Using Visualization to Detect Plagiarism in Computer Science Classes

Transcription

1 Ung Vualzaton to Detect Plagarm n Computer Scence Clae Randy L. Rbler Department of Computer Scence Lynchburg College Lynchburg, Vrgna 241 rbler@lynchburg.edu Marc Abram Department of Computer Scence Vrgna Tech Blackburg, Vrgna 2461 abram@c.vt.edu Abtract Th paper ntroduce a number of general method for vualzng commonalty n et of text fle. Each vualzaton multaneouly compare one fle n the et to all other fle n the et. Thee vualzaton, whch can be computed n O(n) tme and pace, are explaned and then appled to the problem of detectng plagarm n large computer cence clae. A cae tudy preented and ample vualzaton are provded. Fnally, a new nteractve tool that can be ued to produce and manpulate thee vualzaton preented. 1. Introducton Computer cence ntructor uually requre tudent to wrte ther program ndependently. Therefore, any nontrval programmng agnment hould produce a number of unque oluton equal to the number of tudent ubmttng program. Unfortunately, ome tudent mght not oberve the rule regardng plagarm, and copy all or part of another tudent program and ubmt t a ther own work. To reduce the chance of beng detected, they ometme change the name of varable or make other cometc change, but often, n large clae where the perceved probablty of detecton low, they make only the mot mnor alteraton. Perhap f better detecton technque were avalable th type of plagarm could be effectvely deterred. Th paper decrbe the producton of vualzaton that how how mlar one fle to the other fle n a et of fle. When appled to plagarm detecton, they how how mlar one tudent program to all the other program ubmtted. Thee one-to-many fle vualzaton can be computed n O(n) tme. To ee how every fle n a et compare to every other fle n a et requre n uch vualzaton, rather than the O(n 2 ) operaton requred for parwe comparon. Whle thee method are partcularly well uted to plagarm detecton, they are not lmted to that applcaton, and repreent a general method for performng one-to-many comparon among text fle. Vualzaton epecally approprate for plagarm detecton becaue a hgh level of mlarty between document can be caued by factor other than cheatng. So plagarm detecton generally requre two tep. The frt tep to dentfy upcou program, and the econd tep to examne ecton that are mlar. 2. The Categorcal Patterngram (CP) In plagarm detecton, we are attemptng to dentfy equence of character that are preent n more than one program. We defne k a the mnmum equence length of nteret. Every k-length character equence defne an n-gram. N-gram are the equence of n character that ext n a text fle. For example, the frt n-gram of length three n a program that tart wth the word begn, are beg, eg, and gn. We do not nclude pace, carrage return, tab, or other non-prntable character n our n-gram. The ngle fle n a one-to-many comparon wll be referred to a the bae fle. The entre et of fle nvolved n the comparon wll be referred to a the enemble [2]. We defne a categorcal patterngram [9] a a vualzaton that dplay, for each character poton x n the the bae fle, the number of enemble fle that contan the k-length n-gram that begn at that poton. A pont plotted at coordnate (x; y) f the k-length n-gram begnnng at character poton x n the bae fle occur at any locaton n y fle of the enemble A Toy Example Suppoe that we have three fle that contan text a hown n Fgure 1, 2, and 3. A patterngram that ue Toy Fle a a bae fle hown n Fgure 4. The pont n the graph have been labeled wth the correpondng letter n the fle. Th done only a an ad to undertandng how Proceedng of the IEEE Sympoum on Informaton Vualzaton 2 (InfoV') / 2 IEEE

2 Toy: Th a tet. Fgure 1. Toy Fle 2 Compote Categorcal Patterngram (Sequence Length = 4) T h Toy1: Oh ye. Th a tet too. Fgure 2. Toy Fle 1 Fle number 1 T h a t Toy2: Toy2 ha lttle n common wth the other two. Th common. T o y : T h a t e t. Fgure 3. Toy Fle Fgure. CCP of Toy Fle Categorcal Patterngram (Sequence Length = 4) Number of fle contanng n-gram T o y : T h a t e t Fgure 4. Patterngram of Toy Fle the graph generated. In real data the pont are typcally much too dene to allow th type of annotaton. The categorcal patterngram for Toy Fle hown n Fgure 4 depct the one-to-many comparon of Toy Fle to Toy Fle 1 and Toy Fle 2. The frt fve character n Toy Fle produce fve dot at y-coordnate one becaue the bae fle n-gram Toy, oy:, y:t, :Th, and :Th do not occur n the other two fle. The xth, eventh, and eghth character (plotted at x-coordnate, 6, and 7, repectvely) produce pont at y-coordnate three becaue the n-gram that are ntated wth thoe character, Th, h, and, occur n all three fle. Character eght through twelve each ntate n-gram that occur n one other fle n th cae Toy Fle 1, o they are plotted at y-coordnate two. The remanng character belong to equence that do not occur n the other fle and are therefore plotted at y-coordnate one. In real applcaton the equence length of nteret are typcally longer than four character. In addton, the fle are typcally made up of thouand of character, o t not poble to label each pont wth t correpondng letter. However, thee graph are degned mply to how whether or not mlarte ext. Another technque wll be decrbed n ecton that how the n-gram that correpond to the mlarte dplayed n the patterngram. The followng note hghlght ome mportant charactertc of the Categorcal Patterngram (CP). 1. The n-gram that are unque to the bae fle are plotted at y-coordnate one. 2. N-gram n the bae fle that are common to many other fle are plotted wth a large y-coordnate. In plagarm detecton, thee n-gram correpond to character equence that naturally occur frequently wthout any collaboraton between tudent. 3. The equence that may be ndcaton of plagarm are ndcated by pont that have a relatvely mall y-coordnate greater than one. The value of the x- coordnate provde the vewer wth an ndcaton of where the n-gram occur n the bae fle. The locaton of the n-gram n the other enemble fle unavalable n th vualzaton. Perhap the mot mportant nformaton n the graph provded by the preence or abence of cluter of horzontally contguou pont. For example, a contguou ecton of pont plotted at y = 1 ndcate a ecton of code that unque a ecton where the majorty of the pont are plotted at y = 1 alo a good ndcaton of ndependently developed code, partcularly when the value that are not at y =1have relatvely hgh y-coordnate. However, the abence of pont at y-coordnate one for an extended ecton a good ndcaton of plagarm. The domnant y- coordnate n uch a ecton often ndcate the number of collaborator nvolved. Proceedng of the IEEE Sympoum on Informaton Vualzaton 2 (InfoV') / 2 IEEE

3 3. The Compote Categorcal Patterngram The categorcal patterngram defned n the prevou ecton how the degree of mlarty that ext between one fle and the ret of the enemble of fle. The vualzaton preented n th ecton, whch we call the compote categorcal patterngram (CCP), how whch partcular fle are mlar. Each fle agned a number n the range [;n 1], where n the number of fle n the enemble. A pont plotted at (x,y) f the k-length n-gram begnnng at x n the bae fle occur one or more tme n fle y. Fgure, how the CCP for the Toy Fle ung Toy Fle a the bae fle. The followng note hghlght ome mportant charactertc of the Compote Categorcal Patterngram (CCP). 1. Each of the fle n the enemble agned a unque nteger value whch ued a a y-coordnate n the graph. A pont plotted at the y-coordnate aocated wth the bae fle for every character n the bae fle. Th uually appear a a horzontal lne. 2. If a unque equence ext n the bae fle, extendng from character number to character number j, then all the pont plotted n the nterval [; (j k)] wll be at the y-coordnate aocated wth the bae fle. 3. If a equence n the bae fle nterval [; (j k)] matche a equence anywhere n fle m, then (j k +1) pont wll be plotted at y-coordnate m n the nterval [; (j k)]. Th mean that there wll be pont plotted at the y-coordnate aocated wth a matchng fle at the x-coordnate aocated wth the match n the bae fle. 4. If a equence n the bae fle matche multple fle n the data et, then pont wll be plotted at the multple y-coordnate aocated wth the matchng fle. Thee pont wll be plotted n the x nterval correpondng to the poton of the matchng n-gram n the bae fle. 4. Creatng the Patterngram A memory-effcent algorthm that can be ued to produce the patterngram and the compote patterngram ue a hah functon to map the n-gram to nteger, and then ue a hah table to record the number of occurrence and fle d of each k-length n-gram contaned n the enemble. Becaue a trng table ued to hold all the fle n memory, the memory requrement for the hah table ndependent of k, and the total memory requrement proportonal to n, where n the total number of character n all the fle. Addtonally, the ue of recurve hahng by cyclc polynomal [4], provde a contant hahng tme regardle of the ze of k. In practce we have found the memory requrement for the algorthm to be approxmately ten tme larger than the memory requred to tore all the fle n memory. Ung th algorthm, an average Pentum-cla machne can produce all the patterngram for program n an average computer cence cla n jut a few econd.. Pattern Text Vew The pattern text vew provde an addtonal many-to-one vualzaton for text fle. It format the text from the bae fle o that each character that begn a non-unque equence hghlghted, ether ung an underlned font or ung an alternatve font color. The text vew ha a mple relatonhp to the patterngram. If the value of the y-coordnate correpondng to the k-length n-gram begnnng at that character one, then the character not hghlghted. If the correpondng y-coordnate greater than one, then the character hghlghted. Thu, letter begnnng unque n-gram are dtnct from letter begnnng non-unque n-gram. 6. Changng Varable Name Sometme tudent wll change the varable name n a program n an attempt to mak plagarzed code. Th technque wll reduce the effcacy of the patterngram, but wll not elmnate t altogether. However, the effect of varable name change can be defeated through the nerton of a mple preproceng tep known a tokenzaton. The tokenzaton convert all the alphanumerc trng n a program nto ngle character. The pecal character, whch nclude everythng other than the alphanumerc character, are left unaffected. Th produce three effect. The frt that two fle that contan the ame program wth dfferent varable name wll produce dentcal patterngram. The econd effect a reducton n the ze of the data et. The thrd effect that the enemble become more homogeneou, o an ncreae n k wll probably be jutfed. The reult of our cae tudy eem to ndcate that th tranformaton ha lttle advere effect on the vualzaton and ha all the potve effect decrbed. 7. A Cae Study A cae tudy wa performed to ee f thee vualzaton method could be uccefully appled n a claroom tuaton. The tudent had been provded wth a mall ample program and aked to expand t. The ample program contaned approxmately 3 lne of code. The average completed program wa le than 1 lne. Student ubmtted Proceedng of the IEEE Sympoum on Informaton Vualzaton 2 (InfoV') / 2 IEEE

4 ther program va emal, and each program receved wa agned a equental ubmon number Typcal Graph 1 Submon number 23 (k=1) Fgure 6 how the patterngram (k = 1) for ubmon number 23. The majorty of the pont are plotted at y- coordnate one, ndcatng that mot of the ten-character equence n th fle are unque to th fle. There are a number of pont plotted at y-coordnate two, but thee are relatvely evenly dtrbuted. Becaue n-gram that occur frequently are not of nteret here, all pont wth (y >= 1) are plotted at (y = 1). Th provde a better vualzaton of the pont n the nterval 1 to 1, whch the crtcal range n th applcaton. Th graph typcal of a graph aocated wth a program that wa wrtten ndependently. The CCP (k =1)for ubmon number 23 hown n Fgure 7. The old horzontal lne formed by contguou dot at y =23 preent becaue all the n-gram n ubmon 23 mut of coure ext n ubmon 23. CCP wll alway have a horzontal lne of pont at the y-coordnate correpondng to the fle whoe n-gram are ued a the ba for the graph. Fgure 7 alo how a large number of pont plotted n the range x =to x =. Thee pont correpond to the character n the emal meage header. The header contan many of the ame word and phrae, although the header are not dentcal. The effect of the header are vble n all the CCP n th data et. There are other place, uch a near x = 7, where a partcular n-gram from ubmon 23 found to be preent n vrtually all of the graph. Thee matche are caued by common word or phrae whch naturally occur n ndependently developed program or by the ample program that the ntructor provded. Color ha been added to the the dplay to dtnguh between the bae fle (blue), the nfrequently matched n-gram (red), and frequently matched n-gram (green). Fle number Number of fle contanng n-gram Fgure 6. Patterngram of Submon 23 Submon number 23 (k=1) Fgure 7. CCP of Submon 23 Submon number 4 (k=1) 7.2. Submon Showng Collaboraton 1 Fgure 8 and 9 how the CP and CCP for ubmon number 4. Th program nearly dentcal to ubmon number 44, whch wa ubmtted by a dfferent author. The nearly old lne n the CCP from near x = 1 to near x = 2 are a clear ndcaton that the code ecton of thee two program are the ame. Whle the plagarm n Fgure 8 and 9 qute blatant, everal more ubtle ntance of plagarm, uch a that hown n Fgure 1 were alo detected. To further nvetgate the effcacy of th approach, we conducted a number of experment n whch we tred to copy a program from the orgnal ubmon and dgue t preence. We made cope and changed all the varable Number of fle contanng n-gram Fgure 8. Patterngram of Submon 4 Proceedng of the IEEE Sympoum on Informaton Vualzaton 2 (InfoV') / 2 IEEE

5 Submon number 4 (k=1) Reference Fle number Fle number Fgure 9. CCP of Submon 4 Submon number 22 (k=1) 1 1 Fgure 1. CCP of Submon 22 name, we made cope and revered all the lne, we even coped a few a fve lne from one program to another. All of thee attempt were vble n our graph. 8. Statu and Future Drecton We have ncorporated all the vualzaton preented n th paper nto CHITRA, a Unx-baed tool orgnally degned to analyze computer and network performance data and, more recently, nto an nteractve tool called Same, whch run under Wndow98 or WndowNT. For more nformaton, or to download th oftware, vt /lynchburg.edu/rbler r/publc/ame. We are currently nvetgatng other applcaton n whch the patterngram mght be ueful. Thee nclude oftware confguraton management (vualzng how a fle change acro veron), genetc algorthm (vualzng how a populaton of gene converge), and computer performance analy (detectng and dentfyng cyclc behavor). [1] M. Abram. CHITRA9 Uer Manual. Computer S- c. Dept., Vrgna Tech, Blackburg, VA , May <URL: chtra>. [2] M. Abram, N. Dorawamy, and A. Mathur. Chtra: Vual analy of parallel and dtrbuted program n the tme, event, and frequency doman. IEEE Tran. on Parallel and Dtrbuted Sytem, 3(6):672 68, Nov [3] E. H. Ch, P. Barry, E. Shoop, J. V. Carl, E. Retzel, and J. Redl. Vualzaton of bologcal equence mlarty earch reult. In Proceedng Vualzaton 9, page 44 1, Atlanta, GA, Oct [4] J. D. Cohen. Recurve hahng functon for n-gram. ACM Tranacton on Informaton Sytem, 1(3):291 32, July [] M. A. Heart. Tlebar: Vualzaton of term dtrbuton nformaton n full text nformaton acce. In Conference proceedng on Human factor n computng ytem, page 9 66, 199. [6] J. Mothe and T. Dkak. Interactve multdmenonal document vualzaton. In Proc. SIGIR 98, Melbourne, ACM. [7] L. T. Nowell, R. K. France, D. Hx, L. S. Heath, and E. A. Fox. Vualzng earch reult: Some alternatve to querydocument mlarty. In Proc. of SIGIR96, Zurch, Aug [8] R. Rbler, A. Mathur, and M. Abram. Vualzng and modelng categorcal tme ere data. In Sympoum on Vualzng Tme-varyng Data, volume NASA Conference Publcaton 3321, Wllamburg, VA, Jan ICASE and NASA/LaRC. <URL: chtra/doc/- 9vtvdRMA/9vtvdRMA.html>. [9] R. L. Rbler. Vualzng Categorcal Tme Sere Data wth Applcaton to Computer and Communcaton Network Trace. PhD the, Vrgna Polytechnc Inttute and State Unverty, [1] M. Rorvg, M. M. Smth, and A. Uemura. The n-gram hypothe appled to matched et of vualzed japaneeenglh techncal document. In Proceedng of the 1999 Annual Meetng of the Amercan Socety for Informaton Scence, Knowledge: Creaton, Organzaton and Ue, Wahngton D.C., Oct [11] W. Sculln, T. Kwan, and D. Reed. Real-tme vualzaton of world wde web traffc. In Sympoum on Vualzng Tme-varyng Data, volume NASA Conference Publcaton 3321, Wllamburg, VA, Jan ICASE and NASA/LaRC. [12] I. M. Soboroff, C. K. Nchola, J. M. Kulka, and D. S. Ebert. Vualzng document authorhp ung n-gram and latent emantc ndexng. In Proc. NPIV 97, La Vega, Nevada, [13] D. Wu, J. Roberge, D. J. Cork, B. G. Nguyen, and T. Grace. Computer vualzaton of long genomc equence. In Proceedng Vualzaton 93, page 38 31, Atlanta, GA, Proceedng of the IEEE Sympoum on Informaton Vualzaton 2 (InfoV') / 2 IEEE