The United States Military Academy at West Point Orders

西点军校22条军规是您在成功的道路上必须学的一课，不管您有多忙，您的时间有多宝贵，请抽出时间认真阅读思考这些军规.虽然它和成功没有直接的关系，但是认真阅读体会它们，您的人生将会就此改变.人生的整个征途中，下面的西点军校22条军规无时无刻伴随着您，提醒着您，激励着您.当您在生活中、在事业上感到茫然无措、身心俱疲的时候，一定要回来将这些军规多读几遍.

第一条、无条件执行

很少有人可以完全听得进去别人的劝告，可以耐心地接受下达的命令，因为似乎在人们的潜意识里，“叛逆”是一种个性和值得推崇的东西，殊不知“叛逆”同样需要服从的资本.西点军校学子，美国总统艾森豪威尔说过：“任何语言都是苍白的，你唯一需要的就是执行力，一个行动胜过一打计划”.的确，只有在行动中，才会让人感觉到生命的价值，才可以使人变得智慧、勇敢、坚毅和高尚起来.光说不做，什么都无法改变；抱怨着去做，只会更加难过；只做不说才是提升自我的首选.因为，服从本身也是对自我的尊重和肯定.

第二条、没有任何借口

任何带一丝借口的理由都是庸者，愚者自我安慰的掌上名言.“没有任何借口”所表达的内涵远比字面意思要深刻的多.它不仅仅是西点军校对所有学员提出的一个口号，更是我们整个人生需要奉行的一个重要的思想理念和行为准则.它体现的是一种完美的执行能力，一种服从，诚实的态度，一种负责，敬业的精神.而在现实生活中，我们所缺少的正式这种精神.借口成了一面挡箭牌，这本身就是一种不负责任的态度.长此以往有害而无益，因为有各种各样的借口可找，就会疏于努力，不再想方设法去争取成功.你若不想做，会找一个借口；你若想做，会找一个办法.

第三条、细节决定成败

人常说，无论做什么事情只要把握大方向就行了，至于那些细枝末节就不要去管了.殊不知“千里之堤，溃于蚁穴”，小事不注意往往会酿成大问题.西点军校前校长潘莫曾指出：“最聪明的人设计出来最伟大的计划，执行的时候还是必须从小处着手，整个计划的成败就取决于这些细节.”精辟地指出了想成就一番事业必须从简单的事情做起，从细微之处入手.如果说不拘小节拥有的是豁达的人生，那注重细节的人往往会成就非凡的事业

第四条、以上司为榜样

这个世界只在乎你是否达到了一定的高度，而不在乎你是踩在巨人的肩膀上上去的还是踩在垃圾堆上上去的.上去的速度一定程度取决于学习的对象.正所谓人外人山外山，正因为有比自己强的人的存在，才有进步的动力.所以遇到强者是件幸运的事，意味着未来无限精彩的人生.对于值得自己学习的人和事，竭尽全力，这才是正确的人生态度.某种意义上它的背后隐藏的就是成功.

第五条、荣誉原则

所谓荣誉，就是一种向人们展示自己，并可以让人终身受益的资本.是否有荣誉心关系到事情的成败，而荣誉心是由自己的心去决定的.世上不知有多少安于现状的人，只因没有强烈的荣誉观念，所以他们得过且过，每天重复枯燥的生活，永无为目标去努力去激情的感受.其实很多时候，他们自身明明是可以做到的，却被自己丢弃在“平庸的荒漠”.

第六条、受人欢迎

人行于世，总希望和别人和和气气，快快乐乐的相处，某种程度上，这也是做好人办好事的前提.俗话说的好：得人心者得天下.”一个能让人打心眼儿里喜欢的人，可以在社会上左右逢源.人有百种，各有所好.若能做到让人人都喜欢，就不是一般的能力了.世间有这种能力的人可以更快的完成自己的追求，并获得他人的认可，一举双得.

第七条、善于合作

21世纪是充满竞争的世纪，敢于冒险，善于合作，是21世纪对人才规格的基本要求.自己是独立的，但世界是大家的.人生在世，与人沟通和合作是必然的.善于合作，是指在需要互相配合的事情上能够与别人协调一致，做好自己的那个部分.在合作中，要学会乐于助人、虚心请教别人、团结友善、平等待人.养成良好的合作习惯，关系到生活好坏，事业成败.毕竟，个人力量是有限的，只有实现了资源的优化组合，与志同道合者合作，才能实现仅凭自己无法达成的愿望.

第八条、团队精神

一群优秀的人，这样组成的团队是拥有最强生命力和竞争力的团队.而在它背后支撑他们每个成员的巨大力量，便是可贵的团队精神，一种深入灵魂，指引心灵，激人奋进的精神.如果非得说有什么力量是无坚不摧的，那么就是有具备这种精神的最强的团队了，而个体也只有在这样的团队中，才可能发展的更好.

第九条、只有第一

成功是很多人的最终追求，而竞争意识是成功人士的特征之一.只争第一，只当第一，体现常胜者处于不败之地的信心和魄力.对于天生的挑战者来说，第二意味着做的不够，只有第一，才是永恒不变的追求.一个第一，两个第一……永不满足，永不停下脚步.只有这样的胆大和“妄为”，才能有更为持久的竞争力和主动地位.

第十条、敢于冒险

每一年，每一天，时时刻刻，我们每个人都处在一定的风险里，有风险的存在才有预防风险的方法.那些过于保守，缺乏创新意识的人很难成功，而且很多时候，与风险结伴而来的还有机遇.风险越大，机遇越难得.选择还是放弃就成为了人生岔路口的指向标，或功成名就或一败涂地，就看是否有敏锐的眼光和敢于尝试的勇气.

第十一条、火一般的精神

有人说激情是年轻人的专利；也有人说激情是需要培养和刺激的.激情到底是什么？它对我们的人生又有着怎样与众不同的意义.这些虽非只言片语可以回答，但有一点可以肯定；激情，是每个人身上都应具有的精神，是生命对我们的馈赠.让我们也经历一次激情燃烧的岁月吧.

第十二条、不断提升自己

成长，在于每一天的获得和积累；提高，在于自己的学习和努力.西点军校第一任校长，乔纳森?威廉斯曾说过：“不管你有多么伟大，你依然需要提升自己，如果你停止在现有的水平上，实际上你是在倒退.”小到言谈举止，大到人生态度，都离不开主动的提升.成功的路不止一条，成功的标准也不止一个.有勇气不断超越自己，不断超越过去的人，才有可能跻身于成功者的行列.

第十三条、勇敢者的游戏

记得一首名为《勇敢者的心》的诗中这样写道：“用勇气之火去点燃希望之繁星，照亮人生过往中的每一日光阴，只因时间可以摧毁一切懦弱，却埋葬不了一颗勇敢者年轻的心.”一段人生，一场游戏，一个梦.既是游戏，总有输赢，不必为此揪心.只不过，这是一场只有勇敢者才玩得起的游戏.因为只有为梦想而坚持的人，才有胜出的可能，从而成就一段非凡的人生.

第十四条、全力以赴

留有余地还是全力以赴？这是两种不同的处世态度，猛然看去似乎都有道理，事实上要分场合分事情.对于梦想，只需懂得这两句词就够了“把握生命中的每一分钟，全力以赴我们心中的梦，不经历风雨，怎么见彩虹，没有人能随随便便成功……”.

第十五条、尽职尽责

尽职尽责，这不仅是工作原则，也是人生的原则.做事一丝不苟能够迅速培养严谨的品格、获得超凡的智能；它既能带领普通人往好的方向前进，更能鼓舞优秀的人追求更高的境界.无论做何事，务必竭尽全力，因为它决定一个人日后事业上的成败.一个人一旦领悟了全力以赴地工作能消除工作辛劳这一秘诀，他就掌握了打开成功之门的钥匙了.能处处以主动尽职的态度工作，即使从事最平庸的职业也能增添个人的荣耀.

第十六条、没有不可能

将“可能”变成“不可能”的人是十足的懒汉；而将“不可能”变成“可能”的人是自己人生的魔术师.这样的人常常会绝处逢生，再多的艰难困苦，在它面前似乎都只是“摆设”，能以很轻松的心态绕过前行.可能吗？不可能吗？一切都在于你自己是否想去争取，是否想去拥有.

第十七条、永不放弃

在世界上，没有什么东西可以代替坚韧不拔的意志，在拥有这种意志力的人的身上没有所谓的“滑铁卢”，更看不到沮丧的眼泪，不论面对怎样的困境，多大的打击，他们总是埋头苦干，从不轻言放弃.因为他们明白，期待的胜利往往就产生了再坚持一下的努力之中.这是一种可贵的品质，也是成大事立大业者的特征.

第十八条、敬业为魂

工作随着志向走，志向随着工作来.既然工作是志向所指，那必然是要竭尽所能去做的.做一个踏实的工作者，是表现自己的正途，用自身的积极行动去证明你是一个不可或缺的人.要知道，任何一个双手插在口袋里的人都爬不上成功的梯子.只有对事业全身心投入的人，才会获得事业的成功.

第十九条、为自己奋斗

每个人的人生就像一个金字塔，越往上走，你所享受的空间就越大.但大多数人宁可平庸，按部就班的过日子，辛苦地维持现状.只有少数人能在塔里漫步，游刃有余的生活，欣赏塔顶的风光，享受成功的喜悦，这类人就是知道自己为谁而奋斗的人.而当一个人先从自己内心开始奋斗时，他就是个有价值的人.每当想到你所要追求的，动力就会在你的身边.崇高的目标会为你带来无尽的快乐和激情.为自己奋斗，一切都不成问题.

第二十条、理念至上

有什么样的想法就有什么样的选择.有什么样的选择就有什么样的生活.我们虽然不能成为贵族的后代，但是通过我们的努力，我们可以成为贵族的祖先.这是一个充满竞争的时代，要想走向成功卓越，必须学会从起点之前赢得先机.在瞬息万变的各色竞争中，作为智慧之光的想法理念，是制胜的尚方宝剑.理念就是财富，就是资本，所以，拥有了它就拥有了成功的契机.

第二十一、条自动自发

所谓的主动，指的是随时准备把握机会，展现超乎他人要求的工作和表现，以及拥有“为了完成任务，必要时不惜打破成规”的智慧和判断力.我们的人生好比一份考卷，单选，多选，不定项，无其不有.什么是必须做的，什么是可以做的，什么是可做可不做的，都应心知肚明.主动一些，多比别人做一道，就会多学一点，多一份成功的可能.

第二十二条、立即行动

一个人一生的时间是有限的.今天你在二十四小时里无所事事，就意味着在明天的二十四小时里焦头烂额.时间对于每个人都是公平的，浪费自己的时间无异于慢性自杀.而人生里最大的成功就是在最短的时间里达成最多的目标，实现最大的价值.所以，要过有意义的人生，就不要耽搁，立即行动吧！

第二篇：United States Naval Academy American Association for Applied Corpus Linguistics

Truncated First Draft1 Please Do Not Quote Without Permission!!

?2001-2002 William H. Fletcher

fletcher@kwicfinder.com

All Rights Reserved

Concordancing the Web with KWiCFinder William H. Fletcher

United States Naval Academy

American Association for Applied Corpus Linguistics

Third North American Symposium on

Corpus Linguistics and Language Teaching, Boston, MA,

23-25 March 2001

Size and Composition of the Web

The World Wide Web is a wondrous place, with an overwhelming variety of information in countless languages and domains. Just how many webpages there are and how they are distributed by language and content are not easy questions to answer. The Web is constantly growing and changing, and even the best estimates can only approximate its extent and composition. Studies of the nature of the Web echo the story of the blind men and the elephant: each extrapolates from different samples of an ever-evolving entity taken at different times and by divergent means. The most reliable estimates suggest that the number of publicly-indexable webpages in mid-2001 falls in the range of three to five billion (i.e.

thousand million = 109), a number projected to grow to 10-15 billion by mid-decade; others believe these figures too conservative and place the actual numbers at two to three times as many.

These two billion-plus pages are only the visible tip of the iceberg. For a page to be indexable, there must be a valid HTML link to it from another publicly accessible site, which excludes the many pages with restricted access. Far larger is the vast “invisible web” of content in databases, which can only be evoked by entering relevant queries in a text box, and text materials stored in formats which are not typically indexed, such as word processor, Post Script and Adobe Acrobat files.2

Despite the overall size of this corpus, one language, English, continues to predominate. Studies conducted in 2000 by Inktomi and Cyveillance conclude that over 85% of publicly-accessible webpages are in English, but here again even the best-informed estimates vary widely. In the summer of 2001 the Agence de la Francophonie released L5: the Fifth Study of Language and the Internet, based on these studies and the one by Global Reach cited below, complemented by research into the numbers of

webpages in various languages returned by search engines. This report investigates the relative presence of the Romance languages, German, and English among online documents. It shows strong growth among the non-English languages in the proportion of webpages found relative to English, concluding that the number of webpages in each is roughly proportional to the number of Web users with that 1 This rough draft was originally submitted for publication in a volume of selected papers from the conference, a project which has since been abandoned. It will be updated and revised in December 2002, then submitted for publication elsewhere. 2Google.com has recently started indexing other online document formats such as Adobe Acrobat, PostScript, Microsoft Word and PowerPoint, and continues to add more word processor and spreadsheet file types;

unfortunately, incomplete stripping of formatting codes and imperfect reconstruction of text in columns can interfere with word matching and automatic text analysis.

William H. Fletcher Concordancing the Web with KWiCFinder

2

language as native tongue. Data from these and other studies of linguistic diversity on the Web are summarized in this note.3

Historically English-language users and content have overshadowed other languages on the Internet, but the trend away from the preponderance of English seems clear. Statistics compiled by Global Reach illustrate the long-term development. In 1996, four-fifths of the 50 million Internet users were native speakers of English. By September 2001 Anglophones constituted only 43% of the world’s online

population of 503 million. Global Reach expects their share to fall below 30% of the 850 million Web users projected for 2005.4 The anticipated phenomenal growth in this non-Anglophone Web population should spur tremendous expansion of online resources in tongues other than English, particularly the smaller non-Western ones, to the benefit of those who teach, learn, and investigate these languages.

The Web as a Corpus for Language Learning

The abundant and varied texts of the World Wide Web tantalize linguists and language instructors alike: the Web’s ever-expanding, self-renewing machine-readable body of Web pages in scores of languages are easy to retrieve, but they are also challenging to sift through and exploit efficiently. Yet there are compelling reasons to supplement existing corpora with online materials. Once compiled, a corpus represents a snapshot of language usage and issues at the time the content was produced. The great expense of setting up a large corpus precludes frequent replacement or updating, and content can age surprisingly quickly. In contrast, countless new documents appear on the Web daily, so examples of current language usage and contemporary issues abound. In addition, even a large corpus might include few examples if any of a relatively infrequent expression or construction that would not be difficult to

locate online. Furthermore, certain domains or text genres may be underrepresented or missing entirely in an existing corpus. Using the Web as a source one easily can compile an ad-hoc corpus to meet the specific needs of groups of learners or translators. Finally, while off-the-shelf corpora and corpus tools may entail significant fees and often require expensive hardware, the Web is virtually free, and desktop computers to perform the necessary processing are now within the reach of researchers and students alike. 3Percentage of webpages by language. Based on Google.com’s current figures, Alex Franz (2001) reports the following distribution of webpages, round to the nearester full percent: English 58%, Japanese 9%, German 7%, Chinese 6%, French 4%, Russian 3%, Spanish 3%, Italian 2%, Korean 2%, Portuguese 2%, other 4%. The Cyveillance study based on a sampling of 350 million webpages estimates that as of 10 July 2000, 84.7% of

webpages were U.S.-based and the rest “international”; presumably many of the remainer would be in English as well (Moore and Murray 2000). In a telephone interview on 8 November 2000 Julie Keslick of Inktomi indicated that language count was not a primary goal of the January 2000 Web Map study. According to the Inktomi tally, the top ten languages were: English 86.55%, German 5.83%, French 2.36%, Italian 1.55%, Spanish 1.23%, Portuguese 0.85%, Dutch 0.54%, Finnish 0.50%, Swedish 0.36%, Japanese 0.34%. Since the figures add up to about 100%, these languages apparently were the only ones identified. Grefenstette and Nioche (2000) offer a methodologically interesting study to estimate the number of words (not webpages) in a number of Latin-alphabet European languages, but it makes no attempt to estimate numbers for other languages.

4 Number of Internet users by language. Global Reach frequently updates its estimates of the global online population and fully discloses the methodology used to derive them. Its data from September 2001 show the

following percentages of users for the top ten languages: English 43%, Chinese 9.3%, Japanese 9.2%, Spanish 6.7%, German 6.7%, Korean 4.4%, Italian 3.8%, French 3.3%, Portuguese 2.5%, Dutch 2.2%, Other 8.9%. Another interesting source of this data is the Nua Internet “How Many Online” page listed in the bibliography. Current

estimates of number of webservers and users per country (not identical to the number per language) can be found at /daily/TopCountry.html.

William H. Fletcher Concordancing the Web with KWiCFinder

3

While the Web does not constitute a corpus in the classical sense, as a foreign language instructor I use my concordancing Web search agent KWiCFinder (described in detail below) several times a week to access it as if it were one, for developing instructional materials and as well as for my own edification and research, at both the micro- and macro-levels. Let me illustrate how with a number of concrete examples.

To teach grammar or vocabulary, the Web is my primary source of eloquent examples. When I examine examples from the Web, they often force me to refine my understanding of how the language works. For example, textbook chapters on German conjunctions always teach the contrast between aber ‘but,

however’ and sondern ‘but, rather’; I have tired of the small range of examples of sondern I can contrive for my students. When I turned to the Web for variety, a sample of 50 passages with sondern included none of the pattern “not A but B”. In most cases but or even rather were not acceptable translations for sondern; instead, a reformulation with instead (as in this sentence) was most appropriate. (Once again one wonders why textbooks concentrate on the least frequent use and ignore the others?)

The Web also allows me to verify current and possible usages and to obtain a rough indication of their relative prevalence and distribution. For example, I was astonished to encounter the phrase los sesentas ‘the sixties’ in a Latin-American text; I teach and normally would expect los (a?os) sesenta, without the plural marker on the numeral; the former has the overtones of an Anglicism. A series of KWiCFinder searches revealed that this usage is common and locally predominant in Latin America, yet virtually unattested in Spain.

The Web also permits my students and me to confirm and acquire vocabulary not yet found in

dictionaries – nor in off-the-shelf corpora. Once I came across the word privacidad ‘privacy’ in Spanish-language software instructions. My unabridged dictionary from the early nineties was ignorant of this neologism, so I again suspected a blatant Anglicism, but a KWiCFinder search proved that it is indeed used throughout the Spanish-speaking world, even by authoritative sites like IBM and Microsoft. In another case, when I was invited to give a keynote speech on technology in Dutch at a conference in

Belgium, I initially felt insecure: while I have near-native fluency in the language, I have not kept up with the vocabulary of technology. By reading excerpts from webpages that dealt with related topics I could fill in the lexical gaps with minimal effort.

Working with an ad-hoc corpus can help students develop discovery skills and reinforce linguistic

content. For instance, one intermediate German textbook I have used taught the passive voice – formed most frequently with the auxiliary wurde + past participle – right after the subjunctive, which is formed with würde plus infinitive for most verbs. Anglo learners tend to disregard both diacritics and details of form, so some students became confused. To help contrast the two constructions I built a pair of keyword-in-context (KWiC) concordances on wurde and würde plus personal endings from the Web and had the students analyze these constructions in context.5 This enabled them to understand and contrast the building blocks of these two constructions better and to observe dozens of examples of each in action.

Concordancing techniques are also beneficial at the text level. When searching for online documents which will be linguistically accessible to my students, I display the query terms in large chunks of context, up to a couple of hundred words. These long excerpts enable me to evaluate the language and content of the texts quite efficiently. I have learned to skim through excerpts from scores of documents and identify the most appropriate ones quickly. This text-level approach is useful for discerning both 5KWiCFinder supports exporting search reports to an HTML file. Various interactive tools programmed in JavaScript are incorporated into this file, permitting browser-based stand-alone analysis; for details refer to http://miniappolis.com/KWiCFinder/KWiCFinderKWiCFeatures.html.

William H. Fletcher Concordancing the Web with KWiCFinder

4

content and form, i.e. documents on a given topic or from a desired domain as well as those exemplifying a given construction or register. With judicious choice of search terms one can locate texts rich in e.g. past tense forms or subjunctives to serve as readings to reinforce acquisition of structures and forms. Students can also follow this technique to locate relevant online resources for Web-based reports and research projects. Those who do tend to consult a greater variety of sources; those who choose not to often rely on the first few links found on Yahoo or Google.

Approaches to Exploiting the Web as a Corpus A well-known model for finding and using information distinguishes three basic approaches: hunting, or searching directly for specific information, grazing, or using ready-made data sets which are composed and updated by an information provider, and browsing, or finding useful information by

serendipity.(Hawkins 1996) Each of these approaches can serve as a model for corpus building or utilization; a melding of these techniques is most typical-- and most successful.

HUNTING

Due to the Web’s size and lack of organization a search engine provides the most effective entry point for hunting information. There are dozens of general Search engines with world-wide reach, and thousands of others which concentrate on specific geographic regions, knowledge domains, or languages. Since the dawn of Web civilization, Anne Salzman and Doug Mills have sent their ESL students at the University of Illinois on “Grammar Safaris”. With their online assignment sheets as guide and armed only with a web browser, they use a Search engine to track down webpages with examples of the structures they are studying. Then they use the browser’s Find function to locate the examples within the documents and they copy and paste them into a word processor document to bring them to class. According to Salzman and Mills (2001), this approach from the Info-Stone-Age yields plenty of meat for classroom discussion. The hunting model is also being followed to exploit the Web as a corpus for linguistic research. Hans Bickel (2000) reports that investigators at the universities of Basel, Duisburg, and Innsbruck are trolling the Web for examples of regional usage in the various German-speaking countries to complement the material they have gleaned from other sources.6 Other powerful solutions built on Web searching techniques include using the Web to disambiguate natural language confusion sets (Banko and Brill 2001), as a resource for example-based machine translation (Grefenstette 1999), and, building on

Grefenstette's proposed techniques, to resolve prepositional phrase attachment ambiguities for parsing (Volk 2000, 2001).

GRAZING

A hunting party sometimes returns empty-handed, and how much time and effort it will take to bag useful citations is rarely predictable. In contrast to the safari model, Jeremy Whistle (1999) turns his students loose in a ready-to-graze pasture where he controls the kind and quality of the fodder that awaits them. He has selected texts from the “Label France” series published online by the French Ministry of Foreign Affairs. Since these texts are intended for foreigners learning about French civilization and culture, both the language and content is suitable for his students. As government-sponsored instruments of cultural diffusion, the documents entailed no difficulties in obtaining the rights to incorporate them into an offline corpus for desktop use. (The question of developing offline corpora from online documents is addressed 6See the project description “W?rterbuch Nationale Varianten des Deutschen” online at

http://www.germa.unibas.ch/deusem/forsch/Prolex/prolex.de.html

William H. Fletcher Concordancing the Web with KWiCFinder

5

extensively below.) With a search agent like KWiCFinder this approach could easily be implemented in the online-mode: searches could be restricted to a known site or range of sites with appropriate content and language. This extends the very focused and productive grazing model to webpages for which one cannot (or lacks the time to) obtain permission for offline use.

BROWSING

Browsing is central to the Web – indeed, the unplanned discovery of information and insights lies at the heart of learning and research, and both hunting and grazing demand fortuitous finds to succeed. When consciously searching the “World Wide Haystack”, most experienced “hunters” use search-engine hits merely as a point of departure for further browsing; they then typically follow several additional layers of links before reaching their goal. (K?rber 2000) My hard drive preserves scores of documents I have chanced upon while looking for something else online and have saved for possible use in teaching or

research. More frequently I rely on the “applied serendipity” approach described above: sending a search agent to retrieve and excerpt large numbers of documents, then scanning the results to winnow out the chaff and keep the grain. Silvia Bernardini (2001) has written of a systematic approach to increasing the number of serendipitous finds by having students work with a number of different corpora and analytical tools. Jennifer Pearson (2000) stresses that one must guide students to recognize true serendipity, i.e. to determine consciously whether an online document meets the essential criteria of reliability and appropriateness for one’s purposes, in this case to serve as a model for translation7

Search Engines Present and Future

Unless one has already chanced upon suitable pastures for grazing, Search engines remain an essential tool for building any extensive corpus of online documents. The challenges are to ensure that a search yields maximally relevant results and to separate out irrelevant and uninteresting documents efficiently.

The dynamic, increasingly market-driven nature of the Web entails significant challenges and frustrations for efficient online concordancing. The large general-purpose search sites are commercial ventures, set up and maintained at enormous expense. They exist to generate advertising and sales revenues for their owners in exchange for providing a useful service. Merely by coincidence they also can serve serious research purposes, but their owners have no incentive to address the specific needs of academics. In order to maintain or attain profitability, many search sites are evolving into marketing sites: through policies of paid inclusion or paid positioning they can steer searchers away from more relevant results toward their advertisers.

Search engines target the average searcher, whose requirements are quite different from those of a scholar or student. Casual users typically have a well-defined information need such as locating a specific site, finding a valid answer to a question, or finding a well-stocked site meeting their search criteria. In contrast, scholars and teachers must examine and evaluate a range of resources to find the most reliable sources and the most useful texts. Search engines excel at returning large numbers of hits (documents matching one’s query), but not at optimizing their relevance to the searcher’s intent. Frequent changes in document content and “link rot”—the tendency of webpages to move without a forwarding address, or 7In a very revealing study Hildreth (2001) investigated the factors underlying “false positives” in information

seeking, i.e. user satisfaction with poor search results. He concludes: There appears to be little interaction between these two variables [actual search performance, i.e. quality and relevance of results, and user satisfaction]. Searchers may express satisfaction with search results even when the results are far from optimal.”

William H. Fletcher Concordancing the Web with KWiCFinder

6

disappear from the Web altogether—can diminish the usefulness of search results even further.8 Studies of typical user’s search behavior and preferences have strongly influenced the evolution of online searching and suggest what kinds of search engines will thrive in the years to come.9 In general, users show a marked preference for directories with pre-selected links organized by topic or for sites with a natural-language interface such as AskJeeves over full-text search engines like AltaVista or Google. At sites like the latter, 80%-90% of all queries consist of a single word or phrase. While AltaVista supports complex queries with Boolean operators (logical operators like AND, OR, NEAR, NOT) and bracketing, up to 25% of such queries submitted are ill-formed and thus return no results. Users tend to follow up only the first few hits in the search results, calling them up one for one in the same window, then returning to get the next link.

Extrapolating from such studies and from current trends in user figures, it appears that “geek seek” full-text search sites like AltaVista will decline to the benefit of less powerful search engines which offer cleaner, more accommodating user interfaces and higher ranking for the results with the greatest likely relevance. Unfortunately for language professionals, it is precisely the complex queries rooted in the arcane world of Unix and grep that facilitate targeted online linguistic research. AltaVista’s successful challenger Google10 has prospered because its link popularity ranking usually yields relevant results—and because its coverage of the Web is vast and up to date. While it is a full-text search engine, its support for Booleans is limited to AND, OR and NOT; it lacks NEAR, wildcards and bracketing, and its distinction between lower and upper case and between plain characters and those with diacritics is inconsistent. Worse yet, Google’s link popularity ranking works against diversity in the search results. Perhaps the most unsettling trend for linguistic investigation is the development of information retrieval search models and natural language user interfaces. While a boon for novice searchers (and NLP researchers), these approaches will favor the largest languages from the wealthiest countries, excluding those for which linguistic data are already most difficult to obtain.

Genesis and Development of KWiCFinder

After the launch of AltaVista in 1995 I became an intensive search-engine user. Soon I learned how to maximize online search efficiency despite a slow connection: I would get a page of hits, open each in a new window, go back to the search engine for more hits, then evaluate the pages that had loaded in the meantime. Occasionally I would go off leaving a couple of dozen documents to load in my absence for subsequent perusal. Only a small percentage of students and colleagues to whom I tried to teach my multitasking method adopted this approach; the rest continued to express frustration with the large amount of time spent sifting through hits to find relevant webpages. It occurred to me that I could

automate the process of search and retrieval by writing a program to submit the query to AltaVista, then retrieve the pages and save them to disk automatically. This first step, dubbed WebFetch, satisfied my own immediate needs, but had little appeal for my students, since they still had to open and peruse numerous downloaded files. To expedite evaluation of those webpages, I started excerpting the webpages 8The Web Archive “Wayback Machine” launched for public access in October 2001 preserves 10 billion webpages that have changed or vanished.

9Relevant user studies include K?rber 2000; Jansen, Spink and Saracevic 2000; Silverstein, Henzinger, Marais and Moricz 1999.

10The author's webpage http://miniappolis.com/KWiCFinderVSWebKWiC.html details the primary differences between Google and AltaVista.

William H. Fletcher Concordancing the Web with KWiCFinder

7

and producing reports with KWiC display, resulting in KWiCFind, which several volunteers from my students evaluated in the spring of 1997. When our institution finally left Windows 3.1 behind, I reprogrammed it from the ground up for 32-bit Windows, providing the specific enhancements for foreign language users and linguists detailed below. At the 1999 CALICO Conference the reborn KWiCFinder was shown for the first time outside my classroom (Fletcher 1999). The current version of KWiCFinder can be downloaded free of charge at http://miniappolis.com/KWiCFinder/.

AltaVista offers a combination of features that make it the most powerful search engine to support. Unlike many others, it indexes all words, including the frequent “stopwords” ignored by others which may be the focus of linguistic investigation. To a certain extent it allows queries which distinguish upper-case letters from lower-case ones and “special” characters with diacritics from their “plain” counterparts. It even has some language-specific knowledge, for example about the equivalence of ? and ae, ? and ss in German. It provides true world-wide coverage and was the first to offer search by language. Essential for targeted searches, it supports Boolean operators, bracketing, and wildcards, and imposes no limits on the length or complexity of a query. Finally, AltaVista performs literal text matching, without attempting to “second guess” the user’s intent. After having been sold and reorganized several times, AltaVista's market share has diminished significantly, especially in the USA. It lags far behind some rivals in size and freshness of its content, and stands out with the highest percentage of dead links among the major search engines.11 Nonetheless its support for complex queries still makes it a very useful tool.

Daily experience with KWiCFinder and frustration with search engines led me to refine wildcard

matching strategies to reduce false matches. “Wildcards” permit a search term to match likely variants of a given word without the user’s entering each alternate form. For example, the AltaVista wildcard symbol * matches any sequence of zero to five characters, so the search term nation* would match singular / plural forms like nation, nations, as well as derived words like national, nationalism, nationality,

nationalize/ise etc., and labo*r matches both American labor and British labour. Furthermore, AltaVista automatically matches a plain character in a search term with any corresponding accented character, and lower-case letters also match their upper-case counterparts (e.g. a in a search term would match any of aá??à???A???????). These “implicit wildcards” ensure that many paradigmatic and graphic variants of a given word match a single search term, despite the differences introduced by factors like sentence-initial capitalization; required, omitted or misused diacritics; or alternate spellings due to keyboard limitations.

While wildcards increase the efficiency of entering search terms, they can also lead to many irrelevant matches which must be sifted out individually. To address this problem I implemented single-character wildcards and the “sic” option in KWiCFinder. Borrowing from standard concordance practice, I added the wildcard characters ? and % to the inventory to match either one (no more, no less) or zero to one character respectively. KWiCFinder’s “sic” option forces a plain or lower-case character in a search term to match only that exact character. Similarly, to AltaVista’s native NEAR Boolean operator, which requires only that one search term be within ten words on either side of another term, I added BEFORE and AFTER operators, and permitted users to specify a shorter distance between the terms. All these enhancements reduce the likelihood of unwanted matches.

These refinements—single-character wildcard and “sic” matching as well as specifying relative order and degree of proximity of two terms—do come at a significant price. When a query is submitted, it can only be as specific as the search engine’s conventions allow. A more general search may match many 11 tracks numerous search engine developments and statistics.

William H. Fletcher Concordancing the Web with KWiCFinder

8

documents which must be discarded after retrieval and analysis because they do not actually meet the user’s more specific criteria. In one intentionally extreme test I invoked “sic” to seek examples of the German verb form (ihr) fahrt ‘you (plural) go’, far less frequent than either the special-character third-person singular f?hrt or the capitalized noun Fahrt. KWiCFinder had to retrieve over 200 documents matching the search term fahrt according to AltaVista’s criteria to find a single citation of the desired form!12 However, since the program automates the entire process, even in such an extreme case it does use human time very efficiently.

The most efficient searches result from queries which avoid wildcards and specify every alternate search term completely. Nevertheless, entering all desired variants of a given form can be daunting and highly repetitive, especially in languages with richer morphology than English. To transfer this tedious task to the machine, I introduced “tamecards”, a shorthand for generating alternate forms. For example,

KWiCFinder expands the tamecard notation s[iau]ng[,s,ing] to all forms of the verb sing: sing, sings, singing, sang, sung (as well as the nonsense forms sangs, sungs, sanging, sunging, which fortunately yield no false matches). Each of these forms is then submitted to the search engine so that only perfect matches are retrieved. Since derivational and inflectional patterns typically apply to many words, such tamecard formulas can be saved, then pasted in as needed. A further refinement is the “indexed tamecard”, in which every nth field in curly braces corresponds to the corresponding field in other sets of curly braces within the same search term, so that {me,te,se} lav{o,as,a} expands to me lavo, te lavas, se lava. Such shorthand for fully-specified alternate forms would be a boon to searching on sites which do not support wildcards such as Google.

Another pair of KWiCFinder tamecard conventions addresses orthographic inconsistency in compounds which can be written as one word or two, either joined by a hyphen or separated by a space. A hyphen or apostrophe in a search term is expanded to alternate forms with or without a space.13 Consequently, on-line matches any of the interchangeable spellings on-line, on line, or online, and German ich hab’s

matches both ich hab’s and ich habs. This shorthand is particularly useful for contemporary German (as is AltaVista’s lower-case / upper-case equivalence), which now is in a ten-year period of transition to a new spelling. The reforms permanently separate many words formerly written as one, while fusing some former phrases into single words; they also allow individual discretion in breaking up German’s

notoriously long compounds with hyphens, leading to even greater orthographical variation. While the media and most schools are implementing the new spelling, many online sources will continue to reflect traditional orthography for years. With KWiCFinder, the search term kennen-lernen matches both old-style kennenlernen and reformist kennen lernen. This tamecard convention provides a simple means of matching both variants with a single entry.

In addition to these enhancements to query formulation KWiCFinder introduces a further means of narrowing a search, “inclusion” and “exclusion” criteria. These may be words whose appearance on a webpage helps target a specific domain or, alternatively, disqualifies that page from further consideration; these terms are submitted to the search engine as part of the query, but do not appear in KWiCFinder’s search report. Other selection criteria include date, Internet domain (a rough guide to country of origin), as well as host, i.e. a specific Web server, and URL. As exclusion criteria these latter parameters help one filter out unwanted material.

12

13Searching for ihr fahrt OR fahrt ihr would be the efficient way to do this. Like other search engines, AltaVista treats punctuation marks as spaces.

William H. Fletcher Concordancing the Web with KWiCFinder

9

Once launched, KWiCFinder works without further attention, retrieving five to ten documents a minute, excerpting them, and finally producing a search report which displays the key search terms in the amount of context specified by the user, along with information on and links to the source documents. Multiple independent searches can be carried out simultaneously, which is especially beneficial for long

unattended searches. To expedite later review, one can choose to save the original documents on the hard drive in HTML and / or text format. These original texts are then instantly available offline for perusal, editing and reproduction, or for further analysis by a full-featured concordancing program, and they remain accessible even if the online version is changed or removed.

KWiCFinder’s user report options have always offered various ways to set off the keywords from the surrounding text, and allowed a choice between a single report document per search and individual reports per document, practical for rapidly evaluating documents in a more extensive search. Recent stabilization of the XML (eXtensible Markup Language) encoding and XSLT (eXtensible Stylesheet Language Transformation) rendering standards have permitted KWiCFinder to offer an additional highly versatile report format since mid-2000.

XML provides a standard method for tagging structured data in a text file format that can be easily understood by both humans and computers. While HTML offers the page designer (in this case the

KWiCFinder programmer) reasonable control over page appearance, its formatting markup tags furnish no clues to the structure of the information on the page; once an HTML page has been completed, its form is basically set. In contrast, XML has no built-in display formatting, but provides a standard

approach to defining and encoding the structure of the information, essentially as a user-defined database. Consider this simplified snippet of a citation from an XML-encoded search report. Programmer-defined tags identify components as “<precontext>”, “<matchingtext>”, or “<postcontext>”. <cite citeID="8.1.1">

<precontext>

Da das LRZ anfangs mit ?hnlichen Ger?tschaften zu tun hatte

der erste Rechner hie? PERM, natürlich nicht nach dem Erdzeitalter, </precontext>

<matchingtext>

sondern

</matchingtext>

<postcontext>

als Abkürzung für "programmgesteuerte elektronische Rechenanlage München"- k?nnte man hier den ersten Zusammenhang sehen.

</postcontext>

</cite>

All of the data from a KWiCFinder search are stored in this way in an XML file. To generate a useful report, KWiCFinder applies an XSLT “stylesheet” to this database to select which information to display, insert appropriate text labels in the desired language, and format the result as an HTML document for display in its browser window.

The advantages and power of XML encoding becomes clear from the samples of an actual search report accessible via this link. Display form is perfectly separated from search report content, and it can be

modified as needed. To change the display format or the language of the text labels, KWiCFinder merely applies a different stylesheet to the same XML file; there is no need to reanalyze the original documents. With appropriate knowledge of XSLT and browser scripting techniques, an end user could create new report formats or apply other stylesheets to annotate, merge, prune, or restructure XML search reports.

William H. Fletcher Concordancing the Web with KWiCFinder

10

There are numerous instructive examples of these manipulations online (at sites like , and /xml/) and in books, such as Britt and Duynstee (2000); Kay (2000) provides a comprehensive reference to XSLT. While learning to work with these technologies is not a trivial enterprise, the growing commercial enthusiasm for XML promises that this expertise will continue to become more readily available. The ability to perform sophisticated database and report display manipulations in a current-generation browser points the way to a future cross-platform approach to learner concordancing.

WebKWiC

Some searchers have been intimidated by the effort required to download, install, and learn to use KWiCFinder, yet they still can benefit from automation of search and retrieval. To lower the entry

threshold for such users I created WebKWiC,14 a light-weight, fully browser-based JavaScript application. It capitalizes on Google’s “Document from Cache” feature, which serves up a copy of a webpage matching a user’s query from Google’s archives, highlighting instances of the search terms with color codes. WebKWiC retrieves several of these cached pages at a time and adds buttons so the user can navigate easily among citations and windows, greatly enhancing the efficiency of previewing large numbers of documents. WebKWiC also adds a means of entering “special characters” to the user interface and gives certain essential search options greater prominence than does Google’s original page. Google is an ideal partner for an entry-level search agent like WebKWiC. Its straightforward approach to advanced search with “implicit Booleans” is easy to learn, so users either come equipped with or acquire readily transferrable skills. Since Google indexes major non-Western European / non-Roman orthography languages, this approach allows me to meet the needs of a population which KWiCFinder does not support yet.15

Webidence as Evidence

We all know (and may ourselves have voiced) the complaints about online information: there is too much ephemeral content of dubious reliability; journalistic, commercial and personal texts of unknown

authorship and authority abound; assertions are intermingled with and represented as established fact, and details of sources and research methodology are documented haphazardly at best. For linguistic research even more caution is essential for numerous reasons. The Internet domains in a URL (e.g. .ca, .uk, .de, .jp, .com, .edu) are only a rough guide to provenance. In addition, many webpages consist primarily of fragments–titles and captions, supplemented by the occasional imperative (“click here for more

information”, “buy now”). As the lingua franca of the digital frontier, English is both the target and 14

15http://miniappolis.com/WebKWiC/ Recently a couple of alternatives to KWiCFinder and WebKWiC have appeared. Two online pages produce KWiC concordances from search results: WebCORP (.uk), which uses various search engines and provides a number of analytical tools, and WebCONC (http://www.niederlandistik.fu-berlin.de/cgi-bin/web-conc.cgi), which works with Google only. Both offer a distinct advantage: processing takes place on the server, so no software needs to be downloaded, and users with slow connections can concordance large numbers of documents in a relatively short time; neither has the search-engine extensions, language enhancements or reporting flexibility of KWiCFinder. A third possibility, TextSTAT, (http://www.niederlandistik.fu-

berlin.de/taalkunde/software.html), will retrieve and create a simple concordance of a URL entered by the user. Finally, the search agent Copernic (, available in a free “basic” version) performs searches on multiple sites simultaneously and shows one instance from each document of a search term in a very brief context.

William H. Fletcher Concordancing the Web with KWiCFinder

11

source of contamination: non-Anglophones often translate their webpages into Info-Age pidgin English, at the same time fusing creolized Web English into texts in their native tongue. Similarly, while searching for linguistic examples I often have stumbled upon compositions by learners with imperfect mastery of the language (many language courses post student work for peer review) as well as numerous baffling documents that turned out to be machine translated.16 In many online discussion groups, sloppy spelling and careless language appear to be the norm. With its frenetic pace of development, the Web typically values content creation above content perfection and tolerates ill-formed language–after all, those who are upset by it can find relief a click away.

In light of these pitfalls our profession needs to develop “Standards of Webidence” to guide the selection and documentation of online language for linguistic research. We also must understand and beware of the limitations of search engines. In particular, the number of pages matching a query reported by a search engine gives a rough numerical indication at best; comparison of these numbers does not constitute

statistical proof.17 Search engines report the number of pages matching a query, not the actual number of citations on those pages. A single page may contain several alternate usages (as in the los sesentas example above), thus appearing in more than one count. On the other hand, numerous pages may

propagate verbatim a formulation originating in a single document, thus multiplying its frequency, as in copied quotations, song lyrics, aphorisms, anecdotes, and jokes; online forums in which an original

posting and all subsequent comments are repeated in each successive posting; and mirror sites for FADs (frequently-accessed documents). Furthermore, a single site may be responsible for most or all the hits of a query for a spurious or unusual usage.

AltaVista itself warns not to trust its figures: when its servers experience heavy traffic, generating result pages receives priority over producing hit counts, so numbers for the same query easily vary by an order of magnitude over the course of one search session.18 Finally, the fact that a given form or construction can be found on the Web does not amount to proof of its existence in a language: many hapless hapax legomena born of input error or syllable stranding by hyphenation wait on the Web for an unsuspecting searcher to united them with their orphaned siblings.

KWiCFinder facilitates responsible online linguistic scholarship in several ways. It allows one to review large numbers of documents and citations efficiently, with each keyword shown in sufficient context to evaluate its relevance and validity. It can tally the number of instances of each keyword in a document for calculation of its relative frequency. The user can choose to save the documents to a local file to permit further analysis or independent verification of results. It incorporates tools to annotate, classify and delete individual citations or entire documents from a search report. Finally, complementary corpus 16One machine translation was so artless that even the HTML tags were rendered in Spanish, with <CABEZA> and <CUERPO> replacing <HEAD> and <BODY>!

17 It is unclear how many linguists are aware of these limitations. I have seen postings in scholarly fora such as Linguist List citing hit counts from AltaVista as evidence prevalence of a given form over another with no

indication that the poster either has followed up to verify a substantial number of the hits or is even aware of the limitations of this method.

18 The page “AltaVista Advanced Search Tutorial--About the Page Count” cited in the bibliography explains this limitation of AltaVista’s hit counts. Brekke (2000) and Meyers et al. (2001) note this problem as well. I have even received negative hits counts like “We found -40,000,000 results.” To ensure the most accurate counts, follow

AltaVista’s advice by accessing it at off-peak times, e.g. on weekend mornings. Note that specifying “one page per site” does not affect total counts, so it provides no indication of how widespread a given form or construction is.

William H. Fletcher Concordancing the Web with KWiCFinder

12

analysis tools now under development enable one to eliminate unrepresentative, redundant or repetitive documents from further consideration.

Representativeness of the Web as a Linguistic Corpus

In the five years I have been developing and using KWiCFinder it has shown me excerpts from an

estimated quarter million online documents, and I have viewed many of these as complete webpages. My cumulative impression of these documents informs both my concerns about the use of webidence and my conviction that the Web can yield linguistic data which are both useful and reliable.

To proceed from qualitative impressions to quantitative conclusions I compiled a series of pilot corpora with KWiCFinder of Web documents in English for analysis offline. This allowed me to gauge how suitable and representative these texts could be for research or learning and to evaluate techniques to identify webpages with a high proportion of connected text. My goal was to sample language from the Web, not investigate the language of the Web.

A major objective of this pilot study was to develop procedures and tools to automate or expedite identification of the most useful texts. Some steps toward this can be taken when formulating the query, by choosing selection criteria which either exclude a range of texts or favor inclusion of more relevant results. For example, by excluding documents with “copyright” or “all rights reserved” one can screen out commercial and journalistic texts while allowing most academic material to be downloaded. Another simple indicator of likely usefulness is document size: a query to the server can reveal how large a

webpage is before the search agent “decides” to download it. With guidelines for rejecting a page a priori because it is relatively unlikely to contain useful text, the agent could save both bandwidth and processing time.

Web documents typically contain “noise”: headers and footers that identify the document, declare who owns it and reserves rights to it, links to other documents, media and sites (especially advertisers), and other formulaic boilerplate material. Unfortunately HTML provides no consistent way to distinguish such boilerplate elements from the unique textual content of each page.19 Without insight into the structure of a webpage, a search agent has no criteria for excluding any portions of it, and thus must include everything. Obviously, the shorter the webpage, the lower the “signal to noise” ratio, and the less likely it is to contain more than a few sentences of connected text; practical guidelines for a cutoff point are needed. At the other end of the spectrum, the very largest webpages tend to consist of lists and fragments: server logs and statistics, indexes, glossaries, discussion group message headers and

“linketeria” pages. Such webpages can be enormous–up to several megabytes–while pages of that length with coherent text are quite rare.20 Since downloading long documents consumes significant bandwidth, an upper size limit would be useful as well.

For my pilot study I chose as search terms very high frequency words likely to occur in any extended text 19Many websites do use custom templates with comments or element tags which allow one to find page elements like headers, footers, advertisements and contents automatically. While useful for analyzing numerous documents from a single site, parsing heuristics are rarely transferrable from one site to another.

20 There are notable exceptions to this general principle. Once I found all the works of a German author combined into a huge HTML file 4.5 MB in size. Other files can be exceedingly large for the number of words they contain, especially PDF files that incorporate graphics and HTML files exported from Microsoft Word, which can easily be five to ten times the size of a generic HTML file with the same content.

William H. Fletcher Concordancing the Web with KWiCFinder

13

in order to obtain the widest range of documents. For the first 100 texts the query submitted to AltaVista was “the OR of OR a”, i.e. match any document in English containing any of these three words. This round of search results yielded almost pure commercial language of the Web. After tallying up the

collocations, I found that “all rights reserved” was the most frequent trigram (43 times), and “copyright #”21 fell among the top ten bigrams, indicating a high proportion of commercial texts. In a second

sampling I ran a series of queries for a maximum of 1000 documents each for ten high-frequency words.22 Among the 5859 documents these searches yielded were 2277 or 39% duplicates. Since AltaVista had instituted preferential treatment for paying customers early in 2001 and had hardly updated the rest of its database all year, I determined which hosts (webservers) were “overrepresented” in the results and which had yielded the “noisiest” documents. Finally I conducted a third round of searches with twenty high-frequency words, explicitly excluding the overbearing and noisy hosts. To reduce the commercial bias of the sample, these searches were limited to documents last visited by AltaVista before 1 January 2001,

since any clients who paid for their services would have been updated since then. This third round yielded 11,201 documents, 7294 of which were unique.23

After downloading every document matching my queries and eliminating duplicates, I examined those with the most frequent longer n-grams, in this case where n was equal to 20, 12 and 8, to reduce the number of “anomalous” documents further. Since any text with several instances of such an n-gram is repetitive, it deserves close scrutiny. I developed a program to locate files with frequent long n-grams and display their contents for individual examination; this allowed me to determine the source of the redundancy and to judge whether to eliminate them. In a number of cases the repetition came from document-internal navigation links and other boilerplate material; in some transcripts of legal and legislative proceedings repetition of formulaic elements was common. Generally such repetition was deemed minimal in the context of a longer text, so the documents were retained. On the other hand, discussion group threads which repeated the original posting and all subsequent reactions in each

succeeding response were usually marked for elimination. While one could devise a program to weed out the repeated passages in such “fugues on a theme”, the careless language samples rarely appear worth the effort of salvaging.24

After automatically sifting out duplicate and repetitive documents, 7038 remained and the drudgery began: I viewed each of these survivors briefly and eliminated those in which coherent text was not “predominant.” The principle of “predominance” was at best vaguely defined–a major purpose of this exercise was to arrive at guidelines–and since I reviewed about 12 documents per minute I claim no rigorous consistency. I allowed each document a “reasonable amount” of overhead for its size–headers, footers, links, bibliography, lists, non-English text–but not exceeding 20% of content for short

documents, dwindling to about 5% maximum for longer ones. 21

22As a “noise filter” all numbers were mapped onto #, so copyright years are all tallied together. Currently AltaVista limits each search to a maximum of 1000 hits. Since many pages may have changed or become unavailable since AltaVista’s last visit and multiple links in a search report may refer to to the same document, the actual yield of unique webpages was less than 60%.

23Texts were normalized and tokenized by removing all formatting, case distinctions and non-alphabetical characters. The program then compared the file contents and moved duplicates to a separate directory.

24Ironically, search-term ranking algorithms favor such mindless repetition, since it makes a search term very

prominent within a document. Furthermore, in a very short document every word is salient. These two factors raise the “noise level” of search hits considerably.

William H. Fletcher Concordancing the Web with KWiCFinder

14

Since I typically conduct narrowly-defined searches with criteria conceived to limit results to a single domain, I was struck by the variety of content I found. Among the shorter documents–those of a few hundred to a couple of thousand words–commercial and personal text prevailed. At the other end of the scale–up to 60,000 words–legal texts and government proceedings were well represented. The middle range was filled with academic texts–papers, theses, syllabi and course materials, some computer

hardware and programming documentation, other expository prose, drama (including Shakespeare) and fiction, and personal interest pages, as well as a surprising number of religious documents and

commentaries in the Christian, Islamic and Hindu traditions. As expected, it was this middle range that yielded the most useful texts.

The cursory examination of the documents described above led to elimination of roughly 30% of the pages, leaving 4949 documents totaling 5,248,929 words and 34,995,762 bytes. As anticipated, the shortest and longest documents bore the brunt of this pruning. Half of all documents were under 3330 bytes long, and of these about 40% were rejected. Only 10 documents were longer than 100 kb, and more than half of these were deemed primarily nontextual; in fact, no documents over 200 kb were retained. In the range of 5-100 kb, over three-quarters of the documents consisted primarily of text. The optimum size seems to fall around 50 kb, where only 17.8% of documents were rejected. Nevertheless, owing to the far greater number of smaller files, the median size of texts retained was only 3770 bytes!

So what size HTML files are most worth downloading? Unfortunately the file size gives only a rough indication of how much text it contains. Some HTML editors grossly inflate the files (Microsoft Word is the greatest offender here), and included stylesheets and scripts add bulk but no textual content. Stripping out such formatting elements typically reduces files to 40-65% of the HTML size; here again shorter files have greater overhead. This signal-to-noise ratio and the observations in the previous paragraph suggest the following rule of thumb: to maximize the “yield” of coherent text, download HTML files only between 10 and 150 kb in size.

Had we programmed KWiCFinder to follow this rule of thumb in this pilot study, we would have

downloaded only one-third of the final number of files, but that would have yielded a corpus two-thirds of the size of the current one with enormous savings in bandwidth and analysis time. The capability to exclude files below a given size is now on my KWiCFinder “to do” list; currently one can only opt to avoid excessively large files.

Copyright Considerations

Having compiled a corpus of webpages, instructors or investigators may wish to share it with students or colleagues. The daunting effort required to obtain permission from all webpage creators to incorporate their material into instructional or research works–and the typical low response rate for such requests–raises the question whether every Web-based ad-hoc corpus is automatically destined for disposal, not distribution. Clearly including entire webpages without permission in a corpus distributed on CD-ROM would be prohibited. Nevertheless, I would argue that providing a KWiC concordance via the Web of brief excerpts from webpages cached in their entirety on a “corpus server” falls within currently accepted practice. While I lack legal expertise, I do note that for years search engines like AltaVista and Google have included brief KWiC excerpts from documents in their search reports with impunity. In fact, both Google and Web Archive serve up entire webpages and even images on demand from their cache.25 Both 25Indeed, to reduce bandwidth requirements, many large national Internet service providers save copies of webpages and serve them up from their cache rather than fetching the original document whenever requested.

William H. Fletcher Concordancing the Web with KWiCFinder

15

these sites have policy statements suggesting that there is implied consent from the webpage creator to cache and pass on content if the site fails to have a “robots.txt” file prohibiting it and the document lacks a metatag specifying limitations on caching. Web Archive’s FAQ explicitly claims that its archive does not violate copyright law, and it provides a mechanism for copyright holders to request removal of their material from the site.26 Eventually common practice will demand clarification of this legal gray area through legislation or litigation.27 Then my optimism that a Web-accessible corpus derived from online documents retrieved by a search agent in ad-hoc searches can be set up legally may well be vindicated. Meanwhile, our discipline should dare to set precedents and test the limits while scrupulously respecting any restrictions a webpage author communicates via industry-standard conventions (meta-tags, robots.txt).

REFERENCES

Ackermann, E., & Hartman, K. (2000). The Information Specialist’s Guide to Searching and Researching on the Internet & the World Wide Web. Wilsonville, OR: ABF Content.

Agence de la Francophonie. (2001). The Fifth Study on Languages and the Internet. Retrieved 18 October 2001 from the World Wide Web:

http://funredes.org/LC/english/L5/L5contents.html

AltaVista Advanced Search Tutorial - About the Page Count. Palo Alto, CA: AltaVista Company. Verified 11 November 2001 on the World Wide Web:

/adv_search/ast_as_pagecount

Aston, G. (2001). Learning with Corpora. Houston: Athelstan.

Banko, Michele and Brill, Eric. (2001). Scaling to Very Very Large Corpora for Natural Language Disambiguation. ACL 2001. Retrieved 5 October 2001 from the World Wide Web:

/~brill/Pubs/ACL2001.pdf

Bernardini, Silvia. (2000). Systematising Serendipity: Proposals for Concordancing Large Corpora with Language Learners. In Burnard and McEnery, 224-34.

26 This excerpt from the Internet Archive's FAQ asserts that serving cached copies of webpages is legal: Are you violating copyright laws?

No. Like your local library’s collections, our collections consist of publicly available

documents. Furthermore, our Web collection (the Wayback Machine) includes only

pages that were available at no cost and without passwords or special privileges. And if

they wish, the authors of Internet documents can remove their documents from the Wayback Machine at /internet/remove.html.

(/exec/faqsidos/about/faqs.html?index=2)

See also “The Internet Archive Terms of Use, Privacy Policy and Copyright Policy”

/terms/index.html#Copyright_Policy.

27Alternatively, one could reason that a KWiC concordance falls under the fair-use provision of United States copyright law. Crews (2000) and Hilton (2001) both argue for more liberal interpretations of this law than one usually encounters in the copyright policies. of academic institutions.

William H. Fletcher Concordancing the Web with KWiCFinder

16

Bernardini, Silvia. (2000). Serendipity expanded: Exploring new directions for discovery learning. Fourth International Conference on Teaching and Language Corpora, Graz. Abstract retrieved 26 July 2001 from the World Wide Web:

http://www-gewi.kfunigraz.ac.at/talc2000/Dokumente/abstracts/Bernardini%20Abstract.doc

Bickel, H. (2000). Das Internet als Quelle für die Variationslinguistik. In: Annelies H?cki Buhofer (Ed.),Vom Umgang mit sprachlicher Variation. Soziolinguistik, Dialektologie, Methoden und

Wissenschaftsgeschichte. Festschrift zum 60. Geburtstag von Heinrich L?ffler. (pp. 111-124) Tübingen: Francke. Retrieved 12 October 2001 from the World Wide Web: http://www.germa.unibas.ch/seminar/whoiswho/Publikationen/Variationsling.pdf

Brekke, M. (2000). From the BNC toward the Cybercorpus: A Quantum Leap into Chaos? In Kirk, John. M. (Ed.), Corpora Galore: Analyses and Techniques in Describing English. Papers from the Nineteenth International Conference on English Language Research on Computerised Corpora (ICAME 1998) (pp. 227-247). Amsterdam, Atlanta: Rodopi.

Britt, J. & Duynstee, T. (2000). Professional Visual Basic 6 XML. Birmingham UK: Wrox Press.

Burnard, Lou and Tony McEnery (eds) (2000). Rethinking Language Pedagogy from a Corpus Perspective: Papers from the Third International Conference on Teaching and Language Corpora. Frankfurt am Main: Peter Lang. (?od? Studies in Language Vol 2).

Cavaglià, Gabriela and Adam Kilgarriff. (2001). "Corpora from the Web". Fourth Annual CLUCK Colloquium, Sheffield, UK, January 2001. Retrieved 8 November 2001 from the World Wide Web: ftp://ftp.itri.bton.ac.uk/reports/ITRI-01-11.pdf

Grefenstette, Gregory. (1999). The World Wide Web as a Resource for Example-Based Machine Translation Tasks. ASLIB, Translating and the Computer 21, London, Nov 10-11,

1999. Retrieved 12 November 2001 from the World Wide Web:

http://www./research/mltt/publications/Documents/P4

9030/content/gg_aslib.pdf

Hildreth, Charles R. (2001). Accounting for users' inflated assessments of on-line catalogue search performance and usefulness: an experimental study. Information Research, 6/2. Retrieved 22 January 2001 from the World Wide Web: http://InformationR.net/ir/paper101.html

Hilton, James. (2001). Copyright Assumptions and Challenges. EDUCAUSE Review 36/6, November / December, 48-55. Retrieved 12 November 2001 from the World Wide Web:

http://www.educause.edu/ir/library/pdf/erm0163.pdf

Crews, Kenneth D. (2000). Fair Use: Overview and Meaning for Higher Education. Retrieved 12 November 2001 from the World Wide Web: http://www.iupui.edu/~copyinfo/highered2000.html Estadísticas de Internet en el ámbito internacional Madrid: Asociación de Usuarios de Internet. Retrieved 6 November 2001 from the World Wide Web: http://www.aui.es/estadi/internacional/internacional.htm

Evolution & Projections of Online Linguistic Populations. San Francisco, CA: Global Reach Retrieved 6 October 2000 from the World Wide Web: /globstats/evol.html

Fletcher, William H. (1999). Winnowing the Web with KWiCFinder. CALICO, Miami University of

William H. Fletcher Concordancing the Web with KWiCFinder

17

Ohio, Oxford, OH, 5-9 June 1999.

Franz, Alex. 2001. Cite Google.com e-mail Nov

Grefenstette, Gregory & Julien Nioche. (2000) Estimation of English and non-English Language Use on the WWW. RIAO 2000, Paris, 12-14 April 2000. Retrieved 12 October 2001 from the World Wide Web: http://www./research/mltt/publications/Documents/P19137/content/RIAO2000gref.pdf

Global Internet Statistics (by Language). San Francisco, CA: Global Reach Retrieved 6 October 2001 from the World Wide Web: /globstats/index.php3

Godwin-Jones, Robert. (1999). Web Metadata: More Efficient Resource Cataloging and Retrieving. Language Learning & Technology 3, (1): 12-16

Grefenstette, Gregory (1999) The World Wide Web as a Resource for Example-Based Machine Translation Tasks. Retrieved 12 October 2001 from the World Wide Web:

http://www./research/mltt/publications/Documents/P49030/content/gg_aslib.pdf

Grefenstette, Gregory & Julien Nioche. (2000) Estimation of English and non-English Language Use on the WWW. RIAO 2000, Paris, 12-14 April 2000. Retrieved 12 October 2001 from the World Wide Web: http://www./research/mltt/publications/Documents/P19137/content/RIAO2000gref.pdf

Halteren, Hans van (1999). Tekstcorpora en taalgerelateerd onderwijs. STDH Nieuwsbrief, 9 (Feb.): 7-9. Hawkins, Donald T. (1996). Hunting, Grazing, Browsing: A Model for Online Information Retrieval. ONLINE 20(1 January). Retrieved 21 October 2001 from the World Wide Web:

/JanOL/hawkins.html

Hock, Randolph. (1999). The Extreme Searcher's Guide to Web Search Engines. A Handbook for the Serious Searcher. Medford, NJ: Information Today. Supplement and update at

/engines/

Jansen, B. J. (2000) The Effect of Query Complexity on Web Searching Results. Information Research, 6(1). Verified 16 November 2001 on the World Wide Web: http://informationr.net/ir/6-1/paper87.html

Jansen, B. J., Spink, A., & Saracevic, T. (2000). Real life, real users, and real needs: A study and analysis of user queries on the web. Information Processing and Management, 36 (2): 207-227.

Jansen, B. J. & Pooch, U. (2001) A Review of Web Searching Studies and a Framework for Future Research. Journal of the American Society for Information Science and Technology, 52 (3): 235-246.

Johns, Tim F. (2000). Tim Johns English for Academic Purposes Page. Retrieved 18 October 2001 from the World Wide Web: http://web.bham.ac.uk/johnstf/timeap3.htm#revision

Johns, Tim F. (2001). Modifying the Paradigm. Third North American Symposium on Corpus Linguistics and Language Teaching, Boston, MA, 23-25 March 2001.

Kilgarriff, Adam (2001). "Web as corpus". In Paul Rayson, Andrew Wilson, Tony McEnery, Andrew Hardie and Shereen Khoja (eds.) Proceedings of the Corpus Linguistics 2001 conference, UCREL

William H. Fletcher Concordancing the Web with KWiCFinder

18

Technical Papers: 13. Lancaster University, 342-344.

Kay, Michael. (2000). XSLT Programmer's Reference. Birmingham UK:Wrox Press.

K?rber, Sven. (2000). Suchmuster erfahrener und unerfahrener Suchmaschinennutzer im

deutschsprachigen World Wide Web. Ein Experiment. Unpublished master’’s thesis, Westf?lische Wilhelms-Universit?t Münster, Germany. Retrieved 14 September 2001 from the World Wide Web: http://kommunix.uni-muenster.de/IfK/examen/koerber/suchmuster.pdf

Lamy, M. N., Klarskov Mortensen, H. J. & Davies, G. (2000). Using concordance programs in the

modern foreign languages classroom. ICT4LT Module 2.4. Retrieved 1 June 2001 from the World Wide Web: /en/en_mod2-4.htm

Lawrence, S. & C. L. Giles. (1998). Context and Page Analysis for Improved Web Search. IEEE Internet Computing, 2 (4):38-46.

Lawrence, S. & C. L. Giles. (1999). Accessibility of Information on the Web. Nature, 400: 107-109. Summary, commentary, update and download at

Limberg, L. (1999) Experiencing information seeking and learning: a study of the interaction between two phenomena. Information Research 5(1). Retrieved 17 September 2000 from the World Wide Web: http://www.shef.ac.uk/~is/publications/infres/paper68.html

Meyer, Charles Roger Grabowski, Thomas Han, Konstantin Mantzouranis, & Stephanie Moses. (2001). The World Wide Web as Linguistic Corpus. Third North American Symposium on Corpus Linguistics and Language Teaching, Boston, MA, 23-25 March 2001.

McEnery, T. & Wilson A. (1996). Corpus Linguistics. Edinburgh: Edinburgh University Press.

McEnery, T. & Wilson, A. (2000). Corpus linguistics. ICT4LT Module 3.4. Retrieved 1 June 2001 from the World Wide Web: /en/en_mod3-4.htm

Moore, A. & Murray, B.H. (2000). Sizing the Internet. July 10, 2000. Arlington, VA: Cyveillance, Inc. Retrieved 8 October 2000 from the World Wide Web:

/resources/7921S_Sizing_the_Internet.pdf

Nua Internet How Many Online. Dublin: Nua Ltd. Retrieved 8 October 2001 from the World Wide Web: http://www.nua.ie/surveys/how_many_online/index.html and regional subpages.

Pearson, Jennifer (2000). Surfing the Internet: teaching students to choose their texts wisely. In Burnard and McEnery, 235-39.

Petitclerc, Angela (1998). Corpora for Concordances. Paper for course assignment. Montreal: Concordia University. Retrieved 12 March 2000 from the World Wide Web:

http://rkenner.concordia.ca/GSE555_98/A_Petitclerc/corpora.html

Reibold, H. & Neumeier, F. (2000) Findigkeit gefragt. PC Professionell (February) 177-ff. Retrieved 7 July 2000 from the World Wide Web:

http://www.zdnet.de/internet/artikel/scene/200002/suchmaschinen01_00-wc.html

William H. Fletcher Concordancing the Web with KWiCFinder

19

Salzmann, A. & Mills, D. (2000). LinguaCenter Grammar Safari. Retrieved 23 July 2000 from the World Wide Web: http://deil.lang.uiuc.edu/resources/web.pages/grammarsafari.html

Silverstein, C., Henzinger, M., Marais, H. & Moricz, M. (1999). Analysis of a Very Large Web Search Engine Query Log. SIGIR Forum 33(1), 6 -12. Retrieved 11 October 2000 from the World Wide Web: /sigir/forum/F99/Silverstein.pdf

Size of the web: A Dynamic Essay For a Dynamic Medium Retrieved 1 October 2000 from the World Wide Web: http://censorware.org/web_size/

Spink, A. & Xu, J. L. (2000). Selected Results from a Large Study of Web Searching: the Excite Study. Information Research, 6(1). Retrieved 16 October 2000 from the World Wide Web:

http://www.shef.ac.uk/~is/infres/paper90.html

Volk, Martin (2000). Scaling up. Using the WWW to resolve PP attachment ambiguities.

Proceedings of Konvens-2000, Sprachkommunikation, Ilmenau, VDE Verlag, 151-156. Retrieved 12 October 2001 from the World Wide Web:

http://www.ifi.unizh.ch/cl/volk/papers/Konvens2000_Ilmenau.pdf

Volk, Martin (2001). Exploiting the WWW as a corpus to resolve PP attachment ambiguities. Retrieved 12 October 2001 from the World Wide Web:

http://www.ifi.unizh.ch/cl/volk/papers/Lancaster_2001.pdf

Whistle, J. (1999). Concordancing with Students Using an ‘‘Off-theWeb’’ Corpus. ReCALL 11 (2):74-80.

Zanettin, Federico. (2001a) Swimming in Words: Corpora, Translation and Language Learning. In Aston 177-197.

Zanettin, Federico. (2001b) DIY Corpora: the WWW and the Translator. Training the Language

Services Provider for the New Millennium, Porto, Portugal, 25-26 May 2001. Retrieved 22 January 2002 from the World Wide Web: /DIYcorpora.htm