CN105824791A - Reference format checking method - Google Patents

Abstract

The present invention provides a method for checking the format of reference documents, comprising: step 1, expressing the format rules of bibliographical items of references using Schema, wherein the format of bibliographical items of references includes at least one of the following bibliographic items: responsible person, title, Reference type, publisher, publication date, page number; step 2, read each reference, and divide the bibliographic items; step 3, identify the bibliographic items of the references, and extract the identified bibliographic items into XML nodes; where The bibliographic items include at least one of the following: responsible person, title, place of publication, publisher, publication date, etc.; at the same time, it is judged whether the bibliographical bibliographical items include the document type mark, and if not, add the reference according to the bibliographical items The document type mark of the document; step 4, using the format rules for the bibliographical items of the references to verify the bibliographical items.

Description

A Method for Checking the Format of References

technical field

The invention relates to the technical field of text processing, in particular to a method for checking the format of reference documents.

Background technique

In all kinds of papers, it is inevitable to cite previously published references to help readers understand the background knowledge of the paper. Generally, when citing references, it is necessary to provide the author (author), title (title), publisher (publisher (that is, where the work is published), publication page number (publishpage), and publication date (publishyear). However, in conference proceedings, large journals and other documents with a large concentration of papers, each paper will cite a large number of references, so it is difficult to ensure that every paper cites references in the same format.

At present, it is relying on reviewers to review the format requirements while reviewing the paper, and then the editor will review again; but this method of purely relying on manual review is difficult to ensure that there are no omissions.

Contents of the invention

Aiming at the problem that the manual review of the reference format in the prior art is prone to omissions, and it is impossible to ensure that every paper in the anthology or periodicals uses the same rules to quote the references, the technical problem to be solved by the present invention is to provide a The method and system for checking the format of references can automatically check whether the citations of references in the electronic version of the paper conform to the preset rules, so as to ensure the standardization of the format of references, improve efficiency, and prevent omissions.

In order to solve the above problems, an embodiment of the present invention proposes a method for checking the format of reference documents, including:

Step 1. Express the format rules of bibliographical items of references using Schema, wherein the format of bibliographical items of references includes at least one of the following bibliographical items: scribe, title, type of reference, publisher, date of publication, page number;

Step 2. Read each reference and segment the bibliographical items;

Step 3, identifying the bibliographical items of the reference, and extracting the identified bibliographic items into XML nodes; wherein the bibliographic items include at least one of the following: responsible person, title, place of publication, publisher, publication date, etc.; meanwhile, Judging whether the bibliographical item of the reference includes a document type flag, if not, adding the bibliographic type flag of the reference according to the bibliographical item;

Step 4. Verify the bibliographic items by using the format rules for the bibliographical items.

Wherein, the method also includes:

Step 5. When there is an error in the bibliographical item of the reference, modify the bibliographical item; specifically include;

When the error is a missing item, complete the bibliographical item and add punctuation to reorganize to form a standardized reference;

When there are multiple errors, delete the bibliographic item and reorganize with punctuation to form a standardized reference;

When the error is a wrong item, modify it according to the standard format and add punctuation to reorganize it to form a standard reference.

Wherein, said step 2 includes:

Step 21, using ApachePOI to identify the document to extract the reference content;

Step 22. Segment the extracted reference content to obtain bibliographic items, including:

Identify the symbols in the references to determine whether the references include non-half-width symbols, and replace them with corresponding half-width symbols if they are included;

Segment the description items according to the bibliographic symbols.

Wherein, the step 3 includes: using a preset bibliographic identification model to identify the references cited in the text of the paper to extract the bibliographic items of the references, wherein the bibliographic identification model is based on a preset corpus Learned; specifically include:

Step 31, extracting the corpus;

Step 32, using the preset corpus, using the NER algorithm for training to obtain a bibliographic item recognition model;

Step 33, judging whether the reference document includes the reference document type parameter, and if not, judging the type of the reference document by using the bibliographic item of the reference document.

Wherein, the step 33 includes:

Step 331: Build a decision tree for the bibliographic item; specifically include:

Calculate the Gini index Gini, entropy Entropy, and error rate (Error) by the following formula:

$\mathrm{<span\; class="notranslate">\; G</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

$\mathrm{<span\; class="notranslate">\; E.</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

Error＝1-max{p(i)|iin[1,n]}

And calculate the information gain Gain and the information gain rate GainRate

Gain(U,V)＝Entropy(U)-Entropy(U,V))

GainRate(U,V)=Gain(U,V)/Entropy(V)

To determine the root node of the decision tree and the best grouping variables;

Step 332, preprocessing the data. The specific steps include: checking the completeness of the bibliographic items of the references, so as to convert non-numeric and non-name data into digital and name data; If there are missing bibliographical items, fill in the vacant value according to the relevant bibliographical items in the reference; delete the negligible bibliographical items according to the relevance of the bibliographical items; generalize the data;

Step 333, using the decision tree of the references and the preprocessed data to determine the type. In the embodiment of the present invention, the WEKA platform is used for type determination.

Wherein, the step 333 specifically includes:

Step 3331, import the data set to be tested;

Step 3332, acquiring the data to be tested after preprocessing in step 332;

Step 3333, putting the processed data set into different learning schemes for learning and establishing a prediction model to predict unknown instances;

Step 3334, evaluate the predicted result.

The beneficial effects of above-mentioned technical scheme of the present invention are as follows:

With the emergence of a large number of scientific and technological papers, the relevant departments of the state have promoted the standardization and standardization of academic journals, and the format standards of references have been regarded as the rules that authors and editors must abide by. In the process of writing academic papers, the author must learn the standard norms to complete the paper with high quality, and the editors also need to learn the standard norms to complete the checking work of the paper efficiently. Therefore, both authors and editorial staff need a convenient tool to check the standardization of reference formats. Because different types of references have different formats, and the same reference has many bibliographic items, it is inevitable that the author will make mistakes during the writing process. Therefore, there are still a large number of irregularities in the references of academic papers, which is difficult for editors. Personnel increase the difficulty of checking. This topic mainly solves the standardization of reference format, which has high practical value.

1) This research can make the work of checking the format of references more intelligent, reduce errors in the description of references, and improve the efficiency of checking the style of references.

2) A correct understanding of the bibliographic items of the references will facilitate the further exploration and utilization of the references in the future (such as analyzing citation and cited information, evaluating the research level of academic works, and sorting out the research results of relevant authors, etc.).

The embodiment of the present invention can detect the standardization of the format of reference documents, specifically locate the wrong position, and prompt how to correct it, which provides convenience for researchers. The research results of this topic are of great value for improving the quality of digital publishing, promoting the efficient dissemination and utilization of document information, and saving the labor cost of typesetting.

Description of drawings

Fig. 1 is the schematic flow chart of the embodiment of the present invention;

Figure 2 is the code of the XML-based OOXML structure in the Word file;

FIG. 3 is a schematic diagram of the hierarchical relationship among Word document elements;

Figure 4 is the references of 10 format specifications as an example;

FIG. 5 is a schematic structural diagram of a reference decision tree in an embodiment of the present invention;

Figure 6 is a schematic diagram of some records in the ARFF file;

Fig. 7 is the schematic diagram of the conversion in the embodiment of the present invention;

Fig. 8 is 10 reference documents to be tested as example;

Figure 9 is a partial result of bibliographic item identification;

Fig. 10 is the detection result of the reference document to be tested in Fig. 8;

Figure 11 is the XML file generated during the detection process.

detailed description

In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will describe in detail with reference to the drawings and specific embodiments.

The embodiment of the present invention proposes a method for checking the format of reference documents, including:

Step 1. Use Schema to define a standard format template. The purpose is to verify the correctness of the XML files generated by the bibliographic items of the references.

Step 2. Read each reference and segment the bibliographic items. The purpose of using bibliographical segmentation is to make the recognized category as simple as possible and lay a foundation for the next step of identifying bibliographical items. Only by accurately segmenting each bibliographical item can the accuracy of bibliographical item identification be improved.

Step 3. Identify the bibliography description items. On the basis of the segmentation method of reference bibliographic items in step 2, the identification of each bibliographic item of the reference bibliography is carried out, and the description includes the responsible person, title, place of publication, publisher and publication date, etc.

Step 4, extracting the identified bibliographic items into XML nodes.

Step 5, judging the document type flag. The document type marks specified in GB/T7714-2005 are as follows: general book (M), compilation (G), standard (S), periodical (J), computer program (CP), dissertation (D), report (R ), patent (P), database (DB), electronic bulletin (EB), tape (MT), disk (DK), conference proceedings (C), CD (CD), newspaper (N), online network (OL). In step 4), check whether the title and description items identified in step 4) contain the document type flag specified in GB/T7714-2005.

Step 6. Verify based on the standard format template described in step 1. If the document type flag is included, the Schema standard format template of the corresponding document type is called for verification. If the document type flag is not included, the document type is first judged, and then the Schema standard format template of the corresponding document type is called for verification. If the validation passes, the reference is formatted correctly, if not, the reference is incorrectly formatted.

Step 7. Determine the wrong bibliographic item and modify it. Including checking the order of the bibliographic items, finally generating the correct XML instance. The specific design idea is as follows: When the XML file fails to pass the Schema verification, the specific bibliographic items in the XML file that have not passed the verification are extracted. For the bibliographic items that have not passed the verification, it can be summarized into three situations, one is missing items, and the other is There are multiple items, and the other is a wrong item. In case of missing items, complete the bibliographical items and reorganize them with punctuation marks to form references in a standardized format. For multiple items, delete the bibliographic item and add punctuation to reorganize to form a standardized reference. In the case of wrong items, modify according to the standard format, add punctuation marks and reorganize to form references with standard format.

Each step of the embodiment of the present invention is described in detail below.

Step 1. Express the format rules of bibliographical items of references by using codes, wherein the format of bibliographical items of references includes at least one of the following bibliographical items: scribe, title, type of reference, publisher, date of publication, page number.

In the embodiment of the present invention, it can be expressed by XMLSchema language, wherein the bibliography bibliographic item format includes at least one of the following bibliographic items: scribe, title, reference type, publisher, publication date, page number;

The following is an example of the format rules for bibliographical items expressed in XMLSchema language, which is a conference proceedings expressed in XMLSchema

(1) <? xmlversion="1.0"encoding="GB2312"? >

(2)<xs:schemaxmlns:xs="http://www.w3.org/2001/XMLSchema"

(3)xmlns="http://www.w3school.com.cn"

(4) targetNamespace＝http://www.w3school.com.cn"

(5) elementFormDefault="qualified">

(6)<xs:elementname="reference">

(7) <xs:complexType>

(8) <xs:sequence>

(9)<xs:elementname="author"type="xs:string"/>

(10)<xs:elementname="title"type="xs:string"/>

(11)<xs:elementname="type"type="xs:string"/>

(12)<xs:elementname="publish"type="xs:string"minOccirs="0"/>

(13)<xs:elementname="publisher"type="xs:string"/>

(14)<xs:elementname="publish_year"type="xs:string"/>

(15)<xs:elementname="page_number" minOccirs="0"/>

(16)<xs:simpleType>

(17)<xs:restrictionbase="xs:string">

(18)<xs:pattemvalue="(\d{1,4}-)?\d{1,4}"/>

(19)</xs:restriction>

(20)</xs:simpleType>

(21)</xs:element>

(22)</xs:sequence>

(23)</xs:complexType>

(24)</xs:element>

(25)</xs:schema>

Step 2, using the preset library to extract the references cited in the text, and extract the bibliographic items therein; the step 2 includes two parts: the content extraction of the reference bibliography and the bibliographic item extraction of the reference bibliography. Therefore step 2 specifically includes:

Step 21, using Apache POI to identify the document to extract the reference content.

Most of the existing documents are stored in the Microsoft Word format or a Word-compatible format. In a Microsoft Word document, information is stored in the XML-based OOXML (OpenOfficeXML) format. Therefore, Apache POI3.13 can be used to identify documents.

Hereinafter, this application uses an example to illustrate the meaning of the OOXML structure. Edit two paragraphs of text in Microsoft Word2013, which are "identification of bibliographical items in Chinese references" and "thesis". The corresponding XML codes are shown in Figure 2.

In the code, the <w:document> element is the root element of the document, and all other elements are its child elements. Several namespaces are defined behind the element through attributes.

The <w:body> element is where the content of the document resides and is the only required element. It contains many sub-elements, see the OOXML standard for details. Among the many sub-elements, there are three basic elements, namely the <w:p> element, the <w:r> element and the <w:t> element. Among them, the <w:p> element represents a paragraph and is used to define a content that starts on a new line; the <w:r> element represents the content of the sentence level, which can be sentences, mathematical content, smart tags and user-defined tags, etc. , the sentence is the smallest unit that can be styled; the <w:t> element represents the specific text content in the sentence. A schematic diagram of the hierarchical relationship of these elements is shown in FIG. 3 .

Since the position of the references is fixed, after knowing the information in the XML-based OOXML (OpenOfficeXML) format, Apache POI3.13 can be used to identify the document to extract the contents of the references.

Step 22, extracting bibliographic items from the contents of the references.

Since the reference is composed of several bibliographic items, it is necessary to segment the bibliographic items before identifying them. Specifically include:

Symbol normalization step: identify the symbols in the references to determine whether non-half-width symbols are included in the references, and replace them with corresponding half-width symbols if they are included;

Segmentation step: Segment the bibliographic items according to the bibliographic symbols stipulated in the national standard GB/T7714-2005.

In the reference document GB/T7714-2005, all bibliographic symbols are used as prefixes. For example, do not use any signs before the responsible person of the first bibliographic item in the reference; "." is used as the prefix of the title item and the extracted document title item, etc. Through the analysis of GB/T7714-2005, different prefixes are used as separation points to divide the description items.

In the process of sampling and analyzing some graduate dissertations, it is found that there are different types of formatting errors, and the probability model of writing errors can be calculated by using statistical methods. For example, if there is a "." in a reference, "." will not appear in this reference as a separator between bibliographic items; if a reference contains "." as a delimiter between bibliographic items Separator, then "." will not appear as a separator between bibliographic items in this reference.

Step 3. Using a preset bibliographic identification model to identify the references cited in the text of the thesis to extract the bibliographic items of the references, wherein the bibliographic identification model is obtained by learning from a preset corpus.

In the embodiment of the present invention, a named entity recognition method (Stanford NamedEntity Recognizer, NER) of Stanford University based on a conditional random field is used. NER can mark entities according to categories, such as names of people, companies, regions, genes and proteins, etc. NER is equipped with a well-designed feature extractor to identify named entities, and a training model can be obtained after training. In theory, the data used for training, that is, the more manually marked texts, the better the NER recognition effect. The model needs to be retrained to meet new requirements.

Therefore, step 3 specifically includes;

Step 31, extract the corpus; the specific corpus uses the extracted annotated corpus of "People's Daily" in January 1998 and the 2015 Peking University edition of "Overview of Chinese Core Periodicals".

in;

1. The annotated corpus of "People's Daily" extracted in January 1998: Since the "People's Daily" corpus has a large proportion of nouns such as names of people and places, it can be used as a good training corpus.

2. The 2015 edition of Peking University's "Overview of Chinese Core Periodicals"; in addition to names of people and places, it is usually hoped to better identify commonly used journal names and keywords commonly used in some paper titles in papers, so " An Overview of Chinese Core Periodicals in conjunction with the People's Daily.

For example, the title of the graduation thesis or the title of the journal often contains the word "based on", so on the basis of the "People's Daily" corpus, the name of the journal and the statistics that appeared in the 2015 Peking University edition "Overview of Chinese Core Journals" were added. The keywords commonly used in the title, these parts are finally combined together as the system training set of this experiment, which is saved in the testdata.tsv file and used for the closed test of the system.

In addition, the 2015 edition of Peking University's "Overview of Chinese Core Journals" is the post-text references extracted from the graduation thesis of the school, and it is used to form a test set for the open test of the system.

Wherein, the extracted reference content may be as shown in FIG. 4 .

Step 32, using the preset corpus, and using the NER algorithm for training to obtain a bibliographic item recognition model.

NER provides two ways to train the model, namely command line mode and configuration file mode.

In the embodiment of the present invention, configuration files may be used.

Specifically, the name of the configuration file in StanfordNER is austen.prop, and its parameters are modified as shown in Table 1 below

Table 1 austen.prop modification parameter list

Among them, trainFile specifies the data set used for training, and serializeTo specifies the name of the model output after training. Save the modified configuration file and put it in the root directory of the program together with the training data set testdata.tsv, and execute the following command:

java –cpStanford-ner.jaredu.stanford.nlp.ie.crf.CRFClassifier –propausten.prop”

When the execution is successful, ner-model.ser.gz will be generated in the directory, which is the model obtained from the training data.

After obtaining the bibliographic item identification model, the bibliographic item in step 2 can be identified through the bibliographic item identification model.

Since the information in the references may be incomplete, and the type of the reference may be missing, and whether the type of the reference is accurate has an important impact on the subsequent check of the format of the reference, so the embodiment of the present invention may further include:

Step 33, judging whether the reference document includes the reference document type parameter, and if not, judging the type of the reference document by using the bibliographic item of the reference document.

Specifically, step 33 includes:

Step 331: Build a decision tree for the bibliographic items.

Still taking the content of the reference as shown in FIG. 4 as an example, it includes 10 format specifications. It can be seen from Figure 4 that each reference is composed of many bibliographic items, and the composition of bibliographic items is different for different types of references. Through the analysis of 10 references, the bibliographic items and attribute value descriptions of the references are summarized, as shown in Table 2.

Table 2 Bibliographic items and attribute value descriptions of references

The information model of each reference in Figure 4 obtained after data transformation is shown in Table 3:

Information model of each reference in Table 3 and Figure 4

Therefore, the bibliographic item decision tree shown in Figure 5 can be constructed according to Table 3. According to the decision tree in Figure 5, unknown types of documents can be predicted. For example, an existing reference is as follows:

Zhu Gang. New fluid finite element method and forward and reverse mixing problems of impeller machinery. Beijing: Tsinghua University, 1996.

According to the decision tree shown in Figure 5, it can be predicted that it belongs to a dissertation.

To generate the decision tree shown in Figure 5, there are two key issues:

One is how to select a currently best grouping variable from numerous input variables? For example, why should the publisher type be used as the root node of the decision tree? Why choose the author type as the child node of the lower layer instead of other bibliographic items?

The second is how to find an optimal split point from the many values of the grouping variables? For example, the publisher type is a name type, and its attributes include "journals", "educational institutions", "others", and "publishers". Why do you choose "educational institutions" as the split point? After solving these two key problems, a decision tree can be easily constructed.

The concept of "purity" needs to be introduced in the decision tree. There are three kinds of measuring purity methods commonly used, are respectively Gini index (Gini), entropy (Entropy), error rate (Error); Gini index (Gini), entropy (Entropy), error rate (Error) can be calculated by the following formula in the embodiment of the present invention (Error):

Assuming that the attributes of the bibliographic item have n types of different attribute values i (i=1, 2,...,n), the proportion of each type of attribute value p(i) = the number of i-th attribute values/the total value of the attribute The value range of p(i) is [0,1].

$\mathrm{<span\; class="notranslate">\; G</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

$\mathrm{<span\; class="notranslate">\; E.</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

Error＝1-max{p(i)|iin[1,n]}

The above formulas 1-3 of the three purities are the larger the value, the more "impure", and the smaller the more "pure". Practice has proved that the choice of the three formulas has little effect on the final classification accuracy. In the embodiment of the present invention, the entropy formula is also used, and two commonly used attribute selection variables are derived from the entropy formula, which are information gain (Gain) as in formula 4 and information gain rate (GainRate) as in formula 5.

Gain(U,V)＝Entropy(U)-Entropy(U,V))

GainRate(U,V)=Gain(U,V)/Entropy(V)

In information theory, the process of information transmission is realized as a transmission system composed of information source, channel and destination. The information source is the sending end of the information, and the information destination is the receiving end of the information. Taking the above reference type mark prediction as an example, the author type (T1), report number (T2), patent number (T3), publisher type (T4), year volume mark (T5), and page number (T6) are used as input variable, the reference type is marked as an output variable. The decision tree regards the output variable (reference type mark) as the information U sent by the source, and the input variable as a series of information V received by the sink.

Gain(U,V)＝Entropy(U)-Entropy(U,V))

GainRate(U,V)=Gain(U,V)/Entropy(V)

Using GainRate to calculate the above two key issues respectively, the calculation process is as follows:

Take the author type T1 as an example: calculate Entropy(U), Entropy(U|T1), Gains(U,T1), GainsR(U,T1) respectively, including 2 journals, dissertations and books, and reports, 1 each of conference proceedings, patent and standard type documents.

Assuming that the bibliographic item has M different types of attributes, the attribute value u _i (i=1,2,...,M), the proportion of each type of attribute value is p(u _i ), and the author type T1 has N different attributes Value t _1j (j=1, 2, . . . , N).

$\mathrm{<span\; class="notranslate">\; E.</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1.832</span>}$

$\mathrm{<span\; class="notranslate">\; E.</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; T</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1.279</span>}$

Gain(U,T1)=Entropy(U)-Entropy(U|T1)=0.553

GainRate(U,T1)=Gains(U,T1)/Entropy(V)=0.628

That is to say, the information gain rate of the author type (T1) is 0.628, and other bibliographic items are calculated in the same way, and finally the maximum information gain rate of T4 is 1.275, so T4 should be selected as the best grouping variable, which is the decision tree. root node.

There are 4 attributes in the publisher type, which are "journal", "educational institution", "other" and "publisher", so how to choose the split point, the calculation process is similar to the above, after calculation, the "educational institution" is obtained The maximum value of information gain rate is 3.948, so "educational institution" should be selected as the best grouping variable.

It can be seen from the above analysis that the decision tree in the embodiment of the present invention is an intuitive decision analysis method, and its advantages are obvious. The decision tree model is readable and descriptive, which is helpful for manual analysis; and it has high execution efficiency, and can be used repeatedly after only one construction, and can naturally embed the prior knowledge of experts.

Problems such as data inconsistency, data duplication, data containing noise, and high-dimensional data may appear in the bibliographic items of this reference. Therefore, before classifying the bibliographic items, it is necessary to preprocess the data. That is, the step 33 also includes:

Step 332, preprocessing the data.

Specifically, the data preprocessing steps include:

Step 3321, check the completeness of the bibliographic items of the references.

Since there are two types of variables in the decision tree: numeric and named; so the main preprocessing work that needs to be done before constructing the decision tree is to convert non-numeric and non-named data into numeric or named.

In data mining, appropriate attributes are selected from the original data as data mining attributes. The data principles adopted are: as far as possible, give the attribute name and attribute value a clear meaning, remove duplicate data, remove negligible fields, and reasonably select associated fields. . The following describes the preprocessing process in detail.

The original data is the extracted references, and then the bibliographic items of the references are split. Each split bibliographic item can be regarded as the attribute of each record. The following table 4 is a fragment selected from the original data .

Table 4 Raw data records

As can be seen from the above table, after a reference is split, some field values are vacant, and some fields can be ignored, so step 332 may include the following three sub-steps:

Step 3322 , check whether there are missing bibliographic items in the reference documents, and if so, fill the vacant values according to the relevant bibliographic items in the reference documents.

For example, "patent country", "patent number", "report number" in Table 4, and then fill in all the blank values of the selected data. The filling principle of the vacant value is to comply with the type of the existing value of the field. For example, if some values of a field in the existing record are numeric, then the filling values of other vacant values of the field will also be numeric. If some of the field's values are named, the fill-in values of other blank values in this field are also named.

Step 3323, according to the relevancy of the descriptive items, delete the negligible descriptive items. For example, in the reference content shown in Figure 4, the sequence numbers (ie, the previous 1, 2, 3...10) have no effect on the result prediction, but will increase the complexity of the calculation, so they can be deleted. For example, the place of publication field, no matter where the place of publication is, will not affect the type of the final reference, so it can be ignored for the "place of publication" field.

Step 3323, generalize the data. This is due to the fact that in raw recorded data, each field can be summarized into several classes. For example: for the field of "responsible person", the value of the "responsible person" field can be summarized into two categories, one is the name of a specific person, and the other is the name of an organization. The type of the document is predicted according to whether the "responsible person" field is the name of the person or the name of the organization, and has nothing to do with the name of the specific person or the name of the organization. Therefore, the data of the "responsible person" field can be generalized into two types: person name and organization name. And so on, the data of all similar situations will be generalized.

After the above steps, the preprocessed data shown in Table 5 can be obtained.

Table 5 preprocessed data

It can be seen from Table 5 that the preprocessed fields and field values are all in English. Of course, this is only one way of the embodiment of the present invention, and any other form can be used to express the preprocessed fields and field values. Since the WEKA system is used in the embodiment of the present invention to determine the types of references, better calculation results can be obtained by using English fields and field values.

The field value of each field of the embodiment of the present invention is illustrated with the example of Table 5:

The "responsible person" field values are PER.Individual and PER.Group, where PER refers to individuals or groups of people mentioned in the literature, and PER.Individual and PER.Group are subcategories of PER, referring to individuals, groups of people or organizations, respectively.

The field value of "title feature tag" is title_D_tag, title_C_tag, etc. For example, title_C_tag means that the title of the conference proceedings contains the feature tag of "conference collection", and the title of the report generally contains the feature tag of "report"; other types have no value of the feature tag Mark it as no.

The "Publisher" field values are PUB.Press, PUB.Journal, PUB.School, PUB.Institution and NUL, which refer to non-school presses, periodicals, school presses and research institutes respectively, and NUL refers to missing items.

In step 33, a decision tree is constructed through step 331, and after data preprocessing is performed through step 332, it is necessary to determine the types of references. That is, the method also includes:

Step 333, using the decision tree of the references and the preprocessed data to determine the type. In the embodiment of the present invention, the WEKA platform is used for type determination.

The process of data mining in the WEKA platform is as follows:

1) Import the dataset to be tested;

2) Preprocessing the data to be tested (step 332 has been completed);

3) Place the processed data set in different learning schemes for learning and establish a prediction model to predict unknown instances;

4) Evaluate and visualize the predicted results. The following is a detailed introduction to the above four steps.

Therefore step 333 specifically comprises:

Step 3331, import the data set to be tested.

Since the data formats that can be processed under the WEKA platform are CSV and ARFF files, but the most ideal format is ARFF files, so the ARFF format files are used here, and the file format needs to be converted before importing the data set to be tested. The raw data is stored in an EXCEL file, which is first converted to a CSV file and then converted to an ARFF file. Some of the records in the ARFF file are shown in Figure 6.

Step 3332, acquire the preprocessed data of step 332.

Step 3333, select a specific classification algorithm for training and testing classification. In the classification module of the WEKA system, about 50 kinds of classification algorithms are integrated. In the embodiment of the present invention, three classic classification algorithms NativeBays, J48 (decision tree) and ZeroR are selected to classify the test set.

The fourth step evaluates the results of different classification algorithms. There are many methods for evaluating classification accuracy, mainly including cross-validation, holdout, leave-one-out, and back-substitution. The intersection and hold methods are the most commonly used. The leave-one-out method can be regarded as a special case of the crossover method. The back substitution method is generally not used because it evaluates overfitting and leads to high classification accuracy. The visualization of the results can not only visualize the results of a classification, but also visualize the results of a data set. Among them, the visualization of the data set shows a two-dimensional scatter diagram about each pair of attributes, and the output of a certain classification is visualized to show the classification error, tree, cost curve, ROC curve, etc., which are used to evaluate the performance of each learning scheme. .

Although some algorithms have relatively high accuracy in judging the document type flag, it is impossible to achieve 100% accuracy, which will affect the final accuracy of document format error detection.

In order to reduce the prediction error as much as possible, the method of feature marking is adopted in the embodiment of the present invention. That is to say, after the decision tree judges the document type mark, the judgment is made according to the characteristic mark. If the two judgment results are the same, it will be taken as the final result. If the two judgments are not the same, the result of the judgment based on the characteristic mark shall prevail. Table 6 lists the correspondence between various feature marks and reference types.

Table 6 The relationship between feature marks and document types

Step 4: Check the identified bibliographic items of the reference documents by using the format rule codes of the bibliographic items of the reference documents. Specifically, it includes: generating corresponding XML documents for each identified bibliographic item according to the reference type mark, and then using Schema to verify; if the verification is passed, it means that the format of the document is correct; otherwise, it means that the format of the document is wrong.

In the embodiment of the present invention, the XML document of the periodical type generated therein is as follows

(1) <? xmlversion="1.0"encoding="GB2312"standalone="no"? >

(2)<referencexmlns="http://www/w3school.com.cn"

(3)xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"

(4)xsi:schemaLocation="http://www.w3school.com.cnJ_pre.xsd">

(5) <authorauthorLoc="1">Chen Luyao</author>

(6)<titletitleLoc="2">Extraction and logical description of information document structure trust model</title>

(7)<typetypeLoc="2">J_pre</type>

(8) <publishpublishLoc="3">Beijing</publish>

(9) <publisherpublisherLoc="4">Computer application research</publisher>

(10)<publish_yearpublish_yearLoc="5">2015</publish_year>

(11)<volumn_markvolumn_markLoc="6">27</volumn_mark>

(12) <page_numberoage_numberLoc="7">4624-4629</page_number>

(13)</reference>

After the schema template is validated, an error message will be returned if it fails the compiler validation. The error information includes the error type and error description. The error type can roughly determine the problem. If you want to locate the error specifically, you need to combine the error description. Three common error types are summarized below;

(1) cvc-complex-type.3.1. This type of error is that the attribute value in XML does not match the attribute value defined in Schema, for example, the order of tags is reversed.

(2) cvc-complex-type.2.4.a. This type of error is that the logical structure in the XML file does not conform to the Schema specification, for example, there are elements that are not defined in the Schema specification.

(3) cvc-complex-type.2.4.b. This error type is that the content of the XML file is incomplete, such as missing items.

The types of format errors that appear in the references can be attributed to one or more of the above three types of errors. The process of error detection is shown in Table 7.

Table 7 Reference format error detection algorithm

In the algorithm in Table 7, R is the set of reference documents to be tested, and r is a reference document in the R set. ERRORS is a collection of error types that XMLSchema failed to verify references, and Er is the error type corresponding to a reference. After the detection, in order to solve the problem of locating the error item, it is necessary to convert the error description information provided by the compiler into the corresponding position information. An example is used below to illustrate how to perform the conversion, as shown in Figure 7.

It can be seen from Figure 7 that there are three types of bibliographic errors: multiple items, missing items, and out of order. The position number changes corresponding to each case are different. Therefore, Algorithm 2 is designed according to the position number and bibliographic content, as shown in Table 8.

Table 8 Algorithm for locating incorrect items in references

After the algorithm analysis, take the 10 references to be tested in Figure 8 as an example, use this system to carry out normative detection on the 10 references, and the detection results are shown in Figures 9 and 10.

The patent is applied to checking the format of the reference documents after the text, and correcting the incorrect format of the reference documents. The reference information extraction in the present invention adopts extracting from Microsoft Word documents, and is also applicable to extracting references from text files. The following is described by way of example.

Fig. 9 is part of the results of identification of bibliographic items of references. Taking the first reference in Fig. 8 as an example, the first 8 lines in Fig. 9 are the identification results of the first reference. Among them, the first line "J_pre" indicates that the first document lacks the document type flag, and it is predicted to be a journal type through the judgment of the document type sign; the second line indicates that "Chen Luyao" is identified as the author; the third line indicates that the "information document "Extraction and Logical Description of Structural Trust Model" is identified as the title; the fourth line indicates that "Beijing" is identified as the place of publication; the fifth line indicates that "Computer Application Research" is identified as the publisher of the journal type; the sixth line indicates that "2010 " to identify the year of publication; the seventh line means to identify "27" as the volume; the eighth line means to identify "4624-4629" as the page number.

Figure 10 shows the detection results of 10 references. For references with irregular formats, specific error location information is prompted and modification suggestions are given to facilitate modification. Figure 11 is the XML file generated during the detection process.

The present invention has the following beneficial effects:

With the emergence of a large number of scientific and technological papers, the relevant departments of the state have promoted the standardization and standardization of academic journals, and the format standards of references have been regarded as the rules that authors and editors must abide by. In the process of writing academic papers, the author must learn the standard norms to complete the paper with high quality, and the editors also need to learn the standard norms to complete the checking work of the paper efficiently. Therefore, both authors and editorial staff need a convenient tool to check the standardization of reference formats. Because different types of references have different formats, and the same reference has many bibliographic items, it is inevitable that the author will make mistakes during the writing process. Therefore, there are still a large number of irregularities in the references of academic papers, which is difficult for editors. Personnel increase the difficulty of checking. This topic mainly solves the standardization of reference format, which has high practical value.

1) This research can make the work of checking the format of references more intelligent, reduce errors in the description of references, and improve the efficiency of checking the style of references.

2) A correct understanding of the bibliographic items of the references will facilitate the further exploration and utilization of the references in the future (such as analyzing citation and cited information, evaluating the research level of academic works, and sorting out the research results of relevant authors, etc.).

The embodiment of the present invention can detect the standardization of the format of reference documents, specifically locate the wrong position, and prompt how to correct it, which provides convenience for researchers. The research results of this topic are of great value for improving the quality of digital publishing, promoting the efficient dissemination and utilization of document information, and saving the labor cost of typesetting.

The above description is a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (6)

1. a list of references format checking method, it is characterised in that including:

Step 1, by Reference Citation item format convention employing Schema state, wherein said Reference Citation item lattice Formula includes at least one following bibliographical particulars: owner, autograph, list of references type, publisher, publication date, the page number；

Step 2, read each bar list of references, carry out bibliographical particulars cutting；

Step 3, identify Reference Citation item, and the bibliographical particulars that will identify that extracts and becomes XML node；Wherein said bibliographical particulars Including following at least one: owner, inscribe one's name, publish ground, publisher, publication date etc.；Simultaneously, it is judged that this list of references writes Whether record item includes document type mark, without the document type mark then adding this list of references according to bibliographical particulars；

Step 4, utilize described Reference Citation item format convention that bibliographical particulars is verified.

List of references format checking method the most according to claim 1, it is characterised in that described method also includes:

Step 5, when Reference Citation item exist mistake time, bibliographical particulars is modified；Specifically include；

When mistake is lacuna, completion bibliographical particulars the list of references of the restructuring form format specification that puts in the stops；

When mistake is multinomial, delete this bibliographical particulars the list of references of the restructuring form format specification that puts in the stops；

When mistake is wrong item, the reference of the restructuring form format specification that puts in the stops after modifying according to the form of specification Document.

List of references format checking method the most according to claim 1, it is characterised in that described step 2 includes:

Step 21, Apache POI is utilized to be identified document extracting list of references content；

Step 22, the list of references content extracted is carried out cutting to obtain bibliographical particulars, including:

Symbol in list of references is identified, to judge whether list of references includes non-DBC case, if included, It is replaced with corresponding DBC case；

With symbol, bibliographical particulars is carried out cutting according to recording.

List of references format checking method the most according to claim 1, it is characterised in that described step 3 includes: utilize pre- If bibliographical particulars identification model list of references recited in paper word is identified extracting the work of described list of references Record item, presets corpus and carries out what study obtained according to wherein said bibliographical particulars identification model；Specifically include:

Step 31, extraction corpus；

The corpus that step 32, employing are preset, utilizes NER algorithm to be trained obtaining bibliographical particulars identification model；

Step 33, judging whether list of references includes list of references type parameter, if do not included, utilizing list of references Bibliographical particulars judges the type of described list of references.

List of references format checking method the most according to claim 4, it is characterised in that described step 33 includes:

Step 331: construct the decision tree of bibliographical particulars；Specifically include:

By below equation calculate gini index Gini, entropy Entropy, error rate (Error):

$Gini=1-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}p{\left(i\right)}^2$

$Entropy=-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}p\left(i\right)*{\log }_{2}p\left(i\right)$

Error=1-max{p (i) | i in [1, n] }

And calculate information gain Gain and information gain-ratio GainRate

Gain (U, V)=Entropy (U)-Entropy (U, V))

GainRate (U, V)=Gain (U, V)/Entropy (V)

To determine root node and the best packet variable of decision tree；

Data are carried out pretreatment, the most suddenly include by step 332: check the bibliographical particulars integrity of described list of references, It is converted into numeric type, title type with the data by nonnumeric type, non-title type；Search in list of references and whether there is the work lacked Record item, if had, is filled with vacancy value according to bibliographical particulars relevant in list of references；According to the dependency of bibliographical particulars, delete Except the most insignificant bibliographical particulars；Data are generally changed；

Step 333, the decision tree of list of references and pretreated data are utilized to carry out type decision.In the embodiment of the present invention In, use WEKA platform to carry out type decision.

List of references format checking method the most according to claim 5, it is characterised in that described step 333 specifically includes:

Step 3331, import data set to be tested；

Step 3332, obtaining step 332 carry out pretreated testing data；

Step 3333, will process after data set be placed in different Learning Schemes and carry out learning and set up forecast model and predict Unknown example；

Step 3334, to prediction result be estimated.