The PARSEME Shared Task on automatic identification of verbal multiword expressions aims at identifying such expressions in running texts. Typology of verbal multiword expressions, very detailed annotation guidelines and gold-standard data for as many languages as possible will be provided. Since the Prague Dependency Treebank includes Czech multiword expression annotation, it was natural to make an attempt to automatically convert the data into the Shared Task format. However, since the Czech treebank predates the Shared Task annotation guidelines, a prior examination was necessary to determine to which extent the conversion can be fully automatic and how much manual work remains. In this paper, we show that information contained in the Prague Dependency Treebank is sufficient to extract all of the Shared Task categories of verbal multiword expressions relevant for Czech, even if these categories are originally annotated differently; nevertheless, some manual checking and annotation would still be necessary, e.g. for distinguishing borderline cases.
The goal of the PARSEME [
This preference led us to explore the Prague Dependency Treebank (PDT,
[
We believe that for the Czech language, annotation of VMWEs already encoded
in the data of the Prague Dependency Treebank 3.0 (PDT) [
The PARSEME Shared Task identifies six groups of VMWEs: light verb constructions (LVC), idioms (ID), verb particle combinations (VPC), inherently reflexive verbs (IReflV), language specific types and other verbal MWEs (OTH).
All the various types of VMWEs required by the PST are annotated in quite a number of diverse ways in the PDT and the information is spread across several layers of annotation. Thus we first had to relate the PDT annotation to the PST guidelines in order to confirm that the PDT data can be reused for the Shared Task and only then the extraction of all types of VMWEs (relevant for Czech) and their conversion into the PST format could take place.
At the same time (or even more importantly), we were testing the following
four principles for good-quality MWE treebank design published in [
After thorough analysis of the PDT we have concluded that Principles A and B are clearly fulfilled in the PDT due to its explicit MWE annotation. Principle C is also followed thanks to the explicit links between the PDT’s annotation layers. Principle D is, from the PST point of view, followed only partially, since the respective typologies do not match one-to-one.
Thorough inspection of the PDT annotation scheme resulted in an automatic conversion procedure with rules formulated for each of the PST types. Manual checks and some amount of manual annotation is still necessary, even if for only a fraction of the data.
As already explained, the creation of the Czech language data for the PST takes advantage of the existing rich annotation of the PDT, including explicit annotation of VMWEs.
The treatment of verbal idioms (part of the ID category) and LVCs in the PDT
is related to valency, as the valency formalism allows for morphological, syntactic
and semantic description of VMWEs in the treebank [
dostat_se
PRED nevidomý ACT
To sum up, different PST types of VMWEs are obtained from various information sources available at the different layers of annotation in the PDT. See Figure 1 for an illustration of three of them (the annotation view is simplified only to cover MWE-related phenomena); an annotation of the non-verbal MWE “rehabilitacˇní pracovník” (rehabilitation worker) which is not being converted for PST is also shown.
In this section, we describe the PDT-style annotation of the proposed six types of VMWEs recognized in the PST as well as their conversion into the common PST format (Sections 3.1–3.6). Two special aspects are discussed, namely deverbative variants (Section 3.7) and cases of overlapping annotation (Section 3.8). 3.1
In the PDT annotation, LVCs consist of two lexical units: a semantically empty (or “light”) verb and a noun carrying the main lexical meaning of the entire phrase. The nominal part of the LVCs is labeled by the CPHR functor (Compound PHRase). For example: to comePRED into forceCPHR, to undertakePRED preparationsCPHR. LVCs are identified as depicted in Figure 2. 1. Input text Zákon tak vstoupil v platnost.
Law so came into force.
By that the law has come into force.
2. PDT t‐layer vstoupit PRED
<MWEcategory="LVC"> vstoupil Zákon tak vstoupil v platnost. 3. PDT a‐layer Pred has come
into force 4. Output annota on zákon ACT tak podepsat platnost MANN CPHR
Sb Adv vAuxP
Zákon tak #QCor ACT pOlbajtnost
Three more things have to be taken into account: 1. Prepositions, if they are part of the LVC, must be retrieved from the surface syntactic layer, since they are not present on the deep layer. If there is any extra node between a node for a predicate and a node for a CPHR, it is part of the LVC. 2. If reflexive particles are part of the verb lemma (see IReflV in Section 3.4), they also have to become part of the LVC. 3. The CPHR functor is also used for a specific type of phrases with the verb “to be” (it is necessaryCPHR to leave). These phrases, not assumed by the PST guidelines, are excluded by checking the lemma of the verb.
There are 2496 LVCs in the PDT extracted by the above rules. Minor details aside, LVCs as defined for the PST can be identified on the basis of the existing PDT annotation without any additional manual annotation. 3.2
These VMWEs, denoted as ID in the PST guidelines, compose quite a large group containing not only traditional idioms. We have to process it in two steps.
Part of the VMWEs defined as IDs, namely those which are quite fixed idioms, are understood similarly in the PDT and in the guidelines for the Shared Task, e.g.: “házet klacky pod nohy” lit. to-throw sticks under feet (= to put obstacles in one’s way), “brát vítr z plachet” (= to take the wind out of someone’s sails). These verbal idioms (similarly to LVCs) always consist of two nodes in the PDT: the governing verb part and the dependent node (with the DPHR functor = Dependent part of PHRaseme). These idioms can be thus easily extracted by looking for the DPHR functor. The DPHR node represents all other lexical components of the idiom, should there be more than one (lemma of the deep syntactic layer is e.g. “klacky_pod_nohy” or “vítr_z_plachet”), since these are quite fixed expressions in terms of (the impossibility of) insertion or other modification. Even prepositions are part of it and their detection is even easier than with a CPHR. See an example in Figure 3. 1. Input text Odezva na sebe nedala čekat.
Reac on on itself not‐gave wait.
The reac on didn't keep us wai ng.
2. PDT t‐layer dát.enunc
PRED odezva #Neg na_sebe_čekat ACT RHEM DPHR 3. Output annota on
<MWE category="ID"> Odezva na sebe nedala čekat.
didn't keep us wai ng
The other group of VMWEs categorized as ID in PST is not so fixed. VMWEs
from this group do not fulfill the criteria for DPHR annotation in the PDT, but they
still qualify to be an IDin the PST. They have been annotated together with all
other MWEs in PDT 3.0 [
Our approach finds a head in the syntactic tree of such a set. If it is a verb, the MWE is a verbal one (Figure 4). Then other auxiliary nodes (e.g., prepositions) referred to by the annotated content words are added. (The exception is a conjunction introducing the whole phrase: it does not belong to the VMWE.) The resulting VMWE gets the ID mark, unless it overlaps with CPHR or DPHR annotation (see Section 3.8).
We have identified 2107 IDs using either the PDT 3.0 MWE or DPHR annotation. root 2. PDT t‐layer mwe lexeme kroutit PRED verb
3. Output annota on <MWE category="ID"> Nevěřícně krou m hlavou nad legisla vou.
shaking my head nevěřícný hlava legislativa MANN PAT REG 1. Input text Nevěřícně krou m hlavou nad legisla vou.
Disbelievingly I‐shake head over legisla on.
I am shaking my head in disbelief on the legisla on. #PersPron
ACT Verb-particle combinations (VPC) are not present in Czech. A phenomenon similar to VPCs is in Czech realized by verbal prefixes (the result being another single lexical unit, i.e., not a MWE). 3.4
Inherently Reflexive Verbs (IReflV) contain one of two possible clitics in Czech: “se” or “si”, e.g. “bát se” (= to be afraid), “hledeˇt si” (= to mind sth). Such verb is considered a separate lexical unit (different from the verb appearing without the particle if such verb exists at all) and both its parts are represented by just one node at the deep syntactic layer of the PDT, and the node’s lemma matches the PDTVallex lexical unit, which includes the appropriate particle as part of the headword in the lexicon. This annotation was used for exactly the two types qualified as IReflV in the PST guidelines, namely, for the case when the non-reflexive counterpart verb does not exist or when its meaning is markedly changed. Using this annotation, all IReflVs can be extracted from the PDT texts and converted, see Figure 5. 1. Input text Opatření se týká zejména domovníků.
The meassure involves chiefly housekeepers.
týkat_se2. PDT t‐layer3. PDT a‐layer tPýrekád
PRED opatření ACT 4. Output annota on
<MWEcategory="IReflV">
Opatření se týká zejména domovníků.
SObpatření sAeuxT dOobmjovníků involves zejména domovník RHEM PAT zAeujxmZéna
IReflVs should be possible to extract also without the deep syntactic layer; an analytical function of an IReflV reflexive particle should be either AuxT or AuxO on a surface syntactic layer; other values (AuxR, Obj, or Adv) are reserved for reflexive particles used in other than IReflV contexts, e.g. in passive constructions. Suspicious cases (705 verb occurrences) in which the information from the two layers of annotation clashes have been detected by looking for discrepancy between the lemma and the corresponding analytical function and manually checked and corrected when necessary (330 cases). There are some borderline cases where the PDT annotation differs from the PST guidelines; however, these are mainly errors in annotation and not a true difference between the PST and PDT guidelines.
By this approach, 10,266 VMWEs of the IReflV type were extracted from the PDT. The conversion was automatic except for the 705 manually checked occurrences. 3.5
This category (OTH) is specified in the PST guidelines as a VMWE that does not fit into any of the other categories, as described in the previous sections. Namely, it applies to “coordinations of verbs, e.g. to drink and drive, and compound verbs, e.g. to short-circuit, to pretty-print, to voice act”. The second subtype usually results in a one-word expression in Czech, so we need to search only for coordinated verbs.
For this category, the PDT 3.0 MWEs annotation [
PhD‐student is student, as <REFL> suits and befits.
A PhD student is a student, as he should be.
být
PRED doktorand ACT 2. PDT t‐layer student PAT a
CONJ #Gen slušet_se patřit_se ACT RSTR RSTR 3. PDT a‐layer
conjunction verb verb 4. Output annota on <MWEcategory="OTH"> Doktorand je studentem, jak se sluší a patří.
as he should be
PARSEME Shared Task guidelines also recognize other, non-verbal variants of
verbal MWEs, such as relative clauses (heart which he broke), gerunds (heart
breaking), nominal groups (heart-breaking), or adjectival groups (breaking her heart).
In Czech, nominalization is a common way of verbal MWE variation, see [
There is no nominal group annotated as CPHR in the PDT and thus no LVC
variant. There are several nominal MWEs annotated as DPHR, but only seven of
them are made from verbal MWEs. We have picked them manually. During the
PDT 3.0 MWE annotation project [
The situation is quite different for IReflV where many non-verbal lemmas also contain reflexive particles “se” or “si”. These cases qualify themselves as nominal or adverbial variants of inherently reflexive verbs.
To sum up, there are deverbative MWEs in the PST Czech data, however they are not frequent.
We are also preparing other deverbative MWEs using data by an idiom recognizer based on a database, upgraded for deverbatives by Milena Hnátková [5]. 3.8
Since the data for PST are extracted from various pieces of annotation, it can easily happen they are duplicated or that they overlap. All these cases have to be solved properly, as described below. 3.8.1
Coordination Part of a VMWE can be coordinated while the other part is used only once, as in “Ministerstvo poskytuje malým podnikatelu˚m informacˇní služby a poradenskou cˇinnost.” (The ministry provides information services and counselling activities to small businesses.), where two LVCs are present: to provide services and to provide activities. Such a case is correct and both VMWEs should be preserved and marked in the output data, with the verb “provide” being part of both. 3.8.2
Duplicates due to added nodes in the PDT Since a large part of the MWE annotation in the PDT is encoded at the deep syntactic layer, sometimes a VMWE is found that has no direct realization in the surface form of the sentence, although it is present in its deep structure. For example, The measure can be taken for six month at most and only for selected items., which in fact means The measure can be taken for six month at most and the measure can be taken only for selected items. In the PDT, two light verb constructions are annotated and both of them are linked to the same words. This would result in duplicate annotation of the words “measure” “be” and “taken” in the sentence. Such duplicates are detected and removed before the data are exported. 3.8.3
Overlapping different types of VMWEs As described previously, we combine explicit idiomatic annotation (DPHR), explicit light verb annotation (CPHR) and the verbal MWE annotation from PDT 3.0. If they overlap, the type of the MWE (ID or LVC) is always determined by the explicit DPHR/CPHR annotation. If only the PDT 3.0 MWE annotation is present, it always gets ID type as the most probable case; however, this could be checked manually in future.
Whenever IReflV overlaps with any other, usually larger MWE, both are correct and should remain in the output. Other overlaps of different types of VMWEs are not possible due to the source data we work with.
It is yet to be determined what to do with cases where an ID from DPHR and from PDT 3.0 MWE annotation overlaps with different word range. 3.9
After removing the overlaps, there are over 14,000 verbal multiword expressions exported in the PST format. Table 1 presents the numbers of individual types of VMWEs.
VMWE type It can be concluded that due to a well-founded, rich annotation scheme used in the Prague Dependency Treebank, which also conforms to most of the four PARSEME MWE annotation principles, we can almost fully automatically transform the original MWE annotation into the PARSEME Shared Task verbal MWE types. By that, we can extract 14,032 VMWEs.
In the near future, we still want to manually check some borderline cases mentioned above, e.g. whether an isolated verbal PDT 3.0 MWE should be always an ID, or how to solve overlapping annotation of the same type but of a different range. We will include deverbative MWEs from separate automatic lexicon-based annotation. 5
The work described here has been supported by the project PARSEME, LD14117, by the Ministry of Education, Youth and Sports of the Czech Republic, and carried out within the framework of the project COST IC1207 PARSEME. The project used data distributed by the LINDAT/CLARIN repository, supported by the Ministry of Education, Youth and Sports of the Czech Republic (projects LM2010013, LM2015071). We also thank our colleague Milena Hnátková who kindly extracted deverbative variants of VMWEs using her phraseme database and we are working on incorporating this data into our outputs. [5] Milena Hnátková. Znacˇkování frazému˚ a idiomu˚ v Cˇeském národním korpusu s pomocí Slovníku cˇeské frazeologie a idiomatiky. Slovo a slovesnost, 2002. [16] Veronika Vincze, Agata Savary, Marie Candito, Carlos Ramisch, and Fabienne Cap. Annotation guidelines for the PARSEME shared task on automatic detection of verbal multiword expressions, version 6.0, 2016. http: //typo.uni-konstanz.de/parseme/images/shared-task/ guidelines/PARSEME-ST-annotation-guidelines-v6.pdf.