TWI594135B - Plagiarism detecting method of information in english - Google Patents

Description

Plagiarism detection method for English materials

The invention relates to a plagiarism detection method for English materials.

With the demand and popularization of Taiwan's high education, the output of students' papers has also increased substantially. Most of the major academic research results are published in English, but English writing is not easy for Taiwanese students. As time goes by to pass the exam or obtain a degree, the plagiarism is also increasing. In terms of academic theory, plagiarism is not only disrespectful to the original author, nor can it further develop professional knowledge.

However, most of the current plagiarism detection software (such as Turnitin, CopyCatch or EVE2) is a comparison of the paper database or network resources, and only based on a single word (word) comparison, so often synonyms can not It is detected to reduce the accuracy of plagiarism detection.

In view of the above problems, the object of the present invention is to provide a plagiarism detection method for English materials, which can not only use grammatical structure analysis and extract words and phrases in sentences, but also use spoofing as a plagiarism detection unit. And supplemented by semantics to improve the correctness of the comparison using a single word in the prior art.

To achieve the above object, a plagiarism detection method for an English document according to the present invention is implemented on a computer, and the plagiarism detection method includes a grammar processing program, a language recognition program, and a similarity comparison program. The grammar processing program includes a parsing step of separately parsing a source file and a user file to obtain a source word data and a source phrase data, and a user word data and a use. Participant materials. The phrase recognition program includes a set generation step of comparing the source word data with the user word data and comparing the source phrase data with the user phrase data to generate a set of word data and comparing the sources. The phrase material and the user phrase data are used to generate a set of phrase materials. The similarity comparison program performs similarity comparison on the semantics of the aggregated word data and the aggregated document data to calculate the user file by the semantic similarity and order similarity of the aggregated word data and the aggregated phrase data. The degree of plagiarism relative to the source document.

In an embodiment, the source word data or the user word data respectively comprise at least one word, and the source phrase material or the user phrase material respectively comprise at least one phrase.

In one embodiment, the parsing step uses a natural language processing technique.

In an embodiment, in the grammar processing program, a sentence segmentation step, a complete plagiarism detection step, and a root restoration step are further included.

In an embodiment, in the parsing step, the source file and the user file are compared with a dictionary for part-of-speech tagging and phrase capture.

In one embodiment, the words within the aggregated word material correspond to the words within the dictionary.

In an embodiment, in the phrase recognition program, the phrase is used as the comparison unit.

In an embodiment, the words in the aggregate word data include all the words restored by the root of the source file and the user file, and the phrase language in the aggregate document data includes the source file and the user file. The same phrase.

In an embodiment, in the calculation of semantic similarity and order similarity, the phrase is used as the comparison unit.

In an embodiment, in the similarity comparison program, the semantic similarity between the word-to-word, the word-to-word, and the phrase-to-speech is calculated in the aggregate word data and the aggregate document data.

In an embodiment, in calculating the semantic similarity of the word to the word, the word part is calculated according to the part of the word.

In an embodiment, in calculating the semantic similarity of the word to the word, all the words in the set word data are compared with each other to find the maximum value of the similarity of the word to the word.

In an embodiment, in calculating the semantic similarity of a word to a phrase or a phrase to a phrase, the calculation is performed on a word basis.

In an embodiment, in calculating the semantic similarity of the word to the phrase, each word in the set word data is compared with each word in the set phrase data to find the word. The maximum similarity of the word to the phrase.

In an embodiment, in calculating the semantic similarity of the phrase to the phrase, all the words in the collective phrase data are compared with each other to find a maximum similarity of the phrase to the phrase.

In an embodiment, the semantic similarity or order similarity is calculated according to the similar maximum value of the word pair, the similar maximum value of the word to the phrase, and the similar maximum value of the phrase to the phrase.

In one embodiment, when the degree of plagiarism is above a threshold, the user file has plagiarism.

As described above, the plagiarism detection method according to the English document of the present invention includes a grammar processing program, a phrase recognition program, and a similarity comparison program. Among them, the source document and the user file are separately parsed, and the source word data and the user word data are compared and the source phrase data and the user phrase data are compared, and the source document data is compared. The user's phrase data is compared with the semantics, and the degree of plagiarism of the user file relative to the source file can be calculated. Therefore, compared with the prior art, the present invention can not only utilize the grammatical structure analysis and extract words and phrases in sentences, but also use the phrase as the comparison unit of plagiarism detection, supplemented by semantic meaning, thus improving the habit The correctness of the alignment is determined using a single word in the technique.

P01‧‧‧Sentence steps

P02‧‧‧ Complete plagiarism detection step

P03‧‧‧ Grammar Analysis Steps

P04‧‧‧ Root Restoration Step

S01‧‧‧ grammar handler

S02‧‧‧Word recognition program

S03‧‧‧similarity comparison program

FIG. 1 is a flow chart showing the procedure of a plagiarism detection method for English materials according to a preferred embodiment of the present invention.

2 is a flow chart showing the steps of the syntax processing program of FIG. 1.

The plagiarism detection method of the English language according to the preferred embodiment of the present invention will be described below with reference to the related drawings, wherein the same elements will be described with the same reference numerals.

Please refer to FIG. 1 , which is an English text of a preferred embodiment of the present invention. The program flow chart of the plagiarism detection method.

The plagiarism detection method of the English document of the present invention can be implemented on a computer (such as but not limited to a software program). The plagiarism detection method of the English data includes a grammar processing program S01, a phrase recognition program S02, and a similarity comparison program S03.

First, the grammar handler S01 performs syntax processing on a source file and a user file. The source file may include a plurality of files, each file may include at least one sentence, and the user file may also include at least one sentence. Here, the source file refers to a file in the database, and the source thereof may be, for example, but not limited to, a thesis database or a network resource, or an English file obtained by other means. In addition, the user file refers to the file written by the user and is to be detected as a copy of the source file.

The grammar processing program S01 performs sentence cutting on the contents of the source file and the user file, and analyzes and extracts all the words and phrases in the sentence through the grammatical structure, and performs processing operations such as part-of-speech tagging and root-restore. The present invention uses a Natural Language Processing (NLP) technique for grammar processing and analysis, and uses Stanford Parser analysis tools for grammar processing and analysis. The Stanford Parser analysis tool is a grammar parser developed by The Stanford Natural Language Processing Group.

Please refer to FIG. 2, which is a flow chart of the steps of the syntax processing program S01 in FIG. Here, the grammar processing program S01 may include a sentence breaking step P01, a complete plagiarism detecting step P02, a parsing step P03, and a root resting step P04.

First, the sentence step P01 is performed.

In the step S01 of the sentence segment, the source file and the user file are respectively cut into independent sentences using Stanford Parser, and the phrase, word (word, or word) relationship and clause are found from the tree structure. And other grammatical structures.

Next, a full plagiarism detection step P02 is performed.

In the complete plagiarism detection step P02, it is detected in the user file that the sentence in the source file is completely copied. In other words, the sentence in the user file and the sentence in the source file are compared first to find the completely plagiarized part, so as to reduce the complexity of subsequent calculation and processing.

Next, a parsing step P03 is performed.

In the parsing step P03, the source file and the user file are respectively separated. Grammatical analysis is performed to obtain a source word data and a source document data respectively, and obtain a user word data and a user phrase material. Among them, Syntactic Analysis (or Parsing) is a grammar rule and structure that uses a computer to analyze sentences, and then develops in a tree structure. In addition, Stanford Parcer is also used to compare source files and user files with an English vocabulary database (such as WordNet) for word-based tagging and phrase capture. In other words, the parsing step P03 is to use the tree structure generated by Stanford Parser and its dependence relationship after the sentence has been broken and completely plagiarized, and use WordNet as an auxiliary tool to extract the source file and the user file. All words and phrases. Among the part-of-speech tagging, the tokens obtained by Stanford Parcer are converted into WordNet readable form: including four parts of speech: Nouns, Verbs, daggerives and Adverbs. In addition, in the phrase acquisition, the phrase is used as the comparison unit of plagiarism detection, supplemented by semantic meaning, and the tree structure and the dependence relationship generated by Stanford Parser are used respectively to extract all the words.

Finally, the root restoration step P04 is performed.

In the root restoration step P04, the words and phrases extracted in step P03 are used to restore the roots by using WordNet, so as to avoid misjudgment of the same words due to singular plural or tense. After the above four steps, the grammar processing program S01 can be completed, and the source word data and the source phrase data, as well as the user word data and the user phrase data, can be obtained respectively. Here, the source word data or the user word data may respectively contain at least one word, and the source phrase material or the user phrase material respectively includes at least one phrase (or a combination of meaningful English words).

Hereinafter, the above-described grammar processing program S01 will be described in actual English sentences. For simplification of the description, the source file is exemplified by a sentence, for example, "John has turned on his new radio.", and the user file is also exemplified by a sentence, for example, "John switched the new radio on."

After using Stanford Parser for part-of-speech tagging and grammar analysis, and extracting all the words and phrases, the results are as follows: S ₁₁ ={John,has,turned on,his new radio},tree_p ₁₁ ={turned on ,his new radio},T ₁₁ ={John,has}. In addition, S ₁ = {John, switched on, the new radio}, tree_p ₁ = {the new radio}, dep_p ₁ = {switched on}, T ₁ = {John}. Here, S _ni is the i-th sentence of the nth source file, and tree_p _ni is the phrase set extracted from the tree structure in the i-th sentence of the nth source file, T _ni is the nth source file the i-th sentence word set, S _j for the j-th sentence of the user file, tree_p _j for the j-th sentence of the user file, a collection of phrases from the retrieved tree, dep_p _j for the first user files In the j sentence, the set of phrases retrieved from Typed dependencies, and T _j is the set of words in the jth sentence of the user file.

Then the root is restored, and you can get: S ₁₁ ={John,have,turn on,his new radio},tree_p ₁₁ ={turn on,his new radio},T ₁₁ ={John,have}. In addition, S ₁ = {John, switch_on, the new radio}, tree_p ₁ = {the new radio}, dep_p ₁ = {switch_on}, T ₁ = {John}. Here, the source word data is T ₁₁ , the source phrase data is tree_p ₁₁ , the user word data is T ₁ , and the user phrase data is tree_p ₁ plus dep_p ₁ .

Next, referring to FIG. 1, the phrase recognition program S02 is performed. The phrase recognition program S02 includes a set generation step of comparing the source word data with the user word data and comparing the source phrase data with the user phrase data to generate a set word. In addition, the collection generation step also compares the source phrase data with the user phrase data to generate a set of phrase materials. Here, the phrase recognition program S02 still uses the phrase as the unit of comparison.

In the previous program, all the words in the user file have been extracted, which are tree_p ₁ and dep_p _{1 respectively} . Among them, the words in dep_p ₁ have been recognized by WordNet, but tree_p ₁ is simply from the tree structure. Take, it does not make sense for plagiarism. Therefore, through the comparison of the source document and the sentence in the user file, the meaningful phrase is filtered, and the word collection and the phrase collection of the two sentences are combined to generate the aggregate word data and the collection document data.

The process is to first select a qualified phrase through a comparison of sentences and a query of a machine-readable dictionary (ie, WordNet). Ineligible phrases need to be disassembled, and the disassembled phrase is re-screened until all the disassembled phrases have been filtered or disassembled into words (words). Among them, the words that cannot be paired (different) in the source file and the user file are disassembled, and the identifiable words in the disassembled words are classified into the collective word data, and the pieces obtained by disassembling are disassembled. After the words are paired, the pairing is successfully put into the collection document data. If there is no matching, then the WordNet query is entered. If it can be queried, it will be included in the income collection word data. If the query is not re-disassembled, it will not be included in the collection document data. When the phrase cannot be found in WordNet, the phrase recognition program S02 is completed. Among them, the words in the collection word data contain The source file and all the words in the user file that have been restored by the root, and the phrase in the collection phrase data contains at least the same phrase in the source file as the user file. Here, the words in the collection word data correspond to the words in the WordNet dictionary. The so-called relative refers to words that can be queried in the WordNet dictionary and are identifiable or meaningful.

The English sentence of the above source file and user file is still used, and is obtained by the syntax processing program S01: S ₁₁ ={John,have,turn on,his new radio},tree_p ₁₁ ={turn on,his new radio},T ₁₁ ={John,have}. In addition, S ₁ = {John, switch_on, the new radio}, tree_p ₁ = {the new radio}, dep_p ₁ = {switch_on}, T ₁ = {John}.

After the process of comparison, screening, disassembly, etc. of the phrase recognition program S02, the results are as follows: S ₁₁ = {John, have, turn_on, his, new radio}, S ₁ = {John, switch_on, the , new radio}, collection word data = {John, have, turn_on, his, John, switch_on, the}, and collection phrase data = {new radio, new radio}.

Finally, please refer to FIG. 1 again to perform the similarity comparison program S03.

The similarity comparison program S03 is based on the aggregate word data and the aggregate document data, and the similarity comparison is performed from the semantic aspect to calculate the semantic similarity and order similarity of the aggregate word data and the collective document data. The degree of plagiarism of the user's files relative to the source file. Among them, in the calculation of semantic similarity and order similarity, the phrase is used as the comparison unit. In addition, the semantic similarity between the word-to-word, the word-to-speech and the phrase-to-speech is calculated in the aggregate word data and the collective document data. In addition, the calculation of the semantic similarity or the order similarity is respectively based on the similar maximum value of the word pair, the similar maximum value of the word to the phrase, and the similar maximum value of the phrase to the phrase, which will be respectively described below.

In calculating the semantic similarity of the word w to the word w, it is calculated according to the part of speech of the word w. In addition, all the words w in the aggregate word data are compared with each other to find the maximum similarity of the word pair. Because in WordNet, nouns and verbs form a hierarchical structure with the above-mentioned semantic relationship, while adjectives and adverbs cannot be presented in a hierarchical structure. Therefore, when calculating the semantic similarity between words, two The part of speech, and adopt different methods of similarity calculation. If the two words are nouns or verbs, the Path-based measure is used to calculate the similarity. If the two words are adjectives or adverbs, the Gloss-based measure is used to calculate the similarity.

The invention is based on the hierarchical structure of nouns and verbs in WordNet. Combine PATH (Rada, et al., 1989) and WUP (Wu & palmer, 1994) to calculate the semantic similarity of nouns or verbs, while adjectives or adverbs use VECTOR (Patwardhan, 2003) to calculate semantic similarity. Therefore, the following description incorporates the full text of the papers of PATH, WUP and VECTOR into the disclosure of the present invention. The details and formulas are as follows.

The PATH method is calculated using the shortest path length in WordNet using two words, as shown in Equation 3-1:

The WUP method considers the depth of two words and the depth of their smallest common parent node (LCS) in WordNet to calculate word similarity, as in Equation 3-2:

The VECTOR method uses the annotations of the two words w ₁ and w ₂ in WordNet to form the vectors v1 and v2, and then uses Cosine (cosine) to calculate the similarity of the words, as shown in Equation 3-3:

Therefore, overall, the semantic similarity between words and words (hereinafter referred to as status A, status B, and status C) can be calculated as shown in Equations 3-4 to 3-6, where len(w ₁ ) is The length of the word w ₁ in the collection word data:

In addition, in calculating the semantic similarity of words to phrases, the calculation is based on words. Wherein, each word in the aggregate word data is compared with each word in the aggregate document data to find a maximum value of the similarity of the word to the phrase. Due to the collection The phrase p in the data is generated in the phrase recognition program S02, simply due to the alignment of two sentences, and is not queried in WordNet. Therefore, it is necessary to compare all the phrases in the collection phrase data. All the words in it, find out the maximum similarity among them.

Suppose the phrase p={p_t ₁ , p_t ₂ ,...,p_t _N ,...}, where p_t _N is the Nth word of the phrase p, therefore, the words and phrases (hereinafter referred to as The similarity calculation method of condition D and condition E) can be as shown in formula 3-7~3-10: sim ( w,p )=max _N sim ( w,p _ t _N ) (3-7)

In addition, in calculating the semantic similarity of the phrase to the phrase, the calculation is also based on the word. Among them, all the words in the collective document data are compared with each other to find the maximum similarity of the phrase to the phrase. Since the words p ₁ and p ₂ in the aggregate document data are generated by the alignment of the two sentences in the phrase recognition program S02 and are not queried in WordNet, it is necessary to compare the collection pieces one by one. In the corpus, all the words of all the words are used to find the maximum similarity among them.

Suppose the phrase p ₁ ={p ₁ _t ₁ , p ₁ _t ₂ ,..., p ₁ _t _M ,...}, where p ₁ _t _M is the Mth word of the phrase p ₁ , slice p ₂ ={p ₂ _t ₁ ,p ₂ _t ₂ ,...,p ₂ _t _N ,...}, where p ₂ _t _N is the Nth word of the phrase p ₂ , therefore, the slice The calculation formula of the language and the phrase (hereinafter referred to as the condition F) is as shown in 3-11~3-14, where len(p ₁ ) is the length of the phrase p ₁ :

Then, the aggregate word data and the aggregate document data need to be sorted in order and the repeated items are removed to form a joint U, and then the source file and the user file are semantically compared and aligned.

Suppose this union U={JTP ₁ , JTP ₂ ,...,JTP _f ,...,JTP _F },S _ni ={TP ₁ ,TP ₂ ,...,TP _g ,...TP _G }, and S _j ={TP' ₁ , TP' ₂ ,...,TP' _h ,...TP' _H }, and then refer to the sentence similarity calculation method defined by Li, et al. (2006), At the same time, the synonym substitution and the change of word order are considered. The overall similarity between S _ni and S _j is calculated _by combining Semantic Similarity and Order Similarity. The calculation steps are as follows: First step - Calculate the sentence semantic similarity as follows: the union U for all the JTP _f sentence S _ni all TP _g for _{comparison: sem 1 = {max g sim} (JTP 1, TP g), max g sim (JTP 2, TP g ),...max _g sim ( JTP _f ,TP _g )} (3-15)

All JTP _{f in the} union U is compared with all TP _h in the sentence S _j : sem ₂ ={max _h sim ( JTP ₁ ,TP' _h ) , max _h sim ( JTP ₂ ,TP' _h ),... Max _h sim ( JTP _f ,TP' _h )}(3-16)

Finally, the semantic similarity of two sentences needs to be calculated using cosine similarity:

The second step - calculating the sentence order similarity is as follows: in the calculation of sem ₁ and sem ₂ , each of g and h which produces the greatest similarity needs to be recorded. If the similarity is 0, the order is 0, thereby Can form order ₁ and order ₂ , and finally, the order similarity of the two sentences is:

Finally, the third step is performed - the overall similarity of the sentences is calculated as follows: Here, different weights are given, and the overall similarity of the sentences is calculated in combination with the above two similarity calculation formulas: Sim ( S _ni , S _j ) = αSim _sem +(1- α ) Sim _order (3-19), where α ≦1, and Li, et al. (2006) believe that α should be greater than 0.5, so it can be between 0.5 and 1 (0.5≦ α ≦1 ), users can customize according to their needs.

When the degree of plagiarism is higher than a threshold value τ, the present invention can determine that the user file has plagiarism. In other words, if Sim(S _ni , S _j )>τ, the user file will be treated as a plagiarized source file.

Hereinafter, the procedure of the similarity comparison program S03 will be described using the above example.

After the previous procedures S01 and S02, S ₁₁ = {John, have, turn_on, his, new radio}, S ₁ = {John, switch_on, the, new radio}, set word data = {John, have, turn_on, His, John, switch_on, the}, collection phrase data = {new radio, new radio}, and union U = {John, has, turn_on, his, new radio, switch_on, the}.

Therefore, all possible comparisons of words and phrases are shown in the following table:

By comparing the sentences of the source file with the contents of the union U, there will be a comparison condition (conditions A~F). Among them, the situation A, the situation B and the situation C are the comparison of the words and the words respectively, the situation D and the situation E are the comparison of the words and the words respectively, and the condition F is the comparison between the words and the words. .

If the part of speech cannot be directly given 0, it needs to be handled individually. In the case of condition D, Sim(John, new radio)=max(Sim(John,new),Sim(John,radio))=max(0,Sim(John,radio))=Sim(John,radio) ).

Then compare the source file sentences with the union U to get the following comparison table:

Here, the largest number in the column is recorded, and sem ₁ = { _{1, 1} , 2, ₁ , 2, _1, 0} is obtained, and the position record of the column (Row) when the row maximum number is generated is generated. Get up and get order ₁ ={1,2,3,4,5,3,0}.

Next, the user file sentence is compared with the union U to obtain the following comparison table:

Here, record the largest number in the line, assuming Sim (have, switch_on) = 0.2 (can customize other numbers), you can get sem ₂ = {1, 0.2, 1, 0, _{2, 2,} 1} And record the position of the column when the row has the largest number, and get order ₂ ={1,2,2,0,4,2,3}.

Finally, calculate the overall similarity of the sentence to get:

Overall similarity Sim ( S ₁₁ ,S ₁ )= α ×0.79931+(1- α )×0.60114

Assuming α = 0.6 (may be other numbers), the threshold τ is, for example, 0.7 (other numbers can be customized according to requirements), then Sim(S11, S1) = 0.60.79931 + 0.40.60114 = 0.720042 > τ (0.7), therefore, it is determined that the user file has plagiarism relative to the source file.

In summary, the plagiarism detection method according to the English document of the present invention includes a grammar processing program, a phrase recognition program, and a similarity comparison program. Among them, the source document and the user file are separately parsed, and the source word data and the user word data are compared and the source phrase data and the user phrase data are compared, and the source document data is compared. And the user's phrase data, and then similarity comparison in terms of semantics, thereby calculating the degree of plagiarism of the user file relative to the source file. Therefore, compared with the prior art, the present invention can not only utilize the grammatical structure analysis and extract words and phrases in sentences, but also use the phrase as the comparison unit of plagiarism detection, supplemented by semantic meaning, thus improving the habit The correctness of the alignment is determined using a single word in the technique.

The above is intended to be illustrative only and not limiting. Any equivalent modifications or alterations to the spirit and scope of the invention are intended to be included in the scope of the appended claims.

S01‧‧‧ grammar handler

S02‧‧‧Word recognition program

S03‧‧‧similarity comparison program

Claims (16)

A plagiarism detection method for English data is implemented on a computer, and the plagiarism detection method comprises: a grammar processing program, comprising a grammar analysis step, respectively performing a parsing analysis on a source file and a user file To obtain a source word material and a source phrase material, and a user word material and a user phrase material respectively; a language recognition program includes a set generation step, which compares the source word Data and the user word data and the source document data and the user phrase data to generate a set of word data, and compare the source document data with the user phrase data to generate a collection of phrase data, wherein the words in the collection word data include the source file and all the words restored by the root in the user file, and the phrase language in the collection phrase data includes the source The file is the same phrase in the user file; and a similarity comparison program is based on the set word data and the set of phrase data and is similar in terms of semantics and order Comparing, the plagiarism of the user file relative to the source file is calculated by analyzing the aggregate word data and the semantic similarity and order similarity between the source document and the user file and the user file . The method for detecting plagiarism of English materials as described in claim 1 wherein the source word material or the user word data respectively comprise at least one word, the source phrase material or the user phrase material Contain at least one phrase respectively. The plagiarism detection method of the English document as described in claim 1 of the patent application, wherein the grammar analysis step uses a natural language processing technique. The plagiarism detection method of the English document as described in claim 1 , wherein the grammar processing program further comprises a sentence break step, a complete plagiarism detection step and a root root restoration step. For example, in the plagiarism detection method of the English document mentioned in the first paragraph of the patent application, in the grammar analysis step, the source file and the user file are compared with a dictionary for the part-of-speech tagging and phrase 撷take. The method for detecting plagiarism of English materials as described in claim 5, wherein the words in the collective word data correspond to words in the dictionary. For example, in the plagiarism detection method of the English document mentioned in the first paragraph of the patent application, in the phrase recognition program, the phrase is used as the comparison unit. For example, in the plagiarism detection method of the English document mentioned in the first paragraph of the patent application, in the calculation of the semantic similarity and the similarity of the order, the phrase is used as the comparison unit. The method for detecting plagiarism of English materials as described in claim 7 of the patent application, wherein in the similarity comparison program, the word-to-word, the word-to-word, The semantic similarity of words to phrases and phrases. For example, in the plagiarism detection method of the English materials mentioned in the ninth application patent range, the semantic similarity of the calculated words to the words is calculated according to the part of speech of the words. The plagiarism detection method for the English materials mentioned in claim 9 wherein, in calculating the semantic similarity of the words to the words, all the words in the collective word data are compared with each other to Find the maximum similarity between a word and a word. For example, in the plagiarism detection method of the English materials mentioned in the scope of the patent application, in the semantic similarity of the calculated words to the phrase or the phrase to the phrase, the calculation is based on the words respectively. For example, in the plagiarism detection method of the English materials mentioned in claim 11, wherein in calculating the semantic similarity of the words to the words, each word and collection phrase data in the collective word data is Each word in the comparison is compared to find the maximum similarity of the word to the phrase. For example, in the plagiarism detection method of the English document mentioned in claim 13, wherein in calculating the semantic similarity of the phrase to the phrase, all the words in the collective phrase data are compared with each other to Find out the maximum similarity of the phrase to the phrase. The method for detecting plagiarism of English materials as described in claim 14 wherein the semantic similarity or the similarity of the order is based on the similar maximum value of the word pair and the word-to-word. The similarity maximum, and the similar maximum of the phrase to the phrase. For example, the plagiarism detection method of the English document mentioned in the first paragraph of the patent application, wherein when the plagiarism is higher than a threshold, the user file has plagiarism.