Research on the Construction Method of an Assembly Knowledge Graph for a Biomass Heating System

Abstract

In the complex process of assembling biomass heating systems, traditional paper documents and construction process card management methods have weak information correlation and take a long time for information retrieval, which seriously restricts the assembly efficiency and quality. Moreover, the assembly process involves numerous components and complex processes, making it difficult for traditional management methods to cope with. To address this issue, a knowledge graph-based assembly information integration method is proposed to integrate scattered assembly information into a graph database, providing pathways for accessing assembly information and assisting on-site management. The biomass heating system assembly knowledge graph (BAKG) adopts the top-down method construction. After the construction of the upper schema layer, the 3DXML file was parsed, the XML.dom parser in Python3.7.16 was used to extract the equipment structure information, and the RoBERTa-BiLSTM-CRF model was applied to the named entity recognition of the assembly document, which improved the accuracy of entity recognition. The experimental results show that the F1 score of the RoBERTa-BiLSTM-CRF model in entity recognition during the assembly process reaches 92.19%, which is 3.1% higher than that of the traditional BERT-BiLSTM-CRF model. Moreover, the knowledge graph structure generated by the equipment structure data based on 3DXML file is similar to the equipment structure tree, but is more clear and intuitive. Finally, taking the second-phase construction process records of a company as an example, BAKG was constructed and assembly information was stored in the Neo4j graph database in the form of graphs, which verified the effectiveness of the method.

1. Introduction

The biomass heating system uses biomass as the main fuel, and conducts combustion heating in biomass boilers, including large-angle skirted fuel conveyors, membrane wall biomass molded fuel boilers, evaporative heat exchangers, economizers, ceramic multi-tube dust collectors, bag filters, and other large mechanical equipment. Biomass energy has the characteristics of being environmentally friendly, low cost, and carbon neutral. Faced with the dual pressures of energy shortage and environmental degradation, governments around the world attach great importance to the development and utilization of biomass resources [1]. Improving the efficiency of building biomass heating systems is an urgent problem that needs to be solved in the current development and utilization of biomass energy.

Biomass heating system assembly is a complex mechanical product assembly [2]. The general procedure for installing the mechanical equipment it contains is as follows: installation preparation—equipment unpacking inspection—foundation inspection, acceptance and treatment—equipment moving in place—equipment installation (alignment, leveling and centering)—one grouting—precise leveling and alignment—two grouting—test run acceptance [3,4]. Different from the assembly activities of the manufacturing workshop, in addition to mechanical engineering, the assembly process of the biomass heating system will also involve civil engineering, electrical, water supply and drainage, and other projects, and the site environment is complex, which brings greater challenges to the site management of the assembly process. Traditional management methods are usually realized by paper documents, such as assembly process documents, two-dimensional drawings, construction process cards, etc. Construction personnel and management personnel usually need to flip back and forth between various information to find information. For example, in the equipment installation stage, it is necessary to check the two-dimensional drawings to learn the tolerance information and positioning information. It is also necessary to review the assembly process documentation to process information. These documents are stored separately, and it is difficult to connect them with each other, thus information islands are formed. Therefore, the auxiliary assembly by means of information technology and the retrieval path can greatly improve the information retrieval efficiency and management level during the installation process, which has important engineering significance for improving the assembly efficiency and assembly quality of the biomass heating system.

With the digitalization demand in the assembly field, in order to improve assembly efficiency, many information management methods have emerged in recent decades to optimize assembly efficiency and quality. From the perspective of assembly process, Zhou Q et al. divided the product data integration model into a design data model and a process design model based on the process data model of EBOP(electronic bill of process) [5]. Peng Q et al. solved the problem of processing information transfer between different systems based on an IDEF1x assembly process information model [6]. Jinfeng Ld et al., based on the expression form of processing information based on an inter-process model, combine parts, process documents and the assembly process according to the attribute relationship between processes to achieve an accurate expression of process information [7]. Many scholars also focused on the problem of assembly sequence planning. Dini G [8] designed the interference matrix, assembly contact matrix and assembly sequence connection matrix of the assembly to generate the mathematical evaluation model of the assembly sequence, which can carry out assembly sequence planning for the assembly of different numbers of parts. Pedraza [9] introduced the genetic algorithm into the assembly sequence planning problem, and the assembly sequence obtained had the minimum number of assembly direction changes and assembly equipment changes. Over time, the increase in assembly parts led to an increasing amount of algorithm computation. Therefore, many heuristic algorithms such as the moth extinguishing algorithm [10], artificial bee colony optimization algorithm [11], and assembly sequence planning model based on a neural network [12,13,14] have emerged.

The expression methods based on the assembly process information model mostly describe the process information of the assembly process with only geometric models as the sole data carrier. Although this method can fully express the process information of the assembly process, the process information is still an unstructured natural language description and does not provide a convenient search path or operating system. Operators still need to spend more effort to retrieve information. On the other hand, due to the fixed process steps in the biomass heating system assembly and the need to consider factors such as lifting, site, personnel, and energy in addition to assembly, if the assembly sequence planning problem is studied, the mathematical model established will be extremely large and difficult to apply to the complex and variable actual assembly process. Therefore, these methods are not suitable for biomass heating systems. In this study, it is considered that it is an effective way to improve the assembly efficiency and quality of biomass heating system to establish assembly knowledge database of biomass heating system, collect the process parameters, the assembly sequence, assembly resources, and other relevant information to guide construction personnel installation, assist managers in site management, and quality acceptance. At present, assembly data and knowledge in the manufacturing field are mostly based on traditional relational databases with high redundancy, scattered distribution, and weak relevance, and the correlation, cognition, understanding, and reasoning of data and knowledge are rarely studied from the semantic level [15]. Unlike the previous assembly process information representation methods and traditional relational databases, the biomass heating system knowledge graph developed in this study can realize the knowledge-based description, structured expression, and large-scale instance storage of assembly information, as well as the strong interdependence of assembly data. It no longer limits itself to geometric information as a carrier, but integrates various pieces of information and stores them in the form of a graph database, using its “entity-relationship-entity” and “entity-attribute-value” structured expression forms to simply, intuitively, and accurately express assembly information [15].

The knowledge graph was first formally proposed by Google in 2012 [16], with the original intention of improving the search and enhancing the user search experience. A knowledge graph is essentially a structured semantic network [17], a semantic knowledge base with a directed graph structure, which is used to describe concepts and their interrelationships in the physical world in symbolic form. Based on the graphical representation of linked entities, the links between multiple heterogeneous data sources can be effectively represented [18]. A knowledge graph is usually represented in triplet form, such as $G=(E,R,S)$, where $E=\{e_1,e_2,\cdots ,e_n\}$ is the entity set in the knowledge base, including $n$ different entities; $R=\{r_1,r_2,\cdots ,r_m\}$ is the set of relationships in the knowledge base, which contains $m$ kinds of different relationships. $S\subseteq E\times R\times E$ F stands for the set of triples in the knowledge base. The basic forms of triples mainly include entity 1, relation, entity 2, and concept, attribute, attribute value, etc. According to the application scenarios of the knowledge graph, it can be divided into a general knowledge graph and a vertical knowledge graph [19]. A general knowledge graph refers to a knowledge graph that is applicable to all knowledge fields, which is equivalent to a “structured encyclopedia”. Its main application scenarios are internet search and comprehensive question answering. A vertical knowledge graph, also known as a domain knowledge graph, is formed by extracting a knowledge graph based on entities and relationships containing certain key factors in a specific field. It usually has the characteristics of high professionalism, strong integration, and clear relationship trajectory.

These features of a knowledge graph can more clearly describe the complex process information, sequence, resources and other information in assembly activities, and there have been some research results. For example, the knowledge graph of aviation assembly based on the named entity recognition framework, based on continuous learning contains assembly process information, technical knowledge, industry standards and the internal relations among them in the field of aviation assembly [20]. The product assembly knowledge recommendation is based on a knowledge graph and scene perception [21]. The dynamic adjustment of the assembly process was performed by using knowledge graph inference [22]. The assembly task resource allocation based on the knowledge graph framework supports man–machine collaboration in the workshop [23]. It is a structured representation of assembly process planning based on a knowledge graph [24]. It can be seen that the knowledge graph technology is a feasible method in the assembly field. However, at present, the publicly available knowledge graph is basically the general domain knowledge graph, such as Freebase [25], Wikidata [26], DBpedia [27], etc., and the knowledge graph database for the assembly field is still very lacking [15]. At present, the application of knowledge graphs in the field of assembly is mostly seen in workshop assembly, and the research of knowledge graphs for mechanical equipment assembly is relatively nonexistent.

Therefore, in order to fill the gap in the research of knowledge graph of mechanical equipment assembly, the author takes the assembly of biomass heating system as an example, and proposes a method of automatically constructing a knowledge graph of biomass heating system assembly, which provides a reference for knowledge graph and information management in the field of mechanical equipment assembly.

This research focuses on the mining and processing of assembly process information, as well as the integration of different kinds of information, and strives to structures the multi-source heterogeneous data contained in the assembly process of biomass heating system, so as to realize the organization and management of knowledge. In this paper, a new automatic construction method is proposed, which is based on the result of semi-structured data identification; it combines the unstructured semantic data with it to realize the automatic construction of a knowledge graph.

The rest of this article is organized as follows: In the Section 2, we review the relevant work. In Section 3, we outline our approach. The assembly information extraction model is described in detail. The Section 4 describes the construction of an assembly knowledge map of a biomass heating system. In Section 5, the experimental results of this method are given. Finally, the paper summarizes the full text and puts forward some suggestions for future work.

In order to allow readers to better understand the manuscript, the definitions of the abbreviations used in the manuscript are given in Table 1.

2. Related Work

Knowledge graphs can be divided into general knowledge graphs and specialized knowledge graphs according to their application fields. Existing general knowledge graphs have shown advantages in supporting many applications, such as Freebase, Reverb, and Microsoft’s Probase. Kim [28] proposed a knowledge representation framework based on semantic hypergraphs to support knowledge sharing in product development. Z. Liu [29] et al. proposed an Entity Duet Neural Ranking Model (EDRM) that introduces knowledge graphs into neural search systems, achieving a natural combination of entity-oriented search and neural information retrieval. The above methods or models significantly improve the application of knowledge graphs in product information. However, these resources have a weak connection to the field of study in this research.

In professional fields such as manufacturing and construction, knowledge graph-based representation has been widely applied to solve engineering management problems. Rasmussen [30] et al. demonstrated the use of knowledge graphs to complete and manage a typical construction project task, providing engineers with recommendations to support decision-making when changes occur; Zhao and Smidts [31] integrated various knowledge modules into a system, generated a knowledge graph, and applied it to the development of a knowledge base for reactor operators involved in severe accidents in nuclear power plants; Shi [32] et al. utilized fusion technology to integrate spacecraft information, constructed a full lifecycle knowledge graph, and designed a human–machine interface for spacecraft rolling, bearing fault diagnosis based on the theory of function behavior structure. Although these methods have achieved good results, their application scenarios are quite different from the present study.

Entity extraction is one of the most critical steps in building a knowledge graph, primarily used to extract concepts from unstructured and semi-structured data. With the development of hardware devices, entity extraction methods based on deep learning neural networks have become mainstream. Neural networks can self-learn features from datasets and perform better than rule-based, dictionary-based, or statistical machine learning-based methods. Various neural network-based methods are used for entity extraction tasks. Colbert et al. [33] proposed a unified architecture for deep neural networks FNN/CNN-CRF, which can automatically learn internal feature representations from large amounts of data without the need for manual optimization, and without the need for traditional task-specific feature engineering. Compared with other machine learning-related methods, this model has better performance and less computational complexity. Yang et al. [34] proposed an IDCN-BiLSTM-CRF model based on news text and professional medical text data structures. The IDCNN module operates in parallel with the BiLSTM module, utilizing features with additional entity information and contextual features to improve the accuracy of named entity recognition. Qu et al. [35] proposed the BERT-BiLSTM-CRF model for identifying security hazard entities and the RESPCNN-ATT-based model for extracting security hazard relationships, realized entity recognition, and relationship extraction of power safety hazard records.

The above deep neural network methods are all used to process unstructured data. However, in industrial scenarios, large amounts of data are stored in semi-structured formats (such as relational databases, CSV files, etc.) and exist in other semi-structured formats (such as XML, JSON, etc.), which require mapping languages and engines to transform, integrate, and feed the data back into a knowledge graph. Giuseppe F. et al. [36] released the SeMi (Semantic Modeling macro) tool, which can semi-automatically construct large-scale knowledge graphs from (semi) structured data (XML, JSON, CSV). Wu and Weld [37] trained an extraction model on semi-structured data, and then applied the extraction model to extract attributes from unstructured data. However, these tools and methods have not been applied in industrial scenarios, and most of the current research is focused on the field of natural language processing, with little research on semi-structured data in industry.

Therefore, the focus of this study is how to construct a knowledge graph suitable for the construction process of biomass heating systems, in order to carry out subsequent research. The work of this study is dedicated to maximizing the use of semi-structured data in industrial scenarios, with a focus on equipment structures identified from semi-structured data, adding assembly process information on this basis, and combining semi-structured data with unstructured data for entity extraction and relationship extraction to ensure the quality and accuracy of the knowledge graph in the professional field.

3. Overview of BAKG Construction Methods

The construction of BAKG requires the selection of a suitable construction approach in the first place. Currently, there are two construction methods for knowledge graphs. One is the top-down construction method, and the other is the bottom-up construction method [38]. The former first constructs the model and rules, and then fills the data into the knowledge base; the latter first extracts knowledge, then builds models and rules based on knowledge, and then adds them to the knowledge base. BAKG belongs to the professional domain knowledge graph; the scope of knowledge and concepts involved are relatively fixed or relatively controllable, and the accuracy of the professional domain knowledge graph is higher. Therefore, in order to ensure the construction efficiency and map quality of BAKG and prevent concept confusion, the top-down construction method is chosen in this study.

The top-down approach requires first identifying the modules and structures that the BAKG contains; at the same time, some concepts in the assembly process of biomass heating system are defined to cover the assembly process semantics, assembly resources semantics, and equipment structure. In this study, BAKG has the following four definitions:

• 1.. BAKG is a structured semantic database based on a triplet architecture regarding the assembly process, in the form of <Entity-Relationship-Entity>. BAKG has three modules in total. The first module is the assembly structure information model (Definition 2), which contains the components of the entire system and is used to describe the BOM information of the system and the composition relationship of the components. The second module is the assembly process information model (Definition 3), which includes specialized knowledge from assembly processes. The third module is the assembly resource information model (Definition 4).

• 2.. The assembly structure information model is an objective carrier for implementing functions, used to represent the hierarchical semantic relationships and membership semantic relationships between the various components that make up the product, including three types of relationships, Has, HasSub, and HasPart, as well as four entities: final assembly body, device units, component units, and part units. At the same time, Has is defined to represent the composition relationship between the assembly body and each device unit; HasSub represents the composition relationship between device units and component units; HasPart represents the compositional relationship between component units and part units.

• 3.. The assembly process information model is a combination of process information and solid models used to represent assembly operations, including placement, loading, fitting, insertion, and tightening.

• 4.. The assembly resource information model is a non-product assembly element involved in the assembly process, used to ensure assembly quality, limit connection deformation during assembly, and ensure accuracy and interchangeability. It includes three aspects: installation preparation, precautions, and process requirements. Installation preparation includes five entities: location preparation, organizational preparation, technical preparation, foundation preparation, and equipment acceptance. The process requirements include five entities: welds, axis deviation, strength, distance, and specifications.

Secondly, the construction process and method framework need to be determined. The process and method framework proposed in this paper is shown in Figure 1. Based on this framework, the assembly information can be represented structurally and uniformly. The architecture includes three parts: pattern layer construction, assembly information entity recognition, and BAKG construction. These three sections are briefly described below.

• 1.. Construction of pattern layer

The pattern layer of BAKG is based on concepts inherent to the assembly process, the on-site assembly processes of biomass heating systems, and the construction of biomass heating system equipment. Firstly, the equipment is in a tree-like structure, divided into four levels from overall to local: final assembly body, device units, component units, and part units, to construct the assembly structure pattern layer. For example, the biomass heating system is the final assembly body, which includes device units such as furnaces, evaporators, and economizers. The furnace also includes component units such as dampers, membrane walls, and grate units. The air damper is composed of part units such as air damper blades, air damper bearings, air damper shafts, and the connecting rods of the adjustment mechanism.

At the same time, during the assembly process, various components are interconnected, and some components require lubrication or other preparation work before assembly. There are also precautions to be taken during installation to ensure the accuracy and safety of the assembly process. After installation, continuous adjustment and inspection are carried out according to the corresponding process requirements. An assembly process pattern layer is built based on the assembly resource information model and the assembly process information model defined in Section 3. The correlation between concepts in the pattern layer is crucial for BAKG. A relationship list corresponding to the 3D category has been deployed. As shown in Table 2, a table of relationship names, domains, and scopes is created.

The BAKG conceptual pattern layer constructed by combining the above two pattern layers and the correlation between concepts is shown in Figure 2.

• 2.. Entity recognition of assembly information

Entity recognition includes parsing of 3DXML files and named entity recognition based on neural network models. The equipment structure from 3DXML files is parsed, including device units and their component units, part units, and identify assembly process data from assembly process documents, such as the entity of installation preparation, precautions, and process requirements during the installation process, and the subject of the implementation of this process information is the unit of the equipment structure.

• 3.. Construction of BAKG

Based on the extracted entities, guided by the concept of the pattern layer, the entities are connected through their relationships and stored in Neo4j using Cypher language to generate the basic carrier for the assembly structure, assembly process, and assembly technology of biomass heating systems, namely BAKG.

4. Assembly Entity Recognition

Accurately and efficiently identifying entities from various data sources is the foundation for effectively constructing a knowledge graph. In this study, a new entity extraction method is proposed, which first analyzes a 3DXML file to extract equipment structure, and at the same time uses a deep neural network to identify the named entities in assembly process documents, and then aligns the entities identified from two data sources to achieve efficient entity recognition. The framework of the entity extraction method is shown in Figure 3. The two parts of this method are described in detail below in Section 4.1 and Section 4.2.

4.1. Semi-Structured Data Knowledge Extraction

The 3DXML file is an open-standard proprietary 3D file format, developed and supported by Dassault Systems (https://cadexchanger.com/3dxml/, accessed on 22 November 2024). It is a zip archive consisting of a BOM file and one or more 3D representations stored in XML or binary format. The 3DXML file generated based on the 3D model is semi-structured data with sufficient assembly structure information. Starting from the <ProductStructure> node, it has four child nodes [39], among which the <Reference3D> reference node and <Instance3D> instance node are directly related to the assembly structure < Instance3D> is both a component of one <Reference3D> and an instance of another <Reference3D>. The ID attribute values of these two different <Reference3D> are stored in the <IsAggregatedBy> and <IsInstanceOf> child nodes of <Instance3D>, respectively.

3DXML is a file based on XML. The XML.dom parser in Python3.7.16 can recognize information in nodes and use each <Instance3D> as an entity node for entity recognition, ensuring the uniqueness and accuracy of entities; At the same time as entity recognition, using <Reference3D>, <IsAggregatedBy>, and <InstanceOf> as transition nodes for relationship extraction greatly improves the efficiency of knowledge extraction. The specific steps of this method will be demonstrated by combining fragments from the 3DXML file. The 3DXML source file fragment is shown in Figure 4, and the specific steps of entity extraction based on the fragment are shown in Figure 5.

Firstly, the biomass heating system element with ID 1 is defined as the final assembly body. Then, the category of the entity mapped to the <Instance3D> node through its child nodes <IsAggregatedBy> and <InstanceOf> is determined. If the ID value stored in <IsAggregatedBy> is 1, it indicates that it is a device unit; If the <Reference3D> element pointed to by the child node <IsAggregatedBy> is referenced by the child element <InstanceOf> of the <Instance3D> element in the device units, it indicates that the node is a component unit; similarly, when the <Reference3D> element pointed to by <IsAggregatedBy> is referenced by the child element <InstanceOf> of the <Instance3D> element in the component units, it indicates that the node is a part unit. At the same time, by referencing the <Reference3D> element to construct the relationships between various entities, the relationship between the final assembly body and the device units is a Has relationship. The device units and component units have a HasSub relationship. The relationship between component units and part units is HasPart.

4.2. Unstructured Data Entity Recognition

Named entity recognition is a fundamental information extraction task in natural language processing, mainly used to extract named entities from unstructured data and classify them. In this study, three types of physical entities, namely device units, component units, and part units, were mainly extracted from the assembly process documents, as well as three types of installation operation specification entities involved in installation preparation, precautions, and process requirements during the installation process.

Firstly, a RoBERTa pre-trained language model is used to express semantic information. Compared to Word2vec, RoBERTa can better solve the problem of semantic confusion. Compared to BERT, RoBERTa can encode at both character and word levels, making its output vector contain richer in semantic information. By using this model to parse text, output a vector sequence of the text sequence as input for the next step.

In NER tasks, the data involved usually has forward and backward correlations, and the network structure of Recurrent Neural Networks (RNNs) has strong advantages in processing sequential data. Theoretically, it can handle long sequences of information, but in practical operation, it performs less than expected. LSTM [40] improves on the RNN model and can effectively solve the problems of gradient explosion and vanishing by selectively retaining contextual information through the use of specially designed gate structures when processing long sequence data in RNN. In order to solve the problem of the current unit of traditional LSTM networks being unable to act as the “input gate” and “output gate” at the next moment, resulting in the loss of partial information of the previous sequence in the entire unit, Graves [41] et al. proposed the BiLSTM model in 2013. This model includes two hidden layers of LSTM network, connecting the export vectors of the forward and backward hidden layers of LSTM to generate a global vector, expanding the backward knowledge acquisition ability, and then generating contextual information of the given input sequence through a neural network. The LSTM unit consists of an input gate mechanism, a forget gate mechanism, and an output gate mechanism, used for storing and forgetting information. The function of the forget gate mechanism is to control the gradient through it. The input gate mechanism determines the important information available in the current state, and the output gate mechanism determines the state of the next hidden layer. The LSTM network structure is shown in Equations (1)–(5); the structure of the improved BiLSTM unit is shown in Figure 6.

(1) $i_t=\sigma \left(W_{xi}{x}_t+W_{hi}{h}_{t-1}+W_{ci}{c}_{t-1}+b_i\right)$

(2) $f_t=\sigma \left(W_{xf}{x}_t+W_{hf}{h}_{t-1}+W_{cf}{c}_{t-1}+b_f\right)$

(3) $c_t=f_t{c}_{t-1}+i_t\tan h\left(W_{xc}{x}_t+W_{hc}{h}_{t-1}+b_c\right)$

(4) $o_t=\sigma \left(W_{xo}{x}_t+W_{ho}{h}_{t-1}+W_{co}{c}_{t-1}+b_o\right)$

(5) $h_t=o_t\tan h\left(c_t\right)$

Among them $\sigma$ is the logical activation function $sigmoid$; $\otimes$ is a dot multiplication operation; $\tan h$ is a hyperbolic tangent activation function; $x_t$ is the unit input, and $i_t$, $f_t$, $o_t$ are the input gate, forget gate, and output gate of the $t$ time; $w,b$ is the weight matrix and bias vector of the input gate, forget gate, and output gate; ${\tilde{c}}_t$ is the state of $t$ time; $h_t$ is the output of $t$ time. The output of BiLSTM is to concatenate the output sequences of two LSTM units as the final feature representation.

Finally, the CRF is used to select the optimal probability as the global optimal sequence result. CRF is a sequence labeling model that performs joint probability analysis on label sequences by considering the relationship between adjacent labels. An assumption that the text sequence and the predicted sequence are the score of the predicted sequence corresponding to the text sequence is shown in Equation (6):

(6)$score(x,y)={\displaystyle \sum _{i=1}^n\left(P_{x_i,y_i}+A_{y_{i-1},y_i}\right)}$

In the formula, $P_{x_i,y_i}$ represents the probability of the word $x_i$ appearing in the label $y_i$, which is the output score. The $A_{y_{i-1},y_i}$ transition score represents the probability of label $y_{i-1}$ transfer $y_i$, in order to simplify the calculation, $y_0$ is set as the start label, which is used to indicate the start of the sequence, and is used to help the model initialize the transition fraction during the decoding process [42]. Normalize the function $score(x,y)$ to obtain a probability distribution representation of $y$, as shown in Equation (7):

(7)$P\left(y\left|x\right.\right)={\displaystyle \frac{e^{score(x,y)}}{{\displaystyle \sum _{\tilde{y}\in Y_x}{e}^{score(x,\tilde{y})}}}}$

In the formula, $\tilde{y}$ represents the true label sequence, $Y_x$ represents all possible label sequences, and $score(x,\tilde{y})$ represents the score of one possible label sequence among all possibilities. The logarithmic likelihood representation of the training samples yields Equation (8):

(8)$\log P\left(y\left|x\right.\right)=score\left(x,y\right)-\log \left({\displaystyle \sum _{\tilde{y}\in Y_x}{e}^{score(x,\tilde{y})}}\right)$

Then, the optimal sequence of the input sequence is obtained by maximizing the likelihood function, as shown in Equation (9):

(9)$y^{*}=\underset{\tilde{y}\in Y_x}{\arg \max } score\left(x,\tilde{y}\right)$

The RoBERTa-BiLSTM-CRF model can fully consider the characteristics of the assembly process in the assembly process document. By inputting the text sequence $x$ into the model for prediction, the sequence $y^{*}$ will be obtained. Figure 7 shows the basic structure of the model.

5. Case Study

In order to verify the effectiveness of the method, experimental research was conducted based on the records (Chinese Language) of the equipment assembly process during the construction of the second phase of a biomass clean energy heating company. This study is an experimental research work based on the Python 3.7.16 programming language. All studies were conducted on an experimental platform configured with Intel (R) Core (TM) i9-10900K CPU @ 3.7 GHz, GTX 3090 GPU, Python 3.7.16, Tensorflow 1.14.0, and Windows 10 operating system.

Based on the TensorFlow platform deep learning framework, the recognition performance of the assembly process entity recognition model is evaluated using evaluation metrics such as Precision (p), Recall (r), and F1 Score (f). Table 3 shows the textual information (partial) of the equipment assembly process. The longest sentence consists of 240 words, so Max_1ength (the maximum length recorded during the assembly process) is set to 256 words.

Other relevant parameters are shown in Table 4.

5.1. Performance Analysis of Entity Recognition Model

Using assembly process documents as corpus markers, the entity recognition model based on RoBERTa-BiLSTM-CRF and BERT-CRF used in this study were compared, BiLSTM-CRF, ALBERT-BiLSTM-CRF. The entity recognition model was compared with BERT-BiLSTM-CRF and applied to assembly process text information to evaluate the effectiveness of entity recognition in assembly process information. The comparison is mainly made from three aspects: F1 value, precision, and recall. The specific calculation formula is shown in Equations (10)–(12). The performance comparison of entity recognition models is shown in Table 5 and Figure 8.

(10) $F1={\displaystyle \frac{2\times Precision\times Recall}{Precision+Recall}}\times 100\%$

(11) $Precision={\displaystyle \frac{TP}{(TP+FP)}}\times 100\%$

(12) $Recall={\displaystyle \frac{TP}{(TP+FN)}}\times 100\%$

In the formula, $Precision$ represents the ratio of the number of correctly predicted positives to the total number of correctly predicted positives, $Recall$ represents the ratio of the number of correctly predicted positives to the total number of actually positive positives in the sample, $F1$ is the harmonic average of $Precision$ and $Recall$, representing the combined effect of the two.

As can be seen from Table 5, compared with model 1, the F1 value of model 3 increased by 2.79% after the introduction of BiLSTM. The reason for the improved recognition effect is that BiLSTM can fully obtain contextual semantic dependency information and improve the recognition effect. Compared with model 2, the F1 value of model 3 increased by 2.02% after the introduction of BERT. The reason for the improved recognition effect is that BERT model has strong feature extraction ability, and the model after BERT pre-training has a better recognition effect. Compared with model 3, it can be seen that the recognition effect of the RoBERTa model is better than that of BERT, and the F1 value is increased by 3.1%. The reason is that RoBERTa has carried out effective optimization and improvement for BERT pre-training tasks, and RoBERTa adopts dynamic mask strategy, which has a stronger semantic information acquisition ability. As BERT’s optimization and improvement model, ALBERT’s performance is obviously inferior to the RoBERTa model and BERT model. This is because BERT and RoBERTa’s parameters at each layer are independent, and they can fully learn the assembly context information and assembly process entity model features. By sharing parameters, ALBERT does not use independent parameters for each layer to learn different information, resulting in insufficient learning of assembly context information and assembly process entity model features.

Figure 8 shows the training results of each round of the five models under the same parameters. As can be seen from Figure 8a, among the five models, RoBERTa-BiLSTM-CRF model had the smallest number of training rounds to reach the optimal F1 value in the fourth round. Meanwhile, by observing Figure 8b,c, it can be found that the F1 value, accuracy rate, and recall rate of RoBERTa-BiLSTM-CRF are relatively stable with little fluctuation, and the performance of each round is better than other models.

5.2. BAKG Storage and Visualization Expression

Through the above process, the assembly information of the biomass heating system is converted into structured triplet data. The Neo4j graph database can store the structured data in the form of a graph to more intuitively and accurately reflect the connection between the assembly information. The beginning and end data in the triplet data of entity–relationship–entity, entity–attribute–attribute value are stored as node types in Neo4j, and the relationship and attribute relationship are stored as edges, thus completing the storage of BAKG data. As an open-source graph database system, Neo4j will store structured data on the graph, with JAVA as the environment to achieve, with complete transaction characteristics. Neo4j is a local database, without a network, and can access the database and other operations, better cope with the complex environment of the site, and the unstable situation of network communication. Because the BAKG is operated directly locally and contains more entities and assembly relations, compared with other graph databases, it can achieve fast access, display the relationships among entities, influence and association, assist manual assembly and management, and facilitate the subsequent application and development of knowledge graph. Therefore, Neo4j is selected to store the data and knowledge of the BAKG.

According to the BAKG construction method, the extracted information is mapped to the pattern layer and stored in Neo4j to complete the BAKG construction. RoBERTa-BiLSTM-CRF is the named entity recognition (NER) model that was used to extract the assembly information of the biomass heating system. The information in Table 3 was taken as an example, and the extraction results were shown in Table 6. The semi-structured data extraction results are shown in Figure 5. After integrating the two pieces of information and storing them in Neo4j, partial BAKG, as shown in Figure 9, is obtained.

6. Conclusions

In this paper, a knowledge graph is introduced into the field of assembly process planning as a knowledge representation method. BAKG’s conceptual schema layer is built through a top-down approach as the framework for BAKG. For the entity data required for constructing BAKG, the 3DXML file containing the model information is first exported from the 3D model, and the xml.dom parser in python is used to parse it and identify the equipment structure data. Then, the RoBERTa-BiLSTM-CRF deep learning model was adopted to carry out a named entity recognition task. This model could fully obtain global context features and local features combined with global context features and local features to improve entity recognition rate, and could identify assembly process data from assembly process documents. Taking a company’s on-site assembly process document as an example, BAKG is constructed by combining the analyzed equipment structure data. The F1 score of the RoBERTa-BiLSTM-CRF deep neural network for the recognition of text entities in the assembly process is 92.19%, which is 3.1% higher than that of the BERT-BiLSTM-CRF model, and the best F1 score is 92.44%. The accuracy rate was 92.83% and the recall rate was 91.56%. The recognition and integration of the two kinds of data make the structure of BAKG clearer and richer in content, which can greatly improve the efficiency of on-site assembly management and guidance, and efficiently use assembly information.

Based on BAKG, the construction efficiency and management level of biomass heating system have been greatly improved, which means that more biomass heating systems will be put into use, and more and more biomass energy such as agricultural and forestry waste will be consumed, which can greatly reduce carbon emissions, mitigate global warming, and provide ecological protection.

However, this method requires the model data to match the text data, otherwise there will be cases where the knowledge will not align, for example, in the model data, the “support beam” is usually named “Channel steel -L7495” after the specification model. At present, the matching rule is adopted in this study to match the “Channel steel -L7495” and “support beam” each other. This method takes a certain amount of time to formulate rules, so the similarity algorithm is planned to be used in the future to combine model data with text data with high similarity, and to use more diverse assembly data for training and learning, so that the model has the ability to recognize and learn the polysemous word, so as to facilitate the matching between model data and process document data. Moreover, the information regarding the assembly process in this study comes from the assembly document, and information regarding the biomass boiler installation industry is not included. Therefore, in future work, we will consider integrating more sources and types of data into BAKG to expand and enrich the capacity of this knowledge graph, and will also apply BAKG to the intelligent inference of the assembly process with real-time dynamic information. At the same time, we will also consider applying the BAKG construction method to the construction process of the cement industry, which also includes the assembly of large-scale mechanical equipment, to give new vitality to construction management in the cement industry.

Author Contributions

Conceptualization, Z.C.; validation, Y.G.; formal analysis, F.W.; resources, Z.C.; data curation, J.A.; writing—original draft preparation, F.W.; writing—review and editing, Z.C. and Y.M. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Conflicts of Interest

Author Yong Gao is employed by the Wuhan Optics Valley Bluefire New Energy Co., Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. General framework for BAKG construction method.

Figure 2. BAKG conceptual pattern layer.

Figure 3. The entity extraction method framework.

Figure 4. 3DXML source file.

Figure 5. Entity recognition and relationship extraction based on 3DXML.

Figure 6. Schematic diagram of BiLSTM structure.

Figure 7. Schematic diagram of RoBERTa-BiLSTM-CRF structure (Chinese is utilized as the model input).

Figure 8. Comparison of entity recognition model performance: (a) F1; (b) Precision; (c) Recall.

Figure 9. Part of BAKG.

Table 1

Abbreviation definitions.

Abbreviations	Original Text
BERT	Bidirectional Encoder Representation from Transformers
RoBERTa	Robustly optimized BERT approach
BiLSTM	Bidirectional Long and Short Term Memory Networks
CRF	Conditional Random Field
LSTM	Long and Short Term
3DXML	3D Extensible Markup Language

Table 2

BAKG relationship table.

Relationship	Domains	Ranges
AssemblyTo	Insert/Load/Place/Fit/Tighten	Component Units/Part Units
HasAssemblyProcess	Component units/Part units	Insert/Load/Place/Fit/Tighten
Has Precautions	Component units/Part units	Precautions
HasPrepare	Final assembly body/Device units/Component units/Part units	Acceptance/Organization/Technology/Foundation/Location
HasRequirement	Component units/Part units	Strength/weld seam/axis deviation
LessThan	Weld seam/Axis deviation	Distance
Satisfy	Weld seam/Strength	Regulations
Has	Final assembly body	Device units
HasSub	Device units	Component units
HasPart	Component units	Parts units

Table 3

Assembly process document (partial).

No	Content
1	The installation of large boiler components should be carried out after the steel frame is installed, the column base is poured again, and the concrete curing is qualified. Before installation, the quality of steel frame welds and concrete should be carefully inspected to ensure accuracy.
2	Transport the lower large components onto the two strip foundations, and then adjust them in place: align the centerline of the lower large component grate with the centerline of the two strip foundations. Make the other positioning dimensions meet the requirements of the drawing. After calibration, it should be firmly welded to the base, and the height of the fillet weld should not be less than 6 mm.
3	After positioning, the centerline of the upper header should be kept horizontal, with a levelness (total length) not exceeding 3 mm, and the rear end should not be higher than the front end. The elevation of the centerline of the upper header and the centerline of the left and right lower headers should be ensured. The elevation deviation of the water level gauges on both sides should not exceed 5 mm. The centerline of the upper container should be kept horizontal, and wedge iron should be inserted in pairs between the upper large component base and the lower large component support beam to solve the problem. After adjustment, the wedge iron should be welded to the base plate.
4	Before positioning the upper large piece on the load-bearing steel frame, the following precautions should be taken: before positioning the upper large piece, the lower large piece should be installed and inspected to create safe conditions for the upper large piece to fall. When the upper large piece falls and is positioned on the load-bearing steel frame, the position must comply with the dimensions and technical requirements of the drawing. Adjust the thickness of the wedge iron to achieve horizontal positioning and elevation of the upper large components.
5	Weld the upper large component lower drag mud plate firmly to the lower large component support beam, with a fillet weld height of 6 mm.

Table 4

Experimental parameters.

Parameters	Value	Parameters	Value
Max_len	256	Epochs	10
Batch_size	16	Drop_rate	0.5
Lstm_dim	128	Learning_rate	10−5
Hidden_dim	768	Optimizer	Adam

Table 5

Comparison of entity recognition model performance.

No.	Model	F1 (%)	Precision (%)	Recall (%)
1	BERT-CRF	86.30	87.13	85.49
2	BiLSTM-CRF	87.07	87.67	86.49
3	BERT-BiLSTM-CRF	89.09	89.69	88.50
4	ALBERT-BiLSTM-CRF	76.04	80.61	71.95
5	RoBERTa-BiLSTM-CRF	92.19	92.83	91.56

Table 6

BAKG entity extraction results.

No.	Device Units	Component Units	Parts Units	Assembly Operation	Technological Requirements	Distance	Regulations
1	Boiler large components	Upper large piece/Steel frame	Column base	Place	Steel frame weld quality/Concrete quality	2 mm	Concrete Quality Control Standards
2		Lower large piece	Strip foundation/Base/Grate	Welding/Place	Foot filling weld seam/Positioning size/Centerline deviation	2 mm/6 mm	Drawing requirements
3		Upper large piece/Lower large piece	Upper container/Base/Support beam/Water level gauge	Adjust/Wedge iron/Welding	Elevation deviation/Centerline levelness/Front end not higher than rear end	5 mm/3 mm
4		Upper large piece/Steel frame	Preparation	Adjust/Place/Wedge iron	Horizontal positioning/Elevation	Precautions	Drawing dimensions/Technical requirements/Acceptance criteria
			Lower large piece is installed			Position
5		Upper large piece/Lower large piece	Lower drag mud board/Support beam	Welding	Foot filling weld seam	6 mm

References

1. Ma, L.L.; Tang, Z.H.; Wang, C.W.; Sun, Y.M.; Lv, X.F.; Chen, Y. Research Status and Future Development Strategy of Biomass Energy. Bull. Chin. Acad. Sci.; 2019; 34, pp. 434-442. [DOI: https://dx.doi.org/10.16418/j.issn.1000-3045.2019.04.008]

2. Liu, J.H.; Wang, J.F. Collaborative Assembly Design and Planning for Complex Products; Huazhong University of Science & Technology Press: Wuhan, Hubei, China, 2014; 247.

3. Fan, Z.F. Mechanical Equipment Installation Engineering Manual; Metallurgical Industry Press: Beijing, China, 2004; 948.

4. Liu, Q.S.; Liu, Y.L.; Liu, Y.J. Mechanical Equipment Installation Project; China Architecture & Building Press: Beijing, China, 2007; 483.

5. Zhou, Q.; Fan, Y. EBOP based research on complicated product process data model. Manuf. Autom.; 2008; 30, pp. 15-18.

6. Qiu, P.; Qiao, L.H. The Application of Assembly Process Planning Information Model in Information Integration. Group Technol. Prod. Mod.; 2009; 26, 6.

7. Liu, J.; Ni, Z.; Liu, X. Process information expression and management technology based on 3D process model. Comput. Integr. Manuf. Syst.; 2015; 21, pp. 1756-1763.

8. Dini, G.; Santochi, M. Automated Sequencing and Subassembly Detection in Assembly Planning. CIRP Ann.; 1992; 41, pp. 1-4. [DOI: https://dx.doi.org/10.1016/S0007-8506(07)61140-8]

9. Pedraza, G.; Diaz, M.; Lombera, H. An Approach for Assembly Sequence Planning by Genetic Algorithms. IEEE Lat. Am. Trans.; 2016; 14, pp. 2066-2071. [DOI: https://dx.doi.org/10.1109/TLA.2016.7530396]

10. Abdullah, A.; Ab Rashid, M.F.F.; Ponnambalam, S.G.; Ghazalli, Z. Energy efficient modeling and optimization for assembly sequence planning using moth flame optimization. Assem. Autom.; 2019; 39, pp. 356-368. [DOI: https://dx.doi.org/10.1108/AA-06-2018-091]

11. Rashid, M.A.; Mohamed, N.N.; Rose, A.M. A modified artificial bee colony algorithm to optimise integrated assembly sequence planning and assembly line balancing. J. Mech. Eng. Sci.; 2019; 13, pp. 5905-5921. [DOI: https://dx.doi.org/10.15282/jmes.13.4.2019.13.0469]

12. Suszyński, M.; Peta, K. Assembly Sequence Planning Using Artificial Neural Networks for Mechanical Parts Based on Selected Criteria. Appl. Sci.; 2021; 11, 10414. [DOI: https://dx.doi.org/10.3390/app112110414]

13. Masehian, E.; Ghandi, S. ASPPR: A New Assembly Sequence and Path Planner/Replanner for Monotone and Nonmonotone Assembly Planning. Comput. -Aided Des.; 2020; 123, 102828. [DOI: https://dx.doi.org/10.1016/j.cad.2020.102828]

14. de Giorgio, A.; Maffei, A.; Onori, M.; Wang, L. Towards online reinforced learning of assembly sequence planning with interactive guidance systems for industry 4.0 adaptive manufacturing. J. Manuf. Syst.; 2021; 60, pp. 22-34. [DOI: https://dx.doi.org/10.1016/j.jmsy.2021.05.001]

15. Zhang, D.H.; Liu, Z.Y.; Jia, W.Q.; Liu, H.; Tan, J.R. A Review on Knowledge Graph and Its Application Prospects to Intelligent Manufacturing. J. Mech. Eng.; 2021; 57, pp. 90-113. [DOI: https://dx.doi.org/10.3901/jme.2021.05.090]

16. Singhal, A. Introducing the Knowledge Graph: Things, Not Strings. Available online: https://blog.google/products/search/introducing-knowledge-graph-things-not/ (accessed on 16 May 2012).

17. Paulheim, H. Knowledge graph refinement: A survey of approaches and evaluation methods. Semant. Web; 2017; 8, pp. 489-508. [DOI: https://dx.doi.org/10.3233/SW-160218]

18. Xu, Z.L.; Sheng, Y.P.; He, L.R.; Wang, Y.F. Overview of Knowledge Graph Technology. J. Univ. Electron. Sci. Technol. China; 2016; 45, pp. 589-606. [DOI: https://dx.doi.org/10.3969/j.issn.1001-0548.2016.04.012]

19. Liu, Y.C.; Li, H.Y. Survey on Domain Knowledge Graph Research. Comput. Syst. Appl.; 2020; 29, pp. 1-12. [DOI: https://dx.doi.org/10.15888/j.cnki.csa.007431]

20. Liu, P.F.; Qian, L.; Zhao, X.W.; Tao, B. Continual learning framework of named entity recognition in aviation assembly domain. J. Zhejiang Univ. (Eng. Sci.); 2023; 57, pp. 1186-1194.

21. Gong, Z. Research of Product Assembly Knowledge Recommendation Method Based on Knowledge Graph Bidirectional Fusion. Master’s thesis; Zhejiang University: Hangzhou, China, 2020.

22. Sun, X.M.; Liu, S.M.; Shen, X.W.; Huang, D.L.; Bao, J.S. Digital twin-driven intelligent assembly method for high precision products. Comput. Integr. Manuf. Syst.; 2022; 28, pp. 1704-1716. [DOI: https://dx.doi.org/10.13196/j.cims.2022.06.010]

23. Nagy, L.; Abonyi, J.; Ruppert, T. Knowledge Graph-Based Framework to Support Human-Centered Collaborative Manufacturing in Industry 5.0. Appl. Sci.; 2024; 14, 3398. [DOI: https://dx.doi.org/10.3390/app14083398]

24. Shi, X.; Tian, X.; Ma, L.; Wu, X.; Gu, J. A knowledge graph–based structured representation of assembly process planning combined with deep learning. Int. J. Adv. Manuf. Technol.; 2024; 133, pp. 1807-1821. [DOI: https://dx.doi.org/10.1007/s00170-024-13785-4]

25. Bollacker, K.D.; Evans, C.; Paritosh, P.K.; Sturge, T.; Taylor, J. Freebase: A collaboratively created graph database for structuring human knowledge. Proceedings of the SIGMOD Conference; Vancouver, BC, Canada, 9–12 June 2008.

26. Vrandečić, D.; Krötzsch, M. Wikidata: A free collaborative knowledgebase. Commun. ACM; 2014; 57, pp. 78-85. [DOI: https://dx.doi.org/10.1145/2629489]

27. Lehmann, J.; Isele, R.; Jakob, M.; Jentzsch, A.; Kontokostas, D.; Mendes, P.N.; Hellmann, S.; Morsey, M.; Kleef, P.v.; Auer, S. et al. DBpedia—A large-scale, multilingual knowledge base extracted from Wikipedia. Semant. Web; 2015; 6, pp. 167-195. [DOI: https://dx.doi.org/10.3233/SW-140134]

28. Kim, H. Towards a sales assistant using a product knowledge graph. J. Web Semant.; 2017; 46–47, pp. 14-19. [DOI: https://dx.doi.org/10.1016/j.websem.2017.03.001]

29. Liu, Z.; Xiong, C.; Sun, M.; Liu, Z. Entity-Duet Neural Ranking: Understanding the Role of Knowledge Graph Semantics in Neural Information Retrieval. Proceedings of the Annual Meeting of the Association for Computational Linguistics; Melbourne, Australia, 15–20 July 2018.

30. Rasmussen, M.H.; Lefrançois, M.; Pauwels, P.; Hviid, C.A.; Karlshøj, J. Managing interrelated project information in AEC Knowledge Graphs. Autom. Constr.; 2019; 108, 102956. [DOI: https://dx.doi.org/10.1016/j.autcon.2019.102956]

31. Zhao, Y.; Smidts, C. A method for systematically developing the knowledge base of reactor operators in nuclear power plants to support cognitive modeling of operator performance. Reliab. Eng. Syst. Saf.; 2019; 186, pp. 64-77. [DOI: https://dx.doi.org/10.1016/j.ress.2019.02.014]

32. Shi, H.-B.; Huang, D.; Wang, L.; Wu, M.-Y.; Xu, Y.-C.; Zeng, B.-E.; Pang, C. An information integration approach to spacecraft fault diagnosis. Enterp. Inf. Syst.; 2021; 15, pp. 1128-1161. [DOI: https://dx.doi.org/10.1080/17517575.2019.1709663]

33. Collobert, R.; Weston, J.; Bottou, L.; Karlen, M.; Kavukcuoglu, K.; Kuksa, P. Natural Language Processing (Almost) from Scratch. J. Mach. Learn. Res.; 2011; 12, pp. 2493-2537.

34. Yang, T.; Jiang, D.; Shi, S.; Zhan, S.; Zhuo, L.; Yin, Y.; Liang, Z. Chinese Data Extraction and Named Entity Recognition. Proceedings of the 2020 5th IEEE International Conference on Big Data Analytics (ICBDA); Xiamen, China, 8–11 May 2020; pp. 105-109.

35. Qu, Z.; Zhang, Z.; Liu, S.; Cao, J.; Bo, X. Knowledge-driven recognition methodology for electricity safety hazard scenarios. Energy Rep.; 2022; 8, pp. 10006-10016. [DOI: https://dx.doi.org/10.1016/j.egyr.2022.07.158]

36. Futia, G.; Vetrò, A.; De Martin, J.C. SeMi: A SEmantic Modeling machIne to build Knowledge Graphs with graph neural networks. SoftwareX; 2020; 12, 100516. [DOI: https://dx.doi.org/10.1016/j.softx.2020.100516]

37. Wu, F.; Weld, D.S. Autonomously semantifying wikipedia. Proceedings of the International Conference on Information and Knowledge Management; Lisbon, Portugal, 6–10 November 2007.

38. Liu, Q.; Li, Y.; Duan, H.; Liu, Y.; Qin, Z. Knowledge Graph Construction Techniques. J. Comput. Reaserch Dev.; 2016; 53, pp. 582-600. [DOI: https://dx.doi.org/10.7544/issn1000-1239.2016.20148228]

39. Wang, X.-B.; Ning, T.; Wang, K. Parsing 3DXML File Format and Its Application. J. Graph.; 2010; 31, 33.

40. Zhou, X.X.; Gong, Q.M.; Yin, L.J.; Xu, H.Y.; Ban, C. Prediction of TBM stable section excavation parameters based on BLSTM-AM model. J. Rock Mech. Eng.; 2020; 39, pp. 3505-3515.

41. Graves, A.; Mohamed, A.r.; Hinton, G. Speech recognition with deep recurrent neural networks. Proceedings of the 2013 IEEE International Conference on Acoustics, Speech and Signal Processing; Vancouver, BC, Canada, 26–31 May 2013; pp. 6645-6649.

42. Lafferty, J.D.; McCallum, A.; Pereira, F. Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data. Proceedings of the International Conference on Machine Learning; Williamstown, MA, USA, 28 June–1 July 2001.