Related Linguistic Distractors

Linguistic distractors can again be subdivided into several types but let us first look at a characteristic of our LRL. Firstly, there are languages, which are part of a larger language family - LRLf. Then, there are language isolates LRLi, which have no family, a true isolate doesn't even have closely related varieties. Finally, there are mixed languages LRLm known as Pidgins and Creoles, which often exhibit some grammatical features of their substrate while adhering largely to the lexicon of the superstrate. Depending on this characteristic, a language may or may not have closely related sister languages which can be some of the most distracting distractors. Naturally LRLms are confusable with other LRLms of the same or even a closely related superstrate. The border between language and dialect is fuzzy and a matter of definition. As the famous saying attributed to Weinreich reminds us A language is a dialect with an army and a navy. Thus, searching for an LRL, one should be able to clearly delimit the variety against others or include different standards then delimiting the varieties from others in the same continuum. <TAGS for=”LRLf, LRLm”> If one has achieved that, language family trees (classificatory systems) can help us immediately identify candidate distractors. Likewise, a matter of decision is the upper node at which one considers a family a family (would we take Indo-European [for our purposes surely too large a unit], Indo-Iranian or Iranian as the root for the family of Balochi?). Generally, a closer root for the family with an uncontroversial branching point and consistent with general linguistic typology appears advisable. Online sources for classifications are sites such as Ethnologue, Glottologue, WALS, to name but a few. Here, one may search for the closest languages and then test how close they are. For them to be formidable distractors, they should have the same writing system (otherwise they can be skipped) and very few differences from our target LRL. A second factor is the size or status of the sister language, if it is a large language, one may want to put it on the distractor list even if a little less related than another closer LRL sister. There is no perfect strategy of defining a threshold for a distractor sister, but a simple heuristic may be useful: if the percentage of exact orthographic overlap (and/or overlap in trigrams) between two as large and clean as possible word lists (the distance may be weighed by frequency) considerably exceeds the average accidental overlap with the largest unrelated languages of the web, one should include it. One may also consider overlap in grammatical features (WALS) or phoneme inventories and graphemic systems additionally.

The exact threshold may vary per scenario, but this is not tragic, since one always has the possibility to extend the distractor list later. Again, generally one may rather include too many than too few distractors. </TAGS for=”LRLf, LRLm”><TAGS for=”LRLm”> A side-note is that one may want to consider an LRLm attached to the root node of the superstrate language's family and likewise all other LRLm with the same superstrate. The test on the percentages of overlap may be run as described.</TAGS for=”LRLm”> Historical stages (earlier distinguished varieties) of our target LRL may always be a related linguistic distractor.

Unrelated Linguistic Distractors

After having listed the closest related distractors, one may proceed to unrelated linguistic distractors. Here, the linguistic processes of loaning and borrowing is the most important one for the potential of becoming a distractor. Consequently, Sprachbund and regional proximity are good estimators for candidates, but also former colonial languages. Loanwords may constitute a considerable part of a language's vocabulary. Since the direction of loaning often reflects power, it implies a larger part of words from bigger languages. The opposite direction however is equally relevant, words for local plants, meals, industrial products etc. which the larger language has absorbed. What we are looking for with non-related distractors is languages which are either the (ultimate or intermediate) source of many (loan) words of our target LRL or which have themselves borrowed a considerable number of terms from our target LRL or various mixtures of both. Loans may be mediated (for instance, directly loaned from French, but originating in English). If those terms preserve their original orthography (at least in some of the terms or to large degrees), then they can lead to serious distraction in queries. Especially if the contact is so intense, that even some (frequent) function words such as discourse markers (like amma in some oriental languages) have been loaned. It might also matter if one looks more for formal or informal language. All LRLs usually loan and borrow, albeit for different reasons and in different ways; LRLms maintain some words of their substrate for instance. Again, the question when to include a non-related LRL cannot be answered globally, but some thoughts may facilitate the decision. The larger the distractor is (especially English, which loans into many languages, likewise French, Spanish in South America, Russian in the East, Mandarin Chinese in the Far East) the more likely a distractor it will be since these languages are typically associated with plenty of web content. On the other hand, the level of perseverance of original orthography (in the original alphabet) is also very important. While Japanese has many English loanwords, rendering glass as グラス is hardly producing any English content (aided by the fact that in the Japanese rendering there is no more distinction between r and l, so the same transcription could also stand for the word grass which is however usually not used as an English loan). Here, one may thus measure how many words in a sufficiently large wordlist of the target LRL (if one has one) are English words (intersecting with a big English word list or an online corpus query API). In this case, it may also matter which words are such loans: are they all very low frequency terms, so that there is no big danger excluding target LRL content when excluding content containing them? Based on these characteristics and individual ones, one may decide to list a language as distractor language.

Paralinguistic Distractors

<TAG for=”mostly Latin alphabet based LRLs”>

Finally, there are paralinguistic distractors as we learned during our course. It is hard to preview which LRL will attract which paralinguistic distractors, but based on the spread of the Latin alphabet and the fact that most programming code and mark-up code etc. is written in Latin alphabet, it is clear that the likelihood for paralinguistic distractors is clearly much higher for languages written in the Latin alphabet. Now, a paralinguistic distractor should be a consistent unit if one is to later find strategies to exclude content of this type. We found some possibilities:

1. faulty OCR: Some websites provide text derived from OCR which was neither post-corrected, nor very accurate. In fact, there are tremendous examples, where there seem to be many more wrong characters than correct ones; and the numbers of such documents seem to be on the way to being Big Data. OCR-errors are often systematic, that is the same letter sequence in the same font tends to be misread as the same wrong sequence each time encountered. This leads to the systematic distortion of the original language of the OCR (or if the faulty parts make more than 50% of what is being displayed some language like gibberish). If one now thinks that language change likewise distorts an actual variety to form another one, it becomes intuitively clear that the distorted OCR of a sister language or even an unrelated language can accidentally (if in the process some very frequent function words and some fewer, longer content words are produced) resemble another, if we are unlucky our target LRL. So, faulty OCR is always a bullet point on the list of distractors.

2. glossolalia and pseudo-X: Some comedians imitate languages by producing fake sequences incorporating many characteristic sounds and maybe some words of certain languages. To give a written paraphrase, what language would you guess the following sequence to be Das Gehortung warrende Humpelkatz rimpelt in Ratzfatz? Does this look like German to you? Actually, as on the 10th of October 2019, Google translate would also classify the sentence as German and it probably should. Only two words and two morphemes are German: Das is the German article, Ratzfatz, better ratz-fatz is an onomatopoetic meaning immediately, the word Humpelkatz could be analyzed as limping cat. We hope this made clear that such imitation works also on a written basis. If we take this to be a sequence uttered or written by a comedian to humorously emphasize a rough sounding aspect of German maybe based on its affricates and some other phonotactic properties, then with some actual words produced by accident or intent such a sequence would make a formidable distractor (as also the language classification systems witness). Besides comedians, psycholinguists may use pseudo-words (with a certain permissible orthographic and phonological structure) or even phrases for experiments and some religious activities include speaking in tongues (or glossolalia) although this is seldom written down (but could be as auto generated and separated subtitles to a YouTube video for instance). Despite being much less abundant than faulty OCR, one would want to avoid such content in a serious corpus, which is why it is to be considered a possible, yet rare distractor.

3. encrypted text and orthographic plays: Since antiquity ( ) people have used codes to transmit messages which should not be intercepted. Since then an arms-race between cryptography (the process of encoding messages) and cryptoanalytics (the process of decoding) has taken place and seen the development of many different techniques. Some of these codes may produce text looking like our LRL or worse but extremely unlikely senseless text in our LRL (here, one may remember Chomskys famous sentence: Green colorless ideas sleep furiously, which is asemantical but in principle not agrammatical). Or one could simply write a Latin alphabet-based language in a Semitic style omitting short vowels, which then may accidentally look like another language. Also, one could invent a new orthography (for instance a simpler one for French, which then could resemble a French based creole while not being one). This distractor should also be rather rare, but there are cases where it could be a serious distractor.

4. machine code: programs and men produce all kinds of codes for instance in transmission or machine2machine communication. Again, accidentally, these could at least in parts look like sequences of words in the target LRL and thus become some kind of distractor.

5. abbreviations and acronyms: heavily abbreviated text may also lead to a completely different linguistic appearance. Since there are languages, where an abbreviation must not be marked as such (as with a dot in English), these cases could lead to another distractor. Especially in short message communication (where space is also money) innovative abbreviations have become a substandard (lol, 4u, 2b or not 2b, etc.).

Apart from these, there might be other paralinguistic distractors such as lists of names, but we were neither aware of others, nor did we witness them during the course. The extent to which such paralinguistic distractors play a role is largely dependent on the target language and on connected random factors such as the frequency of certain easily misOCRed ngrams such as ni (for m). It may therefore seem advisable to check these paralinguistic distractors one by one. In a tiny experiment, we took an abstract of a paper in Indonesian, reduced the image size and quality by a Ghostscript command, took a screenshot of it and then let tesseract OCR with English extract text from the resulting png. On the Textfile we ran fastText and found that about only 40% of lines had been identified as Indonesian, 33% as English, 15% as Malay and one line each as Finnish, Russian, Hungarian, Ukrainian and Catalan. So, in all, more than half of the lines were misclassified, while especially English and Malay, one close sister language and a contact language had larger proportions. Badly OCRed sister languages may be especially prone to become a distractor.

</TAG for=”mostly Latin alphabet based LRLs”>

With a list of distractors at hand, only two things are missing before starting the search.

URL Guessing

Explained in a simplified way, to access a certain content, a user can type the address in a Browser and thereby asks the server registered under this name for data which it sends back. This is also the first mechanism by which to find a particular web page: direct address input. Now, if one does not know a page, one can guess names. LRL communities might have names such as language-name.country-code-top-level-domain, for instance for German deutsch.de. This may or may not work, in case it doesn't either content in another language is accessed there, a provider has reserved the name and advertises it there or the address cannot be reached. This way of obtaining LRL content is both cumbersome and has a low probability of success. This is because there are numerous possible combinations of subdomains, top-level domains and hostname letters to compose a URL (in various alphabets and ways such as Punycode) and since LRL content may only be present on some files (thus subadresses on a certain server) this again opens many possibilities.

Link-Hopping, Surfing

The second possibility to search the internet is via so called hyperlinks. With any starting point in the internet, one can search and follow the links on that page and go to others. For LRLs this strategy is especially useful when pages are interlinked. There has been a lot of research on the topology of the internet when symbolized by a graph where pages are nodes and hyperlinks edges. One famous contribution ( ) assumes a bow-tie model. Another famous topological property could be the small world property ( ) where few pages function as linking hubs which connect many groups of loosely interlinked pages (which rarely link to pages outside their groups). Finally in search engine research so-called hubs and authorities ( ) are being distinguished where hubs by and large correspond topologically to those in small-world graphs whilst authorities are pages which provide high quality content. Should the webpages of our target LRL be located in disconnected components of the WWW, meaning such groups of pages which are not interlinked with the main core of the internet, this could make them considerably harder to find. Firstly, because then there would be no way to find them through hyperlink hopping (surfing) from pages in the core. Secondly, search engines - the third way of searching the web - would be somewhat less likely to index these pages and thus they may not be retrievable through them. However, we found during the course that most pages in LRLs (of various sizes) were usually somehow connected to the core. Furthermore, we realized that at least some proportion of the disconnected components could be fresh pages, which have not yet been filled with content. Thus at least some disconnected pages are irrelevant. Yet, we cannot exclude that some pages in LRLs are found in disconnected components and as a consequence are only accessible through URL guessing or informants.

Search Engines

The third and presumably most well-known way of searching content on the web is through search engines. A description of the process is problematic for various reasons, mainly because the exact functioning of search engines is their business secret and subject to change at least of parameters at least every now and then (for otherwise people would manipulate webpages in absurd ways in order to come up on top of certain customer-loaded searches). In a nutshell, a search engine periodically sifts through the web (or portions deemed relevant) and generates (or updates) a so-called index, that is a database where addresses are stored along with some features. Now, because this is a secret it is not clear what these features are, but certain words (or ngrams) and their frequencies as well as the number and sources or targets of incoming and outgoing links should be involved almost certainly. On the basis of these indices, search engine queries which the user sends to the web interface of the search engine are being answered. Thus, for each query a search engine receives, an algorithm produces those result pages which according to the features of the index are most relevant to that query. Thus, one does usually not search the internet when using a search engine but only those portions of it known and relevant to the search engine. Generally, large search engines such as above all Google (and Bing) seem to have the largest indices. Further below, we describe how to retrieve LRL documents through search engines by composing linguistically informed queries. A side-note is that semantic web technologies (and search engines) could play an ever more important role if considering projects such as Babelnet which connect languages, some of them small, through semantic web technology. For instance, could a tag for a small language’s name be connected in many meaningful ways to different types of content for it. Meta-linguistic ontologies such as OLiA can already be used to obtain features and characteristics of a potential target language and are used also for smaller languages (such as Dzongkha or Yucatec Maya or Fon) and in interlinked lexical resources.

Dark Webs

There are alternative webs which some call darkwebs. The onion-web is the largest of those and one of the earliest ( ). It was intended as a place where political dissidents could voice their opinions if officially oppressed and actually parts of this web are used today for such purposes. Facebook also has a presence there. However, since in these kinds of webs, users who host or view pages are rather anonymous as secured through the technical underpinnings of the system, which at the same time make it slower than the WWW, much criminal activity is also to be found there. Since LRLs can be oppressed, content may be located in a darkweb. The way to find such content is through darkweb search engines or Hidden Wiki lists with thematically ordered entry-points, and through surfing. URL-guessing is rather impossible, since darkweb URLs are usually long codes, not registered and intentionally chosen as in the WWW (or Surface Web as darkweb surfers may call it) and seldomly meaningful. Legally and ethically, the use of texts found on a darkweb may present a greater challenge than those from the Surface Web. One has to think about their potentially threatening contents, the risk one may put their authors or oneself to using them, the unclaimed or unclear copyright situation and quotability (reproducibility) since contents appear and disappear at high rates in this segment of the web.

Social Media

One must mention Social Media such as Facebook, VKontakte and Twitter, which come with their own search engines and where accounts prompt users into areas, which the bigger search engines are not supposed to index or offer publicly. For these reasons, some content may only be accessible while logged in into an account in a Social Network and consequently looking for content in LRLs may involve Social Media Search. While as of summer 2019, Facebook closed its semantic graph search where one could for instance query all accounts of Hindi-speakers living in Canada and working as teachers, the current search functionalities typically still support searches using individual characteristics such as the school users attended (provided they input this information). However, the search is no more semantic but now combines query terms and is no more fundamentally different from conventional search engines.

Deep Web

A last point to mention is the so called Deep Web, which refers to content which is generated dynamically from the contents of a database only in the moment, the user onsite submits a form or performs a related interactive activity. Such content cannot be indexed by larger search engines unless they mimic the onsite query behavior which would clearly be unfeasible, given for instance all possible queries on a train company’s website for the next trains at every point in time, as well as arrival and departures times at every station. It has been estimated that considerable portions of the WWW are potentially Deep Web contents and they comprise also genuine LRL content.

0. Looking for resources on known sites and BootCatting

The first place to look for content may be one of the larger sites and projects which have focused on LRLs, such as the above-mentioned ones (DoBeS, An Crúbadán, LCC, etc.). However, their content should be thoroughly checked for noise.

Secondly, using BootCat and feeding it with a limited number of terms may be a good start. The results can be URLs or even directly a corpus which BootCat draws from those. The result must then be checked manually and purified, which can in the ideal case provide a larger basis for manual queries or a second and third BootCat round. For some LRLs however, BootCat may not be able to retrieve valid content.

1. Single Term Querying

For a first approach to querying search engines we can take single query terms and compose an evaluative spreadsheet where we note for instance a) how many of the first 10 pages returned for a certain query have been in the target language and b) how many results have been returned. We can then annotate different statistical or semantic or other linguistic properties of query terms. We query some very frequent function words (avoiding terms accidentally or by loaning overlapping with one of the big languages of the WWW or our distractors), some intermediately frequent terms, some content words (even if our wordlist features no explicit frequency information, we can, by universality, be quite certain that a not too unusual and short function word is very frequent, whereas a longer content word, which is not general such as thing, animal, machine should be infrequent). We can try words from different genres, different registers, regions etc. Afterwards, we can look at the list and try to draw inferences on the effect of the assumed characteristics on the evaluation (a and b or a [weighted] product of both). Some properties are unsurprising and can be hypothesized a priori such as function words return more results than content words or that specialized terms return less than more general ones which at times will be used as anaphora for the former. However, the current composition of content of our target LRL on the web is what has to be characterized. Guiding questions such as does a considerable proportion of the assumed content contain pages with folklore content could be tested by using concurrent vocabulary and seeing if result numbers or precision increase. These questions may be very individually dependent on the current time, situation and other circumstances of the LRL community at hand.

2. Multiple Term Querying

After having evaluated single terms, where some may have been found useful, the next proposed step is in combining terms. Generally, the first term can be thought of as constituting a certain amount of results and each subsequent term as eliciting a subset of the previous results, so that in principle on average the number of results decreases with the number of query terms. A function word is then a good first query term but so-called stop words should be avoided.

Stop words are usually very frequent function words which big search engines simply ignore if they appear in queries. The reason is that those would simply elicit too many documents as they appear in quasi all of them (think for instance of the English article). This touches upon another issue, the settings for search. For stop words to be relevant, a search engine must have a list of them and in turn a setting for searches in the target language. Other language dependent settings may also exist and can crucially influence the results in ways again hidden in the business secret.

Often queries contain terms, which have not been indexed (that is they may be contained on the target website but not in the search engine index as a feature) or are simply not present in any target site. Now, it is the secret search engine algorithm which decides how to prioritize imperfect results. For instance, if you query 5 terms and no site is found containing all 5 in the index, should a result page which has only 3 of 5 terms but has a very good hub-score (or authority score) be prioritized over one which has 4 out of 5 terms but appears less connected and important?

This is an additional process to be thought of when querying more than one term. Apart from this, one can make another table (spreadsheet) and start combining terms systematically for their characteristics. Now, the possibilities for combination are manifold, combining a function word with a content word, a general with a specialized word etc. This allows hypotheses to become more flexible. Likewise, interpretation becomes more complex. The benefit is however, that we can get a better idea of which genres, registers etc. are especially fruitful in our scenario. What we found during the course was that if one combines content words which are too unrelated, since they do not occur together naturally, dictionaries or word lists may surface.

3. Queries with Operators

Operators are characters or character sequences which the algorithm of a search engine will treat in a predefined way when seeing them in a query. Among the most well-known and used operators may be the Google's quotation marks and minus operators, where the operator forces the search to look for the occurrence of a sequence which may contain spaces as such (“an apple searches pages which contain exactly this phrase, not such which contain apple somewhere, an being an ignored stop word). The - operator excludes. If we formulate a query and add - terms, we thereby exclude pages which contain the minusterm. This can be used monolingually for word sense disambiguation in queries, for instance searching for bank being interested in the use of the term as riverbank, one may formulate

bank river water -financial -money -business

in order to exclude pages where the dominant sense of the word bank which one can naturally not exclude from this query is the financial institution. In parallel, we can exclude other languages (big WWW languages and our distractors). Since minusterms apply only if any page at all is present in the result set, which has them, one can theoretically add a large number of them to any query. Note that search engines limit the number of maximally allowed search terms (either in tokens or as a certain bit encoding size).

4 General Remarks

Compiling a corpus from internet sources is work-intense. At least if one aims at clean corpora with very low amounts of noise or none at all. Some sections of the internet such as member only content or certain content within social media platforms or sections of the internet not indexed by a search engine are obviously not retrievable via a/that search engine. Thus, manual content search can always produce additional content for target LRLs as long as such content exists. In doing so, we found a profound knowledge of technical and linguistic underpinnings of documents on the web useful. For instance, the search for certain document types (txt, pdf, etc.) partly benefitted from different query term selection and composition strategies (in opposition to html content, we must not expect menu-items in plaintext or pdfs for instance). The personalization options of the search engine (and so does the filter bubble) modified the results partly crucially. Generally, we found search engines to be richer in content for languages spoken on the territories close to their core language than the others (English for Google, Russian for Yandex, Mandarin Chinese for Baidu). Some content was blocked from certain regions.

Nogai (Cemre Koc)

My scenario was about the search for Nogai, a Turkic language spoken by approximately 70,000 speakers in southwestern European Russia (Caucasus). The first step of my search was downloading a Nogai wordlist from crubadan.org which contained over 200 words and word bigrams in .txt-format. Then, I formulated queries by gathering and combining random words out of the downloaded txt-file. At first, I used multiple search engines but later changed only to Google which had given me 7 results in the first run (apparently Nogai newspapers and poems) whereas the others none. However, I noticed that the retrieved newspapers and poems could easily be written in other Turkic languages of their own language branch or neighboring languages like Kumyk, Karachi-Balkar or Bashkir. The main problem was to distinguish newspaper articles and poems as PDFs between Kumyk newspaper (2), Karachi-Balkar and Nogai (3), since their writing systems and lexica significantly overlap. To tackle such a low resource challenge facing the large similarity of Turkic languages in their language family branches ( : 81), it was necessary to collect unique linguistic characteristics such as affixes or cognates in their forms which made it easier to discern Nogai from the other languages (distractors) and to discard unwanted content (noise). The exclusion o f words where the same form appeared in a distractor language (for example уьй (house)) lead to higher numbers of results in the Nogai language (first page from 1/10 to 9/10 hits). Using combinations of unique Nogai words also provided an overall higher number of hits. While мылтык (riffle) has zero hits on the first page in Google the combination with the words сары тамбыз (august) provides three hits, which include a novel of a Nogai writer (Isa Kapaev). Furthermore, it is important to mention that the co mbination of words of one category or topic was advisable as then the number of hits increased. Moreover, the filetype operator used for searching PDFs only influenced the results positively. In summary, I found 30 newspapers and over 30 children’s books in the Nogai language in PDF format by using combinations of unique Nogai words and a filetype PDF operator which is a considerable corpus for such a small language and as such for Nogai to the best of our knowledge unprecedented.

Maori (Marc D. Rahn)

My search scenario was restricted to pdfs and concerned with Maori, a Polynesian language spoken by roughly 160.000 people in New-Zealand. To start, I looked for frequency based Maori wordlists via Google and found one . From this list, I selected a smaller subset for manual work based on the following criteria:

1. The words should be high ranking, that is frequent, so as to elicit as much content as possible.

2. Since shorter words have a higher chance of accidental overlap with another language, I preferred slightly longer words (slightly less frequent). The most frequent word in Maori, "Te" ("The") was not very suitable, as it could easily appear in other languages (which it does, for example in French "you", Dative/Accusative).

3. I also preferred words with a peculiar arrangement of letters, as well as words with diacritic symbols, further reducing the chance of a random match with another language.

For instance, the sixth most frequent word "ngā", is a perfect candidate: It is very common (meaning "and"), yet not too short, could come up in any kind of text and features a peculiar combination of letters and even a diacritic symbol.

The next step in my preparation was to think of distracting languages, especially English. The Web is full of English content, especially for a country like New Zealand where it is the main language. Additionally, I wanted to avoid mixed texts and teaching resources. Therefore, I picked the words "the, and, with" to use for the exclusion of English content (blacklist terms). Despite the fact that those terms are so-called stop words, that is they are ignored by Google when searching, when excluding them they are apparently not. Lastly, as I had discovered with a few similar queries for other languages beforehand, a lot of linguistic resources do come up, both educational and scholarly. To exclude these results, I excluded the words "language" and "status" from all of my results. To then obtain my actual search results, I picked a number of terms on the Maori wordlist that fit my cr iteria (ngā, kua, rā, haere and tētahi) and started searching for them one by one, using the blacklist terms at the same time. I used Google for these queries, because it simply is the biggest search engine, and there is, to my knowledge, no comparable spe cialized search engine for the region of New Zealand or Polynesia. In addition, Google allows for a multitude of search settings: I changed the search language to English, the region to "New Zealand" and turned off personalized results. For example, one of my search queries would look like this:

ngā -the -and -with -language -status filetype:pdf

For a first impression, I chose to estimate the percentage of correct findings (PDF's in only Maori) on the first result page. A count of total results was performed manually by counting all result pages (which is more accurate than the results estimate).

The query given above returned 54 results in total (Last checked: 21.10.2019). Of the first 10 results, 9 were completely in Maori, one was a short list mainly in Maori but with English words appearing, and one was a broken link. The other queries and their results are given here below, see .

It is relatively straightforward to see that a combination of the techniques described above yields the best results. When, however, the techniques are used one by one, the selection of search terms and blacklist terms becomes more significant. Especially when stop words cannot be used (as with larger languages), when the region cannot be filtered for, or when both applies, a combination of terms can still lead to good results. In contrast, a single word query without blacklisted terms is likely to yield quite noisy results even if filtering for the right region. However, longer words with diacritics have led to cleaner results in all circumstances within this scenario.

In both scenarios, pdfs played a crucial role, but this is of course not necessarily so. Filetypes however may not only play an important role for the text type one may find (with presentation formats barely promising connected text for instance). The way in which one chooses terms may also differ depending on the filetype or type of web page, where informal language may for instance be associated with blog entries and comments more than with formal documents such as laws and constitutions. The two scenarios showed that the internet can provide texts even for small languages if one knows well how to find and distinguish them. They have provided many approaches for factors, both linguistic and paralinguistic, which play a role in manually querying content in LRLs. We also analyzed the lexical overlap between the languages involved and found that trees generated from the similarity matrices of lexical overlap in one case roughly reflected genealogy. Compare also who found that language genealogy and language contact (often correlating well with geographical proximity) influence similarity (on various linguistic levels with a hint towards the lexicon). shows a Neighbor Joining Tree from lexical overlap in our corpus for Maori and distractors (mainly Wikipedias, filtered for most frequent English noise). It coincides almost perfectly with linguistic genealogy. For Nogai, the findings differed (with more assumed noise placing Russian in the middle). While Russian is unrelated to Nogai, Bahasa Indonesia is a distant relative of Maori, English not. Despite that, both were clearly distinguishable in the maximal overlap they displayed with any of the other languages which suggests that it could be the case, that only relatively close sister languages play a crucial role as immediate distractors (which the Nogai corpus data roughly supports) whereas distant cousins and contact languages can have a similar degree of lexical similarity, lower than the sister languages.

n-grams are sequences of n characters or words which occur in sequence in texts, where n stands for a number.

The common use of loanword and borrowing in English is at least partly synonymous, see Oxford Advanced Learner’s Dictionary (22.10.2019).

Here, we would like to clarify that if one searches for transliterations, too or for multiple renderings if a language has more than one principle writing system (such as Serbian) the strategy is different. To increase clarity of our method, we assume however that the search scenario involves only one (main) writing system. If your scenario involves more, you can break it down into one scenario per writing system.

Analogously, one may also guess IPs, which browsers also understand, but in guessing which IPs link to content in a certain LRL, only non-linguistic clues are useful for which we have no good description.

This is just a quickly retrieved example, we do not want to advertize or promote contents hosted on that site.

We do not treat FTP and SMTP protocols here.

This refers also to pages, which we knew and found not by search engine queries or surfing.

Actually, people try this through reverse-engineering. The concurrent field is Search Engine Optimization.

Note, that the number-of-results estimate which some search engines provide is based on a quick precomputation and can deviate from the actual number of results.

Likewise, non-stop word-function words appear unuseful towards the end of a query.

https://tereomaori.tki.org.nz/content/download/2031/11466/file/1000 frequent words of Māori- in frequency order.doc

Compare also the classifier, we published on https://github.com/ArminHoenen/URLCoFi

http://www.cirp.gal/corga/

https://invokeit.wordpress.com/frequency-word-lists/; we also intersected all 59 wordlists from 2016 from this project and found through generation of a Neighbour Joining Tree from a distance matrix that the data by and large reflected genealogical relationships if the alphabets were the same and found corroboration that Vietnamese and Indonesian had higher levels of noise.