Handwriting Detection Model Based on Four-Dimensional Vector Space Model

Transcription

1 Journal of Matheatics Research; Vol. 10, No. 4; August 2018 ISSN E-ISSN Published by Canadian Center of Science and Education Handwriting Detection Model Based on Four-Diensional Vector Space Model 1 School of Translation Studies, Jinan University, China 2 School of Electric & Inforation, Jinan University, China Lin Li 1, Xiuteng Duan 2 & Yutong Li 1 Correspondence: Lin Li, School of Translation Studies, Jinan University, Zhuhai, China. Received: April 4, 2018 Accepted: April 17, 2018 Online Published: May 28, 2018 doi: /jr.v10n4p32 Abstract URL: Handwriting detection is ainly used in the criinal investigation. We can use four-diensional vector space odel to build a odel for handwriting detection. This article selects feature quantities such as word frequency, language style, average word length, and sentence structure fro the texts and quantizes the, transforing the into relations between vectors. After quantifying and noralizing the features in an author s article in advance, we can obtain a standard reference vector. Then we do the sae processing on the target text database, and copare it with the standard reference vector in ters of the odulus value and the included angle. Then we could estiate whether the author is the owner of database value. The siulation result shows that the odel is ore accurate and the author of particular texts can be obtained. Keywords: Vector Space Model, handwriting detection, noralized processing 1. Introduction With the popularity of e-ails, obile phone essages, and social networking essages, ore and ore anonyous texts inforation have becoe involved aterials. Identifying the origin of the texts is crucial in the detection of criinal cases. Fro view of functional linguistic, language reflects the inner world of huan beings at all ties. The world we know is realized through language. Language identification has also becoe a highly regarded detective ethod. This gave birth to handwriting detection, that is, atching articles and their authors through specific features in the article. At present, there are few researches on handwriting detection odels. Most of the focus on single-field research and are usually difficult to apply universally. Vector Space Model(VSM), proposed by Gerard Salton and McGill in 1988, using vectors to represent texts, and take weights of the features of texts as coponents. Through the calculation of word frequency and diension reduction of the vectors, this odel can copute the siilarity of texts. Li Xuelei&Zhang Dongo established a odel which correlates the text characters, the ter frequency, the hypertext arkup language tag inforation in the web, and seantic analysis for the question sentences to calculate an adjustable ter frequency weighting paraeter and to increase the separability of feature words vector. (Xuelei Li & Dongo Zhang, 2003) Turney also used a vector-based odel of seantic relations to attain a score of 56% on soe ultiple-choice questions which are fro SAT test. (Turney, Peter D., 2006) This paper proposes a handwriting detection odel based on four-diensional vector space odel, which can be applied to any types of text to solve the proble of author identification in criinal investigation. 2. Handwriting Detection Model Based on Four-diensional Vector Space Model The vector space odel is an algebraic odel that applies to inforation filtering, inforation selection, indexing, and assessent of relevance. It siplifies the processing of text content into operations in vector space, and expresses the siilarity of language features with spatial siilarity. We easure the siilarity of texts by calculating the siilarity of the vectors. In this article, we also select the feature quantities in the texts and quantify the. 2.1 Selecting Feature (1) Selecting the word with highest frequency in each text as one of the recognition criteria. Because the word with highest frequency of each essage also reflects a person s writing habits, in addition to representing the subject of this article. Even the word frequency of one person s several articles settles in a range of frequency fluctuations. (Huiling Wang & et al., 2001) (2) Language style is also an indispensable criteria of great iportance. It is undeniable that each person s writing style is very different. What we choose here is the expression of exclaatory sentences. We calculate the frequency of sentences 32

2 that express the strong eotion of persons as a criterion. (3) The average word length is also an iportant evidence on whether an article is written by a particular people. Soe people prefer to use short, inforal words, while others prefer the long ones. At the sae tie, we also tend to use long and foral words while writing soe foral docuents. Therefore, the average word length is a very good easure to judge the author s identity, educational level, etc. (Qiang Li & Jianhua Li, 2006) (4) Sentence structure is also selected as one of the criteria for judgent. Here we use the frequency of link verbs as the research object. This is because the sophistication and difficulty of writing sentences in the articles not only represents a person s educational level, but also deeply reflects one s writing habits. Even everyone has his habitual structure of sentences, and there is a proportion of each structure type. 2.2 Establishing Feature Database Extracting the Highest Frequency and Establishing a Database In the processing of word frequency feature, only the value of the highest word frequency is considered. In this way, we can obtain the writer s writing habits (soe people have wordy writing style, while others cherish words like gold). Therefore, the highest word frequency can be used as indicator of observation and analysis. (1) The initial data is siply a bunch of text database, we wrote the progra through JAVA, and analyze the highest word frequency in each article, establishing a new database. (2) We copute the arithetic average of each highest word frequency in the above new database, and finally obtain the arithetic average as a coponent of the standard paraeter vector. The forula is as f av = (1) (3) We extract the f ax and f in in any articles,then subtract fro f av, coparing two absolute values. The result is as If f ax f av > f in f av As a result, f ax deviates fro the highest average frequency; If not, f in deviates fro the highest average frequency. f i The results above can be used as the calculation of the dynaic range in the odel Analyzing the Style of Article Sentences and Building a Quantitative Database Language styles are various. Usually the nuber of exclaations and the frequency of odal particles can represent the writing style of a person best. When quantifying, we deal with the ainly by looking for the nuber and frequency of exclaatory sentences. (1) The sae as the quantification of the highest word frequency database above, the initial data is also just a siple text library. We analyzed the frequency of exclaations and odal particles in each text through JAVA prograing to for a new database. (2) We copute the arithetic average of each frequency figure in the new database above, and finally obtain the arithetic average as a coponent of the standard paraeter vector. Assuing the total nuber of texts is, then the forula is as y av = (2) (3) We extract y ax and y in fro the database, and then subtract fro y av, coparing two absolute values. The result is as If y ax y av > y in y av As a result, y ax deviates fro the highest average frequency; If not, y in deviates fro the highest average frequency. y i The results above can be used as the calculation of the dynaic range in the odel. 33

3 2.2.3 The Statistics of Average Word Length and Its Quantification The average word length often reflects a person s educational level (the foral words are usually longer). Therefore, the average word length can also be a indicator to identify a person s identity. (1) We use JAVA to write the progra based on the initial corpus, analyzing the average word length in each text to for a new database. (2) We copute the arithetic average of the average length of each word in the new database above, and finally obtain the arithetic average as a coponent of the standard paraeter vector. The forula is as L av = (3) (3) We extract the L ax and L in in any eails, then subtract fro L av, coparing two absolute values. The result is as If L ax L av > L in L av As a result, L ax deviates fro the highest average frequency; If not, L in deviates fro the highest average frequency. L i The results above can be used as the calculation of the dynaic range in the odel The Analysis and Quantification of Sentence Structure Sentence structure is also a tool to reflect a person s writing habits, any people like to use the subject-linking verbpredicative structure or subject-verb-object structure. Here we choose the the subject-linking verb-predicative structure to quantify its proportion. (1) The original data is still the text library, we use the progra written by JAVA prograing language to obtain linking verbs in each article, so as to get the proportion of the subject-linking verb-predicative syste structure stateents, constituting a new database. (2) We copute the arithetic average of the proportion in the new database above, and finally obtain the arithetic average as a coponent of the standard paraeter vector. Assuing the total nuber of texts is, then the forula is as S av = (4) (3) We extract S ax and S in fro the database, and then subtract fro S av, coparing two absolute values. The result is as If S ax S av > S in S av As a result, S ax deviates fro the highest average frequency; If not, S in deviates fro the highest average frequency. S i The results above can be used as the calculation of the dynaic range in the odel. 2.3 Weight Calculation of Feature Quantities Since the four characteristic quantities have different effects on identifying the author of the text, we need to quantify the proportion of the effect each characteristic quantity. The weight calculation forula is as W ik = t f ik is the frequency at which the feature appears in the text; And the denoinator is the noralization factor. t f ik [log( n k+0.01 )] (5) (t f ik ) 2 [log( n k+0.01 )] 2 34

4 2.4 Four-diensional Space Vector Forula Reference Vector: tolerance scope: I = (W f k favi, W yk yavi, W Lk L avi, W S k S avi ) (6) The result of vector quantification is: I = (W f k favi, W yk yavi, W Lk L avi, W S k S avi ) (7) I s = (W f k favi, W yk yavi, W Lk L avi, W S k S avi ) (8) After that, we proceed as follow. Firstly, we calculate the difference between odules: If 0 < I I s < I, the odulo value satisfies the requireent. Secondly, we calculate the values of included angle: 3. Siulation Results I I s 0 < a I I s < I ( f av + f i, y av + y i, L av + L i, S av + S i ) I f av + f i, y av + y i, L av + L i, S av + S i P = ( f av + f i, y av + y i, L av + L i, S av + S i ) (10) We use texts with ore than 1000 words for siulation to ensure accuracy of the odel. Therefore, we selected three sets of saple, each with a saple size of 15. (Yiping Zeng & Xiaowen Zhu, 2006) 3.1 Data Results The digital data obtained after processing is shown in the following table (sheets 4-1 to 4-3): Author 1: saple Average The Exclaation The proportion of word highest sentence subject-linking length frequency proportion verb-predicative structure 1 text /385= /46=0 20/46= text /393= /89=0 14/89= text /1433= /138=0 64/138= text /4863= /370=0 237/370= text /1515= /128=0 36/128= text /2328= /192= /192= text /2417= /175= /175= text /633=0.7 0/107=0 36/107= text /241= /41=0 13/41= text /690= /89=0 22/89= text /1133= /152=0 48/152= text /429= /44=0 14/44= text /227= /39= /39= text /196= /30=0 8/30= text /329= /53=0 19/53=0.36 average sheet 4-1 (9) 35

5 Author 2: saple Average The Exclaation The proportion of word highest sentence subject-linking length frequency proportion verb-predicative structure 1 text /458= /65= /65= text /262= /48= /48= text /710= /88=0 41/88= text /133= /32=0 8/32= text /232= /51=0 26/51= text /279= /33=0 9/33= text /408= /62=0 23/62= text /395= /53=0 17/53= text /329= /49= /49= text /913= /24=0 5/24= text /458= /136= /136= text /387= /35=0 19/35= text /96= /16=0 2/16= text /1364= /160= /160= text /145= /19=0 3/19=0.16 average sheet 4-2 Author 3: saple Average The Exclaation The proportion of word highest sentence subject-linking length frequency proportion verb-predicative structure 1 text /276= /39=0 15/39= text /286= /34=0 19/34= text /141= /24=0 4/24= text /264= /23=0 7/23= text /179= /29= /29= text /1671= /185=0 91/185= text /246= /39= /39= text /152= /23=0 8/23= text /727= /77=0 36/77= text /381= /53=0 39/53= text /639= /77=0 72/77= text /116= /25=0 10/25= text /207= /46=0 11/46= text /380= /66=0 26/66= text /210= /32=0 13/32=0.41 average average sheet The Average Value of the Characteristic Quantities and the Maxiu Offset The first set: f av1 = ; L av1 = 5.23; y av1 = ; S av1 = 0.36 f 1 = ; L 1 = 1.56; y 1 = ; S 1 = 0.20 I = (W f k , W Lk 5.23, Wyk , W S k 0.36) (11) The second set: f av2 = ; L av2 = 5.31; y av2 = ; S av2 = f 2 = ; L 2 = 1.00; y 2 = ; S 2 = I = (W f k , W Lk 5.31, Wyk , W S k 0.328) (12) 36

6 Third set: f av3 = ; L av3 = 5.77; y av3 = ; S av3 = f 3 = ; L 3 = 1.69; y 3 = ; S 3 = Model Testing At first, we need to copute the weight; The result of noralization is as follow Word frequency weight = Sentence structure weight = Average word length weight = Language style weight = I = (W f k , W Lk 5.77, Wyk , W S k 0.426) (13) There are three databases above, we use the first database to validate the odel: We extract a text randoly fro the author s article database to calculate the vector: The first step, validate the odulus: Then: I s = ( , 2.038, 0, ) (14) I = ( 0.011, 1.67, , 0.116) (15) Therefore:0 < = < So the odulus fit the odel The second step: the coparison of the angles Maxiu deviation: We calculate the included angle between I and P. The result is as rad. We calculate the included angle between I and I s. The result is as rad. It satisfies 0 < < Therefore, we get the result. This article is written by author 1. I = (16) I = ( , , , ) (17) I = (18) I s = ( , 2.038, 0, ) (19) I s = (20) P = ( , 2.193, , ) (21) The odel validity is copleted. this odel can be used for handwriting analysis. P = (22) 37

7 4. Conclusion The vector space odel turns the articles into vectors and the concept of our odel is relatively siple. When the nuber of texts is high and the data is coplete, the results of odel are ore accurate. Our odel can be used to analyze the handwriting of different people, rather than just be used under a certain situation. By establishing a specific database, we can apply it in every situation, which shows that the odel is flexible and reliable. We need to copute a lot while using this odel, so there are still soe disadvantages needed to be optiized, but the odel is of great significance and it provides realistic guidance for criinal cases bases on e-ail. References Xuelei Li, & Dongo Zhang (2003). A Text Categorization Method Based on VSM. Coputer Engineering, 29(17), Turney Peter, D. (2006). Siilarity of seantic relations. Coputational Linguistics, 32(3), , MIT Press. Huiling Wang, Rou Song, & Weichang Dai (2001). The Research of Text Categorization on Style. The 6th national conference on coputational linguistics. Qiang Li, & Jianhua Li (2006). Method of Filting Reactionary Text Based on Vector Space Model. Coputer Engineering, 32(10), 4 5. Yiping Zeng, & Xiaowen Zhu (2006). Application of Coputational Methods to the Study of Stylistics in China. Journal of Fujian Noral University (Philosophy and Social Sciences Edition), 1, Copyrights Copyright for this article is retained by the author(s), with first publication rights granted to the journal. This is an open-access article distributed under the ters and conditions of the Creative Coons Attribution license ( 38