Tag: TG_DOC_TEXT

Text relevance (Maxfreq is the frequency of the most frequent word that makes sense of the length of the document).

Priority bonus, priority 7 - text priority. The binary factor, matters 0 for all monosyllabic requests, and the value of 1 for almost all two or more words, except for a very small number of answers for which there is not a single link that has passed quorum, and the text also did not pass the quorum.

Stript priority for TR is a text priority - there are all the words of the request somewhere in the document (while they pass contextual restrictions on the request, for example, both words DB in one sentence).

PHRASE priority for TR is a text priority - there are all the words of the request in a row in the document.

The presence of an accurate phrase (request text) in the header (more precisely, in the first sentence of the document). Contextual restrictions and feet are taken into account exactly as in TRP2, i.e. Factor [8] Minors Factor [5]

There was a plot that passed the quorum in which all the word positions are designated as those who have the relevance of Best_relev (title or Meta Keywords).

Long document (the longer the document, the greater the value of the factor).

Hitweigt is a variant of textual relevance, in which the weights of all hits are considered equal (i.e., they do not take into account the allowances for title and the proximity of words). In this case, the corresponding hits must be restricted by the syntactic sorcerer, i.e. We can assume that the TRHITW factor is 0 and only when Softandok is 0

Long text without links.

Coincidence of thematic spectra of request and document. Request themes-the result of work ((http://wiki.yandex-team.ru/evgenijjkroxalev/subquery Rules of the sorcerer Subquerysearch)) The subject of the document is taken from Yandex-Catalog

The factor about the number of Refines. In the queries, there is a feature of user refines ('' word that is faced with a percentage sign '). According to the idea, this means something like 'it would be good if the word in the document was'. The only famous ((http://staff.yandex-team.ru/gulin Andrey Gulin)) the valuable use of this feature is a request [ %official %site name of the film]. This feature is unknown to users, because Not described in any documentation. It is planned that it will disappear from the tongue of requests, but in the sorcerer the words with the priority of User_refine will remain. The factor indicates how much the maximum user_refine was simultaneously found in the framework of a single hit in the quorum. It is believed that there are from 0 to 3 (if> 3, then it is believed that 3). This number is waved in the half interval [0.1)

The number for which some linseed factors are multiplied (namely, factors number 6, 7, 47, 66), if text relevant 0, and there are few links

In textual relevance, Lemma coincides.

DSSM model, trained for reformulations, in the document uses relevant to the request of the proposal

TR divided by a cube of the number of words in a request and transformed by a standard REMAPTR.

The language of the document is Russian.

Simple BM25 in text.

Simple BM25 in text and links at the same time.

All the words of the request are in the text + links.

Simple BM25 in pairs of words - we take all pairs of words of the request and consider the number of their entry into the text of the document. In the quality of the weight of the pair we use the sum of the scales of words. It does not work if there is a stop-word in the request

BM25 from the number of sentences in the document in which it occurs.

BM25 according to only in the heading.

BM25 according to only with High Rel-bots ('significant', with the allocation (<b> ITP)).

The document has no TR.

the presence of pairs of words in the exact form

the number of sentences in which there are many words in the exact form

the presence of words in the header in the exact form

BM25 in the exact form

Simple BM25 in the exact form.

the presence of pairs of words taking into account synonyms (> = txtpair)

the number of sentences in which there are many words taking into account synonyms

the presence of words in the header, taking into account synonyms

BM25 taking into account synonyms

Simple BM25 taking into account synonyms.

The relative frequency of the words in the links (1 - the words of the request are often found in links, 0.3 - rarely); More precisely, the value of this factor is pessimized provided: TR = 0 && LR = 0 & (there is not a single link with all the words of the request) && (did not pass the quorum) && (at least one pair of words of the request is found in the text)

There is an exact form of all words of the request in the text/lincers

There is a lemma of all the words of the request in the text/lincers

The document passed Softand on the restrictions of the syntactic sorcerer. Only for documents with textual relevance. For monosyllabic requests, always 1.

The quality of the text. It is considered a rather complex formula

Text quality (classifier Alekseeva)

Document length in sentences

Document type - HTML

Document from porn kitski

A document from a commercial clay. Not used (depreded)

Fast document

The page title contains commercial vocabulary. Not used (depreded)

Commercial page (Classifier Savina)

The document does not have all the words of the request (with an accuracy to a synonym)

The percentage of the words of the request in the document (with an accuracy to a synonym)

The document has all the words of the request (with an accuracy to a synonym)

Tr by pairs of words in the reverse order

TR by pairs of words of the request through one word in texts

The percentage of all the words of the request in the text (with an accuracy to the form)

The document has all the words of the request (with an accuracy to the form)

The length of the page text in the words tlen = map (number of words, 1/400), where map (x, y) = x*y / (1 + x*y)

The length of the maximum coincidence of forms in the text and request

Weight of maximum coincidence of forms in the text and request

The length of the maximum coincidence in the lemma in the text and request

The weight of the maximum coincidence by lemma in the text and request

Options for relevant factors taking into account the feet of words

Options for relevant factors taking into account the feet of words

Options for relevant factors taking into account the feet of words

Options for relevant factors taking into account the feet of words

Options for relevant factors taking into account the feet of words

TR of the best passage - how high -quality snippet

Tr with discount for suggestions

Weight of maximum coincidence of forms in the text and request

Weight of maximum coincidence of forms in the text and request

The page has a about 'payment SMS'.

Stage of the page

Pornography of the page

BM25 for the user region for localized queries, for the unflapped in Cuba, is a country. The texts of the queries sent for the regions can be viewed in Relev_regions.txt in the sorcerer

The weight of the words of the request that is in the text

The weight of the words of the request that is in the text and links

We consider text features, believing that the page title is attributed to each of its proposal, i.e. The distance between the word from Title and any other word 1 sentence. Len is the maximum attitude of words from the request of the text met in some sentence (with attributed Title) in relation to the length of the request. Example [Harms Circus Vertunov] for ((http://wiki.yandex-team.ru//h.yandex.net/?http%3A%2F%2FWWWWIKILIVRES.info%2FWIKI%2F%25D0%25A6%25D %25b8%25D1%2580%25D0%25D0%25A %25BC%25D1%2581%of this document))

The ratio of the amount of IDF words in a sentence+Title to all words.

The same as Jokerlen, in the exact forms

The same as Jokerweight, in the exact forms

equals 2 * NastyContent

The poetry of the document

The maximum poetry of the quatrain

Document language - English

The request is fully covered by two exact groups consisting of an exact Match of the words of a contract in a row ((http://wiki.yandex-team.ru/poiskovajaplatform/tr/coveragebygroups about coating in groups))

There is a group consisting of an Exact Match of the words of the request that covers the request (possibly with a pass, addition or replacement of a word)

The share of the request, covered by the longest group consisting of any hits (including word forms and synonyms). Possibly with a pass, addition or replacement of a word

The language of the document is Cyrillic

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

The date of the document that is registered on the page is remarkable

The document has textual relevance

BM25, where the selected segments of the request act as 'words'

'Weight' of the segments of the request in the text

An indicator of the unnaturalness of the text from the point of view of the Russian language. The number of bad pairs of words in the text, transferred to the segment [0.1] according to the Z/(Z+10) formula

The proportion of bad steam among all found in the table: Z/(X+1), where Z is the number of bad couples in the text, and X is (http://wiki.yandex-team.ru/evgenijgrechnikov/testsynonimizers of 2000-navigable )) steam

The number of Latin letters in the text (not counting the markings), driven into [0.1] formula n/(n+100)

Previous factors - fixed

Previous factors - fixed

Previous factors - fixed

Previous factors - fixed

The number of words in the text (the word is what the lemmeter selected) is displayed in [0.1] according to the formula x/(x+a)

The number of words of the Russian language in the title

The average length of the word

The percentage of the number of words inside the tag <a> .. </a> from the number of all words

The percentage of the number of words outside the tags (outside the brackets <>) from the number of all words

The percentage of the number of words, which are 200 the most frequent words of the language, from the number of all words of the text

The number used in the text 500 of the most popular words of the language, divided by 500

Logarithm of average geometric probabilities of trigrams in the text. (the probability of a trigram - the number of its meetings in the text, divided by the number of all trigrams) is displayed in [0.1] according to the formula -x (x+a)

Logarithm of the average geometric conditional probabilities of trigrams. The conditional probability of a trigram is its probability, divided by the probability of a bigram from the first two words

The difference between the current date and the date of the document defined by the dates, 1 - the date of the document is equal to the current, 0 - the document of 10 years or more, if the date is not defined, equal to 0. Attention! ((1 - dateraage)*60)^2 = age of the page In days.

The maximum number of forms in all words of the request is max in all words of the request request_form_dl_lov/64

The sum of the number of forms balanced by the scales of words - the amount in all words of the request of the number_form_dly_lov/64*weight_lov; REMAP species x/(1 + x).

The unwarmed amount of the number of forms is the amount in all words of the request of the number_form_dl_lov/64/number_lov_

Analogues of the factors of the same name, the weight of the word = 1

Analogues of the factors of the same name, the weight of the word = 1

Analogues of the factors of the same name, the weight of the word = 1

Analogues of the factors of the same name, the weight of the word = 1

The share of different parts of speech in the text. The share of numerals (among all words that managed to recognize part of the speech)

The share of particles

The share of pronoun adjectives

The proportion of pronoun nouns

The share of verbs

The share of words that can be both masculine nouns and nouns of the feminine, but not of the middle kind, among all nouns (examples: 'hummingbirds' are an example of an indefinite kind that can be determined in two ways, 'Alexander' is homonymy).

The size of the largest text segment (from the factor [18] puretext)

DSSM model with early binding, trained on reforming and learned by ASR hypotheses of musical requests for Alice

A model trained on a PRS-Law PRS to predict BERT, trained on sinsig_ce with threshold value 0.5, using a chain of regions to the country

DSSM model with early binding, trained on reforming and learned on musical requests for Alice

Cunning BM25 in a sliding window. The size of the window is set in sentences. 'Jokers' are used for headlines and the beginning of the document. Morphological proximity and structure of the text are taken into account. The weight of the window fades with the removal from the beginning of the document.

The factor about that, a good snippet can turn out.

Simple BM25 in pairs of words - we take all pairs of words of the request and consider the number of their entry into the text of the document. Weight = 1. It does not work if there is a stop-word in the request

Logarithm of the number of shingles on which this document is not unique

Logarithm of the number of shingles on which this owner of the document is recognized as the author

The average weight of non-ugly shingles of this document

The language of the document corresponds to the language language

On the Porn Advertising page

BM25 with different parameters for different fields, including an incoming anchortekst. The weight of the text of the links included on the page is normalized depending on Delta Page Rank links

The size of the minimum piece of text, including all the words of the request found in the document. Not used now. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAformula/tekushhiekomponenty/ymw Read more))

Buettcher, Clarke and Lushman factor (modified) ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAformula/tekushichiekomponenty/bclm more)))))))))

Unigramal language model. Language is modeling according to the document, smoothed out by the general linguistic model. When building a model, the document uses information on which field of the document met the word request (Title, Head or Plain Text)

Calculates the coating of the request with letter trigrams of the document header

Calculates the heading of the heading of the document header with letter trigrams

He considers the sum of the following species: the sequence of words of the request more than two, met in one sentence; It is normalized for the length of the document.

Coincidence of the thematic spectrum (according to DMOZ) request and document. The theme of the request is determined ((http://wiki.yandex-team.ru/jandekspoisk/zarubezhnyjjinternet/dmozqueryClassifier1 The rule of the sorcerer Dmoztheme))

Coincidence of the thematic spectrum (according to DMOZ) request and document. The theme of the request is determined by the best result ((http://wiki.yandex-team.ru/jandekspoisk/zarubezhnyjjinternet/dmozqueryClassifier1 Rules for the sorcerer DmozTheme)) The subject of the document is determined by the automatic classifier

Evaluates the minimum distance between the pairs of words of the request, taking into account the remoteness of the pair from the beginning of the document (Minimal Pair Size with Attenuation). Steles are understood to mean all consistent bigrams of the words of the request. Thus, the number of vapor is equal to the number of words in a request reduced by 1. Accordingly, the factor makes sense for requests consisting of more than one word. (Http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaformula/ Tekushhiekomponenty/MPSA MPSA))

It differs from BCLM in that the weights of all words are considered the same. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAformula/tekushhiekomponenty/bclm2 BCLM2))))))))))))

Text relevant based on the language model, taking into account the absolute position. We go along the text with a window of 20 words, build a language model on each window (that is, the distribution of probabilities in the words of the Russian language) and calculate the probability of generating a request. For removal from the beginning of the document, we finish the model.

Modification of the BCLM2 factor, lightweight for use in tulle. The main difference is that BCLMLite does not use absolute displacements of words relative to the beginning of the document. Instead, the factor works with the usual positions of the type <number of the_prising, position_v_production>. At the same time, the proximity between the words is taken into account only inside the sentence. (Http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaFormula/tekushichiekomponenty/bclmlite bclmlite)))))))))))))

Fixed YMWFull. It differs from the previous version only by behavior on 2 -word queries. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAformula/tekushhiekomponenty/ymw Read more))

Binary factor, every word of the request is in the text or in the links

'Country praets' (AUXQC)

'Country praets' (AUXQC)

Page - '404' (share of tokens '404' in relation to the total number of tokens on the page)

BM25, in which the weight of the word is machine -like

The factor evaluates as the words of the request is grouped with each other in the text of the document without taking into account their order. ((http://wiki.yandex-team.ru/sergejjkrylov/queryWordCohesionTR Description))

Number of letters in the AUX segment

The number of spaces in the AUX segment

The number of commas in the Content segment

Page is a store. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAformula/tekushhiekomponenty/opisanijafaktorov#SSHOP Description)). Not used (depreded)

Logarithm of the number of shingles in the document added by the owner of the site as original texts in ((http://wiki.yandex-team.ru/jandekspoisk/jekosistema/marketingPr/webmasters/plan/vtorcontect of originality plugin)). It does not participate in the formula, it is needed to disconnect the takes

The average filtered number of sources of authorship of the document. It does not participate in the formula, it is needed to disconnect the takes

Dispersion of IDF words,

The language of the document corresponds to the country's request

The share of the segments of the request present in the text

The language of the document is one of the permissible for Turkey (Turkish, English, German, French, Arabic, Azerbaijani) or the document has zero length. In the search stage is calculated only for Isrealgeolocal requests.

Variation of Temo ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaformula/tekushhiekomponenty/DBM25 dBM25), cm.

BM25FDPR with standardization on the average length of the document, depending on the language of the document. ((http://wiki.yandex-team.ru/bm25frework test results.))

The popularity of the language of the document. Number from 0 to 1. (http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaformula/tekushhiekomponenty/languaguaguagepopalarity)))))))

The amount of factors QueryDownerClicksFRC and BM25FDPRFIXED with scales 0.358449 and 0.184922, respectively. '565' in the name of the factor does not need to be perceived literally, it is Legashi or a typo.

The factor evaluates the differences in the positions of words in the heading from the posterity in the request

Variation of Temo ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaformula/tekushhiekomponenty/DBM25 dBM25), cm.

A pool of logs is marked with BERT trained on Sinsig. DSSM model is trained on this pool using BaseregionChain

The degree of commerce page title. Not used (depreded)

BM25 page title by its text

BM25 page title by texts of links to it

The share of unique trigrams in the trigrams of links

Share of unique trigrams of links in trigrams header

The greasy of the page

BM25 in texts and links with special. Libra in the level of coincidence (shape, lemma, synonym)

The weight of the words of the request that is in the text in the exact form

The weight of the words of the request that is in the text with an accuracy to lemma

The weight of the words of the request that is in the text

Maximum amount weight of the words of the request in the window of 50 words

The share of unique trigrams of a footer fragment in trigrams of links

The share of unique trigrams of links among a fragment of trigrams of a footer

Variation of Temo ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaformula/tekushhiekomponenty/DBM25 dBM25), cm.

Variation on the topic BM25

Variation on the topic BM25

Variation on the topic BM25

DBM separately by numbers

DBM separately by geo-objects of request

DBM separately on the noun

Evals the correspondence of the positions of words in the sentences of the document to the positions of words in the request.

The document contains a name from the request.

The document has a direct link to the file

The document has a link to filehosting

The proximity of the words of the request to the most difficult word.

The document contains user review/comment

The time of creating a document with an accuracy of 1.0 is the current month, 0- 10 years ago and older. Temporarily disconnected

The time for updating the document with an accuracy of 1.0 is the current month, 0- 10 years ago and older. Temporarily disconnected

The function of the credibility of the distribution of years in the document. Temporarily disconnected

The arithmetic mean position of dates in the document. Temporarily disconnected

BM with attenuation in the text of catalog links.

The share of the words of the document from the segments with Score> 2.

TEXTBM25 is considered in the title by the text of the name of the user region - similar to the factor 268.

BCLM for Annotation index, doc text and links.

Finetuned reformulations DSSM to commercial clicked bargain odd-like target from visit log

The factor according to the name from the original request is considered according to the contents of the document. Algorithm: Chain0wcm

Request and documentary navigation model.

The factor on the text of the request and heading (Title) of the document, assessment of the compliance of numerical ranges in words-markers

The factor according to the name from the original request is considered according to the contents of the document. The minimum window size, which includes all the words of the request. It is normalized for the number of words in the request.

Factor for name from the original request text of the document. Algorithm Cosinematchmaxpredical.

The factor for all the name from the original request Aggregation on all extensions. Type of aggregation for extensions: the greatest value of the factor; It is considered according to the contents of the document. Algorithm: Chain0wcm

The factor for all the name from the original request Aggregation on all extensions. Type of aggregation for extensions: the greatest value of the factor; It is considered according to the contents of the document. The minimum window size, which includes all the words of the request. It is normalized for the number of words in the request.

The factor for all the name from the original request Aggregation on all extensions. Type of aggregation for extensions: the greatest value of the factor; The text of the document. Algorithm Cosinematchmaxpredical.

DSSM CLOSE DISCOUNT according to data specific for Alice

The factor for telephone attributes Tel_Full from the original request text of the document. Algorithm for aggregation of the scales of words BOCM15. Normalization coefficient 0.01.

The prediction of the total time spent to the end of the session, provided that this pair is implemented by the request-document

The prediction of the contribution of this pair request-document to the timetable

The prediction of the percentage of the length of the track, which will be lost subject to the implementation of this pair of the request

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BM15 in the group of streams 2. The maximum value of the factor for extensions.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3. The maximum value of the expansion factor.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BM15FLOGW0 for Urlu and Title. The maximum value of the expansion factor.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: CosinemaxMatchprediction in text and Title. The maximum value of the expansion factor.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BM15FLOG by the Stream group 1. The average balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) for extensions.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BM15FLOGW0 for Urlu and Title. The average balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) for extensions.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: Minwindowsize in text. The average balanced values ​​of the expansion factor.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: ORIGINALREQUARY ORIGINALREKETRACTRENEXACT for a group of streams for bag factors (text, Title, annotation streams).

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: CosinemaxMattcg bag.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3. The minimum balanced value of the factor for the expansion top.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: Chain0wcm in text. The average balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) according to the expansion top.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3. The average balanced values ​​of the factor for the expansion top.

Bm15K01 factor over hits from Title

Bocm15K001 factor over hits from Title

Bm11Norm16384 factor over hits from Text

Bocm11Norm256 factor over hits from Text

CosineMatchMaxPrediction factor over hits from Text

Bm15FLogK0001 factor over hits from FieldSet1 stream

Bm15FLogK0001 factor over hits from FieldSet2 stream

BclmWeightedFLogW0K0001 factor over hits from FieldSet3 stream

Bm15FLogW0K00001 factor over hits from FieldSetUT stream

Chain0Wcm factor over hits from Body

PairMinProximity factor over hits from Body

MinWindowSize factor over hits from Body

DSSM model trained on clicks.

DSSM model trained on clicks.

DSSM model trained on clicks.

DSSM model trained on clicks.

DSSM model trained on clicks.

DSSM model trained on clicks.

DSSM model trained on clicks.

DSSM model trained on clicks.

Neural model of content quality for medical subjects

Neural model of content quality for medical subjects (for ex -)

Neural model of content quality for financial and legal topics

Neural model of content quality for financial and legal topics (for exposures)

Linguistic boosting factor. Type of extensions: Requestwithregionname. BM11 in the text and the Title of the Document

Linguistic boosting factor. Type of extensions: Requestwithregionname. Cosinematchmaxprediction on the text and dump title

Linguistic boosting factor. Type of extensions: Requestwithregionname. Factor: BM15 in the group of streams 1.

Linguistic boosting factor. Type of extensions: Requestwithregionname. Factor: BM15 in the group of streams 2.

Linguistic boosting factor. Type of extensions: Requestwithregionname. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3.

Linguistic boosting factor. Type of extensions: Requestwithregionname. Chain0WCM factor on the text of the document

Neural model of content quality for SOS topics

Neural model of content quality for SOS subjects (for ex -)

Prediction of the time of the session, provided that this pair is requested by the request-document

Request-document model Sinsiga.

The factor for the original request. It is considered according to the heading of the document. The algorithm for aggregation of words weights is BCLMMIXPLAIN: a linear mixture of annotation BCLM weights and balanced Positionless weights of the word, then the former meters are aggregated through BM15. Normalization coefficient 10^(-5).

The factor for the original request. It is considered according to the heading of the document. CMMATCHTOP5AVGMATCHVALUE algorithm.

The factor for the original request. It is considered according to the heading of the document. The degree of coating of the words of the request is accurate to the form (without synonyms).

The factor for the original request. It is considered according to the heading of the document. The weight of the hit is multiplied by 1/ (1 + the position of the word in the sentence) an algorithm for aggregation of the scales of words: BM15. Normalization coefficient 0.5.

The factor for the original request. It is considered according to the contents of the document. The algorithm for aggregation of words weights is BCLMMIXPLAIN: a linear mixture of annotation BCLM weights and balanced Positionless weights of the word, then the former meters are aggregated through BM15. Normalization coefficient 10^(-5).

The factor for the original request. It is considered according to the contents of the document. Algorithm Cosinematchmaxpredical.

The factor for the original request. It is considered according to the contents of the document. Algorithm Allwcmweightedpredical.

The factor for the original request. It is considered according to the contents of the document. Algorithm for aggregation of the scales of words BOCM15. Normalization coefficient 0.01.

The factor for the original request. It is considered according to the contents of the document. Algorithm: Querypartmatchsumvalueany.

The factor for the original request. It is considered according to the contents of the document. The degree of coating of the words of the request is accurate to the form (without synonyms).

The factor for the original request. It is considered according to the contents of the document. The degree of covering the words of the request in the exact form.

The factor for the original request. It is considered according to the contents of the document. Libra Agnregation algorithm: BM15Maxannotation normalization coefficient 0.01.

DSSM model trained on clicks. Takes bigrams into account.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Factor: AVGPERGRAGRAMMAXVALEANY in the Stream group 5. The average balanced values ​​of the factor for the expansion top.

DSSM model trained on clicks. Takes bigrams into account. Embeddings for documents are computed offline.

A minimum of gradients according to the Bigramm LogdwellTime model.

Maximum from gradients according to the Bigramm Logdwelltime model.

The second central moment (dispersion) from gradients according to the Bigramm Logdwelltime model.

The third central moment from gradients according to the Bigramm Logdwelltime model.

Linguistic boosting factor. Type of extensions: QFUF. Factor: BCLMWEIGHTEDFLOGW0_K0.001 FIELDSET3. The average balanced values ​​of the TOP-10 factor by extensions.

Linguistic boosting factor. Type of extensions: Querytotext. Factor: by minwindowsize according to the contents of the document. The average balanced values ​​of the expansion factor.

Linguistic boosting factor. Type of extensions: Querytotext. Factor: Minwindowsize according to the contents of the document. The average balanced values ​​of the TOP-10 factor by extensions.

Linguistic boosting factor. Type of extensions: QFUF. Factor: BM15FLOGW0_K0.0001 on Ural and the heading. The maximum value of the expansion factor.

Linguistic boosting factor. Type of extensions: QFUF. Factor: BCLMWEIGHTEDFLOGW0_K0.001 FIELDSET3. The average balanced values ​​of the expansion factor.

Linguistic boosting factor. Type of extensions: Querytotext. Factor: PairminProximity according to the contents of the document. The average values ​​of the expansion factor.

Linguistic boosting factor. Type of extensions: Querytotext. Factor: Bocm11_norm256 according to the text of the document. The average values ​​of the expansion factor.

Linguistic boosting factor. Type of extensions: QFUF. Factor: COSINEMATCHMAXPRECTION on the text of the document. The maximum value of the expansion factor.

Linguistic boosting factor. Type of extensions: QFUF. Factor: BM15FLOG_K0.001 according to Fieldset1. The average balanced values ​​of the factor with a quadratic weight in the top 10 in terms of factor value by extensions.

Linguistic boosting factor. Type of extensions: QFUF. Factor: Bocm11_norm256 according to the text of the document. The maximum value of the expansion factor.

Linguistic boosting factor. Type of extensions: QFUF. Factor: BM15FLOGW0_K0.0001 on Ural and the heading. The average balanced values ​​of the expansion factor.

DSSM model trained on clicks, target=OneClicks/Clicks. Takes bigrams into account.

DSSM model trained on clicks, target=QueryDwellTime stream value. Takes bigrams into account.

The normalized amount of the scales of the words of the request that met in the text of the document or links to it.

The normalized amount of the scales of the words of the request that Equal_by_String in the text of the document or links to it.

The normalized amount of the scales of the words of the request that met in the text of the document.

The normalized amount of the scales of the words of the request that met in the links to the document.

The normalized amount of the scales of the words of the request that Equal_by_String in the links to the document.

The normalized scales for the IFILTRETRATIONMODEL words of the request that met in the text of the document or links to it.

The normalized scales for the IFILTRETRATIONMODEL Words of the request, which are Equal_by_String in the text of the document or links to it.

The normalized scales for the IFILTRETRATIONMODEL Words of the request, which Equal_by_lemma in the text of the document or links to it.

The normalized scales for the IFILTRETRATIONMODEL words of the request that met in links to the document.

The normalized scales for the IFILTRETRATIONMODEL Words of the request, which Equal_by_String in the links to the document.

Document DSSM model Language Classifier Rus.

Document DSSM model Language Classifier Eng.

Document DSSM model Language Classifier Other.

DSSM Prediction to determine Alice's irrelevant answers

BM25FDPR with standardization on the average length of the document, depending on the language of the document. Only texts are used.

Query-MainContentKeywords similarity, target: logDwellTime

DSSM model trained on CTRs without miner.

DSSM model trained on click odd pool

DSSM model trained on click personalization pool

DSSM model trained on click triangle pool

DSSM model trained on clicks without miner (with no-clicks and AM-hard negatives). Takes bigrams into account.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3. The maximum balanced value of the factor.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15FLOG in the group of streams 2. The maximum balanced value of the factor.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: ORIGINALREQUENTFRACTFRACTION OF THE FIELDSETBAGOFWORDS Stream.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: MixmatchweightedValue by Stream Querydwelltime. The maximum balanced value of the factor is normalized for the total weight.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15 according to Stream Title. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) normalized for total weight.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3. The minimum value of the factor for the expansion top.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BCLMWEIGHTEDFLOGW0 in the Stream group 3. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) normalized for the total weight.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15FLOG in the Stream group 1. The maximum balanced value of the factor.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15FLOG in the Stream group 1. The total balanced value of the factor is normalized for the total weight.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: Bag AnnotationMatChavgvalue by Stream LongClicksp.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15FLOG for the Stream group 1. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) for expansion top extensions, standardized for the total weight of the expansion top.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15FLOG in the Stream group 1. The minimum balanced value of the factor for the expansion top extensions normalized for the maximum weight by the expansion top.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: PairminProximity according to Stream Body. The maximum balanced value of the factor is normalized for the total weight.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: BM15FLOG for the Stream group 1. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) normalized for total weight.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: SIMPLECLIC SIMPLECLICS bag.

Linguistic boosting factor. Type of extensions: XFDTSHOWKNN. Factor: CosinemaxMatch bag according to Title Stream.

DSSM model trained on clicks without miner (with no-clicks and am_hard negatives 50/50 and then on am_hard negatives only). Takes bigrams into account.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. Factor: Norm256 by stream BOCM11. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}).

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. Factor: Minwindowsize by Stream Body. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) by the expansion top, normalized for the total weight according to the expansion top.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. Factor: Minwindowsize by Stream Body. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) normalized for total weight.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. Factor: Norm256 by stream BOCM11. The total balanced values ​​of the factor multiplied by weight (\ frac {\ sum w_i * (w_i * f_i)} {\ sum w_i}) according to the expansion top.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. The minimum expansion weight.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. The arithmetic mean of expansion weights.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. The total weight of the extensions.

Linguistic boosting factor. Type of extensions: Querytotextbyxfdtshowknn. Factor: ORIGINALREQUENTFRACTFRACTION OF THE FIELDSETBAGOFWORDS Stream.

Neural document model for finding unexpected tin

The initial request with the removal of verbs. It is considered according to the heading of the document. The algorithm for aggregation of the scales of words: BM15. Normalization coefficient 0.1.

The initial request with the removal of verbs. It is considered to be composational stream, consisting of an tokenized Url and a title of a document. The algorithm for aggregation of the scales of words: BM15FLOGW0. Normalization coefficient 0.0001.

The initial request with the removal of verbs. It is considered according to the contents of the document. The minimum window size, which includes all the words of the request. It is normalized for the number of words in the request.

The result of the use of a neural model, trained to distinguish long clicks from other events, the input of the model is the ambassadors and bigram meters, calculated by text streams (Title, Body, URL).

Linguistic boosting factor. Type of extensions: QFUFFILTEDBYXFONSE (QFUF, filtered on the DSSM models Xfonese). Aggregation on all extensions. The greatest value of the factor. Into aircraft association of the URLs, Title, Body, Correctedctr, Longclick, OneClick, Browserpagerank, Splitdwelltime, SampleperiodDayFrc, SimpleClick, Yabarvisits, Yabartime. The algorithm for aggregation of the scales of words: BM15FLOG (BM15 Aggregation of Logarithm of Construction of Words). Normalization coefficient 0.001.

Linguistic boosting factor. Type of extensions: QFUFFILTEDBYXFONSE (QFUF, filtered on the DSSM models Xfonese). Aggregation on all extensions. The greatest value of the factor. Rebelled association of streams Title, Body, LongClick, LongClicksp, OneClick. The algorithm for aggregation of the scales of words: BCLMWEIGHTEDFLOGW0. Normalization coefficient 0.001.

Linguistic boosting factor. Type of extensions: QFUFFILTEDBYXFONSE (QFUF, filtered on the DSSM models Xfonese). Aggregation on all extensions. The greatest value of the factor. It is considered to be composational stream, consisting of an tokenized Url and a title of a document. The algorithm for aggregation of the scales of words: BM15FLOGW0. Normalization coefficient 0.0001.

Linguistic boosting factor. Type of extensions: QFUFFILTEDBYXFONSE (QFUF, filtered on the DSSM models Xfonese). Aggregation on all extensions. The greatest value of the factor. It is considered according to the heading of the document. The algorithm for aggregation of the scales of words: BM15. Normalization coefficient 0.1.

Linguistic boosting factor. Type of extensions: QFUFFILTEDBYXFONSE (QFUF, filtered on the DSSM models Xfonese). Aggregation by TOP-10 (by the value of the factor) extensions. A suspended sum of the Libra of factors. Normalized for the total weight of extensions. Into aircraft association of the URLs, Title, Body, Correctedctr, Longclick, OneClick, Browserpagerank, Splitdwelltime, SampleperiodDayFrc, SimpleClick, Yabarvisits, Yabartime. The algorithm for aggregation of the scales of words: BM15FLOG (BM15 Aggregation of Logarithm of Construction of Words). Normalization coefficient 0.001.

Linguistic boosting factor. Type of extensions: QFUFFILTEDBYXFONSE (QFUF, filtered on the DSSM models Xfonese). Aggregation by TOP-10 (by the value of the factor) extensions. A suspended sum of the Libra of factors. Normalized for the total weight of extensions. It is considered according to the contents of the document. The minimum window size, which includes all the words of the request. It is normalized for the number of words in the request.

The factor for the filtered original request: the DSSM state from the request is calculated without words to the initial request, after which the threshold is cut off. Into aircraft association of the URLs, Title, Body, Links, Correctedctr, LongClick, OneClick, Browserpagerank, Splitdwelltime, SampleperiodDayFrc, SimpleClick, Yabarvisits, Yabartime. The algorithm for aggregation of the scales of words: BM15FLOG (BM15 Aggregation of Logarithm of Construction of Words). Normalization coefficient 0.001.

The factor for the filtered original request: the DSSM state from the request is calculated without words to the initial request, after which the threshold is cut off. It is considered to be composational stream, consisting of an tokenized Url and a title of a document. The algorithm for aggregation of the scales of words: BM15FLOGW0. Normalization coefficient 0.0001.

DSSM model trained on cross language CTRs using serp similarity hard miner.

For all words of the request, the weight is calculated by the Query-Mutation method (the distance between the requests in nash and there is no word). The sum of the scales of the words found in the title is taken, divided by the sum of the scales of all words.

For all words of the request, the weight is calculated by the Query-Mutation method (the distance between the requests in nash and there is no word). Maximum weight is taken among words absent in the title of the document.

The result of the use of a neural model, trained to distinguish long clicks from other events, the input of the model is the ambassadors and bigram meters calculated by text streams (Body, URL).

It is considered as (80-x) where X is the return of the document in the clock (continuously). Uses the data of the Robotaddtime dates

It is considered as (10-x) where X is the return of the document in days (continuously). Uses the data of the Robotaddtime dates

The difference between the current date and the date of the document, determined by the Robotaddtime, 1 - the date is equal to the current, 0 - the document of 10 days or more, or the date is not determined

Linguistic boosting factor. Type of extensions: XFONESEKNN (closest to the DSSM models trained to predict XFDTSHOW of extension). Aggregation on all extensions. The greatest balanced value of the factor. It is normalized for the maximum weight of expansion. Into aircraft association of the URLs, Title, Body, Links, Correctedctr, LongClick, OneClick, Browserpagerank, Splitdwelltime, SampleperiodDayFrc, SimpleClick, Yabarvisits, Yabartime. The algorithm for aggregation of the scales of words: BM15FLOG (BM15 Aggregation of Logarithm of Construction of Words). Normalization coefficient 0.001.

Linguistic boosting factor. Type of extensions: QuerytotextByxfoneKnn (Querytotext extensions of Xfoneeseknn extensions). Aggregation by TOP-10 (by the value of the factor) extensions. A suspended sum of the Libra of factors. Normalized for the total weight of extensions. It is considered according to the contents of the document. The minimum window size, which includes all the words of the request. It is normalized for the number of words in the request.

Linguistic boosting factor. Type of extensions: QuerytotextByxfoneKnn (Querytotext extensions of Xfoneeseknn extensions). Aggregation on all extensions. A suspended sum of the Libra of factors. Normalized for the total weight of extensions. Rebelled association of streams Title, Body, LongClick, LongClicksp, OneClick. The algorithm for aggregation of the scales of words: BCLMWEIGHTEDFLOGW0. Normalization coefficient 0.001.

DSSM model that predicts the logarithm of the longest click on the Serpa. As negative examples, select Urla from past requests of the same user, and the maximum time between requests is no more than 7 minutes (super -cords for reformulations)

DSSM model with early binding, trained in reformulations, which predicts the logarithm of the longest click on the Serpa.

Neural network value for contexts of query hits in document text. Predicts relevance-all-8-years. Uses formula ussr-dump-20190719 prs-20190720 all-8-years [t > 0.25] CrossEntropy 20k 0.25 -S 0.8 -Z 1 predictions for learning.

DSSM model trained on a reformal pool, which in the request, in addition to the request itself, receives 4 extensions of the XFDT with the largest weight

A model trained to predict an assessment of the USSR-DUMP-20190719 PRS-20190720 ALL-8-YEARS [T> 0.25] Crossentropy 20K 0.25 -s 0.8 -z 1.

Neuron document model for finding unexpected tin (for ex -)

A model trained to predict an assessment of the USSR-DUMP-20190719 PRS-20190720 ALL-8-YEARS [T> 0.25] Crossentropy 20K 0.25 -s 0.8 -z 1 and an educational study on assessments of relevance.

Max percent of query words in DirectOffer

Mean percent of query words in DirectOffer

Min percent of query words in DirectOffer

Max percent of DirectOffer words in query

Mean percent of DirectOffer words in query

Min percent of DirectOffer words in query

Max percent of query words in DirectOffer prefix (query length)

Mean percent of query words in DirectOffer prefix (query length)

Min percent of query words in DirectOffer prefix (query length)

Max percent of query words in DirectOffer prefix (2X query length)

Mean percent of query words in DirectOffer prefix (2X query length)

Min percent of query words in DirectOffer prefix (2X query length)

Max percent of query words in DirectTitle

Mean percent of query words in DirectTitle

Min percent of query words in DirectTitle

Max percent of DirectTitle words in query

Mean percent of DirectTitle words in query

Min percent of DirectTitle words in query

Max percent of query words in DirectTitle prefix (query length)

Mean percent of query words in DirectTitle prefix (query length)

Min percent of query words in DirectTitle prefix (query length)

Max percent of query words in DirectTitle prefix (2X query length)

Mean percent of query words in DirectTitle prefix (2X query length)

Min percent of query words in DirectTitle prefix (2X query length)

Max percent of query words in DirectInfo

Mean percent of query words in DirectInfo

Min percent of query words in DirectInfo

Max percent of DirectInfo words in query

Mean percent of DirectInfo words in query

Min percent of DirectInfo words in query

Max percent of query words in DirectInfo prefix (query length)

Mean percent of query words in DirectInfo prefix (query length)

Min percent of query words in DirectInfo prefix (query length)

Max percent of query words in DirectInfo prefix (2X query length)

Mean percent of query words in DirectInfo prefix (2X query length)

Min percent of query words in DirectInfo prefix (2X query length)

Percent of query words in result track name

Percent of result track name words in query

Percent of query words in result album name

Percent of result album name words in query

Percent of query words in result artist name

Percent of result artist name words in query

Percent of query words in wizard title

Percent of wizard title words in query

Percent of query words in wizard track name

Percent of wizard track name words in query

Percent of query words in wizard album name

Percent of wizard album name words in query

Percent of query words in wizard artist name

Percent of wizard artist name words in query

Percent of query words in wizard track lyrics

Percent of wizard track lyrics words in query

Min percent of query words in documents title

Mean percent of query words in documents title

Max percent of query words in documents title

Min percent of documents title words in query

Mean percent of documents title words in query

Max percent of documents title words in query

Min percent of query words in documents snippet

Mean percent of query words in documents snippet

Max percent of query words in documents snippet

Min percent of documents snippet words in query

Mean percent of documents snippet words in query

Max percent of documents snippet words in query

Mean percent of query words in title

Min percent of query words in title

Max percent of title words in query

Mean percent of title words in query

Min percent of title words in query

Max percent of query words in title prefix (query length)

Mean percent of query words in title prefix (query length)

Min percent of query words in title prefix (query length)

Max percent of query words in title prefix (2X query length)

Mean percent of query words in title prefix (2X query length)

Min percent of query words in title prefix (2X query length)

Max percent of query words in headline

Mean percent of query words in headline

Min percent of query words in headline

Max percent of headline words in query

Mean percent of headline words in query

Min percent of headline words in query

Max percent of query words in headline prefix (query length)

Mean percent of query words in headline prefix (query length)

Min percent of query words in headline prefix (query length)

Max percent of query words in headline prefix (2X query length)

Mean percent of query words in headline prefix (2X query length)

Min percent of query words in headline prefix (2X query length)

Confidence of item selector in current gallery

Max percent of query words in ItemName

Mean percent of query words in ItemName

Min percent of query words in ItemName

Max percent of ItemName words in query

Mean percent of ItemName words in query

Min percent of ItemName words in query

Max percent of query words in ItemName prefix (query length)

Mean percent of query words in ItemName prefix (query length)

Min percent of query words in ItemName prefix (query length)

Max percent of query words in ItemName prefix (2X query length)

Mean percent of query words in ItemName prefix (2X query length)

Min percent of query words in ItemName prefix (2X query length)

Max percent of query words in ItemDescription

Mean percent of query words in ItemDescription

Min percent of query words in ItemDescription

Max percent of ItemDescription words in query

Mean percent of ItemDescription words in query

Min percent of ItemDescription words in query

Max percent of query words in ItemDescription prefix (query length)

Mean percent of query words in ItemDescription prefix (query length)

Min percent of query words in ItemDescription prefix (query length)

Max percent of query words in ItemDescription prefix (2X query length)

Mean percent of query words in ItemDescription prefix (2X query length)

Min percent of query words in ItemDescription prefix (2X query length)

Confidence of item selector by name in current gallery

Confidence of item selector by number in current gallery

Fraction of hosts with Russian language

The result of the aggregated tin classifier is used on average to determine tin queries

Fraction of hosts with Russian language

Fraction of hosts with Russian language

The value of the DP classifier is used to filter on average

The result of the text classifier of long views via VowPal Wabbit

The result of the aggregated tin classifier is used on average to determine tin queries

Previous factors - fixed

The result of the Suggestive text classifier via VowPal Wabbit

The result of the text classifier of porn viapal wabbit

The result of the text classifier of tin according to Vowpal Wabbit

Image porn classifier output

Fraction of hosts with Turkish language

Fraction of hosts with Russian language

The value of the DP classifier is used to filter on average

Feature from utracker

The result of the text classifier of long views via VowPal Wabbit

There is an exact form of all words of the request in the text/lincers

The document has all the words of the request (with an accuracy to a synonym)

The result of the aggregated tin classifier is used on average to determine tin queries

The share of the request, covered by the longest group consisting of any hits (including word forms and synonyms). Possibly with a pass, addition or replacement of a word

Previous factors - fixed

BM25 - like factor, but with custom coefficients

The result of the Suggestive text classifier via VowPal Wabbit

Linguistic boosting factor. Type of extensions: QFUF. Aggregation on all extensions. The greatest value of the factor. It is considered according to the heading of the document. The algorithm for aggregation of words weights is BCLMMIXPLAIN: a linear mixture of annotation BCLM weights and balanced Positionless weights of the word, then the former meters are aggregated through BM15. Normalization coefficient 10^(-5).

The percentage of the words of the request in the document (with an accuracy to a synonym)

The result of the text classifier of porn viapal wabbit

The result of the text classifier of tin according to Vowpal Wabbit

The document has all the words of the request (accurate to synonym) - Top Average

Ratio of word_1 hits to all hits - top average

Ratio of word_2 hits to all hits - top average

Ratio of word_3 hits to all hits - top avg

Distance from average position of word 1 to position 1 - top avg

Distance from average position of word 1 to position 1

Distance from average position of word 2 to position 2

Min of distances of average position of word_i to prosition i

Max weighted variance of position of word_i for all i

Ratio of EQUAL_BY_STRING hits of word_1 to all hits of word_2

Ratio of EQUAL_BY_STRING hits of word_2 to all hits of word_2

Ratio of EQUAL_BY_STRING hits of word_3 to all hits of word_3

Ratio of EQUAL_BY_STRING hits to all hits for all words

Ratio of word_1 hits to all hits

Ratio of word_2 hits to all hits

Ratio of word_3 hits to all hits

Average distance between EQUAL_BY_STRING hits of two max-weighted words in one break

Max prefix length(in words) of exact(by position and form) match

Average distance between word_1 and word_2 in a break

Max average distance between word_i and word_i+1 in a break

Max prefix length(in words) of exact(by position) matches

Image porn classifier output

Fraction of hosts with Turkish language

Fraction of hosts with Russian language

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Feature from utracker

Bocm plain used in fastrank for all ann hits

Bocm double used in fastrank for all ann hits

BfExact used in fastrank for all ann hits

ImageDbm25 used in fastrank for all ann hits

LargestSyInexactGroup used in fastrank for all ann hits

Sum of weights for found words in indexkey/inv

Linguistic boosting factor. Type of extensions: XFIMGCLICS. Aggregation on all extensions. The greatest value of the factor. It is considered according to the heading of the document. The degree of coating of the words of the request is accurate to the form (without synonyms).

Linguistic boosting factor. Type of extensions: XFIMGCLICS. Aggregation on all extensions. The greatest balanced value of the factor. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation on all extensions. The greatest balanced value of the factor. It is considered according to the heading of the document. The BCLMPLANEPROXIMITY15W0SIZE1 algorithm: uses BCLM with free weighing if there are several words, if the word is one, then the sum of hits is used as a type of coincidence. Normalization coefficient 0.001.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation by TOP-10 (by the value of the factor) extensions. Nimenest, balanced meaning of the factor. Normalized for the total weight of extensions. It is considered according to the heading of the document. The algorithm for aggregation of words weights is BCLMMIXPLAIN: a linear mixture of annotation BCLM weights and balanced Positionless weights of the word, then the former meters are aggregated through BM15. Normalization coefficient 10^(-5).

Linguistic boosting factor. Type of extensions: QFUF. Aggregation on all extensions. The greatest balanced value of the factor. It is normalized for the maximum weight of expansion. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: XFDTSHOW. Aggregation on all extensions. The greatest balanced value of the factor. It is normalized for the maximum weight of expansion. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

The normalized amount of the scales of the words of the request that met in the text of the document.

The normalized amount of the scales of the words of the request that Equal_by_String in the text of the document.

The normalized amount of the scales of the words of the request that Equal_by_String in the annotations to the document.

The normalized scales for the IFILTRETRATIONMODEL words of the request that met in the text of the document.

The normalized scales for the IFILTRETRATIONMODEL words of the request that met in anotages to the document.

The normalized scales for the IFILTRETRATIONMODEL Words of the request, which Equal_by_String in the text of the document.

The normalized scales for the IFILTRETRATIONMODEL Words of the request, which Equal_by_String in the annotations to the document.

The value of the DP classifier is used to filter on average

The result of the text classifier of long views via VowPal Wabbit

The result of the aggregated tin classifier is used on average to determine tin queries

Previous factors - fixed

The result of the Suggestive text classifier via VowPal Wabbit

The result of the text classifier of porn viapal wabbit

The result of the text classifier of tin according to Vowpal Wabbit

Image porn classifier output

Fraction of hosts with Turkish language

Fraction of hosts with Russian language

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: the greatest balanced value of the factor; It is normalized for the maximum weight of expansion. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation by TOP-10 (by the value of the factor) extensions. Type of aggregation for extensions: the smallest balanced value of the factor; The maximum weight of the extension. It is considered according to the heading of the document. The algorithm for aggregation of words weights is BCLMMIXPLAIN: a linear mixture of annotation BCLM weights and balanced Positionless weights of the word, then the former meters are aggregated through BM15. Normalization coefficient 10^(-5).

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation by TOP-10 (by the value of the factor) extensions. Type of aggregation for extensions: an abstract by square of expansion weight, multiplied by the value of the factor; normalized for the total weight of extensions. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: the greatest balanced value of the factor; It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: the greatest value of the factor; It is considered according to the heading of the document. The weight of the hit is multiplied by 1/ (1 + the position of the word in the sentence) an algorithm for aggregation of the scales of words: BM15. Normalization coefficient 0.01.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: the greatest value of the factor; It is considered according to the heading of the document. The degree of covering the words of the request in the exact form.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation by TOP-10 (by the value of the factor) extensions. Type of aggregation for extensions: the smallest balanced value of the factor; It is considered according to the heading of the document. The degree of coating of the words of the request is accurate to the form (without synonyms).

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: the greatest balanced value of the factor; normalized for the total weight of extensions. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: an abstract by square of expansion weight, multiplied by the value of the factor; normalized for the total weight of extensions. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation on all extensions. Type of aggregation for extensions: the greatest balanced value of the factor; It is normalized for the maximum weight of expansion. It is considered according to the heading of the document. Algorithm Cosinematchmaxpredical.

Linguistic boosting factor. Type of extensions: Kinopoisksuggest (extensions of the textual orgate to text saddles). Aggregation by TOP-10 (by the value of the factor) extensions. Type of aggregation for extensions: the smallest balanced value of the factor; normalized for the total weight of extensions. It is considered according to the heading of the document. The average weight of the anntations among those in which the request was an accurate tuning.

Document from porn kitski

Fast document

Commercial page (Classifier Savina)

The page has a about 'payment SMS'.

Stage of the page

Pornography of the page

On the Porn Advertising page

The poetry of the document

The maximum poetry of the quatrain

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

An indicator of the unnaturalness of the text from the point of view of the Russian language. The number of bad pairs of words in the text, transferred to the segment [0.1] according to the Z/(Z+10) formula

The proportion of bad steam among all found in the table: Z/(X+1), where Z is the number of bad couples in the text, and X is (http://wiki.yandex-team.ru/evgenijgrechnikov/testsynonimizers of 2000-navigable )) steam

The number of Latin letters in the text (not counting the markings) driven into [0.1] formula n/(n+100)

The number of words in the text (the word is what the lemmeter selected) is displayed in [0.1] according to the formula x/(x+a)

The number of words of the Russian language in the title

The average length of the word

The percentage of the number of words inside the tag <a> .. </a> from the number of all words

The percentage of the number of words outside the tags (outside the brackets <>) from the number of all words

The percentage of the number of words, which are 200 the most frequent words of the language, from the number of all words of the text

The number used in the text 500 of the most popular words of the language, divided by 500

Logarithm of average geometric probabilities of trigrams in the text. (the probability of a trigram - the number of its meetings in the text, divided by the number of all trigrams) is displayed in [0.1] according to the formula -x (x+a)

Logarithm of the average geometric conditional probabilities of trigrams. The conditional probability of a trigram is its probability, divided by the probability of a bigram from the first two words

The share of different parts of speech in the text. The share of numerals (among all words that managed to recognize part of the speech)

The share of particles

The share of pronoun adjectives

The proportion of pronoun nouns

The share of verbs

The share of words that can be both masculine nouns and nouns of the feminine, but not of the middle kind, among all nouns (examples: 'hummingbirds' are an example of an indefinite kind that can be determined in two ways, 'Alexander' is homonymy).

The size of the largest text segment (from the factor [18] puretext)

Number of letters in the AUX segment

The number of spaces in the AUX segment

The number of commas in the Content segment

BM25 page title by its text

The greasy of the page

Page - '404' (share of tokens '404' in relation to the total number of tokens on the page)

Long text without links.

The language of the document is Russian.

Logarithm of the number of shingles on which this document is not unique

Logarithm of the number of shingles on which this owner of the document is recognized as the author

Logarithm of the number of shingles in the document added by the owner of the site as original texts in ((http://wiki.yandex-team.ru/jandekspoisk/jekosistema/marketingPr/webmasters/plan/vtorcontect of originality plugin)). It does not participate in the formula, it is needed to disconnect the takes

The average weight of non-ugly shingles of this document

The average filtered number of sources of authorship of the document. It does not participate in the formula, it is needed to disconnect the takes

The document contains user review/comment

The document has a direct link to the file

The document has a link to filehosting

The share of the words of the document from the segments with Score> 2.

The quality of the text. It is considered a rather complex formula

Text quality (classifier Alekseev)

Document length in sentences

Document type - HTML

Document language - English

The language of the document is Cyrillic

The popularity of the language of the document. Number from 0 to 1. (http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayaformula/tekushhiekomponenty/languaguaguagepopalarity)))))))

The function of the credibility of the distribution of years in the document. Temporarily disconnected

The arithmetic mean position of dates in the document. Temporarily disconnected

Max cosine similarity between document and user history with clicks dwelltime > 30sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime > 60sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime > 120sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime > 180sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime > 360sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime <= 30sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime <= 60sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime <= 120sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime <= 180sec, by realtime user_history

Max cosine similarity between document and user history with clicks dwelltime <= 360sec, by realtime user_history

Avg by top-5 maximum cosine similarity between document and user history with clicks dwelltime > 30sec, by realtime user_history

Avg by top-5 maximum cosine similarity between document and user history with clicks dwelltime > 60sec, by realtime user_history

Avg by top-5 maximum cosine similarity between document and user history with clicks dwelltime > 120sec, by realtime user_history

Avg by top-5 maximum cosine similarity between document and user history with clicks dwelltime > 180sec, by realtime user_history

Avg by top-5 maximum cosine similarity between document and user history with clicks dwelltime > 360sec, by realtime user_history

Avg by top-10 maximum cosine similarity between document and user history with clicks dwelltime > 30sec, by realtime user_history

Avg by top-10 maximum cosine similarity between document and user history with clicks dwelltime > 60sec, by realtime user_history

Avg by top-10 maximum cosine similarity between document and user history with clicks dwelltime > 120sec, by realtime user_history

Avg by top-10 maximum cosine similarity between document and user history with clicks dwelltime > 180sec, by realtime user_history

Avg by top-10 maximum cosine similarity between document and user history with clicks dwelltime > 360sec, by realtime user_history

Avg by top-15 maximum cosine similarity between document and user history with clicks dwelltime > 30sec, by realtime user_history

Avg by top-15 maximum cosine similarity between document and user history with clicks dwelltime > 60sec, by realtime user_history

Avg by top-15 maximum cosine similarity between document and user history with clicks dwelltime > 120sec, by realtime user_history

Avg by top-15 maximum cosine similarity between document and user history with clicks dwelltime > 180sec, by realtime user_history

Avg by top-15 maximum cosine similarity between document and user history with clicks dwelltime > 360sec, by realtime user_history

Stript priority for TR is a text priority - there are all the words of the request somewhere in the document (while they pass contextual restrictions on the request, for example, both words DB in one sentence).

The presence of an accurate phrase (request text) in the header (more precisely, in the first sentence of the document).

The document passed Softand on the restrictions of the syntactic sorcerer. Only for documents with textual relevance. For monosyllabic requests, always 1.

Text relevance (maxfreq is the frequency of the most frequent word that makes sense of the length of the document).

The weight of the words of the request that is in the text.

Logdwelltime by the VEB model DSSM, trained in reformulations. It is also used in the ranking of ether.

Win (click longer than 60 seconds) on the DSSM video model trained in reformulations.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation on all extensions. The greatest value of the factor. It is considered according to the contents of the document. The degree of covering the words of the request in the exact form.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation by TOP-10 (by the value of the factor) extensions. Nimenest, balanced meaning of the factor. Normalized for the total weight of extensions. It is considered according to the contents of the document. The algorithm for aggregation of words weights is BCLMMIXPLAIN: a linear mixture of annotation BCLM weights and balanced Positionless weights of the word, then the former meters are aggregated through BM15. Normalization coefficient 10^(-5).

The factor about the proximity of the request and text of the document. It differs from BCLM in that the weights of all words are considered the same. It is also used in the ranking of ether.

DBM (BM25 with machine-like words) exclusively in numbers.

Simple BOCM gluing Links.

BOCM for gluing Links, calculated only on the first sentences with hits and all forms of entering are considered equivalent.

The best BOCM among all links, such as Title (analogue of TOCM), calculated as follows: only sentences with hits are considered and all forms of entry are considered equivalent.

The new DBM only in terms of gluing links (differs from DBMNUMBERS [157] only constants and completely clogs it).

Coating trigrams of Title trigrams. It is also used in the ranking of ether.

Coating trigrams of a query trigrams of Urla.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation on all extensions. The amount by the square of the expansion, multiplied by the value of the factor. Normalized for the total weight of extensions. It is considered according to the heading of the document. PerwordCmmaxMatchMin algorithm: At least according to the maximum of the CMMAXMATCH weight abstracts.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation by TOP-10 (by the value of the factor) extensions. A suspended sum of the Libra of factors. Normalized for the total weight of extensions. It is considered according to the heading of the document. The degree of covering the words of the request in the exact form.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation on all extensions. The sum of the scales of factors. The number of extensions. It is considered according to the contents of the document. The sum of the scales of words, balanced by the weight of the annotation, is normalized for the sum of the scales of words. Only annotations are calculated, on which the sum of words of words is more than 80%.

Linguistic boosting factor. Type of extensions: QFUF. Aggregation by TOP-10 (by the value of the factor) extensions. A suspended sum of the Libra of factors. Normalized for the total weight of extensions. It is considered according to the contents of the document. The maximum weight of the annotation, the prefix of which contains the words of the request in the same order (with an accuracy to the form).

The average value of the webtrdocquorum factor is average.

The average value of the Bocmfull factor is average.

The average value of the Bestfirsthitsentencetocm factor is average.

The average value of the Bocmfull factor in PRS

The average value of the Bestfirsthitsentencetocm factor in PRS

The mid -sequential deviation of the Bocmfull factor in PRS

The mid -sequential deviation of the Titlerigramsinquery factor in PRS

COF-T Variations Pyrson Dul factor Titlecovering in PRs

CoEF-T of Pieron Variation for the Titlerigramsinquery factor in PRS

Resid for Webtrdocquorum Factor in PRS

Resid for Titlecovering Factor in PRS

FRACT for the Titlerigramsinquery factor in PRS

Logdwelltime by the VEB model DSSM. It is also used in the ranking of ether.

Logdwelltime by the VEB model DSSM.

The similarity of the T2q vehicles of the Title and the request a la Klakhman, trained by clicks

DSSM with PlayerDepth Target on the deep click pool video. It is also used in the ranking of ether.

DSSM with PlayerDepth Target on the deep click of a video at L2 Stages.

The factor for the original request. It is considered according to the contents of the document. Algorithm: AVGPERGRAMAMAVGVALueany.

The factor for the original request. It is considered according to the contents of the document. The BCLMPLANEPROXIMITY15W0SIZE1 algorithm: uses BCLM with free weighing if there are several words, if the word is one, then the sum of hits is used as a type of coincidence. Normalization coefficient 0.01.

The factor for the original request. It is considered according to the contents of the document. Algorithm for aggregation of the scales of words BOCM15. Normalization coefficient 0.01.

The factor for the original request. It is considered according to the contents of the document. The degree of covering the words of the request in the exact form.

The factor for the original request. It is considered according to the heading of the document. The degree of coating of the words of the request (all types of hits).

Document from porn kitski

Document from porn kitski

A document from a commercial clay. Not used (depreded)

A document from a commercial clay. Not used (depreded)

Fast document

Fast document

The page title contains commercial vocabulary. Not used (depreded)

The page title contains commercial vocabulary. Not used (depreded)

Commercial page (Classifier Savina)

Commercial page (Classifier Savina)

On the page there is about 'Payment SMS '.

On the page there is about 'Payment SMS '.

Stage of the page

Stage of the page

Pornography of the page

Pornography of the page

On the Porn Advertising page

On the Porn Advertising page

The poetry of the document

The poetry of the document

The maximum poetry of the quatrain

The maximum poetry of the quatrain

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

Show how much the text is unnatural from the point of view of the Russian language. Assessment of how much the text of the document can be considered as a generated synonymizer or automatic. ((http://wiki.yandex-team.ru/jandekspoisk/kachestvopoiska/obshayAfermula/tekushhiekomponenty/antispam?v=1il#h58953-2 more))

An indicator of the unnaturalness of the text from the point of view of the Russian language. The number of bad pairs of words in the text, transferred to the segment [0.1] according to the Z/(Z+10) formula

An indicator of the unnaturalness of the text from the point of view of the Russian language. The number of bad pairs of words in the text, transferred to the segment [0.1] according to the Z/(Z+10) formula

The proportion of bad steam among all found in the table: Z/(x+1), where Z 342 200 223 The number of bad couples in the text, and X 342 200 223 number ((http: //wiki.yandex- Team.ru/evgenijjjgrechnikov/testSynonimizers 2000-navigable)) steam

The proportion of bad steam among all found in the table: Z/(x+1), where Z 342 200 223 The number of bad couples in the text, and X 342 200 223 number ((http: //wiki.yandex- Team.ru/evgenijjjgrechnikov/testSynonimizers 2000-navigable)) steam

The number of Latin letters in the text (not counting the markings) driven into [0.1] formula n/(n+100)

The number of Latin letters in the text (not counting the markings) driven into [0.1] formula n/(n+100)

The number of words in the text (the word is what the lemmeter selected) is displayed in [0.1] according to the formula x/(x+a)

The number of words in the text (the word is what the lemmeter selected) is displayed in [0.1] according to the formula x/(x+a)

The number of words of the Russian language in the title

The number of words of the Russian language in the title

The average length of the word

The average length of the word

The percentage of the number of words inside the tag <a> .. </a> from the number of all words

The percentage of the number of words inside the tag <a> .. </a> from the number of all words

The percentage of the number of words outside the tags (outside the brackets <>) from the number of all words

The percentage of the number of words outside the tags (outside the brackets <>) from the number of all words

The percentage of the number of words, which are 200 the most frequent words of the language, from the number of all words of the text

The percentage of the number of words, which are 200 the most frequent words of the language, from the number of all words of the text

The number used in the text 500 of the most popular words of the language, divided by 500

The number used in the text 500 of the most popular words of the language, divided by 500

Logarithm of average geometric probabilities of trigrams in the text. (the probability of a trigram - the number of its meetings in the text, divided by the number of all trigrams) is displayed in [0.1] according to the formula -x (x+a)

Logarithm of average geometric probabilities of trigrams in the text. (the probability of a trigram - the number of its meetings in the text, divided by the number of all trigrams) is displayed in [0.1] according to the formula -x (x+a)

Logarithm of the average geometric conditional probabilities of trigrams. The conditional probability of a trigram is its probability, divided by the probability of a bigram from the first two words

Logarithm of the average geometric conditional probabilities of trigrams. The conditional probability of a trigram is its probability, divided by the probability of a bigram from the first two words

The share of different parts of speech in the text. The share of numerals (among all words that managed to recognize part of the speech)

The share of different parts of speech in the text. The share of numerals (among all words that managed to recognize part of the speech)

The share of particles