Expletives in Universal Dependency

Gosse Bouma∗◦ Jan Hajic†◦ Dag Haug‡◦ Joakim Nivre•◦ Per Erik Solberg‡◦ Lilja Øvrelid?◦ ∗University of Groningen, Centre for Language and Cognition †Charles University in Prague, Faculty of Mathematics and Physics, UFAL ‡University of Oslo, Department of Philosophy, Classics, History of Arts and Ideas •Uppsala University, Department of Linguistics and Philology ?University of Oslo, Department of Informatics ◦Center for Advanced Study at the Norwegian Academy of Science and Letters

Abstract language that has expletives into a language that does not use expletives (Hardmeier et al., 2015; Although treebanks annotated according to the Werlen and Popescu-Belis, 2017). The ParCor guidelines of Universal Dependencies (UD) now exist for many languages, the goal of co-reference corpus (Guillou et al., 2014) distin- annotating the same phenomena in a cross- guishes between anaphoric, event referential, and linguistically consistent fashion is not always pleonastic use of the English pronoun it. Loaiciga´ met. In this paper, we investigate one phe- et al.(2017) train a classifier to predict the dif- nomenon where we believe such consistency ferent uses of it in English using among others is lacking, namely expletive elements. Such syntactic information obtained from an automatic elements occupy a position that is structurally parse of the corpus. Being able to distinguish ref- associated with a core argument (or sometimes an oblique dependent), yet are non-referential erential from non-referential noun phrases is po- and semantically void. Many UD treebanks tentially important also for tasks like question an- identify at least some elements as expletive, swering and . but the range of phenomena differs between Applications like these motivate consistent and treebanks, even for closely related languages, explicit annotation of expletive elements in tree- and sometimes even for different treebanks for banks and the UD annotation scheme introduces a the same language. In this paper, we present dedicated dependency relation (expl) to account criteria for identifying expletives that are ap- plicable across languages and compatible with for these. However, the current UD guidelines are the goals of UD, give an overview of exple- not specific enough to allow expletive elements to tives as found in current UD treebanks, and be identified systematically in different languages, present recommendations for the annotation of and the use of the expl relation varies consid- expletives so that more consistent annotation erably both across languages and between differ- can be achieved in future releases. ent treebanks for the same language. For instance, 1 Introduction the manually annotated English uses the expl relation for a wide range of constructions, Universal Dependencies (UD) is a framework for including clausal extraposition, weather verbs, ex- morphosyntactic annotation that aims to provide istential there, and some idiomatic expressions. useful information for downstream NLP applica- By contrast, Dutch, a language in which all these tions in a cross-linguistically consistent fashion phenomena occur as well, uses expl only for ex- (Nivre, 2015; Nivre et al., 2016). Many such ap- traposed clausal arguments. In this paper, we pro- plications require an analysis of referring expres- vide a more precise characterization of the notion sions. In co-reference resolution, for example, it of expletives for the purpose of UD treebank anno- is important to be able to separate anaphoric uses tation, survey the annotation of expletives in exist- of pronouns such as it from non-referential uses ing UD treebanks, and make recommendations to (Boyd et al., 2005; Evans, 2001; Uryupina et al., improve consistency in future releases. 2016). Accurate translation of pronouns is another challenging problem, sometimes relying on co- 2 What is an Expletive? reference resolution, and where one of the choices is to not translate a pronoun at all. The latter sit- The UD initiative aims to provide a syntactic uation occurs for instance when translating from a annotation scheme that can be applied cross- linguistically, and that can be used to drive se- pronoun, and there, identical to the non- mantic interpretation. At the clause level, it dis- proximate locative pro-adverb), (ii) nonref- tinguishes between core arguments and oblique erential (neither anaphoric/cataphoric nor dependents of the verb, with core arguments be- exophoric), and (iii) devoid of any but a vac- ing limited to subjects (nominal and clausal), ob- uous semantic role. As a tentative definition jects (direct and indirect), and clausal comple- of expletives, we can characterize them as ments (open and closed). Expletives are of inter- pro-forms (typically third person pronouns est here, as a consistent distinction between exple- or locative pro-adverbs) that occur in core tives and regular core arguments is important for argument positions but are non-referential semantic interpretation but non-trivial to achieve (and therefore not assigned a semantic role). across languages and constructions. Like the UD definition, Postal and Pullum(1988) The UD documentation currently states that emphasize the vacuous semantics of expletives, expl is to be used for expletive or pleonastic but understand this not just as the lack of semantic nominals, that appear in an argument position of a role (iii) but also more generally as the absence of predicate but which do not themselves satisfy any reference (ii). Arguably, (ii) entails (iii) and could of the semantic roles of the predicate. As exam- seem to make it superfluous, but we will see that it ples, it mentions English it and there as used in can often be easier to test for (iii). The common, clausal extrapostion and existential constructions, pre-theoretic understanding of expletives does not cases of true clitic doubling in Greek and Bul- include idiom parts such as the bucket in kick the garian, and inherent reflexives. Silveira(2016) bucket, so it is necessary to restrict the concept characterizes expl as a wildcard for any element further. Postal and Pullum(1988) do this by (i), that has the morphosyntactic properties associ- which restricts expletives to be pro-forms. This is ated with a particular grammatical function but a relatively weak constraint on the form of exple- does not receive a semantic role. tives. We will see later that it may be desirable It is problematic that the UD definition relies to strengthen this criterion and require expletives on the concept of argument, since UD otherwise to be pro-forms that are selected by the predicate abandons the argument/adjunct distinction in favor with which it occurs. Such purely formal selec- of the core/oblique distinction. Silveira’s account tion is needed in many cases, since expletives are avoids this problem by instead referring to gram- not interchangeable across constructions – for ex- matical functions, thus also catering for cases like: ample, there rains is not an acceptable sentence (1) He will see to it that you have a reservation. of English. Criteria (ii) and (iii) from the defini- tion of Postal and Pullum(1988) may be hard to However, both definitions appear to be too wide, apply directly in a UD setting, as UD is a syntac- in that they do not impose any restrictions on the tic, not a semantic, annotation framework. On the form of the expletive, or require it to be non- other hand, many decisions in UD are driven by referential. It could therefore be argued that the the need to provide annotations that can serve as subject of a raising verb, like Sue in Sue appears input for semantic analysis, and distinguishing be- to be nice, satisfies the conditions of the definition, tween elements that do and do not refer and fill a since it is a nominal in subject position that does thematic role therefore seems desirable. not satisfy a semantic role of the predicate appear. In addition to the definition, Postal and Pul- It seems useful, then, to look for a better defi- lum(1988) provide tests for expletives. Some of nition of expletive. Much of the literature in the- these (tough-movement and nominalization) are oretical linguistics is either restricted to specific not easy to apply cross-linguistically, but two of languages or language families (Platzack, 1987; them are, namely absence of coordination and in- Bennis, 2010; Cardinaletti, 1997) or to specific ability to license an emphatic reflexive. constructions (Vikner, 1995; Hazout, 2004). A theory-neutral and general definition can be found (2) *It and John rained and carried an umbrella in Postal and Pullum(1988): respectively. (3) *It itself rained. [T]hey are (i) morphologically identical to pro-forms (in English, two relevant forms The inability to license an emphatic reflexive is are it, identical to the third person neuter probably due to the lack of referentiality. It is less immediately obvious what the absence of coordi- (5) It seems that she came (en) nation diagnoses. One likely interpretation is that (6) Hij betreurt het dat jullie verliezen (nl) sentences like (2) are ungrammatical because the He regrets it that you lose verb selects for a particular syntactic string as its ‘He regrets that you lose’ subject. If that is so, form-selection can be consid- (7) It is going to be hard to sell the Dodge (en) ered a defining feature of expletives. Finally, following Postal and Pullum(1988), we It is fairly straightforward to argue that this con- can draw a distinction between expletives that oc- struction involves an expletive. Theoretically, it cur in chains and those that do not, where we un- could be cataphoric to the following clause and so derstand a chain as a relation between an expletive be referential, but in that case we would expect it and some other element of the sentence which has to be able to license an emphatic reflexive. How- the thematic role that would normally be associ- ever, this is not what we find, as shown in (8-a), ated with the position of the expletive, for exam- which contrasts with (8-b) where the raised sub- ple, the subordinate clause in (4). ject is a referential pronoun. (4) It surprised me that she came. (8) a. *It seems itself that she came It is not always possible to realize the other ele- b. It seems itself to be a primary meta- ment in the chain in the position of the expletive. physical principle For example, the subordinate clause cannot be di- But if it does not refer cataphorically to the ex- rectly embedded under the preposition in (1). traposed clause, its form must also be due to the Whether the expletive participates in a chain or construction in which it appears. This construc- not is relevant for the UD annotation insofar as it tion therefore fulfills the criteria of an expletive is often desirable – for the purposes of semantic even on the strictest understanding. interpretation – to give the semantically active el- ement of the chain the “real” dependency label. 2.2 Existential Sentences For example, it is tempting to take the comple- ment clause in (4) as the subject (csubj in UD) to Existential (or presentational) sentences are sen- stay closer to the semantics, although one is hard tences that involve an intransitive verb and a noun pressed to come up with independent syntactic ev- phrase that is interpreted as the logical subject of idence that an element in this position can actu- the verb but does not occur in the canonical sub- ally be a subject. This is in line with many de- ject position, which is instead filled by an exple- scriptive grammar traditions, where the expletive tive. There is considerable variation between lan- would be called the formal subject and the subor- guages as to which verbs participate in this con- dinate clause the logical subject. struction. For instance, while English is quite re- We now review constructions that are regularly strictive and uses this construction mainly with the analyzed as involving an expletive in the theoret- copula be, other languages allow a wider range of ical literature and discuss these in the light of the verbs including verbs of position and movement, definition and tests we have established. as illustrated in (9)–(11). There is also variation with respect to criteria for classifying the nominal 2.1 Extraposition of Clausal Arguments constituent as a subject or object, with diagnostics In many languages, verbs selecting a clausal sub- such as agreement, case, and structural position ject or object often allow or require an expletive often giving conflicting results. Some languages, and place the clausal argument in extraposed po- like the Scandinavian languages, restrict the nom- sition. In some cases, extraposition of the clausal inal element to indefinite nominals, whereas Ger- argument is obligatory, as in (5) for English. Note man for instance also allows for definite nominals that the clausal argument can be either a subject in this construction. or an object, and thus the expletive in some cases (9) Det sitter en katt pa˚ mattan (sv) appears in object position, as in (6). Also note that it sits a cat on the-mat in so-called raising contexts, the expletive may ac- ‘A cat sits on the mat’ tually be realized in the structural subject position (10) Es landet ein Flugzeug (de) of a verb governing the verb that selects the clausal it lands a plane argument (7). ‘A plane lands’ (11) Il nageait quelques personnes (fr) (16) La branche s’ est cassee´ there swim some people The branch SE is broken ‘Some people are swimming’ ‘The branch broke.’ Despite the cross-linguistic variation, existential In all of these cases, it is clear that the reflexive el- constructions like these are uncontroversial cases ement does not receive a semantic role. In (15), of expletive usage. The form of the pronoun(s) is dosp´ıva´ ‘mature’ only takes one semantic argu- fixed, it cannot refer to the other element of the ment, and in (16), the intended reading is clearly chain for formal reasons, and no emphatic reflex- not that the branch broke itself. We conclude that ive is possible. these elements are expletives according to the def- inition above. This is in line with the proposal of 2.3 Impersonal Constructions Silveira(2016). By impersonal constructions we understand con- structions where the verb takes a fixed, pronomi- 2.5 Inherent Reflexives nal, argument in subject position that is not inter- Many languages have verbs that obligatorily select preted in semantics. Some of these involve zero- a reflexive pronoun without assigning a semantic valent verbs, such as weather verbs, which are tra- role to it: ditionally assumed to take an expletive subject in (17) Pedro se confundiu (pt) Germanic languages, as in Norwegian regne ‘rain’ Pedro REFL confused (12). Others involve verb that also take a semantic ‘Pedro was confused’ argument, such as the French falloir in (13). (18) Smejeˇ se (cs) (12) Det regner (no) laugh REFL it rains ‘he/she/it laughs’ ‘It is raining’ There are borderline cases where the verb in ques- (13) Il faut trois nouveaux recrutements (fr) tion can also take a regular object, but the seman- it needs three new staff-members tics is subtly different. A typical case are verbs like ‘Three new staff members are needed’ wash. That there are in fact two different interpre- Impersonal constructions can also arise when an tations is revealed in Scandinavian by the impossi- intransitive verb is passivized (and the normal se- bility of coordination. (19) is grammatical unless mantic subject argument therefore suppressed). seg is stressed.

(14) Es wird gespielt (de) (19) *Han vasket seg og de andre (no) It is played He washed REFL and the others ‘There is playing’ ‘He washed himself and the others’ In all these examples, the pronouns are clearly From the point of view of our definition, it is clear non-referential, no emphatic reflexive is possible that inherent reflexives (by definition) do not re- and the form is selected by the construction, so ceive a semantic role. It may be less clear that these elements can be classified as expletive. they are non-referential: after all, they typically agree with the subject and could be taken to be 2.4 Passive Reflexives co-referent. It is hard to test for non-referentiality In some Romance and Slavic languages, a pas- in the absence of any semantic role. In particular, sive can be formed by adding a reflexive pronoun the emphatic reflexive test is not easily applicable, which does not get a thematic role but rather sig- since it may be the subject that antecedes the em- nals the passive voice. phatic reflexive in cases like (20).

(15) dosp´ıva´ se drˇ´ıve (cs) (20) Elle s’est souvenue elle-memeˆ mature REFL earlier she REFL-is reminded herself ‘(they/people) mature up earlier’ ‘She herself remembered. . . ’ In Romance languages, as shown by Silveira Inherent reflexives agree with the subject, and thus (2016), these are not only used with a strictly their form is not determined (only) by the verb. passive meaning, but also with inchoative (anti- Nevertheless, under the looser understanding of causative) and medio-passive readings. the formal criterion, it is enough that reflexives are pronominal and thus can be expletives. This is also The counts and proportions for specific con- the conclusion of Silveira(2016). structions in Table1 were computed as follows. Extraposition covers cases where an expletive co- 2.6 Clitic Doubling occurs with a csubj or ccomp argument as in The UD guidelines explicitly mention that “true” the top row of Figure1. This construction occurs (that is, regularly available) clitic doubling, as in frequently in the Germanic treebanks (Dutch, En- the Greek example in (21), should be annotated glish, German, Norwegian, Swedish), as in (22), using the expl relation: but is also fairly frequent in French treebanks, as in (23). (21) pisteuoˆ oti einai dikaio na to I-believe that it-is fair that this-CLITIC (22) It is true that Google has been in acquisi- anagnorisoumeˆ auto (el) tion mode (en) we-recognize this (23) Il est de notre devoir de participer [. . . ] (fr) The clitic to merely signals the presence of the full it is of our duty to participate [. . . ] pronoun object and it can be argued that it is the ‘It is our duty to participate . . . ’ latter that receives the thematic role. It is less clear, however, that to is non-referential, hence it is un- Existential constructions can be identified by the clear that this is an instance of an expletive. The presence of a nominal subject (nsubj) as a sib- alternative is to annotate the clitic as a core argu- ling of the expl element, as illustrated in the mid- ment and use dislocated for the full pronoun dle row of Figure1. Existential constructions are (as is done for other cases of doubling in UD). very widespread and span several language fami- lies in the treebank data. They are common in all 3 Expletives in UD 2.1 treebanks Germanic treebanks, as illustrated in (24), but are also found in Finnish, exemplified in (25), where We will now present a survey of the usage of the these constructions account for half of all exple- expl relation in current UD treebanks. In par- tive occurrences, as well as in several Romance ticular, we will relate the constructions discussed languages (French, Galician, Italian, Portuguese), in Section2 to the treebank data. Table1 gives some Slavic languages (Russian and Ukrainian), expl an overview of the usage of and its lan- and Greek. guage specific extensions in the treebanks in UD v2.1.1 We find that, out of the 60 languages in- (24) Es fehlt ein System umfassender sozialer cluded in this release, 27 make use of the expl it lacks a system comprehensive social relation, and its use appears to be restricted to Eu- Sicherung (de) security ropean languages. For those languages that have ‘A system of comprehensive social secu- multiple treebanks, expl is not always used in all rity is lacking’ treebanks (Finnish, Galician, Latin, Portuguese, Russian, Spanish). The frequency of expl varies (25) Se oli paska homma, etta¨ Jyrki loppu (fi) it was shit thing that Jyrki end greatly, ranging from less than 1 per 1,000 ‘It was a shit thing for Jyrki to end’ (Catalan, Greek, Latin, Russian, Spanish, Ukra- nian) to more than 2 per 100 words (Bulgarian, For the impersonal constructions discussed in Sec- Polish, Slovak). For most of the languages, there tion 2.3, only a few UD treebanks make use of is a fairly limited set of lemmas that realize the an explicit impers subtype (Italian, Romanian). expl relation. Treebanks with higher numbers of Apart from these, impersonal verbs like rain and lemmas are those that label inherent reflexives as French falloir prove difficult to identify reliably expl and/or do not always lemmatize systemat- across languages using morphosyntactic criteria. ically. Some treebanks not only use expl, but For impersonal passives, on the other hand, there also the subtypes expl:pv (for inherent reflex- are morphosyntactic properties that we may em- ives), expl:pass (for certain passive construc- ploy in our survey. Passives in UD are marked tions), and expl:impers (for impersonal con- either morphologically on the verb (by the feature structions). Voice=Passive) or by a passive auxiliary de- pendent (aux:pass) in the case of periphrastic 1The raw counts as well as the script we used to col- lect the data can be found at github.com/gossebouma/ passive constructions. These two passive con- expletives structions are illustrated in the bottom row (left Banks Count Freq Lemmas Extraposed Existential Impersonal Reflexives Remaining Bulgarian 3379 0.021 7 12 0.0 82 0.02 2 0.0 32040.95 79 0.02 Catalan 512 0.001 4 0 0.0 0 0.0 0 0.0 512 1.0 0 0.0 Croatian 2173 0.011 11 2 0.0 4 0.0 1 0.0 21610.99 5 0.0 Czech 5/5 35929 0.018 4 0 0.0 0 0.0 0 0.0 35929 1.0 0 0.0 Danish 441 0.004 2 8 0.02 10 0.02 62 0.14 0 0.0 361 0.82 Dutch 2/2 459 0.001 5 321 0.7 120 0.26 6 0.01 0 0.0 12 0.03 English 4/4 1221 0.003 6 380 0.31 724 0.59 9 0.01 0 0.0 107 0.09 Finnish 1/3 524 0.003 9 15 0.03 268 0.51 53 0.1 0 0.0 188 0.36 French 5/5 6117 0.005 26 162 0.03 1486 0.24 27 0.0 33780.55 1064 0.17 Galician 1/2 288 0.01 6 19 0.07 131 0.45 0 0.0 0 0.0 138 0.48 German 2/2 487 0.003 1 114 0.23 287 0.59 21 0.04 1 0.0 64 0.13 Greek 18 0.000 1 0 0.0 6 0.33 0 0.0 0 0.0 12 0.67 Italian 4/4 4214 0.009 22 107 0.03 1901 0.45 589 0.14 3960.09 1218 0.29 Latin 1/3 257 0.001 1 0 0.0 0 0.0 0 0.0 257 1.0 0 0.0 Norwegian 3/3 6890 0.01 8 1894 0.27 1758 0.26 374 0.05 0 0.0 2864 0.42 Polish 1708 0.02 1 0 0.0 0 0.0 6 0.0 1702 1.0 0 0.0 Portuguese 2/3 1624 0.003 1 20 0.01 628 0.39 20 0.01 6720.41 284 0.17 Romanian 2/2 5209 0.002 22 43 0.01 327 0.06 140 0.03 42810.82 418 0.08 Russian 2/3 55 0.000 3 6 0.11 42 0.76 1 0.02 0 0.0 6 0.11 Slovak 2841 0.03 3 0 0.0 0 0.0 0 0.0 2841 1.0 0 0.0 Slovenian 2/2 2754 0.02 2 0 0.0 1 0.0 0 0.0 2297 1.0 0 0.0 Spanish 1/3 503 0.001 2 0 0.0 0 0.0 0 0.0 503 1.0 0 0.0 Swedish 3/3 1079 0.005 6 371 0.34 283 0.26 85 0.08 0 0.0 340 0.32 Ukrainian 94 0.001 4 16 0.17 62 0.66 0 0.0 120.13 4 0.04 Upper Sorbian 177 0.02 1 0 0.0 0 0.0 1 0.01 1760.99 0 0.0

Table 1: Use of expl in UD v2.1 treebanks. Languages with Count < 10 left out (Arabic, Sanskrit). Freq = average frequency for treebanks containing expl. Count and proportion for construction types. and center) of Figure1. The quantitative overview is difficult to distinguish these automatically based in Table1 shows that impersonal constructions oc- on morphosyntactic criteria. In Table1 we there- cur mostly in Germanic languages, such as Dan- fore collapse these two construction types (Reflex- ish, German, Norwegian and Swedish, illustrated ive). In addition to the pv subtype, we further rely by (26). These are all impersonal passives. We on another morphological feature in the treebanks note that both Italian and Romanian also show a in order to identify inherent reflexives, namely the high proportion of impersonal verbs, due to the Reflex feature, as illustrated by the Portuguese use of expl:impers mentioned above and ex- example in Figure1 (bottom right). 2 In Table1 we emplified by (27). observe that the distribution of passive and inher- ent reflexives clearly separates the different tree- (26) Det ble ikke nevnt hvor omstridt banks. They are highly frequent in Slavic lan- it was not mentioned how controversial han er (no) guages (Bulgarian, Croatian, Czech, Polish, Slo- he is vak, Slovenian, Ukrainian and Upper Sorbian). as ‘It was not mentioned how controversial illustrated by the passive reflexive in (28) and the he is’ inherent reflexive in (29). They are also frequent in two of the French treebanks and in Brazilian (27) Si compredono inoltre i figli adottivi it includes also the children adopted Portuguese. Interestingly, they are also found in (it) Latin, but only in the treebank based on medieval texts. ‘Adopted children are also included’ (28) O centraln´ ´ı vyrob´ eˇ tepla se rˇ´ıka,´ Both the constructions of passive reflexives and in- about central production heating it says herent reflexives (Sections 2.4 and 2.5), make use zeˇ je nejefektivnejˇ sˇ´ı (cs) that the most-efficient of a reflexive pronoun. Some treebanks distin- expl guish these through subtyping of the rela- 2The final category discussed in section2 is that of clitic tion, for instance, expl:pass and expl:pv in doubling. It is not clear, however, how one could recognize the Czech treebanks. This is not, however, the case these based on their morphosyntactic analysis in the various treebanks and we therefore exclude them from our empirical across languages and since the reflexive passive study, although a manual analysis confirmed that they exist at does not require passive marking on the verb, it least in Bulgarian and Greek. csubj ccomp expl obj nsubj expl

It surprised me that she came Hij betreurt het dat de commissie niet functioneert He regrets it that the committee not functions

obl nsubj expl nsubj expl Det sitter en katt pa˚ mattan Es landet ein Flugzeug there sits a cat on the-mat There landed a plane

expl nsubj aux:pass expl expl Es wird gespielt Det dansas Pedro se confundiu there is playing there is-dancing Pedro REFL confused Voice=Passive Reflex=Yes

Figure 1: UD analyses of extraposition [(4) and (6)] (top), existentials [(9) and (10)] (middle), impersonal con- structions (bottom left and center), and inherent reflexives [(17)] (bottom right).

‘Central heat production is said to be the Czech, Polish, Portuguese, Romanian, and Slovak most efficient’ use the subtype expl:pv, while the treebanks for French, Italian and Spanish simply use expl for (29) Skozi steno slisim,ˇ kako se through wall I-hear-it, how REFL this purpose. And even though languages like Ger- zabavajo. (sl) man, Dutch and Swedish do have inherent reflex- have-fun ives, their reflexive arguments are currently anno- ‘I hear through the wall how they have tated as regular objects. fun’ Even in different treebanks for one and the same (30) O deputado se aproximou (pt) language, different decisions have sometimes been the deputy REFL approached made, as is clear from the column labeled Banks ‘The deputy approached’ in Table1. Of the three treebanks for Spanish, for instance, only Spanish-AnCora uses the expl It is clear from the discussion above that all con- relation, and of the three Finnish UD treebanks, structions discussed in Section2 are attested in only Finnish-FTB. In the French treebanks, we ob- UD treebanks. Some languages have a substan- serve that the expl relation is employed to cap- tial number of expl occurrences that are not cap- ture quite different constructions. For instance, tured by our heuristics (i.e. the Remaining cate- in French-ParTUT, it is used for impersonal sub- gory in Table 1). In some cases (i.e. Swedish and jects (non-referential il, whereas the other French Norwegian), this is due to an analysis of cleft con- treebanks do not employ an expletive analysis for structions where the pronoun is tagged as expl. these. We also find that annotation within a single It should be noted that the analysis of clefts dif- treebank is not always consistent. For instance, fers considerably across languages and treebanks, whereas the German treebank generally marks es and therefore we did not include it in the empir- in existential constructions with geben as expl, ical overview. Another frequent pattern not cap- the treebank also contains a fair amount of exam- tured by our heuristics involves clitics and clitic ples with geben where es is marked nsubj, de- doubling. This is true especially for the Romance spite being clearly expletive. languages, where Italian and Galician have a sub- expl stantial number of occurrences of marked as 4 Towards Consistent Annotation of Clitic not covered by our heuristics. In French, Expletives in UD a frequent pattern not captured by our heuristics is the il y a construction. Our investigations in the previous section clearly The empirical investigation also makes clear demonstrate that expletives are currently not an- that the analysis of expletives under the current notated consistently in UD treebanks. This is UD scheme suffers from inconsistencies. For partly due to the existence of different descrip- inherent reflexives, the treebanks for Croatian, tive and theoretical traditions and to the fact that many treebanks have been converted from anno- to the other chain element (if present). tation schemes that differ in their treatment of ex- 2. Treat expletives as core arguments and allow pletives. But the situation has probably been made the other chain element (if present) to instan- worse by the lack of detailed guidelines concern- tiate the same relation (possibly using sub- ing which constructions should be analyzed as in- types to distinguish the two). volving expletives and how exactly these construc- 3. Treat expletives as core arguments and forbid tions should be annotated. In this section, we will the other chain element (if present) to instan- take a first step towards improving the situation by tiate the same relation. making specific recommendations on both of these aspects. All three approaches have advantages and draw- Based on the definition and tests taken from backs, but the current UD guidelines clearly favor Postal and Pullum(1988), we propose that the the first approach, essentially restricting the ap- class of expletives should include non-referential plication of core argument relations to referential pro-forms involved in the following types of con- core arguments. Since this approach is already im- structions: plemented in a large number of treebanks, albeit to different degrees and with considerable variation, 1. Extraposition of clausal arguments (Sec- it seems practically preferable to maintain and re- tion 2.1) fine this approach, rather than switching to a radi- 2. Existential (or presentational) sentences cally different scheme. However, in order to make (Section 2.2) the annotation more informative, we recommend 3. Impersonal constructions (including weather using the following subtypes of the expl relation: verbs and impersonal passives) (Section 2.3) 4. Passive reflexives (Section 2.4) 1. expl:chain for expletives that occur in 5. Inherent reflexives (Section 2.5) chain constructions like extraposition of clausal arguments and existential or presen- For inherent reflexives, the evidence is not quite tational sentences (Section 2.1–2.2) as clear-cut as for the other categories, but given 2. expl:impers for expletive subjects in im- that the current UD guidelines recommend using personal constructions, including impersonal expl and given that many treebanks already fol- verbs and passivized intransitive verbs (Sec- low these guidelines, it seems most practical to tion 2.3) continue to include them in the class of expletives, 3. expl:pass for reflexive pronouns used to as recommended by Silveira(2016). By contrast, form passives (Section 2.4) the arguments for treating clitics in clitic doubling 4. expl:pv for inherent reflexives, that is, pro- (Section 2.6) as expletives appears weaker, and nouns selected by pronominal verbs (Sec- very few treebanks have implemented this anal- tion 2.5) ysis, so we think it may be worth reconsidering their analysis and possibly use dislocated for The three latter subtypes are already included in all cases of double realization of core arguments. the UD guidelines,although it is clear that they are The distinction between core arguments and not used in all treebanks that use the expl rela- other dependents of a predicate is a cornerstone tion. The first subtype, expl:chain, is a novel of the UD approach to syntactic annotation. Ex- proposal, which would allow us to distinguish con- pletives challenge this distinction by (mostly) be- structions where the expletive is dependent on the having as core arguments syntactically but not se- presence of a referential argument. This subtype mantically. In chain constructions like extraposi- could possibly be used also in clitic doubling, if tion and existentials, they compete with the other we decide to include these among expletives. chain element for the core argument relation. In impersonal constructions and inherent reflexives, 5 Conclusion they are the sole candidate for that relation. This suggests three possible ways of treating expletives Creating consistently annotated treebanks for in relation to core arguments: many languages is potentially of tremendous im- portance for both NLP and linguistics. While our 1. Treat expletives as distinct from core argu- study of the annotation of expletives in UD shows ments and assign the core argument relation that this goal has not quite been reached yet, the development of UD has at least made it possi- Sharid Loaiciga,´ Liane Guillou, and Christian Hard- ble to start investigating these issues on a large meier. 2017. What is it? disambiguating the differ- scale. Based on a theoretical analysis of exple- ent readings of the pronoun ‘it’. In Proceedings of the 2017 Conference on Empirical Methods in Nat- tives and an empirical survey of current UD tree- ural Language Processing, pages 1325–1331. banks, we have proposed a refinement of the anno- tation guidelines that is well grounded in both the- Joakim Nivre. 2015. Towards a universal grammar for natural language processing. In International Con- ory and data and that will hopefully lead to more ference on Intelligent and Computa- consistency. By systematically studying different tional Linguistics, pages 3–16. Springer. linguistic phenomena in this way, we can gradu- ally approach the goal of global consistency. Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Yoav Goldberg, Jan Hajic, Christopher D Manning, Ryan T McDonald, Slav Petrov, Sampo Acknowledgments Pyysalo, Natalia Silveira, et al. 2016. Universal de- pendencies v1: A multilingual treebank collection. We are grateful to two anonymous reviewers for In 10th International Conference on Language Re- constructive comments on the first version of the sources and Evaluation (LREC), Portoroz, Slovenia, paper. Most of the work described in this ar- pages 1659–1666. European Language Resources ticle was conducted during the authors’ stays at Association. the Center for Advanced Study at the Norwegian Christer Platzack. 1987. The Scandinavian languages Academy of Science and Letters. and the null-subject parameter. Natural Language & Linguistic Theory, 5(3):377–401.

Paul M Postal and Geoffrey K Pullum. 1988. Expletive References noun phrases in subcategorized positions. Linguistic Hans Bennis. 2010. Gaps and dummies. Amsterdam Inquiry, 19(4):635–670. University Press. Natalia Silveira. 2016. Designing Syntactic Represen- Adriane Boyd, Whitney Gegg-Harrison, and Donna tations for NLP: An Empirical Investigation. Ph.D. Byron. 2005. Identifying non-referential it: A thesis, Stanford University, Stanford, CA. machine learning approach incorporating linguisti- Olga Uryupina, Mijail Kabadjov, and Massimo Poe- cally motivated patterns. In Proceedings of the sio. 2016. Detecting non-reference and non- ACL Workshop on Feature Engineering for Machine anaphoricity. In Massimo Poesio, Roland Stuckardt, Learning in Natural Language Processing, Fea- and Yannick Versley, editors, Anaphora Resolution: tureEng ’05, pages 40–47, Stroudsburg, PA, USA. Algorithms, Resources, and Applications, pages Association for Computational Linguistics. 369–392. Springer Berlin Heidelberg, Berlin, Hei- Anna Cardinaletti. 1997. Agreement and control delberg. in expletive constructions. Linguistic Inquiry, Sten Vikner. 1995. Verb movement and expletive sub- 28(3):521–533. jects in the Germanic languages. Oxford University Press on Demand. Richard Evans. 2001. Applying machine learning to- ward an automatic classification of it. Literary and Lesly Miculicich Werlen and Andrei Popescu-Belis. linguistic computing, 16(1):45–58. 2017. Using coreference links to improve Spanish- to-English . In Proceedings of Liane Guillou, Christian Hardmeier, Aaron Smith, Jorg¨ the 2nd Workshop on Coreference Resolution Be- Tiedemann, and Bonnie Webber. 2014. Parcor 1.0: yond OntoNotes (CORBON 2017), pages 30–40. A parallel pronoun-coreference corpus to support statistical MT. In 9th International Conference on Language Resources and Evaluation (LREC), May 26-31, 2014, Reykjavik, Iceland, pages 3191–3198. European Language Resources Association.

Christian Hardmeier, Preslav Nakov, Sara Stymne, Jorg¨ Tiedemann, Yannick Versley, and Mauro Cettolo. 2015. Pronoun-focused MT and cross-lingual pro- noun prediction: Findings of the 2015 DiscoMT shared task on pronoun translation. In Proceedings of the Second Workshop on Discourse in Machine Translation, pages 1–16.

Ilan Hazout. 2004. The syntax of existential construc- tions. Linguistic Inquiry, 35(3):393–430.