McEnery and Wilson (2001: 29) state that "in principle, any collection of more than one text can be called a corpus". Some authors adhere to this broad interpretation as referring to every text collection. The interpretation enables the definition of "corpus" to be expanded to encompass the entire Web (Kilgarriff and Grefenstette 2003) or, for instance, Webbased text collections of particular languages (Scannell 2007).

However, as McEnery and Wilson (2001: 29) rightly note, "in the context of modern linguistics" the term tends to be used in a more narrow sense. They consider four specific connotations of "corpus": sampling and representativeness, finite size, machinereadable form, and standard reference, i.e., wide availability to its potential users. By "sampling and representativeness" they refer to filling the corpus with "samples of a broad range of different authors and genres which, when taken together, may be considered to 'average out' and provide a reasonably accurate picture of the entire language population in which we are interested" (ibid.: 30). Thus, the narrow interpretation of a linguistic corpus can be phrased in the following way: "a finite-size body of machine-readable text, sampled in order to be maximally representative of the language variety under consideration" (ibid.: 32).

Other authors sometimes use other criteria to describe a "prototypical corpus". Gries and Berez (2015), for instance, enumerate four characteristics that "jointly define a prototypical corpus", which are slightly different from McEnery and Wilson's (2001) definition. Besides machine readability and representativeness, a "corpus is meant to be balanced, which means that the sizes of the subsamples (of speakers, registers, varieties) are proportional to the proportions of such speakers, registers, varieties, etc. in the population the corpus is meant to represent". Furthermore, a "corpus contains data from natural communicative settings, which means that at the time the language data in the corpus were produced, they were not produced solely for the purpose of being entered into a corpus, and/or that the production of the language data was as untainted by the collection of those data as possible". Gries and Berez (2015) define linguistic corpus as "a category that contains exemplars that are prototypical by virtue of exhibiting several widely accepted characteristics, but that also contains many exemplars that are related to the prototype or, less directly, to other exemplars of the category by family resemblance links".

Consequently, there is a continuum of particular exemplars, ranging from the more prototypical, which satisfy all the characteristics, to the less prototypical. The British National Corpus can be considered prototypical: it is machine readable, of finite size, representative, balanced, widely available and contains data from natural communicative settings. It is important to note the possibility of deviations from the prototype. That is to say that if a specific text collection does not comply with one or more of the criteria, such as with collections in an understudied language, this does not automatically mean that it should not be treated as a corpus.

There are no universally accepted criteria on how to delimit understudied languages from well-studied ones. For the purposes of this paper, understudied languages are equated with under-resourced ones, which may theoretically be inaccurate but in practice seems fair enough. The point is that there is (or was until recently) no continuous text production in these languages, which have no established literary tradition, although they may have a recent written tradition.

Szymanski (2011: 1) defines "resourcepoor" languages as those that "lack any significant digital presence¹". This study will not assume that materials should necessarily be digitized, because in principle every non-digital unit of text can technically be converted into a digital one, with more or less effort and expense. Szymanski (2011: 8) also notes that "resource-poor languages are not necessarily endangered, under-studied, or minority languages (although they may be)". Probably, under-resourced languages are still always understudied, because without significant resources it is hard to imagine a particular language producing many scientific works. The opposite seems also to be true, because available material almost always attracts scientific attention.

Maxwell and Hughes (2006: 30) use the term "lower-density" languages to refer to under-resourced languages. They draw a distinction between "high-density", "mediumdensity" and "lower-density" languages, attributing to the latter the absolute majority of the world's languages.

For a few languages of the world (such as English, Chinese and Modern Standard Arabic, and a few Western European languages), resources are abundant; these are the high-density Languages. For a few more languages (other European languages, for the most part), resources are, if not exactly abundant, at least existent, and growing; these may be considered medium-density languages. Together, high-density and medium-density languages account for perhaps 20 or 30 languages, although of course the boundaries are arbitrary. For all other languages, resources are scarce. (Maxwell and Hughes 2006: 30)

Assamese is an Indo-Iranian language spoken by almost 13 million people in India².

Written Assamese makes use of Bengali script. The Assamese corpus (https://cqpweb.lancs. ac.uk/ asm_v2, accessed 07-01-2016) was originally gathered by the Institute of Applied Language Sciences at Bhubaneshwar and then integrated in the scope of the EMILLE (Enabling Minority Language Engineering) project at Lancaster University and Sheffield University³.

The Assamese corpus contains about three million tokens from 1,191 texts in total. All texts are divided into categories depending on the topic (e.g., business, education, mathematics), and users can exclude some of these categories by specifying a subcorpus. This corpus does not provide any kind of wordlevel linguistic annotation⁴. Therefore, the unique searchable items in the corpus are exact word forms or their parts. The developers of the interface do not provide a virtual keyboard, which could have been very useful taking into account the specific script.

Ndebele is a language belonging to the Bantu group of the Niger-Congo macrofamily. It is spoken primarily in Zimbabwe by approximately 1.5 million speakers. The corpus of written and spoken Ndebele (http:// www.edd.uio.no/allex/corpus/africanlang.html, accessed 03-01-2016) was developed within the ALLEX (African Languages Lexicon) Project. This corpus contains 691,268 tokens and is not annotated. A user has very few search options: one can use some regular expressions in the query and modify the extended context from 30 up to 1,000 symbols.

Hadebe (2002: 167) mentions that "the corpus consists of both oral and written texts, all transcribed and converted into machinereadable texts". The approximate percentage is 80/20% for written and oral parts, respectively. The oral part of the corpus violates the principle of "natural communicative settings" (see Section 1.1) because "most of the oral material was collected by means of structured and unstructured interviews" (ibid.: 164), specially for inclusion in the corpus. For more information on the process of collecting and elaborating corpus material see the description of the Ndebele corpus in Hadebe (2002).

The Ossetic language belongs to the Iranian branch of the Indo-Iranian subgroup of the Indo-European family. It is spoken by approximately 550,000 people in the Russian Federation and in Georgia.

The written corpus of the Ossetic language (http://corpus.ossetic-studies.org, accessed 02-01-2016) comprises more than 11 million tokens. It is a literary language corpus because it is basically formed of texts from literary journals as well as some works by Ossetic writers of the 20th century. The complete list of texts included in the corpus is provided on the corpus webpage. The texts belong to the Iron dialect of Ossetic, which is the basis for standard Ossetic.

This corpus is annotated; it includes grammatical information about tokens, as well as their translation. The annotation was made automatically and not disambiguated. The main merit of the Ossetic corpus is the powerful search engine⁵, together with the user-friendly interface. One can search by lexeme, word form, translation, or by a particular set of grammatical features. One can include more than one token in the query and indicate distance between them. There are also some options to specify a subcorpus based on genre, period, authors and titles of documents, etc. Finally, a virtual keyboard is provided for non-standard symbols.

The authors of the corpus also provide some useful facilities to process the search results. It is possible to choose the output characters (Cyrillic/transliteration), the output layout (e.g., with or without morphemic annotation), the number of sentences in the expanded context, etc. The user can also sort the list of results by different parameters, including, for instance, the title/year of the document or preceding word form.

Bambara is a Mande language which belongs to the Niger-Congo macro-family. This language is not endangered (though it is still under-resourced), since it is spoken by more than 10 million people in Western Africa, generally in Mali.

The referential corpus of Bambara (http:// cormand.huma-num.fr, accessed 03-01-2016) is annotated and contains both disambiguated and non-disambiguated subcorpora. As of October 2015, the total volume of the corpus amounts to almost three million tokens, while the disambiguated part is considerably smaller: 426,813 tokens. The corpus of Bambara is formed by texts from different sources that represent different genres and dialect zones. The interface of the website allows a user to specify a subcorpus in order to exclude some documents from the search.

The Bambara corpus allows users to search by lemmas, word forms, phrases, symbols, to specify part of speech and to set a specific context to the left, to the right or to the both sides. It is possible to visualize the results by plotting a frequency diagram.

Kalmyk is Mongolic language spoken in the southern part of the Russian Federation. According to Ethnologue (Lewis et al. 2015), there are 80,500 speakers of Kalmyk. The webpage of the Kalmyk corpus (http:// kalmcorpora.ru, accessed 03-01-2016) reports that the situation of endangerment is even graver than it appears, since of these speakers no more than 5,000 are fluent.

As of May 2015, the Kalmyk corpus comprises the total of 8,691,671 words. The corpus is annotated morphologically and semantically, but not disambiguated. Users can use annotation in search queries. They also can limit the entire search to some particular genres, authors or text types (e.g., oral, folklore, poetic).

At this stage, the Kalmyk corpus does not provide much numerical or statistical information. The presentation of search results includes neither numbering nor the total sentences found. The sections about statistics and frequencies on the corpus webpage are empty, although some graphic representations of statistical data about lemmas' and word forms' frequencies can be accessed via the link "Graphs".

Kukanova (2011) expresses an optimistic view of the possibility of obtaining a representative and balanced Kalmyk corpus in future. However, due to the general insufficiency and inadequacy of available text materials, it is difficult to share this opinion.

The corpora of understudied languages differ in numerous parameters. The composition of such corpora does not depend on the ideal theoretical conception of corpus structure, but more fundamentally on the quality, quantity and diversity of available texts in a particular language. This drastically affects the volume of the corpus and the coverage of different registers, genres and dialects. Corpus creators can solve the problem of coverage by intentionally provoking speakers to produce particular kinds of text that are lacking, as with the oral part of the Ndebele corpus (see Section 2.1). Nevertheless, one should be aware that in these cases the texts do not come from natural communication, and they may therefore be inappropriate for inclusion in a corpus.

Table 2 presents a short comparison of the five corpora examined in this Section.

Table 2:
Basic characteristics of the different corpora of understudied languages.

The characteristics 4-8 in Table 2 do not deal with the basic features of a linguistic corpus discussed in Section 1.1. Annotation, disambiguation, variety of search options, possibility of creation of a subcorpus, special facilities to process the results - all these characteristics make a corpus more useful and more suitable for a wide range of research questions, but none of them can transform a simple raw text collection into a linguistic corpus.

All the examined corpora satisfy the criteria of machine readability and finite size (Table 2, parameters 1 and 2). But none of them satisfies the criteria of representativeness and balance. Furthermore, only some authors of documentary corpora discuss, in very restrained and discreet fashion, the probability of the corpus in question being representative and balanced. The parameter of corpus volume (Table 2, parameter 3) does relate to these two criteria, but very indirectly; cf. "typically researchers focus on sample size as the most important consideration in achieving representativeness" (Biber 1993: 243). In fact, a huge range of sociolinguistic information should also be taken into account.

Xiao (2008: 383) states that "national corpora are normally general reference corpora which are supposed to represent the national language of a country". It is assumed that national corpora are usually highly developed and dispose of all conceivable tools and engines for successful research in different subdisciplines of linguistics. The search facilities provided in the corpus interface can therefore be ignored here.

The most influential example of a "large" linguistic corpus is the British National Corpus (BNC; http://www.natcorp.ox.ac.uk, accessed 04-01-2016). Aston and Burnard (1998: 28) state that "the BNC was designed to characterize the state of contemporary British English in its various social and generic uses". The starting point was the notion about an ideal language corpus without regard to the availability of text material. Particular material for inclusion in the corpus was selected later, based on decisions concerning corpus design, structure and predefined target proportions.

The BNC project started with a careful planning stage where the design principles for the corpus were drawn up. These established a number of selection criteria which were then used for identifying suitable texts to be included in the corpus. (http://www.natcorp.ox.ac.uk/corpus/ creating.xml, accessed 04-01-2016)

It was hoped to maximize variety in the language styles represented, both so that the corpus could be regarded as a microcosm of current British English in its entirety, and so that different styles might be compared and contrasted. Each selection feature was divided into classes and target percentages were set for each class. Thus for the selection feature 'medium', five classes (books, periodicals, miscellaneous published, miscellaneous unpublished, and written-to-be spoken) were identified. Samples were then selected in the following proportions: 60 per cent from books, 30 per cent from periodicals, 10 per cent from the remaining three miscellaneous sources. Similarly, for the selection feature 'domain', 75 per cent of the samples were drawn from texts classed as 'informative', and 25 per cent from texts classed as 'imaginative'. (Aston and Burnard 1998: 29)

The 100-million-token British National Corpus complies with the criteria of being representative and balanced for a particular kind of speaker, register, variety, etc. The only significant point where the principle of balance has been intentionally violated is the ratio of the volume of the written corpus to the volume of the oral one.

There is a broad consensus among the participants in the project and among corpus linguists that a general-purpose corpus of the English language would ideally contain a high proportion of spoken language in relation to written texts. However, it is significantly more expensive to record and transcribe natural speech than to acquire written text in computer-readable form. Consequently the spoken component of the BNC constitutes approximately 10 per cent (10 million words) of the total and the written component 90 per cent (90 million words). These were agreed to be realistic targets, given the constraints of time and budget, yet large enough to yield valuable empirical statistical data about spoken English. (http://www.natcorp. ox.ac.uk/docs/URG.xml, accessed 04-01-2016⁶)

There is also a restriction for too-long texts, which are not be included entirely and truncated at least to 45,000 words.

The Russian National Corpus (http:// ruscorpora.ru/en/index.html, accessed 04-012016) contains more than 300 million words. Like the British National Corpus, it can serve as an example of a representative and balanced corpus.

A national corpus (...) is characterized by representative and well-balanced collections of texts. This means that such a corpus contains, if possible, all the types of written and oral texts present in the language (various genres of fiction, journalistic, academic, and business, as well as dialectal and sociolectal, texts). The proportion of text types in the corpus is based on their share in real-life usage at the time of composition. (http://ruscorpora.ru/en/ corpora-intro.html, accessed 04-01-2016⁷)

A more interesting example of a major national corpus is the Eastern Armenian National Corpus (EANC; http://www.eanc.net, accessed 04-01-2016). Armenian is the statutory national language of Armenia, a relatively small country in the Caucasus. Armenian belongs to the Indo-European family and is spoken by about six million people (Lewis et al. 2015). This corpus chooses its collection, so to speak, "semiselectively".

EANC is designed as a comprehensive corpus with the objective to include as many Standard Eastern Armenian texts as practicable. As of March 2009, EANC comprises about 110 million tokens. Overall, we have been guided by the goal of comprehensive representation - all literary, scientific and oral texts available to us have been indexed for search. The only exception to this are certain widely-available texts, such as electronic press and legal documents, whose presence has been limited for the sake of balance among different genres. (http://www.eanc. net/en/composition, accessed 04-01-2016)

The possible assortment of texts in Eastern Armenian, regardless of its official status, seems not to be large enough to allow corpus creators to reject some texts, trading the total volume for approximating to a more representative and balanced internal structure. Armenian is thus what McEnery and Ostler (2000) call a "smaller national language⁸".

Another type of structured collection of linguistic data is the language archive. Language archives normally include different types of material, not only texts, and in that way they differ from language corpora. However, an archive usually includes all available materials, and the principles of sampling and representativeness are not relevant. These properties link archives with corpora of under-resourced languages, but not with major national corpora. Three archives are discussed below as examples.

One of the biggest language archives is the Archive of the Indigenous Languages of Latin America (AILLA; http://www.ailla.utexas.org, accessed 08-01-2016). It contains a wide range of linguistic data, from non-transcribed speech recordings and digitized researchers' field notes to morphologically analyzed texts, usually accompanied by morpheme-to-morpheme glossing and translation. The archive comprises data on more than 300 American indigenous languages, and most of the data are freely accessible. The AILLA webpage interface allows archives to be browsed by language, collection, country, and depositor's name.

Another example of a language archive is the Pangloss Collection (http://lacito.vjf.cnrs.fr/ pangloss/index_en.htm, accessed 08-01-2016). The main goal of this archive is "to contribute to knowledge of endangered languages and cultures, by sharing annotated spoken texts of lesser-studied languages" (Michailovsky et al. 2014: 120).

ENT#91;The Pangloss CollectionENT#93; contributes to the documentation and study of the world's languages by providing free access to documents of connected, spontaneous speech, mostly in endangered or under-resourced languages, recorded in their cultural context and transcribed in consultation with native speakers. The Collection is an Open Archive containing media files (recordings), text annotations, and metadata; it currently contains over 1,400 recordings in 70 languages, including more than 400 transcribed and annotated documents. The annotations consist of transcription, free translation in English, French and/or other languages, and, in many cases, word or morpheme glosses; they are time-aligned with the recordings, usually at the utterance level. A web interface makes these annotations accessible online in an interlinear display format, in synchrony with the sound, using any standard browser. The structure of the XML documents makes them accessible to searching and indexing, always preserving the links to the recordings. (Michailovsky et al. 2014: 119)

This archive is in fact not too far from a language corpus. It is fully digitized and openaccessed. It provides options for searching and indexing, and includes meta-information about recording and transcriptions.

Another language archive that also focuses on collecting audio- and video-ethnographic materials is the Ethnographic E-Research Online Presentation System (EOPAS; http:// eopas.org, accessed 08-01-2016). EOPAS also provides interlinear linguistic analysis for its recordings⁹, making them highly useful in many kinds of research. Currently this archive centers on indigenous languages from Australia and Oceania, and some from South America.

This is an old-fashioned method of representing linguistically analyzed texts in understudied languages. Such collections are usually quite small because of inevitable size limitations imposed by the printed format. For the same reason they are of course not digitized. Nevertheless, they may contain valuable information about language use and provide linguistic analysis of primary data. Very often, this information is not available elsewhere. The process of digitization (the technical details will not be addressed here) can enable the use of these sources in present-day computerbased search algorithms, thereby advancing the usability of printed text collections to the level of language archives or even corpora of understudied languages.

Printed collections of annotated texts are very widespread sources of linguistic data; there is therefore no need to cite particular collections here. Such collections can appear as separate books (for instance, Mayers 1958), as articles in specialized journals (Romero Méndez 2012) or as appendices to grammatical descriptions (Lacrampe 2014). The main object is to provide additional information on real-life language use that complements grammatical description and vocabulary.

This Section presents a brief comparison of the four linguistic research tools examined above: small corpora of understudied languages, large national corpora, language archives and printed text collections. The basic parameter is the policy carried out regarding the selection of materials to be included in the research device. The developers of small corpora of understudied languages cannot afford the luxury of rejecting available texts, because such languages are normally under-resourced. The same is true for language archives that include material of every kind (sometimes not only textual) in order to cover the entire language use. The developers of large corpora, on the contrary, normally have an almost unlimited selection of available texts, so they have no problem leaving some of them beyond the scope of the corpus in order to maintain its representativeness and balance. Interestingly, authors of printed text materials usually have the same luxury of being selective, due to reasons of space.

Table 3 shows a comparison by this and some other parameters.

Table 3:
Basic characteristics of different research instruments

In fact, every group of instruments mentioned in Table 3 is very diverse. For example, small documentary corpora can be annotated or not, can have a more developed search engine or a less developed one, and can of course be bigger or smaller in size. Comparing the groups in Table 3, we can see that small documentary corpora are placed somewhere between major national corpora at one end and language archives at the other, according to their basic characteristics. They share some features with the former and others with the latter. Ostler (2008), for example, includes language archives such as AILLA (Section 3.2) or OLAC (Open Language Archives Community; http://www.language-archives.org, accessed 08-01-2016) among corpora of less studied languages. A terminological trap whereby both small and large corpora are traditionally called "corpora", and archives are called "archives", should not confuse the matter.