NASICON-type solid electrolyte materials named entity recognition dataset|材料科学数据集|实体识别数据集

1.Framework overview.This paper proposed a pipeline to construct high-quality datasets for text mining in materials science. Firstly, we utilize the traceable automatic acquisition scheme of literature to ensure the traceability of textual data. Then, a data processing method driven by downstream tasks is performed to generate high-quality pre-annotated corpora conditioned on the characteristics of materials texts. On this basis, we define a general annotation scheme derived from materials science tetrahedron to complete high-quality annotation. Finally, a conditional data augmentation model incorporating materials domain knowledge (cDA-DK) is constructed to augment the data quantity.2.Dataset information.The experimental datasets used in this paper include: the Matscholar dataset publicly published by Weston et al. (DOI: 10.1021/acs.jcim.9b00470), and the NASICON entity recognition dataset constructed by ourselves. Herein, we mainly introduce the details of NASICON entity recognition dataset.2.1 Data collection and preprocessing.Firstly, 55 materials science literature related to NASICON system are collected through Crystallographic Information File (CIF), which contains a wealth of structure-activity relationship information. Note that materials science literature is mostly stored as portable document format (PDF), with content arranged in columns and mixed with tables, images, and formulas, which significantly compromises the readability of the text sequence. To tackle this issue, we employ the text parser PDFMiner (a Python toolkit) to standardize, segment, and parse the original documents, thereby converting PDF literature into plain text. In this process, the entire textual information of literature, encompassing title, author, abstract, keywords, institution, publisher, and publication year, is retained and stored as a unified TXT document. Subsequently, we apply rules based on Python regular expressions to remove redundant information, such as garbled characters and line breaks caused by figures, tables, and formulas. This results in a cleaner text corpus, enhancing its readability and enabling more efficient data analysis. Note that special symbols may also appear as garbled characters, but we refrain from directly deleting them, as they may contain valuable information such as chemical units. Therefore, we converted all such symbols to a special token <sYm>. Moreover, numerical and terminological abbreviations are also retained, as they can greatly aid in the extraction of knowledge from materials science literature. Finally, the abstract and main text are processed as inputs for subsequent stages, while the metadata such as the title, authors, and references are stored in a database to ensure the traceability of literature source.2.2 Tokenization and labeling.ChemDataExtractor, a chemical natural language processing toolkit, is firstly employed for splitting segments of materials articles into sentences and tokenizing each sentence into individual tokens. Then, for labeling these tokens, eight entity tags of descriptor including Composition, Structure, Property, Processing, Characterization, Application, Feature, and Condition are defined, of which corresponding descriptions and examples are shown in Table S1. Based on the aforementioned definitions, 55 materials science literature are manually annotated by three materials scientists through the EasyData toolkit. To ensure the effectiveness and consistency of the annotations, data labeled by each annotator underwent a joint review by two others.2.3 The details of dataset file.NASICON entity recognition dataset contains 2343 sentences (samples) in total 4857 entities (labels). Note that different sentences are separated by blank lines in the dataset file (nasicon_ner_org.csv). There are three columns in the file:(1) Words: the word lists obtained by splitting sentences with spaces;(2) POS tag: the lists of part-of-speech tagging for each word using NLTK toolkit;(3) Entity tag: the lists of manual annotation based on the eight pre-defined entity types, including the BIO tag.For example, the tags for “B(B’)O6 octahedra” in the file include {“B(B’)O6”, NNP, B-Structure}, {“octahedra”, NN, I-Structure}. Among them, “NNP” and “NN” respectively represent parts of speech, “B-Structure” represents the tag of begin token of entity, and “I-Structure” represents the tags of inside tokens of entity. More details of this dataset are presented in the main text.3.Citation.Please cite our papers related to our NASICON entity recognition dataset if it is helpful to your research:(1) Liu Y, Ge X Y, Yang Z W, Sun S Y, Liu D H, Avdeev M, Shi S Q. An automatic descriptors recognizer customized for materials science literature[J]. J. Power Sources, 2022, 545: 231946.(2) Liu Y, Liu D H, Ge X Y, Yang Z W, Ma S C, Zou Z Y, Shi S Q. A high-quality dataset construction method for text mining in materials science[J]. Acta Phys. Sin., 2023, 72(7): 070701. (in Chinese) [刘悦, 刘大晖, 葛献远, 杨正伟, 马舒畅, 邹喆乂, 施思齐. 高质量的材料科学文本挖掘数据集构建方法[J]. 物理学报, 2023, 72(7): 070701.]