The valency behavior (argument structure) of lexical items is so varied that it cannot be described by general rules and must be captured in lexicons separately for each lexical item. For verbs, lexicons typically describe only unmarked usage-the active form-while natural languages allow for certain regular changes in the number, type and/or realization of complementations (e.g. passivization). Thanks to their regularity, such changes may be described in a separate rule component of the lexicon; however, they are typically seen in many but not all verbs and their applicability to a given lexical unit (verb meaning) is not predictable from its valency alone. In this paper, we describe our initial experiments with using a large morphologically annotated corpus of Czech for determining which diatheses are applicable to a given lexical unit.
Valency refers to the argument structure of lexical units.1
In the Functional Generative Description (FGD), valency
belongs to the so-called tectogrammatical layer [
vyzvat neˇkohoADDR.Acc, aby se uklidnilPAT.aby-Clause ‘to ask somebody to calm down’ vyzvat neˇkohoADDR.Acc na soubojPAT.na+Acc
‘to challenge somebody to a duel’ • apelovat ‘to appeal’
ACTNom ADDRna+Acc PATaby, at’, že 1Whereas the term lexeme roughly corresponds to a dictionary verb item with all its meanings, by a lexical unit (LU) we refer to a verb in a given meaning. See Section 3.1 for more details.
2The frames and examples are taken from Vallex 2.6, http:// ufal.mff.cuni.cz/vallex/2.6/data/html .
apelovat na kolegyADDR.na+Acc, aby práci dokoncˇiliPAT.aby-Clause vcˇas
‘to appeal to his colleagues to finish the work in time’ • apelovat ‘to put emphasis’
v jeho rodineˇ se stále apeluje na morálkuPAT.na+Acc ‘in his family emphasis is always put on morality’ The above examples demonstrate how valency behavior varies even among semantically close lexical units (LUs), both when they belong to the same lexeme and when they belong to different lexemes. It must therefore be captured for each lexical unit of a verb separately in the form of a lexical entry listed in the valency lexicon. On the other hand, certain changes in the valency structure are regular and can be described in the form of rules which can be specified in a separate component of the lexicon. Such changes are typically seen in many but not all verbs and their applicability to a given lexical unit is not predictable from its valency frame alone.
A lexical entry does not list all of its possible forms but only one—usually the structure corresponding to the active form of the verb, which is considered to be its unmarked use—and a list of rules for creating other possible structures (the marked uses). This description is both economical (less space is needed for storing the information about all available realizations of the LU) and linguistically adequate (it captures generalizations which would not be obvious if all possible surface forms were listed).
Valency lexicons are created with many applications in mind: they help to maintain consistency of corpus annotation, provide syntactic and morphological information during parsing and natural language generation, and may even prove useful in word sense disambiguation and machine translation; moreover, lexicon data is consulted by linguists during their theoretical research and provides useful information for students of Czech. All of these tasks involve actual occurrences of the valency patterns in the natural language, and so the unmarked structures from the lexicon need to be converted into all structures that may appear in the actual data.
A rule based approach to creating derived valency
structures has already been used during the annotation of the
Prague Dependency Treebank3 (PDT) [
Because correctness of the underlying PDT data was assumed, the rules were allowed to heavily over-generate. For example, “passive” frames of the verb mít ‘to have’ were generated although, in reality, it does not form passive in Czech. While this is a reasonable strategy for consistency checks of annotated data, other tasks that utilize a valency lexicon would benefit from lists of diatheses applicable to any given lexical unit. Manual annotation provides a number of examples of lexical units occurring in different types of diatheses; however, the size of the tectogrammatically annotated PDT data is too small, so we cannot make any conclusions from the fact that a lexical unit does not occur in a diathesis. Therefore, we are trying to draw evidence from a much larger, automatically morphologically annotated corpus. We have decided to use SYN, a non-referential corpus of 1,300 million automatically morphologically tagged words.
For Czech, [
Regular changes of the valency structure of a lexical unit,
in the English-language literature usually called
alternations, typically allow the speaker to express the same
situational meaning (i.e., propositional content characterized
by the set of situational participants) in different ways
that result in different perspectives from which the
situation is viewed. Alternations have already been studied
extensively for several decades [
Here we focus on diatheses—specific relations stemming from the changes in the linking of situational participants, valency complementations and surface syntactic positions. Diatheses belong to the group of grammaticalized alternations: they are realized by the use of specific morphological and/or syntactical means, including the grammatical category of voice of the verb and the surface forms of the complementations. They relate different surface syntactic structures of a single lexical unit of a verb. They also belong to the group of conversive alternations: the transformation acts as a permutation on the assignment of valency complementations to surface syntactic positions, typically shifting Actor away from the prominent subject position and filling it with some other complementation. 2.1
In this section, we summarize the description of Czech
diatheses as given by [
The unmarked member of the diathesis. The unmarked usage is described in the lexical entry in the lexicon. The verb appears in an active form or as an infinitive; the complementations are realized in the forms specified for them in the lexicon entry. All complementations specified in the entry as obligatory are present on the tectogrammatical layer, although some of them may be elided in the surface realization of the sentence (if their value is either clear from the context or general); inner participants4 that are not specified in the lexical entry must not appear as arguments of the verb, but free modifications may.
1. passive diathesis (periphrastic passive) e.g. Neustále jsem byl neˇkým vyzýván, abych se legitimoval. ‘All the time - I was - someoneInstr - asked to show my ID.’ – ‘I kept being asked to show my ID’ The form of the verb in this diathesis consists of the past participle of the main verb + the verb být ‘to be’ (in a finite or infinite form). The subject slot of the passive construction either remains empty, or it is filled by a complementation which originally filled an object slot (typically that of an Accusative object, but realization through infinitive, clause, genitive, or phrase jako+Acc ‘as something’ is also possible); if the complementation is expressed as a noun phrase, it is turned into the Nominative case. The Actor (which 3See http://ufal.mff.cuni.cz/pdt2.5/ for information about the current version.
4Inner participants are complementations with either of the functors ACTot, PATient, ADDRessee, EFFect or ORIGin. in the active construction fills the subject slot) may be realized either in the Instrumental case, or as a prepositional phrase od+Gen ‘by/from+Gen’. 2. resultative diathesis with the auxiliary verb být ‘to be’ e.g. Jídlo je uvarˇeno. ‘The food is cooked.’ This form of resultative diathesis differs from the periphrastic passive only in meaning, not in the surface form or structure. In many cases, it is not clear which of the two possible readings the speaker has in mind. For example, the sentence okno je otevrˇeno may be interpreted as a case of the resultative diathesis, describing a state, i.e. ‘the window is (already) open’, or as a case of the passive diathesis, describing an event, i.e. ‘the window is (being) opened’. This ambiguity is called event–state homonymy in Czech linguistics. Because it is so common, we assume that the passive diathesis is possible whenever the resultative diathesis is possible and vice versa.
Moreover, Czech also exhibits a competition between past participles and deverbal adjectives. The kind of deverbal adjectives that we have in mind are formed by adding vowel endings to past participles; both the participle and the adjective can then be used to express the resultative meaning, while only the participle can be used to express the passive meaning. On one hand, this interchangeability of the “short” (participle) and “long” (adjectival) forms is often used as a guideline in determining whether a given sentence should be considered resultative—if the participle can be replaced with the adjective, the resultative interpretation is valid. On the other hand, participle forms are sometimes used in purely adjectival meaning, such as in the sentence stále ješteˇ nebyl najeden ‘he still was not full’, which features the word form najeden (past participle of the reflexive verb najíst se ‘satiate oneself, eat so much that one is full’). If we were to read this as a diathesis, this would have to be a case of periphrastic passive or of the resultative diathesis with auxiliary verb být ‘to be’ formed from the sentence najedl se ‘he ate to be full’. However, it is not possible that the same complementation would fill the subject position in both the active and the passive/resultative diathesis. We have to read this sentence as a sentence with the adjective najedený ‘full, satiated’. 3. possessive resultative diathesis e.g. Maminka má jídlo uvarˇeno. ‘The mother has the food cooked.’ In this type of construction, auxiliary verb mít ‘to have’ is used together with the past participle of the main verb.
Note that the conversive aspect is crucial for our theoretical concept of a diathesis. For example, only one of the two possible readings of the example sentence above is considered to be a diathesis: Mamince vcˇera jídlo prˇipravila teticˇka. Maminka má tedy již jídlo uvarˇeno. ‘The aunt has prepared the food for the mother yesterday. Therefore, the mother has the food cooked already.’ This case is considered to be a diathesis, because the Actor of the first sentence (the aunt) has moved away from the subject position (and is not expressed in the resultative variant at all).
Maminka varˇila celé dopoledne a nyní již má jídlo uvarˇeno. ‘Mother has been cooking all morning and now she already has the food cooked.’ This case is not considered to be a diathesis, because the same complementation is corresponding to the subject in both cases.
In practice, however, it is often impossible to
distinguish between the two readings (Panevová et al. [
5. deagentive diathesis (reflexive passive) e.g. Varˇilo se tu pro emigranty. ‘Cooked - reflexive - here - for - emigrants.’ – ‘It was cooked here for emigrants.’ The only surface marks of this diathesis are a verb in the third person (agreeing with the subject in number and gender, or singular neuter for subjectless sentences) and the free reflexive morpheme se. The Actor is not expressed in this kind of construction at all. Rules for forming the deagentive diathesis have almost the same conditions as the rules for forming the passive diathesis (and both types can be applied to almost any frame), and also the ways in which the Patient, Addressee or Effect are moved into the subject position are the same. However, the sets of verbs that allow the two diatheses are different. 6. dispositional diathesis (mediopassive) e.g. Dobrˇeadverb se (miDat) tu hrál tenis. ‘Well - reflexive - to-me - here - played - tennis.’ – ‘For me, this was a good place to play tennis. I enjoyed playing tennis here.’ A characteristic feature of the dispositional diathesis is an evaluative element, usually an adverb such as dobrˇe ‘well’, pomalu ‘slowly’. The verb form is the same as in the deagentive diathesis, i.e. a third person verb agreeing with the subject (or signular neuter in subjectless sentences) + reflexive partice se. However, the Actor may be expressed on the surface in the dative case.
Although people do not have difficulties
distinguishing between the deagentive and the dispositional
diathesis, the difference is hard to grasp for an
automatic procedure when the Actor in the dative is
elided from the sentence (in a sample from the
corpus SYN2005 cited in [
7. causative diathesis e.g. Nechal si od Gesy varˇit. ‘He let - reflexive - by Gesy - cook.’ – ‘He let Gesy cook for him.’ This verbal form roughly corresponds to the English ‘have something done’. For lexicographic purposes, we view the causative diathesis as a separate sense of the verbs nechat/dát ‘let/give’. One reason for this treatment lies in the fact that two Actors appear in the construction - the Actor of the verb nechat/dát and the Actor of the dependent infinitive. 3
Our proposal is primarily formulated for the purpose of the
description of valency in valency lexions of Czech verbs
built within the framework of the Functional Generative
Description (FGD). We are working with two lexicons,
VALLEX 2.65, see [
Both lexicons are divided into two components: a data component and a rule component. 3.1
The data component consists of word entries corresponding to verb lexemes. Lexeme is an abstract twofold data structure which associates lexical form(s) and lexical unit(s). Lexical forms are all possible manifestations of a lexeme in an utterance (e.g. a lemma or a group of lemmas,7 all morphological forms of these lemmas, and their reflexive and irreflexive forms). All lexical forms of a lexeme are represented by its lemma(s).
In the lexicon, each lexical unit (a sense of a verb) is characterized by a gloss (a verb or a paraphrase roughly synonymous with the given sense) and by example(s) (sentence fragment(s) containing the given verb used in the given sense). The core information on valency characteristics of a lexical unit is encoded as (exactly one) valency frame reflecting the unmarked (active) use of the verb. 8
In an ideal model of the lexicon, information on the possible application of diatheses is stored in each lexical unit in a special attribute -diat. This attribute has not been implemented in either of the lexicons yet, but the attribute -rfl (reflexivity) that is present in Vallex overlaps with the proposed attribute -diat to a certain extent. It lists possible syntactic functions of the relexive morpheme se/si. The values of the -rfl attribute are pass for reflexive passives in verbs with accusative complements, pass0 for reflexive passives in intransitive verbs, and cor4 and cor3 for cases where se and si fill the position of an object in accusative (cor4) or in dative (cor3), showing that the subject is performing an action on itself. If the verb allows for any of these constructions with se/si, the possibility has been exemplified with made up examples (the annotators simply converted the examples given for the active diathesis into a passive/reciprocal construction). Many of these 5http://ufal.mff.cuni.cz/vallex/2.6/doc/home.html 6This is the version that has been published as part of the Prague Czech-English Dependency Treebank 2.0.; it is available from http://ufal.mff.cuni.cz/pcedt2.0/publications/ vallex3.xml and can be browsed at http://ufal.mff.cuni.cz/ lindat/PDT-Vallex.html.
7Vallex lexemes comprise perfective, imperfective and iterative variants, as well as spelling variants, so that the lexicon covers almost twice as many lemmas as lexemes. On the other hand, there is a one-to-one correspondence between lexemes and lemmas in PDT-Vallex. 8See the Introduction for more details about valency frames. lexemes7 lexical units (LU) lemmas (L) LUs (separated by lemma) reflexive nonreflexive, 0 occurrences of past participles (tag „Vs“) nonreflexive, some occurrences of Vs, 1 LU nonreflexive, 0 occurrences in sentences with “se” tagged as “P7-X4-*” examples do not sound natural. We hope that our methods will provide some more natural corpus examples. Moreover, we intend to cover other diatheses that the current annotation does not cover.
See Table 1 for counts of lexemes, lemmas and lexical units in both lexicons. Both lexicons are available in machine-tractable XML format and also as humanfriendly web pages. 3.2
The proposed rule component of the lexicon consists of a set of formal syntactic rules determining changes in the mapping of valency complementations onto surface syntactic positions. They make it possible to obtain all possible surface syntactic manifestations of lexical units of verbs (i.e., number of complementations, their types and possible morphological forms).
At present, we use transformational rules formulated for
the purposes of the description of diatheses in PDT-Vallex,
the lexicon of the Prague Dependency Treebank, see [
Due to the size of the lexicon, it is preferable to minimize the necessary manual work involved in augmenting the lexicon with information about applicable diatheses. Moreover, experience suggests that annotators tend to be positively biased towards assuming the applicability of the diatheses. Also, examples given by annotators tend to be contrieved/unnatural. To address these problems we would like to have a semiautomatic method which should, where possible • automatically decide whether a diathesis is applicable, • provide natural corpus examples of the diathesis to be included in the lexicon, and in uncertain cases • provide corpus evidence on the basis of which the annotators can quickly make the decision.
Below we describe such a method in some detail. The
method works by iterating over the frames in three passes.
The first pass is a negative pass which filters out lexical
units where the diathesis is not applicable due to either
grammatical concerns or insufficient corpus evidence. The
second pass is a positive pass where lexical units with
sufficient evidence for applicability are dealt with. In the
final step, corpus evidence is gathered for the remaining
unclear lexical units. This evidence is then presented to the
annotator for a manual decision. If the second or third
phase yields a large number of examples, the automatic
method should also order them so that simple, clear
examples come first. The method of ordering corpus examples
used by [
Due to the difficulties in distinguishing some of the diatheses mentioned above, the proposed semi-automatic procedure only strives to identify cases of the following diatheses: periphrastic passive, possessive resultative, recipient, and deagentive (reflexive passive). 4.1
In the negative pass we use various methods for excluding inapplicable diatheses. In some cases, we exclude whole lexemes (reflexive verbs and lexemes for which no corpus evidence suggesting a possibility of the diathesis was found); the rule-based exclusion, on the other hand, may exclude some lexical units of a lexeme while other proceed into the next phase.
Reflexives. We assume that none of the diatheses is applicable to a lexeme with a reflexive lemma. These cases include reflexiva tantum ( bát se ‘to fear’) and derived reflexives ( šírˇit se ‘to spread (itself)’). This assumption covers 1528 out of 4789 lemmas occurring in Vallex 2.6, and 1590 out of 7116 verb lemmas occurring in PDT-Vallex 2.0. It can be seen from Table 1 that so far this is the most effective step in the negative pass.
We are aware of the fact that this assumption is only approximately valid. According to [9, p. 93], derived reflexives do not form passive (neither periphrastic nor reflexive), but some reflexiva tantum do; [21, p. 43] is only aware of two reflexive verbs that form a periphrastic passive, the reflexiva tantum tázat se ‘to ask’ and obávat se ‘to fear’, and otherwise assumes that reflexive verbs do not form passives. While [23, p. 124] discusses the limited possibility of forming the reflexive passive of reflexiva tantum, she also gives a (made up?) example of a stylistically non-neutral sentence smálo se, až se plakalo ‘it was laughed so much that it was cried’ with reflexive passive of reflexiva tantum.
We have found several other cases where reflexive verbs
form a diathesis:
1. To se lehko pamatuje. ‘This is easy to remember. It is
easy to remember it.’ (derived from pamatovat si ‘to
remember’)
Na to se lehko zvykne. ‘This is easy to get used to. It
is easy to get used to it.’ (derived from zvyknout si ‘to
get used to’)
Na všechno se zvykne. ‘Everything gets used to.
People get used to everything.’ (derived from zvyknout si
‘to get used to’)
This usage is almost idiomatic; the first two
examples are cases of the dispositional diathesis, and the
third seems to be derived from it. We expect that
further research will show that this type of construction
is productive even among reflexive verbs.
2. Prezident Václav Havel je lidmiInstr nejméneˇ oblíben
od té doby, kdy zacˇal prezidentovat. ‘President
Václav Havel is by-the-people least liked since he
started presidenting.’ – ‘President Václav Havel’s
popularity is the least since he became president.’
(derived from oblíbit si ‘get to like’)
The corpus contains many instances of je oblíben ‘is
liked’ which can be easily analyzed as cases of the
verbo-nominal predicate být oblíben(ý) ‘to be liked’,
not as passive. However, this particular sentence also
contains the Actor lidmi ‘by the people’ in the
Instrumental case, which is typical of a passive
construction. One option is to claim that lidmi is a valency
complementation of the adjective (oblíben kým ‘to be
liked by whom’). The other option is to admit that
this is a case of a passive construction, possibly
related to the historical existence of the verb oblíbit ‘get
to like’ without a reflexive particle (as documented in
[
The possibility to form passives of reflexive verbs is certainly an interesting area for further research.
Rule-based exclusion. Some of the diatheses require a
particular grammatical structure to be applicable. It is
therefore possible to exclude frames where this structure
is absent. Here we rely on [
Corpus-based exclusion. We start with a very naive implementation of this step, excluding the applicability of the diatheses for whole lexemes. Applicability of the diatheses formed with the past participle may be ruled out if the past participle is not found in the corpus. Similarly, we may exclude the applicability of the reflexive passive whenever the verb does not appear in the same sentence as the particle se anywhere in the corpus. Table 1 shows that we need to refine these criteria, especially for the exclusion of the reflexive passive.
The mere presence of the se token is not necessarily indicative of the given diathesis. First of all, the se need not be a particle at all, e.g. in the sentence tancˇil se ženou ‘he danced with (his) wife’, the word se ‘with’ is in fact a preposition. (The morphological tagger used to tag the corpus SYN is accurate enough to overcome this ambiguity.) But even as a particle, se can be part of a different grammatical structure, e.g. in the sentence snažil se tancˇit ‘he tried reflexive to dance’ the word se belongs to the reflexivum tantum snažit se ‘to try’, not to the verb tancˇit ‘to dance’. Limiting the search to segments enclosed by punctuation might exclude some genuine examples of diatheses: minule se, pokud si pamatuji, tancˇilo až do rána ‘the last time reflexive, as far as I remember, danced until morning’ – ‘as far as I remember, the last time dancing continued until morning’; we do not want to take the risk of missing some existing evidence already in this phase. Thus, naive corpus search does not suffice to exclude more than a tiny number of verbs (as can be seen from Table 1): auxiliary methods such as (shallow) parsing or at least clause detection are needed. The Prague Dependency Treebank is too small for the purpose of rejecting the applicability of a diathesis, especially if it is rare such as the possessive resultative. (E.g., there are only about 70 instances of the possessive resultative in the whole of PDT.) Corpus SYN, albeit more adequate in size, is not parsed, so a different, inherently less reliable method must be used. We could, for example, base our decision as to whether the se is connected to the relevant verb or not on their distance in the sentence.
Combination of the rule-based and corpus-based method. The rules allow us to identify frames describing structures in which a given lexical unit may appear in a diathesis. These structures can be turned automatically into patterns for corpus search. In general, no significant conclusions can be drawn from the fact that the resulting search does not produce any results: Czech is a pro-drop language, so even semantically obligatory elements can be elided in the actual sentence. Only the dispositional diathesis contains an element that is obligatory on the surface, but the range of possible morphemic realizations of this evaluative element needs to be further researched. 4.2
In the positive pass we search the corpus for evidence showing the applicability of a given diathesis. Especially an occurrence of a past participle is indicative of a diathesis (although concerns about the competition between past participles and adjectives need to be addressed). The three kinds of diatheses with past participle forms that we intend to distinguish—periphrastic passive, possessive resultative and recipient passive—moreover differ in the auxiliary verbs. Therefore we assume that instances of past participles found in the corpus can be assigned to a diathesis with a fair amount of certainty. The situation with the passive constructions built with the reflexive particle se is more complex, but the techniques developed for the first pass will hopefully help here as well.
The automatic method must be able to assign the evidence found to a particular diathesis and to a particular lexical unit (it does not suffice to know that a verb with many meanings appears in the passive diathesis in the given sentence; we are looking for examples which we can desambiguate). Sometimes, the first pass will give us a single candidate. In other instances, we apply the rules to the remaining frames, derive the description of the full structures corresponding to a diathesis, and then search the corpus for patterns with elements that are unique to only one of the candidates. 4.3
Finally, similar methods as in the second phase will be used, but examples with ambiguous status will be output. We expect that the examples will be automatically assigned to a diathesis with high precision. Thus, for each combination of a lexical unit and a diathesis that remain undecided after the previous pass, the system will be able to provide the annotator with a selection of sentences that could be instances of this LU in the given diathesis with high likelihood. The annotator will then either select a couple of examples that demonstrate the applicability of the diathesis, or will decide that the diathesis is not applicable to the given LU. 5
We introduced a (semi-)automatic method for
identifying lexical units that undergo individual diatheses, and we
have discussed some of the difficulties that stand in the
way of a fully automatic procedure. We have also shown
that the question whether a diathesis is applicable to a
lexical unit may be answered in several different ways:
• The least strict measure is the applicability of a rule
for forming the given diathesis. This is a necessary,
yet not sufficient condition. The rules have been
described in detail in [
The research reported in this paper was supported by the grant of the Czech Science Foundation GACˇ R No. P406/12/0557.
The first author was partially supported by the grant SVV-2013-267314.
This work has been using language resources developed and/or stored and/or distributed by the LINDAT-Clarin project of the Ministry of Education of the Czech Republic (project LM2010013).