A logically aware multi-relationship extraction method for scientific literature in the field of materials

Abstract

The present invention relates to a logically perceived multi-relation extraction method for scientific literature in the field of materials, and belongs to the field of natural language processing. First, dynamic keywords are used instead of predefined static patterns to perform named entity recognition on texts such as material composition, performance and performance values ​​in material literature, and the entities are divided into material entity groups, performance entity groups, and performance value entity groups according to the type of the entity; then the quantity logical relationship and position logical relationship between the entities are used to divide the clauses, and the above logical relationships between the entities in the clauses are used to establish mapping relationships between multiple material composition entities and specific performance value entities to complete data extraction. The method of the present invention can extract material composition and specific performance data from the literature, provide data support for subsequent scientific analysis, and does not require pre-labeling of data or writing of complex regular expressions. The threshold for use is low, and a high accuracy and recall rate are achieved at the same time.

Description

Technical Field

[0001] The present invention relates to a logically perceived multi-relation extraction method for scientific literature in the field of materials, and belongs to the field of natural language processing. Background Art

[0002] Extracting data from scientific literature primarily involves two steps in natural language processing: named entity recognition (NER) and entity relationship extraction (ERE). Research in this area, both domestically and internationally, can be categorized into two areas: deep learning and rule-based learning.

[0003] In terms of deep learning, the representative ones are Liu et al. (<Text mining of hypereutectic Al-Si alloys literature based on active learning> , 2021) adopted a combination of active learning to solve the problem, successfully expanding the manually annotated 8,845 material entities to 16,677 material entities. In addition, Wei et al. {<Construction Method of Material Domain Knowledge Graph Based on Natural Language Processing>, 2022} constructed a Bi-GRU-GNN-CRF material entity relationship joint extraction method, which automatically acquired material domain knowledge such as material entities, relationships, and process flows from patents, papers, and other material documents. Experimental results show that this solution has good precision and recall rate in Chinese data, and can effectively improve the knowledge coverage of the material knowledge graph. The knowledge coverage rate of the material domain knowledge graph constructed based on this method reached 80%, which can provide more comprehensive knowledge support for intelligent material research and development.

[0004] In terms of rules, the representative ones are Wang et al. (<Automated pipeline for superalloydata by text mining> ,2022) believe that due to the lack of a large amount of effectively annotated data in specific fields, it is difficult to directly apply machine learning methods to the extraction of material science literature data in specific fields. However, there are cases where features are prominent in specific fields. Therefore, they constructed a method based on regular algorithms for identifying material entities and distance algorithms for extracting data from material science literature, proving the feasibility of extracting data from articles using rules.

[0005] In recent years, scientists believe that with the help of machine learning, as long as there is sufficient data and rule-discovery algorithms, relevant technologies can be used to assist in the discovery of materials. For example, Malik et al. predicted the final products of solid-state reactions through material data, and Dipendra et al. predicted the properties of materials through machine learning. In addition, some scientists have successfully contributed to the research of new materials using reinforcement learning or other methods. However, using deep learning technology to assist in material discovery requires a large amount of data, and some experimental data are mostly only used within the laboratory. Due to confidentiality conditions or other factors, these material data cannot be shared publicly or easily obtained. Therefore, in the process of combining actual materials research with machine learning, a lot of material data is still missing. Generally speaking, machine learning, especially deep learning, is sensitive to data. Therefore, in the absence of material data, it is difficult for machine learning to build a large number of models like other fields. Therefore, how to quickly obtain a large amount of material data has become a bottleneck for researchers in the Materials Genome Project to combine machine learning to predict material properties.

[0006] Extracting data from scientific literature is essentially the process of entity relationship extraction (ERE) in natural language processing. However, current research focuses on ERE in general domains, such as extracting entity relationships between names of people and places. There has been less attention paid to ERE in specific domains, resulting in a limited number of publicly available datasets.

[0007] Although the above studies on extracting data from materials science literature have their own merits, they also have their weaknesses. The deep learning-based approach requires large-scale data annotation, and there are currently no public datasets. The rule-based material data extraction scheme mainly uses regular expressions to extract material entities, and its application scope is relatively limited. At the same time, the existing methods do not fully consider the logical relationship between entities, which leads to errors in the matching relationship between material entities and performance entities. Summary of the Invention

[0008] To solve the above problems, the present invention provides a logically perceived multi-relationship extraction method for scientific literature in the field of materials.

[0009] The technical solution adopted by the present invention is: a logically perceived multivariate relationship extraction method for scientific literature in the field of materials, comprising the following steps:

[0010] (1) Extracting specific material composition entities, performance entities, and performance value entities from material literature based on dynamic keywords;

[0011] (2) Determine the quantity logical relationship and position logical relationship between each entity group;

[0012] (3) Based on the judgment results of the logical relationship, different schemes are selected to construct the relationship between material composition, performance and specific performance values.

[0013] Step 1.1 Perform named entity recognition on relevant entities in sentences of material documents. Use a dictionary-based filter to filter out entities that do not conform to the rules and filter out common words. After completing entity recognition, divide each entity into material component entity group (M), performance name entity group (P), and performance value entity group (V) according to its type.

[0014] Step 2.1: Determine whether the relationship |M|×|P|=|V| exists among the three entities.

[0015] Step 2.2 If |M| × |P| = |V|, divide the clauses according to the positional relationship of the entities in the material composition entity group (M), the performance name entity group (P), and the performance value entity group (V); if |M| × |P| = |V| is not satisfied, continue to determine whether |M| = 1 holds;

[0016] Step 3.1 If |M|×|P|=|V|, after completing the clause division, the clause logical relationship construction algorithm is used to establish the mapping relationship between material composition and specific performance values; if |M|×|P|=|V| is not satisfied, and |M|=1, the greedy matching algorithm is used to establish the mapping relationship between material composition and specific performance values; if |M|×|P|=|V| is not satisfied, and |M|>1, the shortest distance matching algorithm is used to establish the mapping relationship between material composition and specific performance values.

[0017] Furthermore, the clause division in Step 2.2 adopts the minimum matching principle. Under the premise that the quantitative relationship of the three entities satisfies |M|×|P|=|V|, there must be at least one entity of each type. At the same time, it is agreed that all three entities exist in the first clause and the quantitative relationship satisfies |M|×|P|=|V|. The remaining clauses do not need to meet this requirement. The missing content is completed by the first clause, and the clause division is completed in sequence.

[0018] Furthermore, the clause logical position relationship construction algorithm in Step 3.1 specifically includes:

[0019] The material composition entity group (M) and the performance name entity group (P) are combined by Cartesian product to form a new entity group (M×P), and then this entity group is matched one-to-one with the performance value entity group (V) according to the order within the group.

[0020] Furthermore, in Step 3.1, when |M|=1, all performance value entities are greedily associated with the material entity. When |M|>1, the positional relationship between the material entity and the performance value entity is determined, and the nearest entities are selected to combine and establish the relationship between the material entity and the specific performance value.

[0021] The beneficial effects of the present invention are:

[0022] 1. The present invention proposes a logically perceived multi-relationship extraction method for scientific literature in the materials field, a named entity recognition and extraction based on keywords, and a composite material literature data extraction scheme based on the logical relationship between entity quantity and position. It can extract material composition and specific performance data from the literature, providing data support for subsequent scientific analysis.

[0023] 2. The proposed data extraction model based on logical relationships does not rely on complex regular expressions and does not require pre-labeling of data, which lowers the threshold for use. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a multi-relationship extraction method based on logical perception of scientific literature in the field of materials according to the present invention;

[0025] Figure 2 This is a diagram of clause division logic relationship data extraction Figure 1 ;

[0026] Figure 3 This is a diagram of clause division logic relationship data extraction Figure 2 ;

[0027] Figure 4 This is a schematic diagram of data extraction using the greedy matching algorithm;

[0028] Figure 5 This is a schematic diagram of data extraction using the shortest distance matching algorithm;

[0029] Figure 6 It is a schematic diagram of Cartesian product matching;

[0030] Figure 7 It is a schematic diagram of a one-to-one matching connection operation;

[0031] Figure 8 This is an example diagram of the data extraction results of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] Example 1

[0034] like Figure 1As shown in FIG, a logically perceptual multivariate relationship extraction method for scientific literature in the field of materials includes the following steps:

[0035] (1) Extracting specific material composition entities, performance entities, and performance value entities from material literature based on dynamic keywords;

[0036] (2) Determine the quantity logical relationship and position logical relationship between each entity group;

[0037] (3) Based on the judgment results of the logical relationship, different schemes are selected to construct the relationship between material composition, performance and specific performance values.

[0038] The specific steps are as follows:

[0039] Step 1.1 Perform named entity recognition on the relevant entities in the sentences of the material literature, and use a dictionary-based filter to filter out entities that do not meet the rules and filter out common words. For example, using the Cu keyword can extract two entities, Currently and Cu-Mo. However, Currently does not belong to the material component, but is a specific common word. The present invention filters it out by judging whether the entity belongs to a common word, and retains the Cu-Mo material component entity. After completing entity recognition, the material component entity group (M), performance name entity group (P), and performance value entity group (V) are formed according to the type of each entity;

[0040] Step 2.1: Determine whether the relationship |M|×|P|=|V| exists among the three entities.

[0041] Step 2.2 If |M|×|P|=|V|, then divide the clauses according to the positional relationship of each entity in the material composition entity group (M), the performance name entity group (P) and the performance value entity group (V); Figure 2 and Figure 3 The clause division adopts the minimum matching principle. Under the premise that the quantitative relationship of the three entities satisfies |M|×|P|=|V|, the number of entities of each entity must be at least one. At the same time, it is agreed that all three entities exist in the first clause and the quantitative relationship satisfies |M|×|P|=|V|. The remaining clauses do not need to meet this requirement. The missing content is completed by the first clause (Clause 1), and the clause division is completed in sequence.

[0042] Specifically, first determine whether there are missing items in the subsequent clauses, such as Figure 2Clause 2 is missing the performance name entity group (P) compared with Clause 1. Figure 3 In the example, the material component entity group (M) is missing. For universal representation, we first use a mask to fill in the missing components, then take out the missing specific components from Clause 1 and align them to complete them. After the completion is completed, we construct an algorithm based on the logical relationship of the clauses and perform a Cartesian product combination of the material component entity group (M) and the performance name entity group (P), as shown in the following example: Figure 6 As shown, a new entity group (M×P) is formed, and then the entity group is matched one-to-one with the performance value entity group (V) according to the order within the group, as shown in Figure 7 shown.

[0043] After completing the clause division, the clause logical relationship construction algorithm is used to establish the mapping relationship between material composition and specific performance values.

[0044] Step 3.1 If |M|×|P|=|V| and |M|=1, a greedy matching algorithm is used to establish a mapping relationship between material composition and specific performance values; all performance value entities are greedily associated with the material entity. Figure 4 As shown, the 180Hv performance value entity and the 200Hv performance value entity both establish a relationship with the material composition entity Mo-Cu.

[0045] Step 3.2 If |M|×|P|=|V| is not satisfied, and |M|>1, the shortest distance matching algorithm is used to establish the mapping relationship between the material composition and the specific performance value. Determine the positional relationship between the material entity and the performance value entity, and select the nearest entity to combine and construct the relationship between the material entity and the specific performance value. Figure 5 As shown, the distance between the Co-4Cu material entity and the 274Hv performance value entity is 3 (the number of intervening words), and the distance between Cu-Base and 274HV is 6. Therefore, Co-4Cu and 274HV are selected to combine and construct a relationship.

[0046] Example 2

[0047] Using the method in Example 1, the present invention extracted 1220 data from 589 open access documents, such as Figure 8 The following table shows a partial result. 855 of the results were correct, with a precision of 70.4%. The 855 correct results were divided by the year of the source document, with data from 2011-2019 as the training set and data from 2020-2022 as the test set. This dataset was used to compare with existing entity relationship extraction methods. The three evaluation metrics of precision, recall, and F1 score were compared. The specific calculation scheme is as follows:

[0048] First, define four basic elements:

[0049] TP: True Positive, the predicted sample is true and the actual sample is true.

[0050] FP: False Positive, the predicted sample is true, but the actual sample is false.

[0051] TN: True Negative, the predicted sample is false, and the actual sample is false.

[0052] FN: False Negative, the predicted sample is false, and the actual sample is true.

[0053] The accuracy calculation formula is:

[0054]

[0055] The calculation formula for recall is:

[0056]

[0057] The calculation formula of F1 score is:

[0058]

[0059] By comparison, the method proposed in the present invention has better comprehensive effects, as shown in Table 1:

[0060] Table 1 Comparison of entity relationship extraction results of material documents

[0061]

[0062]

[0063] From the various indicators in Table 1, it can be seen that the logically aware multivariate relation extraction method (LOSAMRE) for scientific literature in the field of materials proposed in this paper is the best compared with other methods, among which Super Alloy Digger (<Automated pipeline for superalloy data by text mining> ,2022) is a rule-based material literature entity relationship extraction method similar to this method. As can be seen from the table, the original method can only be used for the ultra-high temperature alloys mentioned in the original text. When extracting content, the comprehensive score of this method is low. CasRel( , 2020)LSTM JointExtraction(<Joint Extraction of Entities and Relations Based on a NovelTagging Scheme> , 2017), BERT EM (<Matching the Blanks:Distributional Similarityfor Relation Learning> , 2019) and BERT SP (<Extracting Multiple-Relations in One-Pass with Pre-Trained Transformers> , 2019) are four machine learning entity relationship extraction schemes. It can be seen from the table that due to the small amount of training set data, CasRel and LSTM Joint Extraction perform poorly, BERT SP Although the accuracy is 0.7% higher than that of this paper, the comprehensive F1 score is not as good as the model proposed in this paper. SP As a machine learning model, it requires pre-labeled data. In the current field of material data extraction, there is a lack of public data sets, which has a certain cost. However, the extraction scheme proposed in this paper can complete data extraction without labeling data. From another perspective, BERT SP The results also show that the content extracted by the invention can also be used as the input of the machine learning model. Therefore, the data extracted by the present invention can also be used as a solution to establish a data set for the materials science literature data extraction model.

Claims

1. A logically aware multi-relationship extraction method for scientific literature in the field of materials, characterized by: The steps include: (1) Extracting specific material composition entities, performance entities, and performance value entities from material literature based on dynamic keywords; (2) Determine the quantity logical relationship and position logical relationship between each entity group; (3) Select different schemes based on the judgment results of logical relationships to construct the relationship between material composition, performance and specific performance values; Specifically include: Step 1.1 Perform named entity recognition on relevant entities in sentences of material documents. Use a dictionary-based filter to filter out entities that do not conform to the rules and filter out common words. After completing entity recognition, divide the entities into material component entity group M, performance name entity group P, and performance value entity group V according to their types. Step 2.1: Determine whether the relationship |M|×|P|=|V| exists among the three entities. Step 2.2 If |M|×|P|=|V|, then divide the clauses according to the positional relationship of each entity in the material composition entity group M, the performance name entity group P, and the performance value entity group V; if |M|×|P|=|V| is not satisfied, then continue to determine whether |M|=1 holds; Step 3.1 If |M|×|P|=|V|, after completing the clause division, the clause logical relationship construction algorithm is used to establish the mapping relationship between material composition and specific performance values; if |M|×|P|=|V| is not satisfied, and |M|=1, the greedy matching algorithm is used to establish the mapping relationship between material composition and specific performance values; if |M|×|P|=|V| is not satisfied, and |M|>1, the shortest distance matching algorithm is used to establish the mapping relationship between material composition and specific performance values.

2. The method for extracting multivariate relationships based on logical perception of scientific literature in the field of materials according to claim 1 is characterized by: The clause division in Step 2.2 adopts the minimum matching principle. Under the premise that the quantitative relationship of the three entities satisfies |M|×|P|=|V|, there must be at least one entity of each type. At the same time, it is agreed that all three entities exist in the first clause and the quantitative relationship satisfies |M|×|P|=|V|. The remaining clauses do not need to meet this requirement. The missing content is completed by the first clause, and the clause division is completed in sequence.

3. The method for extracting multivariate relationships based on logical perception of scientific literature in the field of materials according to claim 1 is characterized by: The clause logical relationship construction algorithm in Step 3.1 specifically includes: The material composition entity group M and the performance name entity group P are combined by Cartesian product to form a new entity group M×P, and then this entity group is matched one-to-one with the performance value entity group V according to the order within the group.

4. The method for extracting multivariate relationships based on logical perception of scientific literature in the field of materials according to claim 1 is characterized by: In Step 3.1, when |M|=1, all performance value entities are greedily established with the material component entity.

5. The method for extracting multivariate relationships based on logical perception of scientific literature in the field of materials according to claim 1 is characterized by: In Step 3.1, when |M|>1, the positional relationship between the material composition and the specific performance value is determined, and the nearest entities are selected and combined to construct the relationship between the material composition and the specific performance value.

Images