An engineering cost text information extraction method based on semantic understanding

Abstract

The application discloses a kind of engineering cost text information extraction methods based on semantic understanding, it is related to information extraction technical field, including, to the context related semantic feature data is eliminated ambiguity, obtains term semantic category label result, term semantic category label result and corresponding context fragment are as node, constructs candidate relationship graph, candidate relationship graph is input hierarchical perception's graph reasoning network, the entity type of node and the relationship type between nodes are identified, according to the entity type of node and the relationship type between nodes, top-down construction bill of quantities tree, obtains engineering cost text information;The application overcomes the problem of homonymy and expression diversity effectively through text information extraction method, realizes from unstructured text to clear structure, semantic explicit end-to-end conversion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information extraction, and in particular to an engineering cost text information extraction method based on semantic understanding. BACKGROUND

[0002] With the deepening of new models such as building informatization, intelligent construction and whole-process engineering consulting, the quantity and complexity of engineering cost documents have been significantly improved. Engineering cost documents cover various types such as bidding documents, bill of quantities, quota instructions, contract terms and change warrants, and carry key business information such as material specifications, construction technology, measurement rules and pricing basis. In order to solve the problems of low efficiency and easy errors caused by manual processing, in recent years, with the development of deep learning methods, semantic understanding methods based on BiLSTM, Transformer architecture and pre-trained language models have been widely used in engineering text analysis, further expanding the depth and breadth of information extraction.

[0003] However, the existing text information extraction method still needs to be improved. First, it only relies on surface lexical matching or general semantic models, which limits the accuracy of entity recognition. Second, the existing information extraction model usually ignores the logical structure and semantic hierarchy of the document, resulting in a lack of computational and business usability of the extraction results. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an engineering cost text information extraction method based on semantic understanding to solve the problems of limited entity recognition accuracy, lack of computational and business usability.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides an engineering cost text information extraction method based on semantic understanding, which comprises,

[0008] Receiving engineering cost documents and preprocessing to obtain structured document data;

[0009] Performing multi-granularity segmentation on the structured document data to obtain character-level text data, an engineering cost field term set and a context fragment;

[0010] Performing semantic anchoring processing on each field term in the engineering cost field term set to obtain semantic anchoring features;

[0011] Performing semantic modeling on the context fragment and the semantic anchoring features to obtain context-related semantic feature data;

[0012] Disambiguation is performed on the context-related semantic feature data to obtain a term semantic category annotation result;

[0013] The term semantic category annotation result and the corresponding context segment are taken as nodes to construct a candidate relation graph;

[0014] The candidate relation graph is input into a hierarchical-aware graph reasoning network to identify the entity type of the nodes and the relation type between the nodes;

[0015] According to the entity type of the nodes and the relation type between the nodes, an engineering quantity list tree is constructed from top to bottom to obtain the engineering cost text information.

[0016] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the preprocessing includes cleaning and standardization.

[0017] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the multi-granularity segmentation is performed on the structured document data to obtain character-level text data, an engineering cost field term set and context segments, and specifically:

[0018] The text content in the structured document data is segmented in character order to obtain the character-level text data;

[0019] An engineering cost field core dictionary is constructed, and a forward maximum matching scan is performed on the character-level text data to identify and extract all matched words to form the engineering cost field term set;

[0020] According to the paragraph mark, the project number sequence and the table boundary recorded in the structured document data, the continuous character-level text data is divided into logically independent semantic units to obtain the context segments.

[0021] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the semantic anchoring processing is performed on each field term in the engineering cost field term set to obtain semantic anchoring features, and specifically:

[0022] Based on the engineering cost field core dictionary, the standard semantic anchor point of each field term in the engineering cost field is queried;

[0023] After the standard semantic anchor point is determined, the semantic alignment operation is performed on the field term and the corresponding standard semantic anchor point to obtain semantic constraint information;

[0024] The semantic constraint information is numerically encoded and combined to obtain the semantic anchoring features.

[0025] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the context segment and the semantic anchor feature are subjected to semantic modeling to obtain context-related semantic feature data, specifically:

[0026] The context segment is serialized and encoded to obtain a context segment sequence vector.

[0027] The context segment sequence vector and the semantic anchor feature are spliced in the feature dimension to form a joint semantic vector.

[0028] The joint semantic vector is input into a multilayer perception machine for nonlinear transformation and deep feature fusion to obtain a dense vector, and the dense vector is taken as the context-related semantic feature data.

[0029] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the context-related semantic feature data is subjected to ambiguity resolution to obtain a term semantic category labeling result, specifically:

[0030] The context-related semantic feature data is input into an ambiguity resolution classifier for forward propagation calculation to obtain a probability distribution vector, and a semantic category set is constructed.

[0031] The elements in the probability distribution vector are mapped back to the semantic category set to retrieve corresponding specific categories.

[0032] The retrieved specific categories are taken as the term semantic category labeling result.

[0033] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the probability distribution vector refers to the context-related semantic feature data input into the ambiguity resolution classifier, and the ambiguity resolution classifier is subjected to linear transformation and nonlinear mapping by each hidden layer of the ambiguity resolution classifier, and then normalized by the Softmax function of the output layer to obtain.

[0034] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding, the term semantic category labeling result and the corresponding context segment are taken as nodes to construct a candidate relationship graph, specifically:

[0035] A graph node is created for the term semantic category labeling result and the corresponding context segment.

[0036] The graph node includes a term semantic category labeling result node and a context segment node.

[0037] According to the original position and hierarchical information recorded in the structured document data, a structure edge is established between the context segment nodes.

[0038] A semantic anchor edge representing a subordinate relationship is established between the term semantic category annotation result node and the context segment node;

[0039] A syntax edge is established for the term semantic category annotation result nodes having a direct syntax dependency relationship and appearing in the same context segment;

[0040] All created graph nodes, structure edges, semantic anchor edges and syntax edges are combined to form a complete candidate relationship graph.

[0041] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding provided in the application, the candidate relationship graph is input into the hierarchical perception graph reasoning network, and the entity type of the node and the relationship type between the nodes are identified, and specifically:

[0042] The candidate relationship graph is input into the hierarchical perception graph reasoning network, the node features of the candidate relationship graph are aggregated and propagated by a graph attention network layer of the graph reasoning network, and an updated node is generated;

[0043] After the updated node is input into an entity type classifier and a relationship type classifier, forward calculation is performed to obtain the entity type of the node and the relationship type of the edge.

[0044] As a preferred scheme of the engineering cost text information extraction method based on semantic understanding provided in the application, the engineering quantity list tree is constructed from top to bottom according to the entity type of the node and the relationship type between the nodes to obtain the engineering cost text information, and specifically:

[0045] Based on the relationship type of the edge, edges belonging to a hierarchical inclusion relationship are screened in the candidate relationship graph to obtain hierarchical relationship edges;

[0046] The root node of the engineering quantity list tree is identified based on the hierarchical relationship edges, and the parent nodes of all nodes are identified in the candidate relationship graph taking the root node as a starting node;

[0047] The nodes are sequentially connected with the corresponding parent nodes from top to bottom to generate the engineering quantity list tree, and the engineering quantity list tree is taken as the engineering cost text information.

[0048] The present application has the beneficial effects that: by introducing a semantic anchoring mechanism, the engineering cost terms are accurately mapped to the standard semantics in the professional context, effectively overcoming the problems of homonymy and expression diversity, combining the deep fusion modeling of context and anchoring characteristics, dynamically generating a semantic representation with business awareness, and then through the ambiguity resolution classifier oriented to the closed category system, realizing high-precision classification of terms to computable business labels, and finally relying on the hierarchical perception graph reasoning network, automatically identifying the entity type and hierarchical inclusion relationship on the basis of preserving the original logical structure of the document, and constructing a complete structure and clear semantics of the bill of quantities tree from top to bottom, realizing the end-to-end conversion from unstructured text to structured and clear semantics. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0050] Figure 1 The flowchart of the method for extracting engineering cost text information based on semantic understanding.

[0051] Figure 2 The flowchart for obtaining engineering cost text information.

[0052] Figure 3 The flowchart for obtaining semantic anchoring features.

[0053] Figure 4 The flowchart for obtaining context-related semantic feature data. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings.

[0055] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0056] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0057] Referring to Figures 1-4 For an embodiment of the present application, the embodiment provides a semantic understanding-based engineering cost text information extraction method, comprising the following steps:

[0058] S1, receiving engineering cost documents and preprocessing to obtain structured document data.

[0059] It should be noted that the received engineering cost documents include PDF format, scanned image format and DOCX format. The global fixed threshold method is applied to the binary operation of the scanned image format engineering cost document. Specifically, the gray value of each pixel in the image is compared with the global fixed threshold value, and the pixels with a gray value higher than the global fixed threshold value are set to white, and the pixels with a gray value lower than or equal to the global fixed threshold value are set to black, to generate a binary image. Perform connected component analysis based on pixel connectivity on the binary image. Specifically, scan the binary image and label the connected black pixel regions, identify all independent connected regions, filter out noise connected regions with small area (less than two to three standard deviations of the average pixel area), abnormal shape or overlapping position, and keep all connected regions that meet the character geometry and distribution characteristics as independent character regions. Match each character region with each character template in the character template library at the pixel level; combine the matching successful character template and the corresponding character into a character sequence according to the spatial coordinate order of the character region in the image; record the position coordinates and character format of each character in the character sequence.

[0060] For engineering cost documents in editable formats such as PDF and DOCX, use the document parsing function to extract text, character format, paragraph structure, table cell data and list level information. After extraction, perform data cleaning. Specifically, remove the header and footer and irrelevant marks in the text, correct the misspelled words generated in the character recognition process, convert full-width characters to half-width characters, and unify the expression format of dates, currencies and numbers to standard format. After integrating the cleaned text, table row and column structure, paragraph structure, list numbering system and character format, perform unified coding to generate structured document data.

[0061] The construction process of the character template library is as follows: high-quality character image samples (including Songti, Simhei, Fanti and Kaishu) covering commonly used fonts of engineering cost documents are collected. The collected character image samples are preprocessed (including gray-scale, binarization, size normalization to a unified pixel size, and gravity center centralization). For each target character in the character coding system, such as all Chinese characters, letters, numbers and punctuation marks in the GB2312 character set, multiple clear and non-deformed instance images are selected from the preprocessed character image samples. The selected target characters and instance images are aligned and averaged at the pixel level to generate a bitmap template of the standard form of the character; the standard character code corresponding to each bitmap template, the font category and the basic structure feature vector are recorded, and all generated bitmap templates are stored to form a complete character template library.

[0062] The global fixed threshold is set according to the gray scale statistical characteristics of the scanned image (engineering cost document), and the value range is between 80 and 220; for example, when the gray scale mean of the scanned image is 80-120, the global fixed threshold can be set in the interval of 60-100; when the gray scale mean of the scanned image is 170-210, the global fixed threshold can be set in the interval of 160-220; when the scanned image gray scale distribution presents a bimodal structure, the global fixed threshold can be set in the middle interval between the two peak values.

[0063] S2, the structured document data is divided into multiple granularities to obtain character-level text data, engineering cost field term set and context fragment.

[0064] S2.1, if necessary, read the text string of the structured document data. Starting from the first character position of the text string, read and separate the characters one by one according to the storage order of the characters in the string. Each separated independent character and the corresponding order index number in the text string are taken as a minimum character unit. All separated and recorded character units are arranged according to the order index number to form a character sequence. The character sequence is the character-level text data.

[0065] S2.2, if necessary, collect and integrate a wide range of professional documents in the field of engineering cost (including engineering quota standards, historical cost list and bidding contract template). Clean and organize the collected professional documents, extract all professional vocabulary, phrases and fixed expressions, and form an initial term candidate list. Remove duplicates and classify the initial term candidate list to form a core dictionary in the field of engineering cost. Specifically, perform duplicate detection based on complete string matching on the initial term candidate list, and merge term entries with identical character sequences into one entry. At the same time, the definition with the highest frequency in the professional documents is merged as the initial standard interpretation. Note that the initial standard interpretation at this time is not the final interpretation, which will be corrected in the subsequent S5 ambiguity elimination process combined with context-related semantic features. After completing the deduplication operation, design a classification framework (including materials and equipment, construction process, cost composition, measurement unit, specification standard, etc. Main categories) according to the knowledge architecture of the field of engineering cost. According to the semantic, common context and role in cost calculation of the deduplicated term entries, they are attributed to the preset classification framework. Integrate all classified term entries, encode and index them in alphabetical order of pinyin, and generate an electronic entry library, i.e. a core dictionary in the field of engineering cost.

[0066] All term entries in the core dictionary of the field of engineering cost are arranged in descending order according to the length of the characters. Starting from the beginning of the character-level text data, the current unmatched character is selected as the starting point in turn; from the starting point, a substring equal in length to the longest term entry is cut from the character-level text data. In the core dictionary of the field of engineering cost, find the term entry that completely matches the substring. If the search is successful, this term entry is identified as a successfully matched word, and the content and position in the character-level text data are recorded. If the search fails, shorten the length of the cut substring and repeat the search process until a matching term entry is found or the length of the substring is reduced to a single character. For the case of a single character that is not successfully matched in the core dictionary of the field of engineering cost, mark this character as an unidentified unit and move the scanning starting point. After traversing the entire character-level text data, all successfully matched words are summarized to form a term set in the field of engineering cost.

[0067] S2.3, if applicable, parse the document format information stored in the structured document data, identify and locate all paragraph markers, item number sequences, and table boundaries. For paragraph markers, take the start position of each paragraph marker as the starting point of a context fragment, and take the end position of the paragraph marker as the termination point of the context fragment, and cut out the continuous character-level text data between the starting point and the termination point to form a context fragment in units of paragraphs. For item number sequences, take the start position of each item number and the corresponding attached text content as the starting point of a context fragment, and take the start position of the next item number or the end position of the current chapter as the termination point of the context fragment, and cut out the continuous character-level text data between the starting point and the termination point to form a context fragment in units of item number entries. For table boundaries, directly cut out the continuous character-level text data contained in each independent table cell to form a context fragment in units of table cells.

[0068] S3, perform semantic anchoring processing on each domain term in the engineering cost domain term set to obtain semantic anchoring features. The semantic anchoring feature flowchart is as shown in Figure 3

[0069] If applicable, traverse each domain term in the engineering cost domain term set; take the domain term as a query keyword and perform an exact match search in the term entry index of the engineering cost domain core dictionary. Specifically, perform a character-level complete comparison between the query keyword and each term entry. When the query keyword is completely consistent with the standard text expression of a certain term entry, it is determined that the match is successful. After the match is successful, read and return the initial standard explanation associated with the corresponding term entry from the engineering cost domain core dictionary. The returned initial standard explanation is identified as the standard semantic anchor point of the domain term in the engineering cost domain. If the match fails, the standard semantic anchor point of the domain term is marked as "undefined". When there are multiple standard explanations corresponding to a domain term in the engineering cost domain core dictionary, the multiple standard explanations associated with the domain term can be taken as standard semantic anchor points respectively.

[0070] ​The text strings of the field terms and the text strings of the standard semantic anchors are respectively vectorized, specifically, the text strings of the historical field terms and the corresponding text strings of the historical standard semantic anchors are collected in pairs from historical engineering cost documents and annotated materials; the text strings of the historical field terms and the text strings of the historical standard semantic anchors are spliced, and the spliced strings are converted into a vocabulary index sequence and an attention mask sequence by a BERT model-specific tokenizer. The vocabulary index sequence and the attention mask sequence are input into the BERT model for training. During the training process, the BERT model performs forward calculation through a multi-layer bidirectional Transformer structure to obtain a text string representation vector of the historical field terms and a text string representation vector of the historical standard semantic anchors, respectively. Based on the text string representation vector of the historical field terms and the text string representation vector of the historical standard semantic anchors, a semantic matching loss function is constructed. The semantic matching loss function can be a mean square error loss function, which realizes the constraint of semantic consistency by minimizing the distance between the vector representations of the matched field terms and standard semantic anchors. The gradient of the semantic matching loss function with respect to the parameters of the BERT model is calculated by a backpropagation algorithm, and the Adam optimizer is used to iteratively update the parameters of the BERT model (the learning rate can be set to ~ , and the batch size can be set to 8-32 according to the size of the training set); the above forward calculation, loss calculation and parameter update process is repeated; when the difference in the semantic matching loss function between adjacent two iteration processes is less than a convergence threshold, the training of the BERT model is completed, and a trained BERT model is obtained.

[0071] The convergence threshold is set to avoid the risk of overfitting and waste of computing resources caused by infinite training process, for example, 1e-5. The value is based on the fact that when the performance of the BERT model no longer improves, the training is stopped.

[0072] The text string of the domain term is converted into a domain term vocabulary sequence, the text string of the standard semantic anchor is converted into a standard semantic anchor vocabulary sequence, and [CLS] and [SEP] markers are inserted at the beginning and end of each sequence respectively using the tokenizer of the BERT model. The corresponding paragraph type sequence and attention mask sequence are created for the domain term vocabulary sequence and the standard semantic anchor vocabulary sequence respectively. The domain term vocabulary sequence, the corresponding paragraph type sequence and the attention mask sequence are taken as one input group; the standard semantic anchor vocabulary sequence, the corresponding paragraph type sequence and the attention mask sequence are taken as another input group; the two input groups are respectively input into the trained BERT model, the embedding layer of the BERT model adds the three sequences of the input group to generate a combined embedding vector; the combined embedding vector is input into the first Transformer encoder layer of the BERT model, the Transformer encoder layer performs multi-head self-attention calculation on the combined embedding vector, sums the calculated result with the combined embedding vector and performs layer normalization; the layer normalization result is input into the feedforward neural network for nonlinear transformation; the nonlinear transformation result is connected with the layer normalization result through residual connection, and layer normalization is performed again to obtain the output representation of the first Transformer encoder layer; the output representation of the first Transformer encoder layer is taken as input and transmitted to the next Transformer encoder layer of the BERT model, and the same multi-head self-attention calculation, residual connection, layer normalization and feedforward neural network processing are repeatedly performed; after the processing of the last Transformer encoder layer, the sequence representation of the domain term and the sequence representation of the standard semantic anchor are output. For the sequence representation of the domain term and the sequence representation of the standard semantic anchor, the hidden state vector corresponding to the [CLS] marker position is extracted, that is, the domain term semantic vector and the standard semantic anchor semantic vector.

[0073] The cosine similarity value between the domain term semantic vector and the standard semantic anchor semantic vector is calculated, and the formula is:

[0074] ;

[0075] wherein, represents the cosine similarity value, represents the domain term semantic vector, represents the standard semantic anchor semantic vector.

[0076] The cosine similarity value is directly taken as the semantic consistency strength score; word segmentation and part-of-speech tagging are performed on the standard semantic anchor point text to obtain a word sequence; the word sequence is input into a dependency syntax analyzer to identify the dependency relationship between the words; according to the dependency relationship between the words, the words representing actions and behaviors are identified from the word sequence and taken as predicates; the dependency relationship between the predicate and the rest of the words (words other than the predicate in the word sequence) is identified, when the dependency relationship is a verb-object relationship, the corresponding word is determined as a process object argument; when the dependency relationship is a subject-predicate relationship, the corresponding word is determined as an agent argument; when the dependency relationship is a modifier-noun relationship, the corresponding word is determined as an attribute modifier argument; when the dependency relationship is a condition-noun relationship, the corresponding word is determined as a condition argument; when the dependency relationship is a parallel relationship, the corresponding word is determined as a same-type parallel argument; at the same time, when multiple argument roles are identified in the standard semantic anchor point text, all the multiple argument roles can be retained; the predicate and the argument roles of the standard semantic anchor point text form a structured semantic framework. For the domain terminology text and the semantic framework, check whether the words in the domain terminology text contain the predicate and the argument roles in the semantic framework, and record the matching successful predicate and argument roles as structured constraints. The semantic consistency strength score and the structured constraints jointly form semantic constraint information.

[0077] The semantic constraint information includes the semantic consistency strength score and the structured constraints. The semantic consistency strength score itself is a scalar value between -1 and 1, which is directly taken as part of the numerical encoding. For the structured constraints, a unique integer identifier is assigned to each predicate and argument role extracted in the semantic framework, and each identifier is mapped to a fixed-dimensional dense vector through a learnable embedding lookup table. The mapping operation is the numerical encoding of the structured constraints. Specifically, the core predicate and argument roles are extracted from the structured constraints; a continuous integer identifier is assigned to each unique predicate in the structured constraints, and a continuous integer identifier is assigned to each unique argument role type, ensuring that all predicate identifiers and argument role identifiers together form a non-repeating integer identifier sequence. A learnable embedding lookup table is created, which has a row number equal to the total number of integer identifiers and a column number equal to the preset feature vector dimension. During the numerical encoding process, when a predicate or argument role needs to be processed, the corresponding row vector parameter is indexed from the embedding lookup table according to the assigned integer identifier, which is the dense vector of the predicate or argument role. The dense vector is taken as the numerical encoding result of the structured constraint.

[0078] The preset feature vector dimension is set according to the semantic complexity of the predicate and argument role and the input requirements of the subsequent semantic modeling network, for example, 128, which is valued according to the semantic matching effect and computational complexity under different dimensions.

[0079] The semantic consistency strength score is concatenated with all dense vectors to form a combined feature vector.

[0080] This process involves collecting and organizing historical semantic constraint information and corresponding true semantic anchoring feature labels generated during the processing of historical engineering cost data. These true semantic anchoring feature labels originate from historical engineering cost documents that have undergone cost review or settlement. These historical engineering cost documents include bill of quantities, quota specifications, tender control price specifications, approved settlement statements, and their corresponding cost review documents. The standard semantic categories, engineering attributes, and semantic roles of engineering cost terminology within these historical engineering cost documents can serve as the basis for true semantic anchoring feature labels.

[0081] Historical semantic constraint information is converted into numerical historical combined feature vectors. These vectors are then input into a multilayer perceptron for training. During training, the mean squared error loss function (MSE) between the semantic anchor features predicted by the MSE and the true semantic anchor features is calculated. The gradient of the MSE loss function with respect to all weight and bias parameters in the MSE is calculated using the backpropagation algorithm. A stochastic gradient descent optimizer is then used to update the weight and bias parameters of the MSE based on the calculated gradients. This iterative process of forward propagation, loss calculation, backpropagation, and parameter update is repeated until the maximum number of iterations (e.g., 200) is reached, resulting in the trained MSE.

[0082] The maximum number of iterations is set based on the convergence characteristics of the loss function and computational resources during the training of the multilayer perceptron; the value is determined by selecting the corresponding number of iterations when the loss of the multilayer perceptron tends to stabilize.

[0083] The combined feature vector is input into the trained multilayer perceptron. The first fully connected layer of the multilayer perceptron performs a linear transformation on the combined feature vector. Specifically, it multiplies the combined feature vector with the weight matrix of the first fully connected layer and adds a bias vector to generate a first intermediate vector. The first intermediate vector is then input into the ReLU activation function. The ReLU activation function performs a non-linear mapping on each element of the first intermediate vector, setting elements with values ​​less than zero to zero and leaving elements with values ​​greater than or equal to zero unchanged, generating an activated first intermediate vector. The activated first intermediate vector is then input into the second fully connected layer of the multilayer perceptron, where the linear transformation and non-linear mapping are repeated, outputting a second intermediate vector. This second intermediate vector is the semantic anchoring feature. Semantic anchoring features are continuous vector representations in the normative semantic space of the engineering cost field. Their effect is to explicitly encode the semantic consistency and role constraints between the original terms and standard semantic anchors into learnable features.

[0084] S4. Perform semantic modeling on the context fragments and semantic anchoring features to obtain context-related semantic feature data. The flowchart for obtaining context-related semantic feature data is as follows: Figure 4 As shown.

[0085] S4.1 It should be noted that, according to the original order of appearance of characters or words in the text, the context segment is sequentially expanded to form an ordered sequence of symbols; each symbol in the symbol sequence is assigned a unique numerical code to form a symbol-level numerical representation; according to the sequential position of the symbols in the context segment, corresponding sequential position information is added to the symbol-level numerical representation, so that the symbol-level numerical representation includes both symbol content information and sequential position information; based on the adjacent relationship of symbols in the ordered sequence, a weighted cumulative operation is performed on the symbol-level numerical representations corresponding to adjacent symbols to generate a continuous numerical sequence, which can reflect the overall semantic order characteristics of the context segment. Weighted summation, averaging, or maximum value selection operations are performed on the continuous numerical sequence to fuse the sequential semantic information of different positions in the continuous numerical sequence, and the resulting multidimensional numerical result forms the context segment sequence vector.

[0086] Perform a dimensionality consistency check on the context fragment sequence vector and semantic anchor features to ensure that the context fragment sequence vector and semantic anchor features meet the concatenation requirements in terms of numerical representation. According to the feature dimension arrangement order, connect the numerical values ​​of each dimension of the context fragment sequence vector and the numerical values ​​of each dimension of the semantic anchor features in sequence along the feature dimension direction to form a joint semantic vector.

[0087] S4.2 It should be noted that the joint semantic vector is input into a retrained multilayer perceptron (which has the same network structure as the multilayer perceptron used to generate semantic anchoring features in step S3, but the two are trained separately for different semantic modeling tasks, with independent parameters and no shared weights), and nonlinear transformation and deep feature fusion are performed to obtain a dense vector, which is used as context-related semantic feature data.

[0088] S5. Perform ambiguity removal on context-related semantic feature data to obtain term semantic category annotation results.

[0089] S5.1 Input the context-related semantic feature data into the ambiguity disambiguation classifier for forward propagation calculation to obtain the probability distribution vector.

[0090] It should be noted that historical context-related semantic feature data is collected and paired with real semantic category labels to form a training set;

[0091] The training set is input into the ambiguity resolution classifier for training. During the training process, the cross-entropy loss between the predicted probability distribution of the ambiguity resolution classifier and the true semantic class label (encoded by one-hot) is calculated. At the same time, the gradient of the cross-entropy loss with respect to all parameters of the ambiguity resolution classifier is calculated by the backpropagation algorithm, and the optimizer (such as Adam) is used to update the parameters of the ambiguity resolution classifier according to the gradient. This process is repeated until the maximum number of iterations (e.g. 150) of the ambiguity resolution classifier is reached, and the trained ambiguity resolution classifier is obtained.

[0092] The maximum number of iterations of the ambiguity resolution classifier is set according to the convergence characteristics of the cross-entropy loss function in the classification task, and is valued according to the trend of the classification loss with the number of iterations.

[0093] The context-related semantic feature data is input into the trained ambiguity resolution classifier. The context-related semantic feature data passes through the first hidden layer and the second hidden layer of the ambiguity resolution classifier in turn, and each layer performs linear transformation followed by ReLU non-linear activation to obtain the activation value for classification. The activation value is multiplied by the weight matrix of the output layer and then added to the bias vector of the output layer to generate the original logic value vector. The original logic value vector is normalized by the Softmax function to obtain the probability distribution vector.

[0094] Although both the multi-layer perceptron and the ambiguity resolution classifier use a feedforward neural network structure containing fully connected layers and non-linear activation functions, there are essential differences between them in terms of technical purpose, network configuration, and training objectives. The multi-layer perceptron aims to fuse and map heterogeneous features into a low-dimensional semantic feature space, and the output layer dimension is a fixed value. The goal of training or optimization is feature learning. The ambiguity resolution classifier is a standard classification network, where the input layer receives a dimension of 512, the first hidden layer is a fully connected layer containing 256 neurons followed by a ReLU activation function, the second hidden layer is also a fully connected layer containing 128 neurons followed by a ReLU activation function, and the output layer has the same number of neurons as the number of semantic classes (e.g. 20 neurons). The Softmax function is used to output the probability distribution, and the goal of training is to maximize the classification accuracy. The multi-layer perceptron and the ambiguity resolution classifier are in a serial relationship, functionally independent, and do not share parameters.

[0095] The maximum number of iterations is to avoid the waste of computing resources or the risk of overfitting caused by the infinite increase of the number of iterations. The value is determined based on the condition that the performance of the ambiguity resolution classifier no longer improves in consecutive multiple iteration periods, and the number of iterations before reaching this condition is taken as an empirical value.

[0096] S5.2、Should be noted that, by analyzing historical engineering cost documents, standard quota library and industry standards, extract high-frequency occurrence, with clear semantic boundaries of the term type, form the initial semantic category list. The initial semantic category list is merged with semantic overlap, and the category is too broad, for example, an initial "cost-related" category is divided into "material cost", "labor cost", "mechanical use fee", "measures project cost", "regulatory fee" and "taxes" and other specific categories, forming a semantic category system. Assign a unique, clear text label to each category in the semantic category system (for example, "material name", "engineering quantity", "unit of measurement", "construction process", "cost index", "fee item"), and all text labels are summarized to form a structured, closed list, which is the semantic category set. It should be noted that the size of the semantic category set needs to be the same as the dimension of the probability distribution vector, and the dimension of the probability distribution vector corresponds one-to-one to the arrangement order of the categories in the semantic category set.

[0097] Because the size of the semantic category set is the same as the dimension of the probability distribution vector, each element in the probability distribution vector has an index corresponding to a specific category in the semantic category set. Traverse all elements in the probability distribution vector, record the index value of the element in the probability distribution vector; use the index value as a lookup key to locate in the semantic category set; get the text label at the specified position of the index value, that is, the specific category. The specific category is used as the term semantic category annotation result.

[0098] S6、The term semantic category annotation result and the corresponding context fragment are used as nodes to construct a candidate relationship graph.

[0099] S6.1、Should be noted that, for each term semantic category annotation result, assign a unique node identifier, and create a graph node data structure, mark the graph node data structure as a term semantic category annotation result node. For each context fragment, assign a unique node identifier, and create a graph node data structure, mark the graph node data structure as a context fragment node. The node identifiers of the term semantic category annotation result node and the context fragment node belong to different namespaces to ensure the global uniqueness of all graph node identifiers in the entire graph structure.

[0100] S6.2, If applicable, extract the logical position information of each context fragment in the engineering cost document (including the section number, the paragraph number, the bullet level and the order in the document flow) from the structured document data. Traverse all the created context fragment nodes, compare the logical position information of any two context fragment nodes. When the logical position information of two context fragment nodes satisfy a certain relationship (one is enough), a directed or undirected connection edge is created between the two context fragment nodes, and is marked as a structure edge.

[0101] It should also be noted that the certain relationship includes two kinds: the first is the direct parent-child hierarchical relationship, that is, the section number or the bullet level of one context fragment node is the direct superior of the other context fragment node; the second is the continuous sequence relationship, that is, two context fragment nodes are in adjacent positions in the document flow and belong to the same logical level. The direct parent-child hierarchical relationship between two context fragment nodes uses a directed edge; the continuous sequence relationship between two context fragment nodes uses an undirected edge.

[0102] S6.3, If applicable, for each term semantic category annotation result node, locate one or more positions of the term text in the character-level text data according to the term text recorded in the term semantic category annotation result node. According to each occurrence position of the term text, determine the minimum context fragment including this position. In the created context fragment nodes, find the context fragment node whose text content completely matches the minimum context fragment. Between the found context fragment node and the term semantic category annotation result node, a directed edge (from the term semantic category annotation result node to the context fragment node) is created, and is marked as a semantic anchor edge.

[0103] S6.4、It should be noted that the complete text content is read from the context fragment node. The complete text content is input into the dependency syntax analyzer, which tokenizes and part-of-speech tags the complete text content string, and divides the complete text content into a sequence of words with part-of-speech tags. The dependency syntax analyzer analyzes the grammatical modification relationship between each pair of words in the word sequence according to the grammar rule library, determines a core word for each word (dependent word), and labels the dependency relationship type, to obtain an analysis result. Specifically, the dependency syntax analyzer automatically determines and assigns a core word for each word, while determining the dependency relationship type (such as “subject-predicate relationship”, “verb-object relationship”, “determiner-noun relationship”, etc.) between the core word and the dependent word. The core word is the central component in the sentence that grammatically governs or modifies other words. For example, in the phrase “pouring concrete”, the dependency syntax analyzer will determine that “pouring” is the core word (verb) and “concrete” is the dependent word (noun, as the object of “pouring”), and label them as “verb-object relationship”; the analysis result is presented in a set form, and each element in the dependency relationship set records a complete dependency relationship triple, which includes the index of the dependent word, the index of the core word, and the dependency relationship type between them. Traverse the dependency relationship set, and for each dependency relationship, determine whether the core word and the dependent word both exist in the term semantic category labeling result node. The judgment basis is to check whether the text of the core word and the dependent word respectively matches the term text recorded in a certain term semantic category labeling result node. If the core word and the dependent word can both find a matching term semantic category labeling result node, and the two matching term semantic category labeling result nodes are anchored to the same context fragment node, then a directed edge is created between the two term semantic category labeling result nodes. Label this directed edge as a semantic edge. Combine all the created nodes, structural edges, semantic anchor edges, and syntax edges to form a complete candidate relationship graph.

[0104] The construction process of the grammar rule library is as follows: collect and analyze large-scale standard corpus (for example, grammar textbooks) in the target field (such as Chinese); summarize the combination rules and dependency patterns at the word, phrase, and sentence levels in the grammar textbooks, and formalize them into a series of explicit production rules and constraints, such as “a verb can govern a noun as an object” and “an adjective can modify a noun”; at the same time, label the applicable word combination and context conditions for each rule to form a structured rule item set. Encode and store the rule item set in a machine-readable format (such as feature structure, context-free grammar, etc.) to form a grammar rule library for syntax analysis.

[0105] S7、Input the candidate relationship graph into the hierarchical-aware graph reasoning network to identify the entity types of the nodes and the relationship types between the nodes.

[0106] A candidate relationship graph of the project cost is collected and constructed as a training set, including labeled node entity types and edge relationship types; the graph reasoning network includes a graph attention network layer, an entity type classifier, and a relationship type classifier; the architecture of the entity type classifier is sequentially connected by an input layer (dimension 512), a hidden layer, and an output layer, wherein the hidden layer contains 256 neurons and a ReLU nonlinear activation function; the number of neurons in the output layer is equal to the total number of entity types (for example, 10), and is followed by a Softmax activation function.

[0107] The relationship type classifier includes an input layer (dimension 1024, which is obtained by splicing two 512-dimensional node vectors), a hidden layer (256 neurons followed by a ReLU activation function), and an output layer (the number of neurons is equal to the total number of relationship types, for example, 5, followed by a Softmax activation function).

[0108] The graph attention network layer includes two layers of graph attention networks, each layer of graph attention network adopts 8 attention heads, the output dimension of each head is 64, and after splicing, a linear transformation layer is passed through to output a dimension of 512, and each layer is followed by an ELU activation function.

[0109] The number of layers and the dimension design of the graph attention network layer, the entity type classifier, and the relationship type classifier are based on the node semantic complexity, the number of relationship types, and the size of the graph structure in the candidate relationship graph, wherein two layers of graph attention networks are used to cover the first-order and second-order neighborhood information propagation; the multi-head attention mechanism is used to distinguish different semantic relationship perspectives, and the output dimension 512 is used to unify the semantic expression scale.

[0110] The training set is input into the graph reasoning network for training. During the training process, the cross-entropy loss between the predicted entity type probability distribution of the graph reasoning network and the true entity type label is calculated, which is denoted as the node classification loss; at the same time, the cross-entropy loss between the predicted relationship type probability distribution and the true relationship type label is calculated, which is denoted as the edge classification loss. The node classification loss and the edge classification loss are summed to obtain the total loss of the graph reasoning network. The gradient of the total loss with respect to all trainable parameters in the graph reasoning network is calculated by the back propagation algorithm. The Adam optimizer is used to update all trainable parameters of the graph reasoning network according to the gradient. The forward propagation, loss calculation, back propagation, and parameter update are repeatedly performed until the maximum number of iterations (for example, 100 times) corresponding to the graph reasoning network is reached, and the trained graph reasoning network is obtained.

[0111] The maximum number of iterations corresponding to the graph reasoning network is set according to the joint convergence characteristics of the node classification loss and the edge classification loss, and is valued according to the trend of the total loss with the number of iterations.

[0112] For each graph node in the candidate relationship graph (including the term semantic category annotation result node and the context fragment node), a node feature vector is initialized, which is obtained by encoding the semantic information recorded by the graph node (such as term text, semantic category label or context text content). The candidate relationship graph is input into the trained graph reasoning network. The first graph attention network layer of the graph reasoning network calculates the attention coefficient between each graph node and all neighbor graph nodes in the candidate relationship graph, which depends on the graph node feature vector, the edge type connecting the graph nodes, and the hierarchical information inherited from the structured document data. The node feature vectors of all neighbor graph nodes are weighted and summed based on the attention coefficient; the weighted sum result is combined and linearly transformed with the node feature vector of the graph node itself to generate a combined feature vector; after the combined feature vector is activated by a nonlinear activation function, it is output as the output of the first graph attention network layer. The output of the first graph attention network layer is input into the subsequent second graph attention network layer, and the same neighbor feature aggregation based on the attention coefficient, combination transformation and nonlinear activation process are repeated. After the iteration of all graph attention network layers, each graph node obtains multi-hop neighbor information, edge type information and document hierarchical information to form a new node. The new node is input into the entity type classifier integrated in the graph reasoning network, and the entity type classifier performs forward calculation on the new node to output an entity type distribution vector corresponding to different entity types of the new node, and selects the entity type corresponding to the maximum probability value as the entity type of the new node. For the new nodes connected to each other in the candidate relationship graph, the corresponding node representation vectors are extracted and spliced to form an edge vector; the edge vector is input into the relationship type classifier integrated in the graph reasoning network, and the relationship type classifier performs forward calculation on the edge vector to output a relationship type distribution vector, and selects the relationship type corresponding to the maximum probability value as the relationship type of the edge and labels it.

[0113] The technical purpose of the entity type classifier is to determine the specific category of the entity represented by a single node in the graph in the field of engineering cost, for example, to determine whether a node belongs to "material", "engineering quantity" or "comprehensive unit price". The technical purpose of the relationship type classifier is to determine the semantic relationship category represented by the edge connecting two nodes in the graph, for example, to determine whether an edge belongs to "hierarchical inclusion", "attribute belongs to" or "process dependence".

[0114] S8, according to the entity type of the node and the relationship type between the nodes, a bill of quantities tree is constructed from top to bottom, and engineering cost text information is obtained, and the specific acquisition process is as shown in Figure 2

[0115] ​S8.1、Should be noted that the traversal of all edges in the candidate relationship graph, read each edge corresponding to the relationship type; the relationship type of each edge and the pre-defined hierarchical containment relationship type are compared and judged; when the relationship type of the edge is consistent with the hierarchical containment relationship type, the edge is marked as a hierarchical relationship edge; after the traversal and comparison of all edges are completed, the hierarchical relationship edge set is formed.

[0116] Traverse each edge in the hierarchical relationship edge set, count the number of times each graph node in the candidate relationship graph appears as a "child node" (i.e. the end point pointed to by the edge); identify the graph nodes with zero occurrence times as top-level nodes without parent nodes; based on the entity type of each top-level node, select the top-level node with the highest priority of entity type as the root node of the bill of quantities tree, for example, the entity of the "project name" or "total project" type is preferred to the entity of the "divisional engineering" type, and the entity of the "divisional engineering" type is preferred to the entity of the "sub-item engineering" type.

[0117] After determining the root node, find all edges in the hierarchical relationship edge set that have the root node as the "departure endpoint", and the "child node" connected by the edge is the child node of the root node. Record the root node as the parent node of the child node.

[0118] The priority of the entity type is set according to the business scope covered by different entities in the engineering cost business and the position in the bill of quantities hierarchical system; the entity type with a larger business coverage range has a higher priority, for example, the entity type representing the overall engineering and project level has a higher priority than the lower-level entity types such as divisional engineering and sub-item engineering.

[0119] Create an empty tree data structure and add the root node of the bill of quantities tree as the root node of the tree data structure. Traverse all graph nodes whose parent nodes have been identified, starting from the root node of the bill of quantities tree, recursively add each graph node as a child node under the corresponding parent node. During the addition process, maintain the parent-child hierarchical relationship between the graph nodes; store the original information (including the term text and semantic category of the term semantic category annotation result node, and the text content of the context fragment node) in each graph node as an attribute in the tree data structure; the tree data structure completely retains the topological structure based on the hierarchical containment relationship in the candidate relationship graph and the semantic and text information of all graph nodes. This tree data structure is the bill of quantities tree. The bill of quantities tree is a structured, analyzable, and directly usable engineering cost text information for cost calculation and analysis.

[0120] The embodiment also provides a computer device suitable for the method for extracting text information of engineering cost based on semantic understanding, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the method for extracting text information of engineering cost based on semantic understanding proposed in the above embodiment.

[0121] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0122] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to realize the method for extracting text information of engineering cost based on semantic understanding proposed in the above embodiment. The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0123] To sum up, by introducing a semantic anchoring mechanism, the engineering cost terms are accurately mapped to the standard semantics in the professional context, effectively overcoming the problems of homonymy and expression diversity, combining the deep fusion modeling of context and anchoring characteristics, dynamically generating semantic representations with business awareness, then through the ambiguity resolution classifier oriented to the closed category system, realizing high-precision classification of terms to computable business labels, finally relying on the hierarchical perception graph reasoning network, on the basis of preserving the original logical structure of the document, automatically identifying the entity type and hierarchical inclusion relationship, and constructing a complete structure and clear semantics of the bill of quantities tree from top to bottom, realizing the end-to-end conversion from unstructured text to structured and clear semantics.

[0124] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for extracting engineering cost text information based on semantic understanding, characterized in that: include, Receive engineering cost documents and preprocess them to obtain structured document data; The structured document data is segmented into multiple granularities to obtain character-level text data, a set of engineering cost terminology, and contextual fragments, specifically: The text content in structured document data is segmented according to character order to obtain character-level text data; Construct a core dictionary for the field of engineering cost, and perform positive maximum matching scans on character-level text data to identify and extract all successfully matched words, forming a terminology set for the field of engineering cost. Based on the paragraph marks, item number sequences, and table boundaries recorded in the structured document data, continuous character-level text data is divided into logically independent semantic units to obtain context fragments; Semantic anchoring processing is performed on each term in the terminology set of the engineering cost field to obtain semantic anchoring features, specifically: Based on the core dictionary of engineering cost, query the standard semantic anchor points of each term in the field of engineering cost; After determining the standard semantic anchor points, perform semantic alignment operations between the domain terms and the corresponding standard semantic anchor points to obtain semantic constraint information; Semantic constraint information is numerically encoded and combined to obtain semantic anchoring features; Semantic modeling is performed on context fragments and semantic anchor features to obtain context-related semantic feature data; Perform ambiguity removal on context-related semantic feature data to obtain term semantic category annotation results; Using the semantic category annotation results of terms and their corresponding context fragments as nodes, a candidate relationship graph is constructed. The candidate relationship graph is input into a hierarchical graph reasoning network to identify the entity type of the nodes and the relationship type between the nodes; Based on the entity type of the nodes and the relationship type between the nodes, a bill of quantities tree is constructed from top to bottom to obtain the project cost text information.

2. The method for extracting engineering cost text information based on semantic understanding as described in claim 1, characterized in that: The pretreatment includes cleaning and standardization.

3. The method for extracting engineering cost text information based on semantic understanding as described in claim 2, characterized in that: The step of semantically modeling context fragments and semantic anchoring features to obtain context-related semantic feature data specifically involves: Serialize and encode the context fragment to obtain the context fragment sequence vector; The context fragment sequence vector and the semantic anchoring feature are concatenated along the feature dimension to form a joint semantic vector; The joint semantic vector is input into a multilayer perceptron for nonlinear transformation and fusion with deep features to obtain a dense vector, which is then used as context-related semantic feature data.

4. The method for extracting engineering cost text information based on semantic understanding as described in claim 3, characterized in that: The process of disambiguating context-related semantic feature data and obtaining term semantic category annotation results specifically involves: Context-related semantic feature data is input into the ambiguity disambiguation classifier for forward propagation calculation to obtain the probability distribution vector and construct a set of semantic categories. Map the elements in the probability distribution vector back to the semantic category set to retrieve the corresponding specific category; The specific categories retrieved are used as the semantic category annotation results for the terms.

5. The method for extracting engineering cost text information based on semantic understanding as described in claim 4, characterized in that: The probability distribution vector refers to the vector obtained by inputting context-related semantic feature data into the ambiguity elimination classifier, performing linear transformation and nonlinear mapping on the context-related semantic feature data in each hidden layer of the ambiguity elimination classifier, and then normalizing it using the Softmax function of the output layer.

6. The method for extracting engineering cost text information based on semantic understanding as described in claim 5, characterized in that: The step of constructing a candidate relationship graph by using the term semantic category annotation results and corresponding context fragments as nodes is as follows: Create a graph node for each term semantic category annotation result and its corresponding context fragment; The graph nodes include term semantic category annotation result nodes and context fragment nodes; Based on the original location and hierarchy information recorded in the structured document data, structural edges are established between context fragment nodes; Establish semantic anchor edges representing the subordinate relationship between the term semantic category annotation result node and the context fragment node; Syntactic edges are established for semantic category annotation result nodes of terms that have direct syntactic dependencies and appear in the same context segment; Combine all created graph nodes, structural edges, semantic anchoring edges, and syntactic edges to form a complete candidate relation graph.

7. The method for extracting engineering cost text information based on semantic understanding as described in claim 6, characterized in that: The process of inputting the candidate relationship graph into a hierarchical perceptual graph reasoning network to identify the entity type of nodes and the relationship type between nodes specifically involves: The candidate relationship graph is input into the hierarchical perception graph reasoning network. The graph reasoning network aggregates and propagates the node features of the candidate relationship graph through the graph attention network layer to generate updated nodes. After the updated node is input into the entity type classifier and the relation type classifier, forward computation is performed to obtain the entity type of the node and the relation type of the edge.

8. The method for extracting engineering cost text information based on semantic understanding as described in claim 7, characterized in that: The process involves constructing a bill of quantities tree from top to bottom based on the entity type of the nodes and the relationship type between the nodes to obtain the project cost text information. Specifically: Based on the relationship type of the edges, edges belonging to the hierarchical inclusion relationship are filtered in the candidate relationship graph to obtain hierarchical relationship edges; The root node of the bill of quantities tree is identified based on the hierarchical relationship edge, and the parent node of all nodes in the candidate relationship graph is identified with the root node as the starting node. Connect each node to its corresponding parent node sequentially from top to bottom to generate a bill of quantities tree, which is then used as the text information for project cost.

Images