Some Mistakes Of Linguistic Corpususers

The first error is related to the calculation of the word frequency in the diachronic perspective.

It is necessary to correctly calculate the tokens’ frequency. Let’s calculate, for example, the tokens’ frequency of the ancient Russian word перст / finger. Using the corpus search results, we can get the following data: 141 tokens before 1800and 1069 tokens after 1900. We understand it can’t be really correct because an ancient word can’t be used more frequent recently. It is necessary to calculate the relative frequency based on the absolute quantity using a certain formula. According to this formula, the relative frequency is equal to: the tokens quantity should be divided by words quantity in the text and multiplied by 1 000 000. There are 12 636 732 tokens available before 1800. We get the frequency of 10. There are 208 953 262 tokens available after 1900. We get the frequency of 0, 05.

Let's consider the frequency of the ancient Russian word завет / testament . Using the corpus search results, we can get the following data. There are 1 095 tokens available in 623 documents with words volume 225 402 521 after 1900. According to the above formula, we get the frequency 12. There are 748 tokens in 421 documents with words volume 116 306 013 before 1900. We get the frequency 6.

The second error is related to the grammatical search. D. Beiber and S. Conrad explored corpus materials in teaching of English grammar and proved their effectiveness.

According to their opinion, there are three types of important results for teaching grammar: 1) information about the frequency; 2) fixing comparisons; 3) associations between grammatical structures and words (Beiber, Conrad, 2012). Let's search, for example, propositions containing predicates of Future tense form in combination with time interval circumstances containing preposition до / before, until. Using the main corpus of National Corpus of Russian Language, we get the following search result – one proposition. There is one example containing an expression Ябудуписать / I will write in combination with time interval circumstance including the preposition до / before [(1)]. Thetotalsearchvolumeis115,645documents, 23,803,881propositions, 283,431,966words.

(1)Ларка, это письмо я буду писать до тех пор,  пока не получу ответа на первое [НКРЯ]/ Larka, I'llwrite thisletter until Igetananswertothefirstone(NCRL).

There are no examples with Future tense predicates in combination with time interval circumstance including preposition к/ by(котъезду, к приезду/ the departure, by the arrival) . There is one example with Future tense predicate появится / will appear in combinationwithtimecircumstance к празднованию / to the celebration , expressing the precedence meaning.

(2) Кпразднованию 850-летияМосквывбулочных появится батонсинтригующимназванием «Веторон» [НКРЯ] / To celebration of the 850th anniversary of Moscow, loafs with an intriguing title "Vetoron" will appear in the bakeries(NCRL).

We tried to specify parameters. We used the searches with three words: an auxiliary verb will + a verb + a preposition with a noun , as well as other combinations. It is impossible to believe that in Russian Future tense predicates are so rarely used with the preposition до /to . Only as a result of numerous tests we managed to correctly set the search parameters: the first word is a verb, an indicative mood, the future tense + the second word is a preposition до / to . We found 27,999 tokens. There are three examples with Future tense predicates combined with timeintervalcircumstances containing preposition до / before, until.

(3)Этот процесс затянется до октября , и уже ясно, что с чеченской стороны его подпишет избранный президент. Это сделает документ ещё более значимым[НКРЯ] / This process will drag on until October , and it is already clear that on behalf of Chechen side the document will be signed by the elected president. This will make the document even more significant (NCRL).

While you are working with the linguistic corpus you can not completely trust it. It is necessary to evaluate the received data based on your language competence, to correctly select and set the search parameters (Hunston, Francis, 2000).

The main task of the linguistic corpus is not only to contain texts, but to contain correctly annotated texts. For example, a morphologically annotated corpus includes a morphological indication of speech parts for all words, which allows the user to quickly find a word of a desired speech part. However, over time, the corpus volume increased significantly, which led to drop in search performance. The main problem was that different grammatical analyzes attributed to the same word began to mix due to morphological homonymy. So, for example, the word берет means a hat and is considered as an inanimate masculine noun. Thesameword берет means takes andis also considered as an indicative mood of the verb totake PresentSimpletense, 3-d person. Due to these difficulties it is impossible to obtain search results for a demand "verb of the masculine gender". You will encounter the same difficulties while searching the word of plural number книжки / books . Unremovedhomonymy could bring the user a lot of unpleasant surprises. Even in the search for a prefix " re -" the corpus gives out the word пе-ре (re) -ц (means a pepper), although it has nothing to do with this prefix.

The linguistic corpus can be imperfect, so the user needs to remember and understand its features and limitations. There can always be omissions in the corpus, because in any live language, as a result of language play or word creation; there are lexemes, word forms and new meanings, which are called occasionalisms.

In the oral speech subcorpus you can find many examples of occasional or innovative word forms that are used to attract the interlocutor’sattentionand make statement informal or ironic. In this case, the annotation is more transparent and clear so that the user can easily find the necessary word form. Therefore, when drawing up a subcorpus, it is necessary to take into account the potential users and to facilitate interaction with the information source.

The "buyer" of the corpus, taking advantage of the "alien" corpus, consciously takes risks. The user takes risks that the corpus annotation may be imperfect. The corpus creators do certain compromise between theoretical knowledge and the possibilities of computer realization.

Therefore, the user should, on the one hand, be cautious about the search results received, on the other hand, the user should not forget that "not everything that seems, at first glance, is a mistake of the corpus".

The third mistake is lexical homonyms. For example, it is necessary to be cautious when you meet all lexical and grammatical homonyms. You shouldn’t forget that unintelligible homonymy can produce ambiguous results (Kutter, Kantner, 2012).

There is a searchexampleof aword лук/ an onion inoral speech corpus on NCRL. As a result of search we have received the following examples:

(4) Ну/ тамбылтакойбоевойбизнесмениз ВеликихЛук / ДмитрийМатвеев / которыйизбиралсядепутатомобластногособрания / онбольшевсехпообещалвложитьденег / онисталлидером [НКРЯ] / Well, there was a businessman from Velikiye Luk / Dmitry Matveyev / who was elected as a deputy of the regional assembly / he promised to invest more money / he became the leader(NCRL).

(5) Аэтосамое… алук почёминтереснотогда? [НКРЯ] / And this is the most interesting... how much is the onion then? (NCRL).

(6) Знаете/ чтотакоеарбалет? Усиленный лук. Стрела его бьёт далеко/ пробивает рыцарские доспехи [НКРЯ] / Doyouknowwhatacrossbowis? Reinforced crossbow. His arrow beats far / knight armor pierces (NCRL).

In the example (4) we see the proper name, the location name ВеликиеЛуки / Velikiye Luk. In next example (5) the word лук means an onion. The word лук means a crossbow in the last example (6). You have to decide what is the most convenient to you.

6.4. The fourth mistake is to search for a semantic features combination. Also, you need be careful and attentive with the corpus. If you "play" with semantic features search, you can get ambiguous results.

For example, if you specify semantic feature частитела / "human body parts" and отрицательныечерты / "negative" features, you can get the following examples:

(7) Икогдаясалатготовилвотиз брюшексёмги / ятожебралморковьоттуда [НКРЯ] / And when I cooked salad with salmon stomach / I also took carrots (NCRL).

(8) Что означает « перст указующий » - знаменитый жест Древней Руси? [НКРЯ] / What does " finger pointing " mean - the famous gesture of Ancient Russia? (NCRL).

In the example (7) we see the word брюшко means thestomach. The word перстуказующий means finger pointing in last example (8). You have to decide if there is a human body part with negative feature among these examples. The "negative meaning" of the words in these examples is questionable.

The linguistic corpus user should not rely entirely on the corpus search results because there will be "noise" or context shortage to some extent.

• Example of "noise": There are a large number of texts in NCRL, in which homonymy has not been removed. Therefore, in the search results of one word form, we meet contexts with a homonymous token. For example, if you are looking for a verb of imperative mood рой / dig, you will also get a noun пчелиныйрой / swarm (for example, "bee swarm") into search results.

• Example of "shortage": if we perform a search for word forms or if we incorrectly create a search for lexical and grammatical forms, the results will not show you the possible word forms, but only a part of them. Let's try to find the use of the quotation рукописинегорят/ "manuscripts do not burn." The search gives us 51 examples of this quotation. A lexico-grammatical search with the distance between words from 1 to 4 gives us 60 citations in the texts. Therearesuchinterestinginterpretationsamongthem.

(9) Рукописи, конечно, негорят, новосстанавливатьихохкактрудно[НКРЯ] / Manuscripts do not burn , of course, but it is difficult to restore them (NCRL).

(10) Я, какдурачок, сиделиплакалнадними, неверя, что рукописиивпрямьнегорят [НКРЯ] / I, like a fool, sat and cried over them, not believing that the manuscripts really do not burn (NCRL).

(11) Настоящие рукописи не только не горят, но и не публикуются! [НКРЯ] / These manuscripts are not only burning , but they are not being published either! (NCRL).

An example of competent corpus use in phonetic research is the works of A. Piperski and A. Kukhto (Piperski, Kukhto, 2016). In their article, the authors automatically analyze the selectedsubcorpusincluding poems by ten poets from National Corpus of Russian language. They performed their search for word forms with stress variability. The word forms quantity lies in the interval from 30 to 200 word forms for different speech parts. In the article, a quantitative measure is proposed for estimating the overall variability of stress, independent of the corpus size.