Opinion Extraction, Summarization and Tracking in News and Blog Corpora

Transcription

1 Opo Extracto, Suarzato ad Trackg ews ad Blog Corpora Lu-We Ku, Yu-Tg Lag ad Hs-Hs Che Departet of Coputer Scece ad Iforato Egeerg atoal Tawa Uversty Tape, Tawa {lwku, Abstract Huas lke to express ther opos ad are eager to kow others opos. Autoatcally g ad orgazg opos fro heterogeeous forato sources are very useful for dvduals, orgazatos ad eve goverets. Opo extracto, opo suarzato ad opo trackg are three portat techques for uderstadg opos. Opo extracto es opos at word, setece ad docuet levels fro artcles. Opo suarzato suarzes opos of artcles by tellg setet polartes, degree ad the correlated evets. I ths paper, both ews ad web blog artcles are vestgated. TREC, TCIR ad artcles collected fro web blogs serve as the forato sources for opo extracto. Docuets related to the ssue of aal clog are selected as the experetal aterals. Algorths for opo extracto at word, setece ad docuet level are proposed. The ssue of relevat setece selecto s dscussed, ad the topcal ad opoated forato are suarzed. Opo suarzatos are vsualzed by represetatve seteces. Text-based suares dfferet laguages, ad fro dfferet sources, are copared. Fally, a opoated curve showg supportve ad osupportve degree alog the tele s llustrated by a opo trackg syste. Itroducto Watchg specfc forato sources ad suarzg the ewly dscovered opos are portat for goverets to prove ther servces ad for copaes to prove ther products (Dave et al., 2003 ad Moraga et al., 2002). Opo extracto detfyg copoets whch express opos s fudaetal for suarzato, trackg, ad so o (Ku, L, Wu ad Che, 2005). At docuet level, Webe, Wlso ad Bell (200) recogzed opoated docuets. Pag, Lee, ad Vathyaatha (2002) classfed docuets by overall setets stead of topcs. Dave s (2003) ad Hu s (2004) researches focus o extractg opos of revews. However, a docuet Copyrght 2006 Aerca Assocato for Artfcal Itellgece ( All rghts reserved. cossts of varous opos. Rloff ad Webe (2003) dstgush subjectve seteces fro objectve oes. K ad Hovy (2004) propose a setet classfer for Eglsh words ad seteces, whch utlzes thesaur. However, teplate-based approach eeds a professoally aotated corpus for learg, ad words thesaur are ot always cosstet setet. Hu ad Lu (2004) proposed a opo suarzato of products, categorzed by the opo polarty. Lu, Hu ad Cheg (2005) the llustrated a opo suarzato of bar graph style, categorzed by product features. evertheless, they are both doa-specfc. Webe et al. (2002) proposed a ethod for opo suarzato by aalyzg the relatoshps aog basc opoated uts wth a docuet. Extractg opos o products (Hu ad Lu, 2004) s dfferet fro that o ews or wrtgs. For these kds of artcles, ajor topc detecto s crtcal to expel o-relevat seteces (Ku, L, Wu ad Che, 2005) ad sgle docuet suarzato s ot eough. Pag, Lee ad Vathyaatha (2002) showed that ache learg approaches o setet classfcato do ot perfor as well as that o tradtoal topc-based categorzato at docuet level. Iforato extracto techologes (Carde et al., 2004) have also bee explored. A statstcal odel s used for setet words too, but the experet ateral s ot descrbed detal (Takaura et al., 2005). The results for varous etrcs ad heurstcs also deped o the testg stuatos. ews ad blog artcles are two portat sources of opos. The wrtg of the forer s coparatvely foral to that of the latter because blog artcles expressg persoal opos of the wrters are ofte wrtte a casual style. Because o queres are posed beforehad, detectg opos s slar to the task of topc detecto at setece level. Besdes dstgushg betwee postve ad egatve opos, detfyg whch evets correlated wth whch opos are also portat. Ths paper proposes a ajor topc detecto echas to capture a cocepts ebedded plctly a relevat docuet set. Opo suarzato further retreves all the relevat seteces related to the ajor topc fro the

2 docuet set, deteres the opo polarty of each relevat setece, ad fally suarzes postve seteces ad egatve seteces. Suares ad setet scores are fally used by the opo trackg syste. We wll eploy docuets of dfferet sources ad dfferet laguages to deostrate the perforace of opo extracto, suarzato ad trackg. Corpus Descrpto Three sources of forato are collected for the experets: TREC corpus (Text REtreval Coferece), TCIR 2 corpus ad artcles fro web blogs. TREC corpus s Eglsh, whle the other two are Chese. Two Chese aterals are aotated for the ter-aotator agreeet aalyss ad the experet of opo extracto. All of the are the used opo suarzato. Opo suares about aal clog of these three sources are used as llustratos. Data Acqusto The frst corpus used s the test bed of ovelty track TREC 2003 (Soboroff ad Hara, 2003). There are 50 docuet sets 2003 TREC ovelty corpus, ad each set cotas 25 docuets. All docuets the sae set are relevat. I TREC corpus, set 2 ( cloe Dolly sheep ) s take as a exaple. It dscussed the feasblty of the gee clog ad the perspectves of the authorty. The secod corpus s TCIR. Che ad Che (200) developed a test collecto CIRB00 for Chese forato retreval TCIR 2. The test collecto cossts of 50 topcs ad 6 of the are opoated topcs. Total 92 docuets relevat to the sx topcs are chose as trag data ths paper. Docuets of a addtoal topc aal clog of TCIR 3 are selected fro CIRB0 ad CIRB020 docuet collectos ad used for testg. Blog s a ew rsg couty for expressg opos. To vestgate the opos expressed blogs, we retreve docuets fro blog portals by the query aal clog. The ubers of docuets relevat to aal clog three dfferet forato sources are lsted Table. Source TREC TCIR BLOG Quatty Table. ubers of docuets for opo suarzato Aotatos To buld up trag ad testg sets for Chese opo extracto, opo tags at word, setece ad docuet levels are aotated by 3 aotators. We adopt the taggg forat specfed the paper (Ku, Wu, L ad Che, 2005). There are four possble values say, postve, eutral, egatve ad o-setet, for the opo tags at three levels. TCIR ews ad web blog artcles are aotated for ths work. Iter-aotator Agreeet At frst, the agreeet of aotatos s evaluated. The agreeets of tag values at word, setece ad docuet levels are lsted Tables 2, 3 ad 4, respectvely. Aotators A vs. B B vs. C C vs. A Ave Percetage 78.64% 60.74% 66.47% 68.62% All agree 54.06% Table 2. Agreeet of aotators at word level Aotators A vs. B B vs. C C vs. A Ave Percetage 73.06% 68.52% 59.67% 67.% All agree 52.9% Table 3. Agreeet of aotators at setece level Aotators A vs. B B vs. C C vs. A Ave Percetage 73.57% 68.86% 60.44% 67.62% All agree 52.86% Table 4. Agreeet of aotators at docuet level Agreeets of data fro ews ad blogs are lsted Table 5 for coparso. SOURCE TCIR BLOG Level Setece Docuet Setece Docuet Average agreeets of two 53.33% 4.8% 73.85% 64.7% aotators All agree 33.33% 7.65% 6.40% 4.8% Table 5. Agreeets of aotatos of data fro two dfferet sources Table 5 shows that taggg ews artcles has lower agreeet rates tha taggg web blogs. Ths s because blog artcles ay use spler words ad are easer to uderstad by hua aotators tha ews artcles. Fro the aalyses of ter-aotator agreeet, we fd that the agreeet drops fast whe the uber of aotators creases. It s less possble to have cosstet aotatos whe ore aotators are volved. Here we adopt votg to create the gold stadard. The ajorty of aotato s take as the gold stadard for evaluato. If the aotatos of oe stace are all dfferet, ths stace s dropped. A total of 3 docuets, 8 seteces but 0 words are dropped. Accordg to ths crtero, Table 6 suarzes the statstcs of the aotated testg data

3 Postve eutral egatve oopoate d Total Word Setece Docuet Table 6. Suary of testg data Tables 7-9 show the aotato results of three aotators coparg to the gold stadard (.e., the ajorty). O average, a aotator ca otor the opos of the whole to aroud 80.4%. Ths value ca be cosdered as a referece whe we evaluate the perforace of algorths. The decso of opo polartes depeds uch o hua perspectves. Therefore, the forato etropy of testg data should also be take to cosderato, whe coparg syste perforace. Aotators A B C Average Recall 94.29% 96.58% 52.28% 8.05% Precso 80.5% 88.87% 73.7% 80.85% f-easure 86.86% 92.56% 60.99% 80.4% Table 7. Aotators perforace referrg to gold stadard at word level Aotators A B C Average Recall 94.44% 38.89% 90.97% 74.77% Precso 7.20% 74.67% 50.9% 65.35% f-easure 8.9% 5.4% 64.69% 65.67% Table 8. Aotators perforace referrg to gold stadard at setece level Aotators A B C Average Recall 00% 50% 85% 78.33% Precso 7.43% 7.43% 65.38% 69.4% f-easure 83.33% 58.82% 73.9% 72.02% Table 9. Aotators perforace referrg to gold stadard at docuet level The a of ths research s to sulate the opos of the ass. Istctvely, a statstcal odel would be a good choce to detect opoated cotets. I the followg secto, a statstcal algorth s proposed for opo extracto. Opo Extracto The goal of opo extracto s to detect where docuets opos are ebedded. Opos are hdde words, seteces ad docuets. A opo setece s the sallest coplete seatc ut fro whch opos ca be extracted. The setet words, the opo holders, ad the cotextual forato should be cosdered as clues whe extractg opo seteces ad deterg ther tedeces. Therefore, the extracto algorth s bult botto up by detectg setet words at frst, the detfyg the opo polartes of seteces ad fally docuets afterwards. We postulate that the opo of the whole s a fucto of the opos of the parts. That s, a suary report s a fucto of all relevat opoated docuets, the opo of a docuet s a fucto of all the supportve/osupportve seteces, ad the degree of a supportve/osupportve setece s a fucto of a opo holder together wth setet words. Opo scores of words, whch represet ther setet degrees ad polartes, are detered by the proposed forulas. Algorth [Word Level] Setet words are eployed to copute the tedecy of a setece, ad the a docuet. To detect setet words Chese docuets, a Chese setet dctoary s dspesable. However, a sall dctoary ay suffer fro the proble of coverage. We develop a ethod to lear setet words ad ther stregths fro ultple resources. Frst we collect two sets of setet words, cludg Geeral Iqurer (abbrevated as GI) ad Chese etwork Setet Dctoary 2 (abbrevated as CSD). The forer s Eglsh ad we traslate those words to Chese. The latter, whose setet words are collected fro the Iteret, s Chese. Table 0 shows the statstcs of the revsed dctoares. Words fro these two resources for the seed vocabulary our dctoary. Dctoary Postve egatve GI 2,333 5,830 CSD 43,948 Total 2,764 7,778 Table 0. Qualfed seeds The, we elarge the seed vocabulary by cosultg two thesaur, cludg tog2y4c2c2l2 (abbrevated as Cl) (Me et al. 982) ad the Acadea Sca Blgual Otologcal Wordet 3 (abbrevated as BOW). Cl s coposed of 2 large categores, 94 ddle categores,,428 sall categores, ad 3,925 word clusters. BOW s a Chese thesaurus wth a slar structure as Wordet 4. However, words the sae clusters ay ot always have the sae opo tedecy. For exaple, 寬恕 (forgve: postve) ad 姑息 (appease: egatve) are the sae syoy set (syset). How to dstgush ths polarty wth the sae cluster/syset s the ajor ssue of usg thesaur to expad the seed vocabulary ad s addressed below

4 We postulate that the eag of a Chese setet word s a fucto of the coposte Chese characters. Ths s exactly how people read deogra whe they coe to a ew word. A setet score s the defed for a Chese word by the followg forula. Ths equato ot oly tells us the opo tedecy of a ukow word, but also suggests ts stregth. Moreover, usg these equatos, syoys of dfferet polartes are dstgushable whle dog thesaurus expaso. We start the dscusso fro the defto of the forula of Chese characters. c c c + f c () P f c c (2) c + f c Where c ad f c deote the frequeces of a character c the postve ad egatve words, respectvely; ad deote total uber of uque characters postve ad egatve words, respectvely. Forulas () ad (2) utlze the percetage of a character postve/egatve words to show ts setet tedecy. However, there are ore egatve words tha postve oes the seed vocabulary. Hece, the frequecy of a character a postve word ay ted to be saller tha that a egatve word. That s ufar for learg, so a oralzed verso of Forulas (3) ad (4) show as follows s adopted. j c c j j c P c / / c j + f / c j fc / f c j j c (4) c / c + f j c / fc j j j Where P c ad c deote the weghts of c as postve ad egatve characters, respectvely. The dfferece of P c ad c,.e., P c - c Forula (5), deteres the setet tedecy of character c. If t s a postve value, the ths character appears ore tes postve Chese words; ad vce versa. A value close to 0 eas that t s ot a setet character or t s a eutral setet character. S c ( P ) c c (5) Forula (6) defes: a setet degree of a Chese word w s the average of the setet scores of the coposg characters c, c 2,, c p. p S w S (6) c j p j If the setet score of a word s postve, t s lkely to be a postve setet word, ad vce versa. A word wth a setet score close to 0 s possbly eutral or osetet. Cosderg setet scores of words, setet words ca be detected. Wth the setet words extracted, we are able to tell the opo tedeces of seteces ad docuets. f c j (3) [Setece Level]. For every setece 2. For every setet word ths setece 3. If a egato operator appears before, the reverse the setet tedecy. 4. Decde the opoated tedecy of ths setece by the fucto of setet words ad the opo holder as follows. S p S opo holder Swj j (7) Where S p, S opo-holder, ad S wj are setet score of setece p, weght of opo holder, ad setet score of word w j, respectvely, ad s the total uber of setet words p. [Docuet level]. For every docuet 2. Decde the opoated tedecy of ths docuet by the fucto of the opoated tedeces of seteces sde as follows. S d S p j (8) Where S d ad S p are setet scores of docuet d ad setece p, ad s the aout of evdece. If the topc s at type,.e. atcostructo of the da, reverse the setet type. Perforace of Opo Extracto The gold stadard s used to evaluate the perforace of opo extracto at word, setece ad docuet level. The perforace s copared wth two ache learg algorths,.e., SVM ad the decso tree, at word level. C5 syste s eployed to geerate the decso tree. For ache learg algorths, qualfed seeds are used for trag (set A) ad gold stadard s used for testg (set B). o-oralzed oralzed % Verb ou Average Verb ou Average Precso Recall f-easure Table. Perforace of setet word g SVM Testg A B Trag A 92.08% 45.23% B 48.39% 56.87% Table 2. Result atrx of SVM (Precso) C5 Testg A B Trag A 83.60% 36.50% B 0% 4.50% Table 3. Result atrx of decso tree (Precso)

5 As Tables -3 show, the proposed setet word g algorth acheves the best average precso 6.06% of Verb ad ou whle SVM acheves 46.8% (outsde test, average of 45.23% ad 48.39%) whle C5 does eve worse (precso 0% because of a sall trag set). Our algorth outperfors SVM ad the decso tree setet word g. Ths s because the seatcs wth a word s ot eough for a ache learg classfer. I other words, ache learg ethods are ot sutable for word level opo extracto. I the past, Pag et al. (2002) showed that ache learg ethods are ot good eough for opo extracto at docuet level. Uder our experets, we coclude that opo extracto s beyod a classfcato proble. Source TCIR WEB BLOG Precso 34.07%.4% Recall 68.3% 56.60% f-easure 45.42% 8.99% Table 4. Opo extracto at setece level Source TCIR WEB BLOG Precso 40.00% 27.78% Recall 54.55% 55.56% f-easure 46.6% 37.04% Table 5. Opo extracto at docuet level Table 4 ad Table 5 show that the precso rates are low at both the setece ad docuet levels. Ths s because the curret algorth oly cosders opoated relatos but ot relevat relatos. May seteces, whch are orelevat to the topc aal clog, are cluded for opo judget. The o-relevat rate s 50% ad 53% for TCIR ews artcles ad web blog artcles, respectvely. Extractg opos oly s ot eough for opo suarzatos. The focus of opos should also be cosdered. I the followg opo suarzato secto, a relevat setece selecto algorth s troduced ad appled whe extractg seteces for opo suarzatos. The experetal results of opo extracto cosderg relevat relatos are also lsted. Opo Suarzato Tradtoal suarzato algorths rely o the portat facts of docuets ad reove the redudat forato. Ulke the tradtoal algorths, two factors say, the setet degree ad the correlated evets, play the ajor roles of opo suarzato. The repeated opos of the sae polarty caot be dropped because they stregthe the setet degree. However, the redudat reasos why they hold ths posto should be reoved whe geeratg opo suares. Ad eedless to say, opos ust be detected frst for the opo suarzato. All these reasos ake the opo suarzato ore challegg. Opo suarzato as to produce a cross-docuet opoated suary. For ths purpose, we eed to kow whch seteces are opoated ad tell f they focus o a desgated topc. A algorth, whch decdes the relevace degree ad the setet degree, s proposed ths secto. To put ephass o the opoated factor, the vsualzato of opo suares s dfferet fro the tradtoal suares. A text-based suary categorzed by opo polartes s also llustrated ths secto. The, a graph-based suary alog te seres s llustrated by a opo trackg syste. Algorth Choosg represetatve words that ca exactly preset the a cocepts of a relevat docuet set s the a work of relevat setece retreval. A ter s cosdered to be represetatve f t appears frequetly across docuets or appears frequetly each docuet (Fukuoto ad Suzuk, 2000). Such ters for the ajor topc of the relevat docuet set. How to choose the ajor topc s descrbed as follows. We assg weghts to each word both at docuet level ad paragraph level. I the followg forulas, W deotes weghts; S s docuet level whle P s paragraph level. TF s ter frequecy, ad s word cout. I the subscrpts, sybol s the docuet dex, sybol j s the paragraph dex, ad sybol t s the word dex. Forulas (9) ad (0) copute TF*IDF scores of ter t docuet ad paragraph j, respectvely. Forulas () ad (2) deote how frequetly ter t appears across docuets ad paragraphs. Forulas (3) ad (4) deote how frequetly ter t appears each docuet ad each paragraph. W W St TFS t log (9) S Pjt TFPjt log (0) Pj 2 ( WS t ea) () St TH Dsp Dev St WS t ea Dsp (2) St 2 ( WP t ea) j DspPt WP t ea j DevP jt Dsp j Pt (3) (4) A ter s thought as represetatve f t satsfes ether Forulas (5) or (6). Ters satsfyg Forula (5) ted to appear few paragraphs of ay docuets, whle ters satsfyg Forula (6) appear ay paragraphs of few docuets. The score of a ter, defed as the absolute value of Dev us P j t Dev, easures how S t sgfcat t s to represet the a cocepts of a relevat docuet set.

6 DspSt Dsp Pt S, Pj S DevSt DevPjt (5) Dsp > Dsp S, P S Dev > Dev (6) St Pt j Coparg wth Fukuoto ad Suzuk (2000), we odfy the scorg fucto paragraph level. All docuets the sae corpus are cocateated to a bgger oe,.e., the orgal boudares betwee docuets are eglected, so that words whch repeat frequetly aog paragraphs wll be chose. We also use threshold TH to cotrol the uber of represetatve ters a relevat corpus. The larger the threshold TH s, the ore the uber of ters wll be cluded. The value of ths paraeter s traed the experets. The perforace of extractg relevat seteces s good (Ku et al., 2005). I the opo suarzato algorth, seteces that are relevat to the ajor topc ad also express opos are extracted. There are two clues for extracto,.e., cocept keywords ad setet words. The forer deteres the relevace of a setece to the topc, ad the latter detfes the degree of the setece opo. I our experets, cocept keywords are fro a predefed feld (TCIR) or autoatcally extracted fro docuets of the sae topc, That s, set 2 of TREC, topc ZH037 of TCIR, ad aal clog for blog artcles. The opo-oreted seteces are extracted fro topcal set ad ther tedeces are detered. Copared to the setece-level opo extracto algorth last secto, detectg the topcal seteces s the frst step. The overall procedure s show as follows.. For every topcal setece 2. For every setet word ths setece 3. If a egato operator appears earby, reverse the setet tedecy. Every setet word cotrbutes ts setet score to ths setece. 4. Decde the opo tedecy of a setece by the fuctoal coposto of setet words,.e., Forula (7). Setet words are ecessary to decde opo polartes. If the total score of setet words s postve/egatve, the setece s postve/egatve-oreted. Besdes, we also cosder opo operators, e.g., say, preset, show, suggest, etc. If a setece cotas such a opo operator that follows a aed etty wth zero opo score, t s regarded as a eutral opo. Source TCIR Level Setece Docuet Precso 57.80% 76.56% Recall 67.23% 72.30% f-easure 62.6% 74.37% Table 6. Opo extracto results cosderg cocept words As we have etoed, ajor topc detecto s requred for opo suarzato. Table 6 shows the results of St Pjt cosderg relevace relatos together wth setets. TCIR corpus, wth TREC style, cotas cocept words for each topc. These words are take as the ajor topc for the opo extracto. Seteces cota at least oe cocept word are cosdered relevat to the topc. Obvously, the results are uch better tha those Tables 4 ad 5. However, the real applcatos, the ajor topcs are ot avalable. For web blog artcles, words represet the ajor topc ust be selected autoatcally. The algorth for choosg represetatve words s adopted ad opoated seteces are extracted aga. Table 7 shows the experetal results. Source TCIR Blog Precso 38.06% 23.48% Recall 64.84% 50.94% f-easure 47.97% 32.58% Table 7. Opo extracto results cosderg autoatcally extracted topcal words Coparg to Table 4, the precso creases after applyg ajor topc detecto algorth. It cocludes that the relevat setece selecto s a portat ssue the opo suarzato. Totally 29.67% ad 72.43% of o-relevat seteces are fltered out for ews ad web blog artcles, respectvely. The perforace of flterg o-relevat seteces blog artcles s better tha that ews artcles. The result s also cosstet wth the hgher agreeet rate of aotatos blog artcles. Total 5 topcal words are extracted autoatcally fro blog artcles whle ore, 73 topcal words, are extracted fro ews artcles. These all tell that the cotet of ews artcles dverge ore tha that of blog artcles. However, the judget of setet polarty of blog artcles s ot spler (precso 38.06% vs %). The topcal degree ad the setet degree of each setece are eployed to geerate opo suares. We dstgush betwee postve ad egatve docuets. A docuet s postve f t cossts of ore postve-topcal seteces tha egatve-topcal oes; ad vce versa. Aog postve ad egatve docuets, two types of opo suarzatos are proposed, that s, bref ad detaled opo suary. For bref suary, we pck up the docuet wth the largest uber of postve or egatve seteces ad use ts headle to represet the overall suary of postve-topcal or egatve-topcal seteces. For detaled suary, we lst postve-topcal ad egatve-topcal seteces wth hgher setet degree. Exaples of bref ad detaled suares are show Tables 8 ad 9, respectvely. Chese Scetsts Suggest Proper Legslato for Postve Cloe Techology UK Goveret Stops Fudg for Sheep egatve Clog Tea Table 8. Bref Opo Suary of TREC

7 Postve egatve Ahad Reja Al-Jud, Assstat Secretary Geeral of the Islac Orgazato, declared earler that the sear would be aed at sheddg lght o edcal ad legal aspects of the teratoally cotroversal ssue ad seekg to take a stad o t. Dolly the cloed sheep s oly 3, but her gees are already showg sgs of wear ad she ay be susceptble to preature agg ad dsease -- all because she was coped fro a 6-year-old aal, Scottsh researchers say. Table 9. Detaled Opo Suary of TREC Opo Suares of ews ad Blogs Coparg to the opo suary of TREC set 2 (show Table 9), Tables 20 ad 2 lst the opo suares for TCIR ad blog artcles usg the sae topc aal clog. ews ad blog artcles are two a sources for opos. Dfferet sources of artcles erch the cotet. Geerally speakg, ews docuets are ore objectve whle blog artcles are usually ore subjectve. Besdes, the opo holders fro two sources are of dfferet socal classes. The opos extracted fro ews are ostly fro faous people. Istead, the opos expressed blogs ay coe fro a o ae. Lstg opos of dfferet sources parallel provdes vews of the sae publc ssue. The opo suarzato algorth proposed s laguage depedet. Wth ths ethod, opos of dfferet coutres are vsble. Moreover, ths s surely the prototype of the cross lgual opo suarzato. Postve 上述建議來自四名科學家所組成的專家小組, 該小組於一月應英國政府之邀成立, 就複製所衍生的法律與倫理問題提出相關建議 (The above suggesto cae fro a group of four scetsts. The group was fored uder the request of the Brtsh goveret. The group was to provde advces o laws ad theores cocerg clog.) egatve 在複製羊成功的消息宣布之後, 美國總統柯林頓及 生物倫理顧問委員會 斥複製人不道德, 柯林頓禁止使用聯邦經費從事複製人類的實驗, 並要求民間自我克制不作這種研究 (After the aouceet of the success sheep clog, U.S. Presdet Clto ad atoal Boethcs Advsory Cosso reproved hua clog as oral. Clto forbade usg federal fuds for hua clog experets ad asked the geeral publc to refra fro dog such research.) Table 20. Detaled Opo Suary of TCIR Postve 而複製技術如果成熟, 它將會是一種強大有用的工具, 任何工具都可能被善用或誤用, 評價一個工具不能只拿它被誤用的情境去批評它, 因而禁制了它被善用的原始目的與機會, 妥善的立法規範管理似乎才是較理性的作為 (Whe the clog techology reaches aturty, t wll becoe a powerful ad useful tool. Ay tool ca be used for a good cause or sused, so we should ot blatatly crtcze ad dsss a tool for ts possble abuses ad deprve t of ts opportuty to be used a good way. Istead, we should coe up wth sutable regulatos for usg the tool.) egatve 有人反對複製人, 因為違反了上帝的旨意 (Soe people are agast clog hua begs, because t coflcts wth teachgs of God.) Table 2. Detaled opo suary of blog artcles A Opo Trackg Syste Although opo suares ca be geerated by usg opo extracto ad opo suarzato algorths, they ay be dstrbuted dscretely whe relevat docuets are large. Lke a evet, we are ore cocered of how opos chage over te. A opo trackg syste as to tell how people chage ther opos as te goes by. A certa quatty of artcles s ecessary for such aalyss. Because the uber of artcles relevat to aal clog s ot large eough to track opos TCIR corpus, we take the presdet electo the year 2000 Tawa as a llustratg exaple. Opos towards four persos March 2000 are show Fgure. A B C D 2000/03/20: Electo Day Fgure. Opos towards four persos Persos A, B ad C were caddates ad D was the presdet at that te. Perso A was the Presdet elect. The tred fts the opos ths perod ad the opo suares ca tell evets correlated wth these opos. Ths trackg syste ca also track opos accordg to

8 dfferet requests ad dfferet forato sources, cludg ews ageces ad the web. Opo treds toward oe specfc focus fro dfferet expressers ca also be copared. Ths forato s very useful for the goveret, sttutes, copaes, ad the cocered publc. Cocluso ad Future Work Ths paper proposes algorths for opo extracto, suarzato, ad trackg. Dfferet aterals dfferet laguages are expereted ad copared. The ature of ews ad blog artcles s qute dfferet. Copared wth blog artcles, ews artcles ow a larger vocabulary. O the oe had, that akes the relevat setece retreval harder. O the other had, a larger vocabulary helps whe decdg setet polartes. The setet word er es postve ad egatve setet words ad ther weghts o the bass of Chese word structures. The f-easure s 73.8% ad 63.75% for verbs ad ous, respectvely. Experetal results also tell that ache learg ethods are ot sutable for setet word g. Utlzg the setet words ed together wth topcal words, we acheve f-easure 62.6% at the setece level ad 74.37% at the docuet level. Ivolvg topcal words ehaces the perforace of opo extracto. A opo trackg syste provdes ot oly text-based ad graph-based opo suares, but also the tred of opos fro ay forato sources. Opo suares show the reasos for dfferet stads people take o publc ssues. Opo holders are cosdered ths research. Experts or goveret offcers have ore fluece whe expressg opos. However, how opo expressers fluece the setet degree has ot yet bee explored. Idetfyg opo holders s very portat for aalyzg opos. Properly decdg the power of opo holders ot oly tells relable setet degree, but also aswers to the opoated questos. Moreover, the relatos betwee holders ad ther opos are the key to solve ultperspectve probles opos. Experetal resources ad tools ths paper are avalable at Ackowledgeets Research of ths paper was partally supported by atoal Scece Coucl, Tawa, uder the cotract SC E00-00-PAE. Refereces Carde, C., Webe, J., Wlso, T. ad Lta, D. Cobg lowlevel ad suary represetatos of opos for ultperspectve questo aswerg. Proceedgs of AAAI Sprg Syposu Workshop, pages Che, K.-H. ad Che, H.-H. Cross-Laguage Chese Text Retreval TCIR Workshop Towards Cross-Laguage Multlgual Text Retreval. ACM SIGIR Foru, 35(2), pages 2-9, Dave, K., Lawrece, S., ad Peock, D.M. Mg the peaut gallery: Opo extracto ad seatc classfcato of product revews. WWW 2003, pages , Fukuoto, F. ad Suzuk, Y. Evet Trackg based o doa depedecy. SIGIR 2000, pages 57-64, Hu, Mqg ad Lu, Bg. Mg ad Suarzg Custoer Revews. SIGKDD 2004, pages 68-77, Hu, Mqg ad Lu, Bg. Mg Opo Features Custoer Revews. AAAI 2004, pages , 2004 Ku, L.-W., L, L.-Y., Wu, T.-H. ad Che., H.-H. Major topc detecto ad ts applcato to opo suarzato. SIGIR 2005, pages , Ku, L.-W., Wu, T.-H., L, L.-Y. ad Che., H.-H. Costructo of a Evaluato Corpus for Opo Extracto. TCIR 2005, pages , Lu, B., Hu, M. ad Cheg, J. Opo Observer: Aalyzg ad coparg opos o the web. WWW 2005, pages K, Soo-M ad Hovy, Eduard. Deterg the Setet of Opos. Colg, pages , Me, J., Zhu, Y. Gao, Y. ad Y, H.. tog2y4c2c2l2. Shagha Dctoary Press Moraga, S., Yaash, K., Tatesh, K. ad Fukusha, T. Mg product reputatos o the web. ACM SIGKDD 2002, pages , Pag, B., Lee, L., ad Vathyaatha, S. Thubs up? Setet classfcato usg ache learg techques. Proceedgs of the 2002 Coferece o EMLP, pages Rloff, E. ad Webe, J. Learg extracto patters for subjectve expressos. Proceedgs of the 2003 Coferece o EMLP, pages Soboroff, I. ad Hara, D. Overvew of the TREC 2003 ovelty track. The Twelfth Text REtreval Coferece, atoal Isttute of Stadards ad Techology, pages 38-53, Takaura, H., Iu, T. ad Okuura, M.. Extractg Seatc Oretatos of Words Usg Sp Model. ACL 2005, pages Webe, J., Wlso, T., ad Bell, M. Idetfy collocatos for recogzg opos. Proceedgs of ACL/EACL200 Workshop o Collocato Webe, J., Breck, E., Buckly, C., Carde, C., Davs, P., Fraser, B., Lta, D., Perce, D., Rloff, E., ad Wlso, T. RRC suer workshop o ult-perspectve questo aswerg, fal report. ARDA RRC Suer 2002 Workshop