insufficient controllability and reproducibility in enterprise technology requirement identification scenarios. Existing technologies often use direct prompts to output label results or prompt learning based on a small number of examples, lacking structured generation and normalization of semantic evidence for label triggering and label suppression, resulting in evidence drift due to fluctuations in the generated prompt content. At the same time, there is a lack of a process to separately construct semantically enhanced text and semantically intervened text for label pair scoring, making it difficult to stably extract label dependencies. Furthermore, for structured decoding and consistency correction, existing solutions usually use one-time decoding or simple threshold screening, lacking single-point flipping and local recalculation mechanisms based on marginal confidence, and lacking a delayed acceptance update criterion linked to the consistency score of the directed influence matrix. This makes it easy for label state updates to fall into unstable oscillations on boundary samples, or require full recalculation, which increases computational overhead and response latency.

[0005] Therefore, how to provide a method for automatically identifying enterprise technology needs based on a large model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an automatic identification method for enterprise technology requirements based on a large model. This invention constructs semantically enhanced text and semantically intervened text by introducing label-triggered semantic evidence and label-suppressed semantic evidence. It combines a Classifier Trellis structure driven by a directed influence matrix to realize multi-label dependency modeling and decoding inference, and completes label consistency correction based on a single-point flipping and delayed acceptance mechanism with marginal confidence. This achieves stable and accurate identification of enterprise technology requirements and has the advantages of strong label association characterization, high consistency of identification results, and strong controllability of the inference process.

[0007] An automatic identification method for enterprise technology requirements based on a large model according to an embodiment of the present invention includes the following steps: Step 1: Obtain the enterprise's technical requirements text, perform sentence segmentation, noise reduction, and field standardization to obtain standardized requirements text; Step 2: Obtain the label set, and generate a set of semantic evidence for label triggering and a set of semantic evidence for label inhibition based on the large model; Step 3: Generate semantic intervention text and semantic enhancement text based on standardized requirement text, tag-triggered semantic evidence set, and tag-inhibited semantic evidence set; Step 4: Calculate the directed influence strength of each tag pair based on the semantically intervened text and the semantically enhanced text, forming a directed influence matrix. Based on the directed influence matrix, determine the number of Trellis grid layers, the Trellis grid width, the node arrangement, and the set of dependent edges to form the Classifier Trellis structure. Step 5: Train a node classifier based on the Classifier Trellis structure. The node classifier takes the standardized demand text and the parent node label state as input and outputs the node conditional probability. Step 6: Perform forward decoding based on the Classifier Trellis structure to obtain forward node confidence, perform reverse consistency backhaul to obtain reverse node confidence, and fuse forward node confidence and reverse node confidence to obtain marginal confidence; Step 7: Perform single-point flipping and local recalculation on the label based on marginal confidence, update the label status according to the delayed acceptance criterion, and output the enterprise technology requirement identification result when the stability judgment condition is met.

[0008] Optionally, step one specifically includes: Obtain the enterprise's technical requirements text, and perform character normalization processing on the enterprise's technical requirements text. Character normalization processing includes full-width character half-width conversion, whitespace character normalization, and control character deletion. The enterprise's technical requirements text is processed into a sentence sequence. The sentence segmentation is based on a set of delimiters, which includes periods, semicolons, exclamation marks, question marks, and line breaks. Denoising is performed on the sentence sequence to obtain a denoised sentence sequence. The denoising process includes removing hyperlink strings, HTML tags, emojis, and non-semantic noise tags. Perform field standardization on the denoised sentence sequence and concatenate them according to the field order to obtain the standardized requirement text.

[0009] Optionally, step two specifically involves: Retrieve the tag set, which includes tag identifiers and tag names; The large model is defined as a pre-trained generative language model. The large model includes model parameters, lexicalization rules, and generation / decoding rules. The lexicalization rules convert the prompt text into a sequence of lexical units. The generation / decoding rules generate the output sequence of lexical units word by word based on the sequence of lexical units in the prompt text using an autoregressive generation method. For each tag in the tag set, a trigger-generated prompt text is constructed. The trigger-generated prompt text includes the tag name and evidence output format constraints. The trigger-generated prompt text is input into the large model to output a set of trigger phrases and a set of trigger sentences. The set of trigger phrases and the set of trigger sentences constitute the tag-triggered semantic evidence set, which serves as the source of semantic evidence for the semantically enhanced text. For each label in the label set, a suppression generation prompt text is constructed. The suppression generation prompt text includes the label name and evidence output format constraints. The suppression generation prompt text is input into the large model to output a set of suppression phrases and a set of suppression sentences. The set of suppression phrases and the set of suppression sentences constitute the label suppression semantic evidence set, which serves as the source of semantic evidence for the semantic intervention text. Evidence normalization processing is performed on the tag-triggered semantic evidence set and the tag-inhibited semantic evidence set. Evidence normalization processing includes word segmentation normalization, case normalization, punctuation normalization, duplicate item deletion, length threshold filtering, and order fixing.

[0010] Optionally, step three specifically includes: For each tag in the tag set, read the standardized requirement text, the tag-triggered semantic evidence set, and the tag-inhibited semantic evidence set; A semantically enhanced fragment sequence is generated based on the set of semantic evidence triggered by the tags. The semantically enhanced fragment sequence is obtained by splicing together the set of trigger phrases and the set of trigger sentences in the order of arrangement. The standardized requirements text and the semantically enhanced fragment sequence are concatenated in the concatenation order to generate semantically enhanced text, with a single space as the concatenation separator. A semantic intervention fragment sequence is generated based on the set of semantic evidence for suppression of labels. The semantic intervention fragment sequence is obtained by splicing together the set of suppression phrases and the set of suppression sentences in the order of arrangement. The standardized requirement text is subjected to trigger fragment deletion processing. The trigger fragment deletion processing performs text matching and deletes the matched fragments based on the set of trigger phrases and the set of trigger sentence patterns to obtain the intervention base text. The semantic intervention text is generated by concatenating the base intervention text and the sequence of semantic intervention fragments in the concatenation order, with a single space as the concatenation separator.

[0011] Optionally, step four specifically involves: Determine the number of tags in the tag set, and establish a tag index sequence based on the tag identifiers; For each ordered tag pair in the tag index sequence, which contains a first tag and a second tag, read the semantic enhancement text and semantic intervention text corresponding to the first tag, construct a tag scoring prompt text containing the tag name of the second tag, and input the tag scoring prompt text into the large model to obtain the semantic enhancement text score value and the semantic intervention text score value. The intensity of directed influence is determined by the difference between the score of semantically enhanced text and the score of semantically intervened text, and the intensity of directed influence is written into the directed influence matrix according to the label index sequence. Determine the Trellis grid width and the number of Trellis grid layers. The number of Trellis grid layers must satisfy the condition that the product of the Trellis grid width and the number of Trellis grid layers is not less than the number of labels. The outward influence summary value of each label is calculated based on the directed influence matrix. The outward influence summary value is the summary result of the directed influence intensity of all labels except the first label in the label index sequence pointed to by the first label. The nodes are generated and arranged in descending order of the outward influence summary value. The labels are filled into each layer and column of the Trellis grid according to the node arrangement to generate the node arrangement mapping relationship; Determine the set of dependent edges, which satisfies the constraint that the layer index of the parent node is less than the layer index of the child node, and the constraint that the upper limit of the in-degree of the child node is satisfied. The set of candidate parent nodes consists of nodes whose layer index is less than the layer index of the child node. For each child node, select the upper limit number of parent nodes in the candidate parent node set according to the directed influence strength from large to small to establish dependent edges, forming the Classifier Trellis structure.

[0012] Optionally, step five specifically includes: Obtain a set of standardized requirement texts and a set of label status annotations. The set of label status annotations contains the truth value of each label in the label set on each standardized requirement text in the set of standardized requirement texts. Based on the Classifier Trellis structure, the node set, the dependent edge set, and the node arrangement mapping relationship are determined. Based on the dependent edge set and the node arrangement mapping relationship, the parent node set corresponding to each node is determined. The parent node set consists of the parent nodes pointing to the nodes in the dependent edge set and is arranged in ascending order according to the grid layer index corresponding to the node arrangement mapping relationship. For each node in the node set and for each standardized requirement text in the standardized requirement text set, extract the true values ​​of the corresponding tags from the tag state annotation set and form the parent node tag state vector according to the order of the parent node set. The standardized requirement text and the parent node label state vector are concatenated in a fixed order to form the node classification input. The node classification input is then input into the node classifier to output the node conditional probability. The training loss is calculated based on the true values ​​of the labels corresponding to the nodes and the conditional probabilities of the nodes in the label state annotation set, and the node classifier parameters are updated to obtain the set of node classifiers that have been trained.

[0013] Optionally, step six specifically includes: Obtain the standardized requirements text, obtain the Classifier Trellis structure and read the node set, dependency edge set, node arrangement mapping relationship, directed influence matrix, and node classifier set; Based on the node arrangement mapping relationship, a forward decoding sequence is generated by incrementing the grid layer index and the grid column index; For each node in the forward decoding sequence, the set of parent nodes is determined based on the set of dependent edges and the label state vector of the parent node is read. The standardized requirement text and the label state vector of the parent node are concatenated in a fixed order to form the node classification input. The input is given to the node classifier to obtain the node conditional probability. The node conditional probability is used as the forward node confidence. Based on the decision threshold, the node conditional probability is mapped to the node label state and written into the label state vector. Based on the node arrangement mapping relationship, a reverse propagation sequence is generated by decreasing the grid layer index and decreasing the grid column index; For each node in the backpropagation sequence, the set of child nodes is determined based on the set of dependent edges. The confidence of the forward node corresponding to the set of child nodes is read, and the directed influence strength from the node to each child node in the directed influence matrix is ​​read. The confidence of the forward node corresponding to the set of child nodes is weighted and normalized according to the directed influence strength to obtain the confidence of the reverse node. Marginal confidence is obtained by weighted fusion of forward node confidence and reverse node confidence based on fusion weight, with the fusion weight ranging from 0 to 1.

[0014] Optionally, step seven specifically includes: Generate label state vectors based on marginal confidence and decision thresholds; The candidate label set for flipping is selected based on marginal confidence. The candidate label set for flipping consists of labels whose marginal confidence is between the lower threshold and the upper threshold, where the lower threshold is less than the upper threshold. For each tag in the candidate tag set to be flipped, a single-point flip is performed to generate a candidate tag state vector. A single-point flip is to invert the state of the corresponding tag in the candidate tag state vector. The local recalculation range is determined based on the set of dependent edges. The local recalculation range consists of the node corresponding to the single-point flip label and the set of child nodes of the node corresponding to the single-point flip label. Based on the set of node classifiers, the node conditional probability is recalculated for the nodes within the local recalculation range and the marginal confidence is updated. The label state vector is updated according to the delayed acceptance criterion. The delayed acceptance criterion is to accept the candidate label state vector when the consistency score of the candidate label state vector is not less than the consistency score of the current label state vector and the marginal confidence summation value of the candidate label state vector is not less than the marginal confidence summation value of the current label state vector. The consistency score is calculated based on the directed influence matrix and the label state vector. The enterprise technology requirement identification result is output when the label state vector meets the stability judgment condition. The stability judgment condition is that the label state vectors obtained in two adjacent updates are consistent.

[0015] The beneficial effects of this invention are: This invention introduces a set of semantic evidence triggered by tags and a set of semantic evidence suppressed by tags, generated based on a large model, into the automatic identification process of enterprise technical requirements. This transforms the original requirement text into semantically enhanced text and semantically intervened text, thereby explicitly modeling the semantics related to tags at the text level. This ensures that tag discrimination no longer relies solely on implicit vector representations but is based on controllable and reproducible semantic evidence. Simultaneously, by performing uniform sentence segmentation, denoising, and field standardization on the enterprise technical requirement text, the impact of differences in text expression on the recognition results is effectively reduced. This ensures that requirement texts from different sources and with different styles have consistent structured semantic input conditions before entering the recognition process, improving the overall recognition stability.

[0016] In the label reasoning and result correction stage, this invention calculates the directed influence strength between label pairs based on semantically enhanced text and semantically intervened text, and constructs a Classifier Trellis structure accordingly to achieve structured expression and ordered decoding of label dependencies. Marginal confidence is obtained under the fusion mechanism of forward decoding and reverse consistency feedback. Furthermore, through single-point flipping, local recalculation and delayed acceptance criteria based on marginal confidence, the label state is consistently corrected, so that the label set gradually converges under the premise of satisfying dependency constraints. In the complex, multi-label, and strongly correlated enterprise technology demand identification scenario, this invention improves the rationality and overall consistency of label combination, and avoids the problems of irrelevant label interference and conflicting label coexistence. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows a method for automatic identification of enterprise technology requirements based on a large model proposed in this invention. Figure 2 This diagram illustrates the marginal confidence single-point flipping and local recalculation of an automatic identification method for enterprise technology requirements based on a large model proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-2 An automatic identification method for enterprise technology requirements based on a large model includes the following steps: Step 1: Obtain the enterprise's technical requirements text, perform sentence segmentation, noise reduction, and field standardization to obtain standardized requirements text; Step 2: Obtain the label set, and generate a set of semantic evidence for label triggering and a set of semantic evidence for label inhibition based on the large model; Step 3: Generate semantic intervention text and semantic enhancement text based on standardized requirement text, tag-triggered semantic evidence set, and tag-inhibited semantic evidence set; Step 4: Calculate the directed influence strength of each tag pair based on the semantically intervened text and the semantically enhanced text, forming a directed influence matrix. Based on the directed influence matrix, determine the number of Trellis grid layers, the Trellis grid width, the node arrangement, and the set of dependent edges to form the Classifier Trellis structure. Step 5: Train a node classifier based on the Classifier Trellis structure. The node classifier takes the standardized demand text and the parent node label state as input and outputs the node conditional probability. Step 6: Perform forward decoding based on the Classifier Trellis structure to obtain forward node confidence, perform reverse consistency backhaul to obtain reverse node confidence, and fuse forward node confidence and reverse node confidence to obtain marginal confidence; Step 7: Perform single-point flipping and local recalculation on the label based on marginal confidence, update the label status according to the delayed acceptance criterion, and output the enterprise technology requirement identification result when the stability judgment condition is met.

[0020] In this embodiment, step one specifically includes: The system obtains the enterprise's technical requirements text and performs character normalization processing on it. Character normalization processing includes full-width character half-width conversion, whitespace character normalization, and control character deletion. Full-width character half-width conversion converts full-width letters, full-width numbers, and full-width punctuation marks into half-width characters based on a character mapping table. Whitespace character normalization uniformly replaces tab characters, carriage returns, and line feeds with single spaces and compresses consecutive spaces into single spaces. Control character deletion deletes characters whose character encoding values ​​fall within the control character range. The enterprise's technical requirements text is processed into a sentence sequence. The sentence segmentation is based on a set of delimiters, which includes periods, semicolons, exclamation marks, question marks, and newlines. The sentence segmentation scans along the text character sequence from the first character to the last character. When any delimiter in the delimiter set is encountered, the text is truncated to form a sentence fragment. Leading and trailing whitespace characters of the sentence fragments are deleted, and sentence fragments with a length of zero are discarded. The order of the sentence fragments in the text is preserved to form a sentence sequence. The sentence sequence is denoised to obtain a denoised sentence sequence. The denoising process includes hyperlink string deletion, HTML tag deletion, emoji deletion, and non-semantic noise tag deletion. Hyperlink string deletion is performed on consecutive character segments that match "http: / / " or "https: / / ". HTML tag deletion is performed on tag segments in the form of "<...>". Emoji deletion is performed on emoji characters. Non-semantic noise tag deletion is performed on tag segments that consist of repeated symbols and do not carry semantic content. After denoising, whitespace normalization is performed on the sentence again. The denoised sentence sequence is subjected to field standardization processing and then concatenated according to the field order to obtain the standardized requirement text. The field standardization processing performs field segmentation on the denoised sentence sequence to obtain a field set, which consists of field names and field values. The field names are processed to unify capitalization, whitespace normalization, and punctuation. The field values ​​are processed to unify units and formats while retaining the correspondence with the field names. The field order is determined by a preset field name sequence. The field values ​​are read sequentially according to the field order and concatenated. The concatenation uses a single space as a delimiter to generate the standardized requirement text.

[0021] In this embodiment, step two specifically includes: Retrieve the tag set, which contains tag identifiers and tag names. Tag identifiers are used to distinguish different tags in the tag set, and tag names are used to represent the semantic content of the tag and serve as the prompt text content field. The large model is defined as a pre-trained generative language model. The large model includes model parameters, lexicalization rules, and generation / decoding rules. The lexicalization rules convert the prompt text into a sequence of lexical units. The prompt text contains a character sequence and is mapped to a sequence of lexical numbers according to the lexicalization rules. The generation / decoding rules adopt an autoregressive generation method, generating the output sequence of lexical units one by one according to the position of the lexical units and stopping the generation when the termination marker condition is met. For each tag in the tag set, a trigger-generated prompt text is constructed. The trigger-generated prompt text includes the tag name and evidence output format constraints. The evidence output format constraints limit the output content to include a set of trigger phrases and a set of trigger sentences, and limit the set separator and the entry separator. The trigger-generated prompt text is input into the large model and the output word sequence of the large model is read. The output word sequence is decoded into output text and parsed according to the evidence output format constraints to obtain the set of trigger phrases and the set of trigger sentences. The set of trigger phrases and the set of trigger sentences constitute the tag-triggered semantic evidence set. In the tag-triggered semantic evidence set, the trigger phrases are phrase fragments used to enhance the tag semantic evidence, and the trigger sentences are sentence fragments used to enhance the tag semantic evidence. For each tag in the tag set, a suppression generation prompt text is constructed. The suppression generation prompt text includes the tag name and evidence output format constraints. The evidence output format constraints limit the output content to include a set of suppression phrases and a set of suppression sentences, and limit the set separator and the item separator. The suppression generation prompt text is input into the large model and the output word sequence of the large model is read. The output word sequence is decoded into output text and parsed according to the evidence output format constraints to obtain the set of suppression phrases and the set of suppression sentences. The set of suppression phrases and the set of suppression sentences constitute the tag suppression semantic evidence set. In the tag suppression semantic evidence set, the suppression phrases are phrase fragments used to intervene in the tag semantic evidence, and the suppression sentences are sentence fragments used to intervene in the tag semantic evidence. The evidence normalization process is performed on the tag-triggered semantic evidence set and the tag-suppressed semantic evidence set. The evidence normalization process performs word segmentation standardization, capitalization standardization, and punctuation standardization on each phrase and sentence entry, counts the occurrence of entries and deletes duplicates, filters and deletes entries whose length is less than the lower limit or greater than the upper limit based on the length threshold, and performs fixed order on the retained entries according to the preset sorting rules. The normalized tag-triggered semantic evidence set and the normalized tag-suppressed semantic evidence set are output.

[0022] In this embodiment, step three specifically includes: For each tag in the tag set, read the standardized requirement text, the tag trigger semantic evidence set, and the tag inhibition semantic evidence set. The tag trigger semantic evidence set includes the trigger phrase set and the trigger sentence set, and the tag inhibition semantic evidence set includes the inhibition phrase set and the inhibition sentence set. Based on the tag-triggered semantic evidence set, a semantically enhanced fragment sequence is generated. The semantically enhanced fragment sequence is generated according to a preset arrangement rule, which includes arranging the trigger phrase set first and then the trigger sentence set, arranging the trigger phrase set in the order of the items, and arranging the trigger sentence set in the order of the items. The trigger phrase items and trigger sentence items are written into the fragment sequence in sequence according to the preset arrangement rule. The standardized requirement text and the semantically enhanced fragment sequence are concatenated in the following order to generate the semantically enhanced text: the standardized requirement text first and the semantically enhanced fragment sequence second. The concatenation separator is a single space. A single space is inserted between adjacent entries in the semantically enhanced fragment sequence to form an enhanced tail string, which is then concatenated with the standardized requirement text to obtain the semantically enhanced text. A semantic intervention fragment sequence is generated based on the set of semantic evidence for suppression of labels. The semantic intervention fragment sequence is generated according to a preset arrangement rule. The preset arrangement rule includes arranging the set of suppression phrases first and then the set of suppression sentences, arranging the set of suppression phrases in the order of items, and arranging the set of suppression sentences in the order of items. The suppression phrase items and suppression sentence items are written into the fragment sequence in sequence according to the preset arrangement rule. Trigger fragment deletion processing is performed on the standardized requirement text. Trigger fragment deletion processing performs text matching and deletes matching fragments based on the set of trigger phrases and the set of trigger sentences. Text matching searches for trigger phrase entries and trigger sentence entries in the standardized requirement text one by one. The matching method is complete matching. When deleting matching fragments, the matching fragments are replaced with empty strings and the replaced text is normalized for whitespace characters to obtain the intervention base text. The semantic intervention text is generated by concatenating the base intervention text and the semantic intervention fragment sequence in the following order: base intervention text first, semantic intervention fragment sequence second. The concatenation separator is a single space. A single space is inserted between adjacent entries in the semantic intervention fragment sequence to form an intervention tail string, which is then concatenated with the base intervention text to obtain the semantic intervention text.

[0023] In this embodiment, step four specifically includes: Determine the number of tags in the tag set, establish a tag index sequence based on the tag identifier, arrange the tag index sequence according to the preset sorting rule of the tag identifier, and the number of tags is the number of elements in the tag index sequence; For each ordered tag pair in the tag index sequence, the ordered tag pair contains a first tag and a second tag. Read the semantic enhancement text and semantic intervention text corresponding to the first tag. Construct a tag rating prompt text containing the tag name of the second tag. The tag rating prompt text contains a rating object field and a rating output format field. Write the semantic enhancement text or semantic intervention text into the rating object field. Limit the output to a single value and limit the range of values ​​in the rating output format field. Input the tag rating prompt text into the large model to obtain the semantic enhancement text rating value and the semantic intervention text rating value. The directed influence strength is determined based on the difference between the semantically enhanced text score and the semantically intervened text score. The directed influence strength is used to characterize the degree of directed influence of the first label on the second label. The directed influence strength of the first label on the second label is written into the corresponding matrix element position of the directed influence matrix. The matrix row index is consistent with the position number of the first label in the label index sequence, and the matrix column index is consistent with the position number of the second label in the label index sequence. The complete directed influence matrix is ​​obtained by traversing all ordered label pairs in the label index sequence and filling the directed influence matrix. Determine the Trellis grid width and the number of Trellis grid layers. The Trellis grid width is a preset positive integer, and the number of Trellis grid layers is the smallest positive integer that satisfies the condition that the product of the Trellis grid width and the number of Trellis grid layers is not less than the number of labels. The Trellis grid width and the number of Trellis grid layers are used to define the number of columns and layers of the Trellis grid. The outward influence summary value of each label is calculated based on the directed influence matrix. The outward influence summary value is the sum of all matrix element values ​​of a label in the corresponding row of the directed influence matrix, excluding the matrix element values ​​of a label pointing to another label. Nodes are generated in descending order of the outward influence summary value. The node arrangement is a rearranged sequence of the label index sequence. The labels are filled into each layer and column of the Trellis grid according to the node arrangement to generate a node arrangement mapping relationship. The filling order is as follows: the first label in the node arrangement is filled into the first column of the first layer, and the same layer is filled in increments according to the column index. After reaching the width of the Trellis grid, the filling is switched to the first column of the next layer and continues until the number of labels is filled. The node arrangement mapping relationship includes the mapping record of label identifier to grid layer index and grid column index. The dependency edge set is determined, which satisfies the constraint that the layer index of the parent node is less than the layer index of the child node and the constraint that the in-degree of the child node is at an upper limit. The candidate parent node set consists of nodes whose layer index is less than the layer index of the child node. For each child node, the directed influence strength from each parent node to the child node is read in the candidate parent node set. The nodes are sorted in descending order of directed influence strength and the parent nodes with the upper limit of in-degree are selected to establish dependency edges. The dependency edge set and the node arrangement mapping relationship together constitute the Classifier Trellis structure.

[0024] In this embodiment, step five specifically includes: Obtain a set of standardized requirement texts and a set of label status annotations. The set of standardized requirement texts consists of standardized requirement texts arranged by text identifiers. The set of label status annotations contains the label truth value of each label in the label set on each standardized requirement text in the set of standardized requirement texts. The label truth value is binary, with 0 representing an invalid label and 1 representing a valid label. The node set, dependency edge set, and node arrangement mapping relationship are determined based on the Classifier Trellis structure. The node set is obtained by mapping the labels in the label set, and there is a one-to-one correspondence between the node identifier and the label identifier. The node arrangement mapping relationship includes the mapping records from the node identifier to the grid layer index and the grid column index. Based on the dependency edge set and the node arrangement mapping relationship, the parent node set corresponding to each node is determined. The parent node set consists of the parent nodes pointing to the nodes in the dependency edge set, and is arranged in ascending order according to the grid layer index corresponding to the parent node. When the grid layer index corresponding to the parent node is the same, it is arranged in ascending order according to the grid column index corresponding to the parent node. The arrangement order of the parent node set is a fixed arrangement order. For each node in the node set and for each standardized requirement text in the standardized requirement text set, read the corresponding label truth value from the parent node set from the label status annotation set, and write it into the position of each element of the vector in a fixed arrangement order of the parent node set to form the parent node label status vector. The dimension of the parent node label status vector is equal to the number of parent nodes contained in the parent node set. The standardized requirement text and the parent node label state vector are concatenated in a fixed order to form the node classification input. The fixed concatenation order includes the standardized requirement text first and the parent node label state vector second. The parent node label state vector is concatenated into the standardized requirement text in the form of a numerical sequence to form the node classification input. The node classification input is input into the node classifier to output the node conditional probability. The node conditional probability is the probability value of the node label truth value being 1. The training loss is calculated based on the true values ​​of the labels corresponding to the nodes and the conditional probabilities of the nodes in the label state annotation set, and the node classifier parameters are updated. The training loss is obtained by accumulating the standardized requirement texts in the standardized requirement text set one by one. The node classifier parameters are updated iteratively using gradient descent until the training termination condition is met, and the set of node classifiers that have been trained is output.

[0025] In this embodiment, step six specifically includes: Obtain the standardized requirement text, obtain the Classifier Trellis structure and read the node set, dependency edge set, node arrangement mapping relationship, directed influence matrix, and node classifier set. The node set is obtained by mapping the label set and corresponds one-to-one with the node classifier set. A forward decoding sequence is generated based on the node arrangement mapping relationship. The forward decoding sequence is sorted in ascending order according to the grid layer index and then in ascending order according to the grid column index within the same grid layer index. For each node in the forward decoding sequence, the set of parent nodes is determined based on the set of dependent edges and the parent node label state vector is generated by ascending the grid layer index corresponding to the parent node. The standardized requirement text and the parent node label state vector are combined in a fixed concatenation order to form the node classification input. The fixed concatenation order includes the standardized requirement text first and the parent node label state vector last. The node classification input is input into the node classifier to obtain the node conditional probability. The node conditional probability is used as the forward node confidence and written into the forward node confidence set. Based on the decision threshold, the node conditional probability is mapped to the node label state and written into the label state vector. The reverse backhaul sequence is generated based on the node arrangement mapping relationship. The reverse backhaul sequence is sorted in descending order according to the grid layer index and then sorted in descending order according to the grid column index within the same grid layer index. For each node in the backpropagation sequence, the set of child nodes is determined based on the set of dependent edges. The forward node confidence corresponding to the set of child nodes is read, and the directed influence strength of each child node in the directed influence matrix is ​​read. The product of the forward node confidence and the directed influence strength is calculated for each child node in the set of child nodes, and the product results of all child nodes are summed to obtain a weighted summary value. The weighted summary value is normalized to obtain the reverse node confidence and written into the reverse node confidence set. The marginal confidence is obtained by weighted fusion of forward node confidence and reverse node confidence based on fusion weight. The fusion weight ranges from 0 to 1, and a fusion weight value is assigned to each node in the node set. The marginal confidence is written into the marginal confidence vector according to the node index.

[0026] In this embodiment, step seven specifically includes: A label state vector is generated based on marginal confidence and a decision threshold. The decision threshold is a preset numerical threshold. For each label in the label set, the marginal confidence is read. If the marginal confidence is not less than the decision threshold, the label state is set to a valid state. If the marginal confidence is less than the decision threshold, the label state is set to an invalid state. The labels are written into the label state vector in the order of the label index sequence. The candidate label set for flipping is selected based on marginal confidence. The candidate label set for flipping consists of labels whose marginal confidence is between the lower threshold and the upper threshold. The lower threshold is a preset numerical threshold, and the upper threshold is a preset numerical threshold. The lower threshold is less than the upper threshold. During the selection, the relationship between the marginal confidence and the lower threshold and the upper threshold is judged for each label in the label set, and the labels that satisfy the condition that the marginal confidence is not less than the lower threshold and the marginal confidence is not greater than the upper threshold are written into the candidate label set for flipping. For each tag in the candidate tag set, perform a single-point flip to generate a candidate tag state vector. The candidate tag state vector is obtained by copying the tag state vector. The single-point flip is to invert the value of the vector element corresponding to the target tag in the candidate tag state vector. If the original vector element of the target tag is a valid state, it is flipped to an invalid state. If the original vector element of the target tag is an invalid state, it is flipped to a valid state. Determine the local recomputation range based on the set of dependency edges. The local recomputation range consists of the node corresponding to the target label and the set of child nodes of the node corresponding to the target label. The set of child nodes is determined based on the set of dependency edges. When the parent node in the set of dependency edges is the node corresponding to the target label and the child node is any node, the child node is written into the set of child nodes. Perform node conditional probability recomputation on the nodes within the local recomputation range based on the set of node classifiers and update the marginal confidence. When recomputing, read the label states corresponding to the set of parent nodes for each node within the local recomputation range and form a parent node label state vector. Combine the standardized requirement text and the parent node label state vector in a fixed concatenation order to form the node classification input and input it into the node classifier to obtain the node conditional probability. Write the node conditional probability into the forward node confidence set and perform reverse node confidence calculation and normalization according to the reverse propagation sequence. Fuse the forward node confidence and the reverse node confidence according to the fusion weight to obtain the marginal confidence and update the marginal confidence vector; Update the label state vector according to the delayed acceptance criterion. The delayed acceptance criterion is to accept the candidate label state vector when the consistency score of the candidate label state vector is not less than the consistency score of the label state vector and the sum of the marginal confidences of the candidate label state vector is not less than the sum of the marginal confidences of the label state vector. The consistency score is calculated based on the directed influence matrix and the label state vector. When calculating the consistency score, traverse all the matrix elements of the directed influence matrix and sum the values of the matrix elements that satisfy the condition that the parent label state is the valid state and the child label state is the valid state to obtain the consistency score. The sum of the marginal confidences is the result of summing the elements of the marginal confidence vector; Output the enterprise technology demand recognition result when the label state vector meets the stability determination condition. The stability determination condition is that the label state vectors obtained from two adjacent updates are the same. The consistency of two adjacent updates is judged by comparing the values of each element of the label state vector.

[0027] Embodiment 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the actual business scenario of automatic recognition of enterprise technology demands. This scenario comes from the enterprise technology demand texts continuously accumulated in the technology transfer service platform, the enterprise innovation demand solicitation system, and the industry-university-research docking platform. Such demand texts are usually filled in by enterprise technology leaders, project declaration personnel, or business personnel. The text lengths vary greatly, and the expression methods are significantly different. They contain both relatively standardized technical descriptions and a large number of colloquial and empirical expressions. At the same time, there are common problems such as cross-technology field mixed descriptions, unclear demand priorities, and fuzzy label boundaries. The platform has found in the long-term operation process that traditional methods based on keyword matching or single-model multi-label classification have problems such as unstable recognition results, missing associated labels, and simultaneous appearance of conflicting labels in this scenario, which directly affects subsequent technology matching, expert recommendation, and achievement transformation decisions.

[0028] In this scenario, the system first receives the original technical requirement text submitted by the enterprise. It then performs sentence segmentation, noise reduction, and field standardization on the text, transforming requirement content with different formats and expression habits into standardized requirement text with a consistent structure. After standardization, the system, based on a pre-defined set of tags, calls a pre-trained generative language model to generate a corresponding tag-triggered semantic evidence set and a tag-suppression semantic evidence set for each tag. The tag-triggered semantic evidence set describes semantic segments that should increase the probability of a tag's validity when they appear in the requirement text, while the tag-suppression semantic evidence set describes semantic segments that should decrease the probability of a tag's validity when they appear in the requirement text. Subsequently, the system constructs semantically enhanced text and semantically intervened text based on the standardized requirement text and the two sets of semantic evidence, respectively, ensuring that each tag has a clear source of semantic evidence in subsequent reasoning.

[0029] In the label dependency modeling stage, the system calculates the directed influence strength between label pairs based on semantically enhanced text and semantically intervened text, and writes the directed influence strength into a directed influence matrix. The system automatically determines the number and width of the Trellis grid based on the directed influence matrix, and generates the node arrangement and dependency edge set to construct the Classifier Trellis structure. In this structure, the directional dependency relationship between labels is explicitly expressed, avoiding the subjectivity of manually constructing the label hierarchy. In the model training stage, the system trains the node classifier based on the labeled historical demand text, so that when the node classifier receives standardized demand text and parent node label state as input, it outputs the node conditional probability. In the inference stage, the system performs forward decoding on the Classifier Trellis structure to obtain forward node confidence, and performs reverse consistency backpropagation in combination with the directed influence matrix to obtain the marginal confidence of each label.

[0030] To address the issue of unstable label states in boundary samples, the system further performs single-point flipping and local recalculation based on marginal confidence. It recalculates the node conditional probability only for the affected nodes and their child nodes, and updates the label state vector according to the delayed acceptance criterion, so that the label set gradually converges under the consistency constraint. When the label state vectors obtained from two adjacent updates are consistent, the system outputs the final enterprise technology requirement identification result. Through this process, the system can effectively reduce the phenomenon of unreasonable label combinations and conflicting label coexistence while maintaining inference efficiency.

[0031] To verify the beneficial effects of this invention, a comparative experiment was conducted using enterprise technology requirement text datasets of consistent origin and comparable size. The comparison methods included keyword-based methods, independent multi-label classification models, and structured inference methods without introducing a single-point flipping mechanism. The experiment statistically analyzed label recognition accuracy, label consistency index, average inference time, and manual review and correction rate. The results are shown in the table below: Table 1 Comparison of the effects of automatic identification of enterprise technology needs

[0032] The comparison results show that, without significantly increasing inference time, the present invention achieves significant improvements in both label recognition accuracy and label consistency. In particular, in samples involving cross-technical requirements, the proportion of conflicting labels appearing simultaneously has decreased significantly. The substantial reduction in the manual review and correction rate indicates that the enterprise technical requirement identification results output by the present invention are more in line with the cognitive expectations of business personnel and technical experts, reducing the cost of manual intervention. When continuously processing requirement texts in batches, the fluctuation range of the system's identification results is significantly smaller than that of the comparison method, demonstrating high stability and reproducibility.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatically identifying enterprise technology requirements based on a large model, characterized in that, Includes the following steps: Step 1: Obtain the enterprise's technical requirements text, perform sentence segmentation, noise reduction, and field standardization to obtain standardized requirements text; Step 2: Obtain the label set, and generate a set of semantic evidence for label triggering and a set of semantic evidence for label inhibition based on the large model; Step 3: Generate semantic intervention text and semantic enhancement text based on standardized requirement text, tag-triggered semantic evidence set, and tag-inhibited semantic evidence set; Step 4: Calculate the directed influence strength of each tag pair based on the semantically intervened text and the semantically enhanced text, forming a directed influence matrix. Based on the directed influence matrix, determine the number of Trellis grid layers, the Trellis grid width, the node arrangement, and the set of dependent edges to form the Classifier Trellis structure. Step 5: Train a node classifier based on the Classifier Trellis structure. The node classifier takes the standardized demand text and the parent node label state as input and outputs the node conditional probability. Step 6: Perform forward decoding based on the Classifier Trellis structure to obtain forward node confidence, perform reverse consistency backhaul to obtain reverse node confidence, and fuse forward node confidence and reverse node confidence to obtain marginal confidence; Step 7: Perform single-point flipping and local recalculation on the label based on marginal confidence, update the label status according to the delayed acceptance criterion, and output the enterprise technology requirement identification result when the stability judgment condition is met.

2. The method for automatic identification of enterprise technology requirements based on a large model according to claim 1, characterized in that, Step one specifically involves: Obtain the enterprise's technical requirements text, and perform character normalization processing on the enterprise's technical requirements text. Character normalization processing includes full-width character half-width conversion, whitespace character normalization, and control character deletion. The enterprise's technical requirements text is processed into a sentence sequence. The sentence segmentation is based on a set of delimiters, which includes periods, semicolons, exclamation marks, question marks, and line breaks. Denoising is performed on the sentence sequence to obtain a denoised sentence sequence. The denoising process includes removing hyperlink strings, HTML tags, emojis, and non-semantic noise tags. Perform field standardization on the denoised sentence sequence and concatenate them according to the field order to obtain the standardized requirement text.

3. The method for automatic identification of enterprise technology requirements based on a large model according to claim 1, characterized in that, Step two specifically involves: Retrieve the tag set, which includes tag identifiers and tag names; The large model is defined as a pre-trained generative language model. The large model includes model parameters, lexicalization rules, and generation / decoding rules. The lexicalization rules convert the prompt text into a sequence of lexical units. The generation / decoding rules generate the output sequence of lexical units word by word based on the sequence of lexical units in the prompt text using an autoregressive generation method. For each tag in the tag set, a trigger-generated prompt text is constructed. The trigger-generated prompt text includes the tag name and evidence output format constraints. The trigger-generated prompt text is input into the large model to output a set of trigger phrases and a set of trigger sentences. The set of trigger phrases and the set of trigger sentences constitute the tag-triggered semantic evidence set, which serves as the source of semantic evidence for the semantically enhanced text. For each label in the label set, a suppression generation prompt text is constructed. The suppression generation prompt text includes the label name and evidence output format constraints. The suppression generation prompt text is input into the large model to output a set of suppression phrases and a set of suppression sentences. The set of suppression phrases and the set of suppression sentences constitute the label suppression semantic evidence set, which serves as the source of semantic evidence for the semantic intervention text. Evidence normalization processing is performed on the tag-triggered semantic evidence set and the tag-inhibited semantic evidence set. Evidence normalization processing includes word segmentation normalization, case normalization, punctuation normalization, duplicate item deletion, length threshold filtering, and order fixing.

4. The method for automatically identifying enterprise technology requirements based on a large model according to claim 1, characterized in that, Step three specifically involves: For each tag in the tag set, read the standardized requirement text, the tag-triggered semantic evidence set, and the tag-inhibited semantic evidence set; A semantically enhanced fragment sequence is generated based on the set of semantic evidence triggered by the tags. The semantically enhanced fragment sequence is obtained by splicing together the set of trigger phrases and the set of trigger sentences in the order of arrangement. The standardized requirements text and the semantically enhanced fragment sequence are concatenated in the concatenation order to generate semantically enhanced text, with a single space as the concatenation separator. A semantic intervention fragment sequence is generated based on the set of semantic evidence for suppression of labels. The semantic intervention fragment sequence is obtained by splicing together the set of suppression phrases and the set of suppression sentences in the order of arrangement. The standardized requirement text is subjected to trigger fragment deletion processing. The trigger fragment deletion processing performs text matching and deletes the matched fragments based on the set of trigger phrases and the set of trigger sentence patterns to obtain the intervention base text. The semantic intervention text is generated by concatenating the base intervention text and the sequence of semantic intervention fragments in the concatenation order, with a single space as the concatenation separator.

5. The method for automatically identifying enterprise technology requirements based on a large model according to claim 1, characterized in that, Step four specifically involves: Determine the number of tags in the tag set, and establish a tag index sequence based on the tag identifiers; For each ordered tag pair in the tag index sequence, which contains a first tag and a second tag, read the semantic enhancement text and semantic intervention text corresponding to the first tag, construct a tag scoring prompt text containing the tag name of the second tag, and input the tag scoring prompt text into the large model to obtain the semantic enhancement text score value and the semantic intervention text score value. The intensity of directed influence is determined by the difference between the score of semantically enhanced text and the score of semantically intervened text, and the intensity of directed influence is written into the directed influence matrix according to the label index sequence. Determine the Trellis grid width and the number of Trellis grid layers. The number of Trellis grid layers must satisfy the condition that the product of the Trellis grid width and the number of Trellis grid layers is not less than the number of labels. The outward influence summary value of each label is calculated based on the directed influence matrix. The outward influence summary value is the summary result of the directed influence intensity of all labels except the first label in the label index sequence pointed to by the first label. The nodes are generated and arranged in descending order of the outward influence summary value. The labels are filled into each layer and column of the Trellis grid according to the node arrangement to generate the node arrangement mapping relationship; Determine the set of dependent edges, which satisfies the constraint that the layer index of the parent node is less than the layer index of the child node, and the constraint that the upper limit of the in-degree of the child node is satisfied. The set of candidate parent nodes consists of nodes whose layer index is less than the layer index of the child node. For each child node, select the upper limit number of parent nodes in the candidate parent node set according to the directed influence strength from large to small to establish dependent edges, forming the Classifier Trellis structure.

6. The method for automatically identifying enterprise technology requirements based on a large model according to claim 1, characterized in that, Step five specifically involves: Obtain a set of standardized requirement texts and a set of label status annotations. The set of label status annotations contains the truth value of each label in the label set on each standardized requirement text in the set of standardized requirement texts. Based on the Classifier Trellis structure, the node set, the dependent edge set, and the node arrangement mapping relationship are determined. Based on the dependent edge set and the node arrangement mapping relationship, the parent node set corresponding to each node is determined. The parent node set consists of the parent nodes pointing to the nodes in the dependent edge set and is arranged in ascending order according to the grid layer index corresponding to the node arrangement mapping relationship. For each node in the node set and for each standardized requirement text in the standardized requirement text set, extract the true values ​​of the corresponding tags from the tag state annotation set and form the parent node tag state vector according to the order of the parent node set. The standardized requirement text and the parent node label state vector are concatenated in a fixed order to form the node classification input. The node classification input is then input into the node classifier to output the node conditional probability. The training loss is calculated based on the true values ​​of the labels corresponding to the nodes and the conditional probabilities of the nodes in the label state annotation set, and the node classifier parameters are updated to obtain the set of node classifiers that have been trained.

7. The method for automatic identification of enterprise technology requirements based on a large model according to claim 1, characterized in that, Step six specifically involves: Obtain the standardized requirements text, obtain the Classifier Trellis structure and read the node set, dependency edge set, node arrangement mapping relationship, directed influence matrix, and node classifier set; Based on the node arrangement mapping relationship, a forward decoding sequence is generated by incrementing the grid layer index and the grid column index; For each node in the forward decoding sequence, the set of parent nodes is determined based on the set of dependent edges and the label state vector of the parent node is read. The standardized requirement text and the label state vector of the parent node are concatenated in a fixed order to form the node classification input. The input is given to the node classifier to obtain the node conditional probability. The node conditional probability is used as the forward node confidence. Based on the decision threshold, the node conditional probability is mapped to the node label state and written into the label state vector. Based on the node arrangement mapping relationship, a reverse propagation sequence is generated by decreasing the grid layer index and decreasing the grid column index; For each node in the backpropagation sequence, the set of child nodes is determined based on the set of dependent edges. The confidence of the forward node corresponding to the set of child nodes is read, and the directed influence strength from the node to each child node in the directed influence matrix is ​​read. The confidence of the forward node corresponding to the set of child nodes is weighted and normalized according to the directed influence strength to obtain the confidence of the reverse node. Marginal confidence is obtained by weighted fusion of forward node confidence and reverse node confidence based on fusion weight, with the fusion weight ranging from 0 to 1.

8. The method for automatic identification of enterprise technology requirements based on a large model according to claim 1, characterized in that, Step seven specifically involves: Generate label state vectors based on marginal confidence and decision thresholds; The candidate label set for flipping is selected based on marginal confidence. The candidate label set for flipping consists of labels whose marginal confidence is between the lower threshold and the upper threshold, where the lower threshold is less than the upper threshold. For each tag in the candidate tag set to be flipped, a single-point flip is performed to generate a candidate tag state vector. A single-point flip is to invert the state of the corresponding tag in the candidate tag state vector. The local recalculation range is determined based on the set of dependent edges. The local recalculation range consists of the node corresponding to the single-point flip label and the set of child nodes of the node corresponding to the single-point flip label. Based on the set of node classifiers, the node conditional probability is recalculated for the nodes within the local recalculation range and the marginal confidence is updated. The label state vector is updated according to the delayed acceptance criterion. The delayed acceptance criterion is to accept the candidate label state vector when the consistency score of the candidate label state vector is not less than the consistency score of the current label state vector and the marginal confidence summation value of the candidate label state vector is not less than the marginal confidence summation value of the current label state vector. The consistency score is calculated based on the directed influence matrix and the label state vector. The enterprise technology requirement identification result is output when the label state vector meets the stability judgment condition. The stability judgment condition is that the label state vectors obtained in two adjacent updates are consistent.

Images