For centuries, linguists have deliberated on how to represent meaning. Recently, this inquiry has been pursued not only as an intriguing theoretical problem but also due to its practical implications for applications. Here we provide a comparison of two meaning representations rooted in two diferent linguistic traditions and based on diferent theoretical assumptions: the meaning representation used in the family of Prague Dependency Treebanks and the Uniform Meaning Representation. We discuss the possibility and limitations of an automatic “translation” between these two formalisms, focusing esp. on selected deep syntactic phenomena afecting the shape of sentence graphs. Specifically, we concentrate on predicates and their argument structures, lexicons available for both approaches, levels of abstraction, and on coreference. We believe that the mutual inspiration of both approaches can lead to a substantially deeper understanding of language semantics.
tions such as the Minimal Recursion Semantics (MRS) [
While [
comparison of two meaning representations rooted in two diferent linguistic traditions and based on difer- 2.1. PDT-MR: PDT meaning ent theoretical assumptions: the meaning representa- representation tion used in the family of Prague Dependency Treebanks (PDT-MR, references in Sect. 2.1 below) and the Uniform The Prague Dependency Meaning Representation (PDTMeaning Representation (UMR, references in Sect. 2.2 MR) originates primarily in the tectogrammatical layer below). of language description [9, 10, 11], as designed within
The choice of the first formalism is clear from our the theoretical approach of the Functional Generative perspective: The three most prominent PDT treebanks, Description (FGD) [12, 13] and instantiated in the famnamely the original PDT, PDiT (Czech texts with dis- ily of Prague Dependency Treebanks, esp. the Prague course annotation) and PCEDT (Czech portion of parallel Dependency Treebank (PDT) [14, 15] and Prague DisCzech and English texts) represent the most extensive course Treebank (PDiT) for Czech [16, 17], and the paraland well-developed datasets available for Czech (see Sect. lel Prague Czech-English Dependency Treebank (PCEDT) 2.1 for references). Furthermore, the PDT formalism has [18, 19]. been applied not only to Czech and English: a PDT-like PDT-MR is a dependency-oriented complex annotation annotation is available, among others, for Latin texts as scheme covering deep syntax, with predicate-argument well.3 structure forming a core of the dependency representa
Moreover, we are familiar with this approach and tion. It presents also meaning-relevant morphological inpossess well-functioning processing tools (like a graph formation (like tense, number, gender, or (deontic) modality), information structure and discourse relations, including coreference annotation. Fig. 1 exemplifies slightly simplified meaning representation following the PDT-MR guidelines.
3The texts annotated in the PDT style are the Index Thomisticus Treebank (ITTB) and a portion of the Latin Dependency Treebank (LDT); https://itreebank.marginalia.it/view/download.php
One of the main features of the PDT-MR approach is its focus on linguistically structured meaning (rather than on semantics or even pragmatics beyond the language structure). As such, the goal of the PDT-MR annotation of a sentence is to capture especially the lexical choice (content words), the deep syntactic relations among them, the meaning-relevant morphological categories, coreference relations, and information structure. Consequently, the PDT-MR annotation more-or-less directly refers to the annotated text.
získal získat-001 :aspect performance :modal-strength full-affirmative
got --> get :ARG2 :manner :ARG1 :ARG0 V volbách volit-001 :wiki 1992_Estonian_parliamentary_electionQ372557
:aspect activity :modal-strength full-affirmative in elections --> (to) vote podle výsledků
výsledek :wiki voting_resultQ51591359
:refer-number plural according to results --> result :temporal :ARG1 :place :temporal :mod nedělních date-entity :wiki SundayQ132 :weekday neděle Sunday's --> Sunday parlamentních parlament :wiki RiigikoguQ217799 :refer-number singular parliamentary --> parliament v Estonsku country :name Estonsko :wiki EstoniaQ191 :refer-number singular in Estonia --> Estonia hlasů hlas :wiki voteQ1306135 :refer-number plural votes --> (a) vote včerejší včera yesterday's --> yesterday předběžných předběžný preliminary blok Vlast , jehož
blok :name Vlast :wiki IsamaaQ163347 :refer-number singular bloc Homeland, whose --> bloc :quant nejvíce nejvíce most :ARG0-of kandidátem byl kandidovat-001 :aspect performance :modal-strength full-affirmative was nominated as a candidate --> nominate as a candidate :ARG2
:ARG1 prezidentským
prezident :wiki presidentQ30461 :refer-number singular presidential --> president
Lennart Meri
person :name Lennart Meri :wiki Lennart_MeriQ153149 :refer-number singular
Lennart Meri 2.2. UMR: Uniform Meaning sentence level representation.
Representation Besides that, UMR provides a more comprehensible annotation of epistemic modality, and marks temporal and The Uniform Meaning Representation [20, 21, 22] is a se- coreference relations (both intra- and inter-sentential); mantic annotation schema that presents sentence mean- this forms the document-level representation, as illustrated ing while abstracting away from syntax (and thus it is in Fig. 3. designed specifically for cross-lingual applications). UMR also aims to capture quantification and scope for
UMR elaborates the Abstract Meaning Representa- the benefit of logical inference [ 20]; however, this kind tion (AMR) [24, 21] that focuses primarily on predicate- of annotation is not available in the released dataset [22]. argument structures and was developed first for English One of the main UMR features is a looser relation but later applied also to other languages, incl. Czech to syntax and the primary focus on semantics. Conse[25]. This part of the UMR representation, referred to quently, it provides the same representation for all posas the sentence-level representation, captures—in addition sible (syntactic) variants of a statement, including its to predicate-argument structures—esp. representation of restructuring or splitting into more sentences. In fact, multi-word expressions and named entities (including this approach allows for much broader interpretation of their anchoring in the English Wikipedia), and aspect sentences to be represented, compared to the PDT-MR annotation for predicate verbs. Fig. 2 illustrates the UMR approach. nodes linked by coreference arrow(s).
stract depiction of sentence meaning. Following its AMR predecessor, it seeks rather for a logical representation than for a syntactic one. According to the AMR 1.2.6 Specification (dated May 1, 2019), 4 it adopts a simplified, standard neo-Davidsonian semantics [26, 27, 24].
In UMR, two types of nodes are distinguish, as illustrated in Fig. 4. Non-leaves (inner nodes) stand for variables. The variables are instances of concepts, which are represented as leaves in the UMR graph. They represFaigmuereex3am:Tphleesdenotceunmceeinntd-ilceavteels tUeMmpRoraanl nreolatatitoionsn, mofotdhael sent primarily entities (as man, parlament ‘parliament’, dependencies and coreference chains identified in this sen- or blok ‘bloc’) and events (as získat-001 ‘get’, volit-001 tence: ‘(to) vote’, or taste-01) but there are also special keywords (i) The temporal annotation determines mutual temporal re- for entity types (as date-entity in Fig. 2),5 quantity lations for all temporal expressions and all events identified types (e.g., temporal-quantity) and for discourse rein the sentence and relates them to the date of the document lations (as, e.g., conjunction and) and other operators creation (e.g., the variable s5v3, standing for the relative tem- (e.g., more-than). poral expression včera ‘yesterday’, refers to the particular time Having two types of nodes, there are also two types of period before the document were created; further, the event edges in UMR: First, the edges connecting leaf nodes with conceptualized by the predicate získat-001 ‘get’, variable s5z, corresponding variables, representing thus the instance happened after the event identified by the volit-001 ‘vote’, vari- relation. (aibi)leThse5vm,oedtca.l)a.nnotation indicates that the author of the text The second type of edges, those connecting non-leaf is sure that all three events identified in the sentence have nodes represent semantic relations between concepts happened (encoded as the ‘:full-afirmative’ relation). (instantiated as respective variables). (iii) As this sentence is a part of a longer document, the annotation identifies which events and which entities has been taste-01 already mentioned in the document (e.g., the volit-001 ‘vote’ event, variable s5v, is the same event as the one with the s1t today variable s3v mentioned in one of the previous sentences).
meaning representations. However, they difer in the way how graph nodes and edges are used to represent sentence meaning.
PDT-MR. In PDT-MR, the graph reflects deep syntactic structure of a sentence. Its nodes represent content words (or, better to say, their deep syntactic counterparts, see Fig. 1). Edges stand for deep syntactic relations between content words. The only exemptions are (i) the technical root node (containing metadata such as the ID of the sentence) serving as the mother node of the main predicate in a sentence and (ii) special nodes and edges used for the representation of paratactic structures (coordination, apposition). In fact, the PDT-MR structures are trees (when ignoring coreference links); i.e., any lexical content that should be repeated in the sentence structure (calls for “re-entrancy”) is represented as two (or more) s1t2 4https://github.com/amrisi/amr-guidelines/blob/master/amr.md 5These include also abstract entities (as person for Lennart Meri) identifying types of named entities, i.e., expressions referring to concrete persons, institutions, places, etc., see also footnote 15. :ARG1 (s1f / free-04
:ARG1 s1p) :temporal (s1t2 / today))
There is one important feature of the AMR (and UMR) PDT-MR. PDT-MR implements only the very first steps formal representation: it allows for “re-entrancies”, i.e., a of such an abstraction, as represented by the concept of variable can appear as a child of more than one semantic the so-called t-lemma (understood as a meaning counrelation; thus the structure is a directed, rooted graph, terpart of a lexical unit present in a surface sentence) which is typically acyclic, as illustrated in Fig. 4 and in and the sempos attribute (semantic part of speech). For iantsetxatn-cferierenldaltyiown)a.y below (the slashes “/” represent the example, the morphological possessive adjective matčin ‘belonging to mother’ is represented by the same t-lemma (s1t / taste-01 as the morphological and semantic noun matka ‘mother’ :ARG0 (s1p / person :wiki "Edmund_Pope" and it is characterized as a possessive form; similarly, the :name (s1n / name relative possessive pronoun jehož is captured as který ::oopp12 ""EPdompuen"d)") in Fig. 1. However, diferent t-lemmas are supposed for bojovat ‘to fight’, bojování ‘fighting’, boj ‘(the) fight’, and bojovník ‘fighter’. 6
forming the PDT-MR data to UMR presents a challenge due to two main reasons: A. a lack of such information in the PDT-MR data and B. an insuficient definition of these concepts in UMR, as we discuss below.
from PDT-MR to UMR is relatively easy: Each non-root PDT-MR node retains its id (= variable) and its lexical content is moved to a newly created leaf node, connected A. Lack of information in PDT-MR: First, the relevant with the original one by an edge representing the instance information—which lexemes (words) are related to event relation. In this way, the requirements of the formal concepts (verbal predicates) (and how they are related) structure of the sentence-level annotation are secured. In and which rather represent entities—is only very partially addition, nodes within a single sentence that are marked available in the PDT-MR data (in the form of a note in the as coreferential should be merged, as is discussed in Sect. related lexicons, see below). Thus, we also experiment 3.3 (under type A). with external lexical resources (in combination with some
As for the opposite transformation (from UMR to heuristics or manual annotation).
PDT-MR), the situation is more complicated: First, each Selected types of derivational information are stored UMR leaf-node and the respective variable node must in the MorfFlex dictionary [29, 30], which covers genbe merged to a single PDT-MR node, with the variable eral Czech morphology. Even more, this information is serving as the unique identifier of the node and with the formalized so it can be used to identify base lexemes auconcept having the role of the content word. Further, tomatically. However, the lexicon is focused mainly on nodes with re-entrancies must be split (each of the origi- (inflectional) morphology; thus, the derivational infornal relations going to a diferent clone of the child node, mation might not be entirely complete in some cases. and the children will be linked as coreferential). In this There is one more language resource very relevant for way, the adjusted sentence-level graph meets the PDT- this task, namely DeriNet, the Lexical Network of WordMR formal requirements. However, there is no straight- Formation Relations in Czech [31, 32]. It has a form of a forward way how to transfer the information stored in network where nodes represent Czech lexemes and edges the UMR document-level representation. correspond to derivational links, i.e., relations between derivatives and their base lexemes. Thus DeriNet can 3.2. Events: Predicate and its argument reveal a relationship among, e.g., bojovat ‘(to) fight’, bostructure jování ‘fighting’, boj ‘(the) fight’, bojovník ‘fighter’, bojující ‘fighting (Adj)’, bojiště ‘battlefield’, bojovný ‘fighting 3.2.1. Concepts vs. content words as nodes (Adj)’, bojůvka ‘(storm) troop’, but also zabojovat ‘to fight UMR. One of the crucial distinctions UMR works with shortly’, odboj ‘resistance movement’, odbojář ‘resistance is the conceptual distinction between entities (objects), fighter’, zbojník ‘brigant’, souboj ‘combat’, and many othstates (properties), and events (processes). The aim is to ers; all of them share some aspects of the base lexeme abstract from morphological characteristics, i.e., whether bojovat ‘(to) fight’. the given concept appears in the surface text as a morpho- 6The underlying Functional Generative Description [12, 13] stipulogical noun, adjective, or verb. For example, think of a lates that derived surface forms that preserve their semantic part of driving event as a concept represented by the drive pred- speech are represented by their base words, e.g., bojovat ‘to fight’, icate (drive-01 in the PropBank lexicon [28]); then, drive, bojování ‘fighting’ and boj ‘(the) fight’ (being semantic verbs) share adnridviintsgaarsgwumelelnatssdtrriuvcetruarere(trheeprlaesstenotneedabsysothmisebporeddyicthaatet ttnhiooeunsn,ac(moagneetlrneaxtrinycoatulocnbo).onHjcoevopnwtíekavn‘edfigrh,tihtneurts’h,swehPohuDiclThd dhisaatcvale,aosthsnielfieysdaliammseistereemdpanrenustmeicnbtearacts as ARG0 of the driving event). of types are covered in this way.
However, neither MorfFlex nor DeriNet provide an For example, in the UMR approach, so-called stative information on the type of derivation—whether a partic- verbs (as, e.g., love) are treated diferently than verbs ular derivative is an event nominal, agent noun, property, denoting processes (as, e.g., run or damage). According place, tool, etc. Thus, some heuristics must be applied to the Specification, these verbs indicate events only if to identify the particular type of derivation, as it has an packaged as predication (and non-events in modification impact on the argument structure of the derivative. or reference packaging). Consequently, their annotations
To start with the simplest and most systematic class of in a main clause and in a relative clause difer: For exdeverbal nouns, we focused on nouns ending with -ní/-tí ample, in My cat loves wet food, the verb love denotes an ifrst. In total, 1,690 such nouns are identified in the PDT- event of loving; in My cat, that loves wet food, is beautiful, MR data (source: the PDT-Vallex lexicon, see Sect. 3.2.2). the verb love is packaged as a modification, thus it is not We combined this information with the derivational in- considered as an event (with all the consequences for formation stored in DeriNet and in MorfFlex. annotation). This distinction, however, is questionable
We learned that even for this type of nouns, the deriva- for Czech and Latin, where statives represent a blurred tion information is not complete in these resources in all category; thus, operative criteria for their identification cases. For example, ambiguous dojetí ‘arrival’ or ‘emo- cannot be applied (in contrast to English, where stative tion’ is only shown as derived from dojet ‘arrive’ in Der- verbs exhibit specific syntactic behavior). iNet; in addition, MorfFlex identifies also dojmout ‘touch; In any case, the criteria proposed in the Specification afect’ as the base verb. With the help of DeriNet and are not directly applicable to Czech and Latin data in MorfFlex, we were able to process 1,668 -ní/-tí nouns and many cases. This leads to the need to specify the boundidentify the base verb lexemes. Seven of the remaining 22 ary diferently for those two languages. Tentatively, all nouns have an explanatory note in PDT-MR identifying concepts represented as predicate verbs and concepts the base verb, reducing the number of unidentified nouns used in predication are considered events in Czech and to 15. Latin UMR data.8 As such, they are characterized by
The second necessary step in the noun-to-verb conver- valency frames (∼ rolesets) stored in the PDT-Vallex lexision process is to identify the relevant sense among all con (see Sect. 3.2.2). Further, abstract predicates, selected senses of the identified base verb (as stored in the PDT- implicit rolesets (Sect. 3.2.3), and reified relations (Sect. Vallex lexicon), which is a necessary step allowing us to 3.2.4) are treated as events in Czech and Latin UMR data. capture the event argument structure of the derivative.
This task still needs to be completed. 3.2.2. PropBank lexicon vs. PDT valency lexicon B. Unclear boundary between entities and events in UMR: Furthermore (and even more importantly), UMR does not suficiently define the crucial boundary between entities and events. The UMR 0.9 Specification (dated August 8, 2022) simply states: 7 “[E]vent identification is based on a combination of semantic type and information packaging [33]. Semantic type refers to the diference between entities (or, objects), states (or, properties), and processes; this can be thought of as a categorization of things in the real world. Information packaging (also called discourse function or information structure), on the other hand, characterizes how a particular linguistic expression “packages” the semantic content.” However, the Specification suggests that it is (at least to some extent) the English grammar and English wordformation processes that are used as criteria to set the boundary (which, of course, contradicts the basic UMR principles).
guidelines.md#part-3-1-umr-concepts Core argument structure
tation of predicate-argument structure as used in AMR, relying on the PropBank lexicon (called “PropBank Frame Files”) [34]. This lexicon stores predicates (mainly verbs, but also nouns and adjectives) subdivided into individual senses, assigned with a set of arguments and their coarsegrained semantic arguments9 [28]. Originally designed for English,10 it has been later used for a number of other treebanks of diferent languages (e.g. for Hindi, Chinese, Arabic and others); see also “IBM Universal Proposition Banks” project11 [35].
The PropBank lexicon uses ARG0 to ARG5 labels to identify semantic roles of arguments, with ARG0 reserved for the Prototypical_Agent and ARG1 for the Prototypical_Patient or Theme12 [36]. In addition, the ARGM label for adjunct-like arguments is being used (with several subtypes, as e.g., location). For example, the 8Thus, stative verbs are treated in the same way as verbs denoting processes, as we cannot set reliable criteria for their identification. 9http://propbank.github.io/ 10http://propbank.github.io/v3.4.0/frames/index.html 11https://universalpropositions.github.io/ 12According to [28], “No consistent generalizations can be made across verbs for the higher-numbered arguments, . . . ”
stores information on candidate mappings and includes supplementary valency information that can be used for future manual extensions of the mappings.
verb base,13 thus identifying its two senses, 1. ‘be located in’ and 2. ‘justified by, made up of’; the first roleset, marked as base-01, describes three arguments: ARG0-PAG: agent basing something somewhere, ARG1-PPT: institution,
ARGM-LOC: for location, where ARG1 is based. The argument labels are preserved across the rolesets of a given predicate (wherever relevant), disregarding the sentence structure (e.g., compare John.ARG0 broke the window.ARG1 to The window.ARG1 broke).
posed mappings of PDT-MR functors to PropBan arguments, a default mapping can be used, as suggested by [46, 47]. Based on introspection,they hypothesized that ACT typically corresponds to the ARG0 argument, PAT is most often ARG1, ADDR is typically ARG2, and so on. For example, the verb živit ‘nourish’ has two actants, PDT-MR. PDT-MR adheres to the original valency the- which by default get the following roles: ory [37], as instantiated in several valency lexicons, ACT (Agent/Causal agent) → ARG0, VALLEX [38, 39], PDT-Vallex [40, 41] (both for Czech), PAT (Entity fed or maintained) → ARG1. EngVallex [42, 43] (for English), and Latin VALLEX v114 In this case, the argument labels agree with the argument [44] (for Latin). The PDT-Vallex and EngVallex serve for specification and argument labels of the English verb annotation of the PDT and PCEDT corpora, while Latin nourish (as provided in the PropBank lexicon), thus the VALLEX v1 was built upon the tectogrammatical layer of default mapping is correct. However, for verbs with more Latin texts annotated in the PDT style. than two arguments, the default mapping is not satisfac
Similarly as the PropBank lexicon, the lexicons of the tory enough. For example, the verb nachystat ‘prepare’ Vallex family provide valency frames (∼ rolesets) for indi- has three actants, which by default get the following vidual predicate (primarily verb) senses. Instead of num- roles: bered arguments, they use five labels for the so-called ac- ACT (Creator) → ARG0, tants (ACT for Actor/Bearer, PAT for Patient, ADDR PAT (Thing made ready) → ARG1, for Addressee, ORIG for Origin, and EFF for Effect). ORIG (Created_from) → ARG3. However, their specification difers from the PropBank The mapping of the first two argument labels seems approach: the first two arguments ( ACT and PAT) have correct (as they agree with the arguments of the verb no specific semantics, ACT being assigned typically to prepare); however, the semantics of the last argument the argument in the subject position (in an active sen- (Created_from) correspond rather to ARG2 of prepare tence), PAT to the argument in the object position (for (thus, the default ARG3 label is inconsistent with Propverbs with the only object position). Only with verbs Bank). with three and more arguments, semantics of individual To evaluate the proposed approach, we compared the arguments plays role. As a consequence, the labels are default mapping and the verb-specific mapping presented not preserved in lexical alternations with diferent syn- by [45] on the available 10,426 functor-argument pairs; tactic structure (e.g., compare Jan.ACT rozbil okno.PAT, the results are in Table 1. The table reveals that the default ‘John.ACT broke the window.PAT’ and Okno.ACT se roz- mapping represents a relatively good approximation for bilo, ‘The window.ACT broke’). the first three actants (reaching accuracy of 86.8%). However, it is a futile attempt to use the proposed mapping for the last two actants. Instead, we suggest to convert them to the general (verb non-specific) UMR roles Source (for ORIG) and Goal (for EFF).
To summarize, the combination of the mapping procedure proposed in [45] (for predicate verbs with the verb-specific mapping available) and the default arguments mapping for ACT, PAT and ADDR actants (for other verbs) can serve as a good starting point for future manual refinement of UMR argument labeling for Czech predicate verbs.
UMR approaches difer in the argument labeling style and in the specification of individual argument roles (manifested esp. in the treatment of the first two arguments). Fortunately, Hajič et al. [45] provide a partial mapping of PDT-MR rolesets to PropBank-based UMR rolestes. Based on existing resources, they have been able to convert automatically and with high certainty about 43% of PDT-Vallex argument labels, so-called functors (out of 42,116 PDT-Vallex functors) to PropBank argument labels. In this way, the core of the PropBank-like lexicon for Czech has been established. The lexicon has a form of a table with verb specific argument mapping (when available) that can be used for the automatic conversion of the 13http://propbank.github.io/v3.4.0/frames/base.html 14https://github.com/CIRCSE/Latin-VALLEX Non-core arguments and adjuncts
for non-core arguments, adjuncts, and other relations beyond the scope of the (core) argument structure.
Table 1 “non-verbal”) predicates. These predicates are used for Accuracy of the default actant to argument mapping. representing the predication of property, possession, and location. They are identified by special labels (serving as PDT-MR → UMR artificial lemmas) equipped with their own rolesets. mapping correct (%) incorrect For example, the abstract predicate have-place-91 is used for a (predicative) location; it has two argument ACT → ARG0 4,355 (82.6%) 918 roles, ARG1 for Theme and ARG2 for Location. This PAT → ARG1 3,829 (92.5%) 310 predicate applies, e.g., in sentences like Brambory.ARG1 OADRDIGR→→AARRGG32 46541 ((8240..46%%)) 18976 jsou ve sklepě.ARG2, ‘The potatoes.ARG1 are in the celEFF → ARG4 0 ( 0.0%) 216 lar.ARG2’. Similarly, the exist-91 predicate represents a thetic location, characterized by a reverse role semantotal 8,699 (83.4%) 1,727 tics (applicable, e.g., for Na obzoru.ARG1 je Sněžka.ARG2, ‘There is the Sněžka mountain on the horizon’).
Another example of constructions that should be treated using abstract predicates are constructions with
UMR adopts rather coarse-grained labels for adver- the copula verb, corresponding to být ‘be’ in Czech and bial modifications; compare, e.g., two general labels sum ‘be’ in Latin. In those cases, the following abstract for temporal relations (temporal, duration) and nine predicates should be used: more fine-grained temporal labels used in PDT-MR (dis- • have-mod-91 tinguishing, e.g., relations like When?, From_when?, (as in cs. Podle čeho soudíte, že v tom jste nejlepší?, To_when?, or Till_when?). Given this, PDT-MR labels ‘Why do you think you are the best at that?’; lat. for adverbial modifications can be (at least tentatively) Vita ipsa brevis est, ‘Life itself is short’), mapped to UMR relations based on a simple translation • have-role-91 table. (as in cs. Vinken je prezidentem společnosti Elsevier
However, UMR introduces also some relations that N. V., ‘Mr. Vinken is a chairman of Elsevier N.V.’; are more specific than those used in PDT-MR, as, e.g., lat. Cato quaestor fuit, ‘Cato was quaestor’), quantity, age, topic, or medium. Thus, in the data • identity-91 converted from PDT-MR, these labels are not identified (as in cs. USA jsou jedinou zemí, kde . . . , ‘The US correctly; instead, more general UMR labels are used, as, is the only country where . . . ’). e.g. mod (when used as a nominal modifier).
3.2.3. Abstract concepts vs. strong lexicalization PDT-MR data that should be represented by abstract predOne of the main goals of the UMR approach is to provide icates is a challenging task. We can indicate a tentative a meaning representation usable for various languages al- list of Czech predicates (e.g., the respective senses of the lowing for cross-linguistic comparability of annotations. verbs mít ‘have’, patřit ‘belong’, vlastnit ‘own’, etc.) as This is supported by the introduction of the concepts of well as other relations (as, e.g., constructions with possesabstract predicates and implicit rolesets.15 PDT-MR, on the sive forms, like Mariina/její taška, ‘Maria’s/her bag’) that other hand, can be characterized as strongly lexicalized express possession with reasonable certainty. However, approach, relying on the predicate structure of individual the cases of location predication and property predica(mostly verbal) predicates. tion are more complex as all their subtypes are typically expressed by the verb být ‘be’ in Czech, which is categorized either (i) as the copula or (ii) as the existential Abstract predicates or so-called substitute verb (subsumed under the single UMR. UMR introduces nine abstract (also referred to as PDT-Vallex entry in the PDT-MR data). Thus, it is not possible to automatically distinguish more subtle senses as required in UMR.
The same holds true for Latin, where the verb sum ‘be’ can correspond to several UMR abstract predicates. 15In addition, UMR employs a set of abstract entities identifying entity types. They serve several purposes: (i) they stand for arguments in case of not overtly present arguments (or arguments present just as pronouns), (ii) they are used for classification of named entities (e.g., Lennart Meri is classified as a person in Fig. 2), and (iii) they provide an identification of structured data as special “en- Implicit rolesets tities” (as, e.g., date-entity, further structured with attributes UMR works with a list of other implicit rolesets that like day, month, year, century, etc.) or “quantities” (as, e.g., conceptualize various linguistic constructions. We can wmoitnhetthaerayt-trqibuuatnetsiqtuya,ntteanmdpournaitl).-quantity-quantity, both distinguish two main types here, rolesets for special linFocusing on the predicate-argument structure in UMR and PDT- guistic constructions and those used for the analysis of MR, we leave abstract entities aside here. structured texts.
A. First, to exemplify the rolesets for special linguistic constructions, we can list the following: • have-degree-91 as in comparison constructions (e.g., Dívka je vyšší než chlapec, ‘The girl is taller than the boy’), • include-91, as in the part-whole relation (e.g., Pro blok Vlast hlasovalo asi 20.5 procenta z celkového počtu 457 319 voličů, ‘About 20.5 percent of the total number of 457,319 voters voted for the Homeland bloc.’) • resemble-91 is used for analogies (e.g., It was like
mud running down the mountain. . . )
Identification of similar constructions in the PDT-MR data requires future examination; we postpone this task to the (near) future.
B. Second, the structured texts can be exemplified with, e.g., cite-91 for citations, hyperlink-91, or street-address91. In PDT-MR, there are special rules for structured text. However, the representation adhere to the language (deep syntactic) structure: it is governed by the general rules for distinguishing between verbal clauses and nonverbal clauses. Thus, we assume that there is only a very limited possibility to automatically convert the PDT-MR data to the UMR-compliant form. 3.2.4. Reification UMR. UMR fully adopts the AMR approach, within which reification is understood as a technique to convert a relation into an (abstract) predicate. This technique makes it possible to put focus on the (original) relation itself, to modify or to negate it. Then, the concepts interlinked by the original relation are converted to the arguments of the resulting abstract predicate.
To illustrate the technique, compare the annotation of the sentence We know the knife that is in the drawer (focusing on the knife, Fig. 5, upper part) and its modification We know the knife was not in the drawer yesterday; as the place relation is modified by the temporal concept and negated, it asks for reification, i.e., it is converted to the have-place-91 predicate, as shown in Fig. 5, lower part.16
According to the AMR 1.2.6 Specification (dated May 1, 2019), “AMR with reification” is considered “real AMR” (with “non-reified relations as semantic sugar”). To put it diferently, following the “reify all the time” principle, it would eliminate almost all relations17 and replace them with abstract predicates. However, as this would be an inconvenient technicality, the AMR Specification Coreference is generally understood as a relation between two or more expressions in a text that refer to the same concept, seen as a mental concept of a realworld entity or event. Such expressions usually form coreferential chains, which make the text(s) coherent. 16The example is borrowed from the AMR Specification, Members of the coreferential chain are connected by an https://github.com/amrisi/amr-guidelines/blob/master/amr. anaphoric relation, i.e., the intra-textual relation that is bimd#reification, and converted to follow the UMR principles. lateral and asymmetric, having an anaphor (a pronoun in 17With the exception of argument relations, relations used in con- most typical case) and its antecedent/postcedent (usually a junctions, and relations or attributes related to abstract entities, content word). Primarily expressions referring to entities, see also footnote 15. prefers non-reified relations when annotating data, aiming to support the corpus consistency—unless reification is needed (i.e., unless focusing on the relation, modifying or negating it). Further, it suggests that the AMR representation “will be normalized into reified form behind the scenes.” PDT-MR. PDT-MR does not allow for negating or modifying a relation itself – the underlying principles suppose that a speaker will overtly mark such communication needs, thus they will express the focus on the relation itself by choosing diferent syntactic structure and/or lexicalization, accompanied it with the relevant information structure.
UMR) relies on the data post-processing within which the AMR/UMR representations are converted into the reified forms, we give up attempts to identify constructions in the PDT data that call for reification and leave them to be handled in the subsequent phases of the project.
but also those referring to events can be interlinked with the anaphoric relation.
PDT-MR. In the PDT-MR approach, all types of coreference are treated in the same way. Each expression in the anaphoric reation is typically represented as a single node in the graph.18 The (node for the) anaphor bears a set of coreference attributes identifying esp. the ID(s) of the antecedent/postcedent node(s), the type of the coreference19 and the type of reference (e.g., specific or generic).
For example, in Fig. 1, the relative pronoun jehož (normalized as který ‘which’) coreferes with blok ‘bloc’ (i.e., it identifies its ID in the respective attribute), which is visualized as the brown arrow (brown color stands for grammatical coreference).
From the technical point of view, there is no diference in treating coreference (and bridging relations) within a sentence and these relations crossing sentence boundaries.
ence relations (two applicable within a single sentence and one for inter-sentential relations); we will briefly sketch them now and comment on the possibilities of the PDT-MR to UMR conversion. 3.3.1. Coreference within a single sentence Re-entrancy of a variable taste-01
taste-01 Edmund
Pope #Cor
today free-04
ARG1 Edmund
Pope
today free-04 Further, UMR employs the idea of inverse roles (already
který závod závod intro d'wuhcoe'd in A'fMactRo)r,y'which are used primaril y'fafcotorrtyh'e annotation of relative clauses (type A below), nominalizations (type B below), and embedded interrogatives (which we leave aside for the time being). The inverse role makes it possible to represent an embedded event as an event modifying one of the arguments; thus it is relevant for the sentence-level representation.
As already discussed in Sect. 3.1, there is a possibility of A. In PDT-MR, a relative clause is represented as a a re-entrancy of a variable within a sentence in UMR, subtree rooted at its verbal predicate, which itself is a exemplified by Fig. 4 representing the sentence Edmund child of (the node for) the modified concept; further, the Pope tasted freedom today. Here the Edmund Pope entity relative expression (a relative pronoun or pronominal serves as ARG0 of the taste-01 predicate and at the same adverb) typically serves as an argument or adjunct in the time as ARG1 of the free-04 predicate. This type is strictly valency structure of the clause predicate.20 At the same limited to intra-sentential relations and applies to the time, there is a coreferential link between the relative exsentence-level representation. pression and the modified concept. For example, consider
In PDT-MR, there are two nodes representing this en- the following sentence: Lidé, kteří bydlí v blízkosti závodu, tity, one in the argument structure of the predicate taste- si stěžují na zdravotní potíže, ‘People who live near the 01 (with the t-lemma Edmund Pope) and one in the argu- factory have been complaining of health problems’. Its ment structure of free-04 (with the special coreferential (simplified) PDT-MR representation is presented in Fig. t-lemma substitute #Cor and the coreference attribute 7 (left). The tree for the relative clause is rooted at the storing the ID of the antecedent node); the relation is node for the verb bydlet ‘live’, which is treated as the visualized as an arrow interconnecting these two nodes, head of the attribute clause (the RSTR functor) modifying see Fig. 6. the expression člověk ‘person’ (standing for lidé ‘people’);
The information provided in PDT-MR is suficient for the relative pronoun který ‘who’ is ACT of the bydlet the sound identification of nodes that should be merged ‘live’ predicate; this pronoun is interconnected with its antecedent člověk ‘person’ with a coreferential link.
When converting to UMR, the node for the relative expression and the one for the modified expression are merged; further, the relation between (the node for) the predicate verb of the relative clause and (the one for) 18The antecedent/postcedent node(s) may stand for the whole subtree(s) it/they govern(s). Further, the PDT-MR annotation schema also allows for exophora, i.e., a type of coreference with a pronoun referring to a situation or reality external to the text; we will leave such cases aside here. 19Apart from grammatical and textual coreference, relations of bridg- 20The relative expression can be embedded more deeply in the sening anaphora are also distinguished, incl. relations such as set- tence (e.g., in the case of the possessive personal pronoun); in such subset, part-whole or function-object. cases, the conversion follows the same principles.
today
For example, let us have the nominal group představitel republiky ‘the representative of the republic’: in UMR, the agentive noun představitel ‘(the) representative’ should be seen as ARG0 of the predicate představovat-003 ‘represent’ (i.e., it is ‘the person who represents the republic’), while republika ‘republic’ serves as ARG1 of the predicate, as the following annotation shows:
21For simplicity, the PDT-MR-like labels are kept in both graphs in Fig. 7 (supposing that they will be converted to the UMR labels in the following steps).
Unfortunately, neither the PDT-MR data nor available external resources allow for a sound identification of nominalizations and their source predicates (as already discussed in Sect. 3.2.1). Thus, only for those entities that can be identified as related to events—i.e., we can identify the source predicate, the type of derivation, and the respective valency frame (∼ roleset) in the PDT-Vallex lexicon identifying the argument structure, as was discussed in Sect. 3.2.1 and 3.2.2—we can modify the source PDT-MR representation to comply with the target UMR principles.
grants Language Understanding: from Syntax to Discourse 2016, pp. 2371–2378. URL: https://aclanthology.org/ Prague Discourse Treebank 3.0, 2022. URL: http: L16-1376. //hdl.handle.net/11234/1-4875, LINDAT/CLARIAH[9] E. Hajičová, Dependency-Based Underlying- CZ Digital Library, ÚFAL, MFF UK, Prague, Czechia.
Structure Tagging of a Very Large Czech Cor- [18] J. Hajič, E. Hajičová, J. Panevová, P. Sgall, O. Bopus, Special Issue of TAL Journal, Grammaires jar, S. Cinková, E. Fučíková, M. Mikulová, P. Pajas, De Dépendence / Dependency Grammars (2020) J. Popelka, J. Semecký, J. Šindlerová, J. Štěpánek, 57–78. J. Toman, Z. Urešová, Z. Žabokrtský, Announc[10] J. Hajič, E. Hajičová, J. Mírovský, J. Panevová, Lin- ing Prague Czech-English Dependency Treebank guistically Annotated Corpus as an Invaluable Re- 2.0, in: Proceedings of LREC 2012, ELRA, Issource for Advancements in Linguistic Research: tanbul, Turkey, 2012, pp. 3153–3160. URL: https: A Case Study, The Prague Bulletin of Mathe- //aclanthology.org/L12-1280/. matical Linguistics (2016) 69–124. doi:10.1515/ [19] J. Hajič, E. Hajičová, J. Panevová, P. Sgall, pralin-2016-0012. S. Cinková, E. Fučíková, M. Mikulová, P. Pa[11] D. Zeman, J. Hajič, FGD at MRP 2020: Prague jas, J. Popelka, J. Semecký, J. Šindlerová, Tectogrammatical Graphs, in: Proceedings of J. Štěpánek, J. Toman, Z. Urešová, Z. Žabokrtský, the CoNLL 2020 Shared Task: Cross-Framework Prague Czech-English Dependency Treebank Meaning Representation Parsing, Association 2.0, 2012. URL: http://hdl.handle.net/11858/ for Computational Linguistics, 2020, pp. 33–39. 00-097C-0000-0015-8DAF-4, LINDAT/CLARIAHURL: https://aclanthology.org/2020.conll-shared.3. CZ Digital Library, ÚFAL, MFF UK, Prague, doi:10.18653/v1/2020.conll-shared.3. Czechia. [12] P. Sgall, Generativní popis jazyka a česká dekli- [20] J. van Gysel, M. Vigus, J. Chun, K. Lai, S. Moeller, nace (Generative Description of a Language and J. Yao, T. O’Gorman, J. Cowell, W. Croft, C.-R. the Czech Declension), Academia, Praha, 1967. Huang, J. Hajič, J. Martin, S. Oepen, M. Palmer, [13] P. Sgall, E. Hajičová, J. Panevová, The Meaning of J. Pustejovsky, R. Vallejos, Designing a uniform the Sentence in Its Semantic and Pragmatic Aspects, meaning representation for natural language proReidel, Dordrecht, 1986. cessing, KI - Künstliche Intelligenz 35 (2021) 343– [14] J. Hajič, E. Bejček, J. Hlaváčová, M. Mikulová, 360. doi:10.1007/s13218-021-00722-w.
M. Straka, J. Štěpánek, B. Štěpánková, Prague De- [21] S. Wein, J. Bonn, Comparing UMR and crosspendency Treebank - Consolidated 1.0, in: Proceed- lingual adaptations of AMR, in: Proceedings of the ings of LREC 2020, ELRA, Marseille, France, 2020, Fourth International Workshop on Designing Meanpp. 5208–5218. URL: https://aclanthology.org/2020. ing Representations (DMR 2023), Association for lrec-1.641/. Computational Linguistics, Nancy, France, 2023, pp. [15] J. Hajič, E. Bejček, A. Bémová, E. Buráňová, 23–33. URL: https://aclanthology.org/2023.dmr-1.3.
E. Fučíková, E. Hajičová, J. Havelka, J. Hlaváčová, [22] J. Bonn, C. Ching-wen, J. A. Cowell, W. Croft, P. Homola, P. Ircing, J. Kárník, V. Ket- L. Denk, J. Hajič, K. Lai, M. Palmer, A. Palmer, tnerová, N. Klyueva, V. Kolářová, L. Kučová, J. Pustejovsky, H. Sun, R. Vallejos Yopán, J. Van GyM. Lopatková, D. Mareček, M. Mikulová, sel, M. Vigus, N. Xue, J. Zhao, Uniform meaning J. Mírovský, A. Nedoluzhko, M. Novák, P. Pajas, representation, 2023. URL: http://hdl.handle.net/ J. Panevová, N. Peterek, L. Poláková, M. Popel, 11234/1-5198, LINDAT/CLARIAH-CZ Digital LiJ. Popelka, J. Romportl, M. Rysová, J. Semecký, brary, ÚFAL, MFF UK, Prague, Czechia. P. Sgall, J. Spoustová, M. Straka, P. Straňák, [23] M. Novák, UMR Visualization, unpublished, 2023. P. Synková, M. Ševčíková, J. Šindlerová, J. Štěpánek, [24] L. Banarescu, C. Bonial, S. Cai, M. Georgescu, B. Štěpánková, J. Toman, Z. Urešová, B. V. Hladká, K. Grifitt, U. Hermjakob, K. Knight, P. Koehn, D. Zeman, Š. Zikánová, Z. Žabokrtský, Prague M. Palmer, N. Schneider, Abstract Meaning RepreDependency Treebank - Consolidated 1.0 (PDT-C sentation for Sembanking, in: Proceedings of the 1.0), 2020. URL: http://hdl.handle.net/11234/1-3185, 7th Linguistic Annotation Workshop and InteropLINDAT/CLARIAH-CZ Digital Library, ÚFAL, MFF erability with Discourse, Association for ComputaUK, Prague, Czechia. tional Linguistics, Sofia, Bulgaria, 2013, pp. 178–186. [16] P. Synková, J. Mírovský, L. Poláková, M. Rysová, URL: https://aclanthology.org/W13-2322.
Announcing the Prague Discourse Treebank 3.0, [25] N. Xue, O. Bojar, J. Hajič, M. Palmer, Z. Urešová, in: Proceedings of LREC-Coling 2024, ELRA, X. Zhang, Not an interlingua, but close: CompariTorino, Italy, 2024, pp. 1270–1279. URL: https:// son of English AMRs to Chinese and Czech, in: Proaclanthology.org/2024.lrec-main.114. ceedings of LREC 2014, ELRA, Reykjavik, Iceland, [17] P. Synková, M. Rysová, J. Mírovský, L. Poláková, 2014, pp. 1765–1772. URL: https://aclanthology.org/
V. Sheller, J. Zdeňková, Š. Zikánová, E. Hajičová, L14-1332/. [26] D. Davidson, The logical form of action sentences, A. Vernerová, E. Bejček, Z. Žabokrtský, VALLEX in: N. Rescher (Ed.), The Logic of Decision and 4.5, 2022. URL: http://hdl.handle.net/11234/1-4756, Action, University of Pittsburgh Press, 1967, pp. LINDAT/CLARIAH-CZ Digital Library, ÚFAL, MFF 81–95. UK, Prague, Czechia. [27] J. Higginbotham, On semantics, Linguistic inquiry [40] J. Hajič, J. Panevová, Z. Urešová, A. Bémová, 16 (1985) 547–593. V. Kolářová, P. Pajas, PDT-VALLEX: Creating a [28] M. Palmer, D. Gildea, P. Kingsbury, The Proposition large-coverage valency lexicon for treebank annoBank: An Annotated Corpus of Semantic Roles, tation, in: Proceedings of The Second Workshop Computational Linguistics 31 (2005) 71–106. doi:10. on Treebanks and Linguistic Theories, volume 9 of 1162/0891201053630264. Mathematical Modeling in Physics, Engineering and [29] J. Hlaváčová, M. Mikulová, B. Štěpánková, Cognitive Sciences, Vaxjo University Press, Vaxjo, Konzistence morfologického slovníku MorfFlex, Sweden, 2003, pp. 57–68.
Jazykovedný časopis / Journal of Linguistics 72 [41] Z. Urešová, A. Bémová, E. Fučíková, J. Hajič, (2021) 855–861. V. Kolářová, M. Mikulová, P. Pajas, J. Panevová, [30] J. Hajič, J. Hlaváčová, M. Mikulová, M. Straka, J. Štěpánek, PDT-Vallex: Czech Valency lexiB. Štěpánková, Morflex CZ 2.0, 2020. URL: http: con linked to treebanks 4.0 (PDT-Vallex 4.0), //hdl.handle.net/11234/1-3186, LINDAT/CLARIAH- 2021. URL: http://hdl.handle.net/11234/1-3499, CZ Digital Library, ÚFAL, MFF UK, Prague, Czechia. LINDAT/CLARIAH-CZ Digital Library, ÚFAL, MFF [31] L. Kyjánek, Z. Žabokrtský, M. Ševčíková, J. Vidra, UK, Prague, Czechia.
Universal Derivations 1.0, A Growing Collection [42] S. Cinková, From PropBank to EngValLex: Adaptof Harmonised Word-Formation Resources, The ing the PropBank-Lexicon to the Valency Theory Prague Bulletin of Mathematical Linguistics 115 of the Functional Generative Description, in: Pro(2020) 5–30. ceedings of LREC 2006, ELRA, ELRA, Genova, Italy, [32] J. Vidra, Z. Žabokrtský, L. Kyjánek, M. Ševčíková, 2006, pp. 2170–2175.
Šárka Dohnalová, E. Svoboda, J. Bodnár, DeriNet [43] S. Cinková, E. Fučíková, J. Šindlerová, J. Ha2.1, 2021. URL: http://hdl.handle.net/11234/1-3765, jič, EngVallex - English Valency Lexicon 2.0, LINDAT/CLARIAH-CZ Digital Library, ÚFAL, MFF 2021. URL: http://hdl.handle.net/11234/1-3526, UK, Prague, Czechia. LINDAT/CLARIAH-CZ Digital Library, ÚFAL, MFF [33] W. Croft, Radical Construction Grammar: Syntac- UK, Prague, Czechia.
tic Theory in Typological Perspective, Oxford Uni- [44] M. Passarotti, B. G. Saavedra, C. Onambele, Latin versity Press, 2001. doi:10.1093/acprof:oso/ Vallex. A Treebank-based Semantic Valency Lexi9780198299554.001.0001. con for Latin, in: Proceedings LREC 2016, ELRA, [34] S. Pradhan, J. Bonn, S. Myers, K. Conger, Portorož, Slovenia, 2016, pp. 2599–2606. URL: https: T. O’Gorman, J. Gung, K. Wright-Bettner, M. Palmer, //aclanthology.org/L16-1414.
PropBank comes of age—larger, smarter, and more [45] J. Hajič, E. Fučíková, M. Lopatková, Z. Urešová, diverse, in: Proceedings of the 11th Joint Confer- Mapping Czech Verbal Valency to PropBank Arence on Lexical and Computational Semantics, ACL, gument Labels, in: Proceedings of the Fifth InterSeattle, Washington, 2022, pp. 278–288. doi:10. national Workshop on Designing Meaning Repre18653/v1/2022.starsem-1.24. sentations (DMR 2024), ELRA and ICCL, Torino, [35] I. Jindal, A. Rademaker, M. Ulewicz, H. Linh, Italia, 2024, pp. 88–100. URL: https://aclanthology.
H. Nguyen, K.-N. Tran, H. Zhu, Y. Li, Universal org/2024.dmr-1.10. proposition bank 2.0, in: Proceedings of LREC 2022, [46] J. Hajič, O. Bojar, S. Cinková, R. Sudarikov, ELRA, Marseille, France, 2022, pp. 1700–1711. URL: Z. Urešová, M. Popel, O. Dušek, Tectogrammathttps://aclanthology.org/2022.lrec-1.181. ical to AMR conversion: current status, un[36] D. R. Dowty, Thematic proto-roles and argument se- published, 2014. URL: https://www.clsp.jhu.edu/ lection, Language 67 (1991) 547–619. doi:10.2307/ workshops/14-workshop/, JHU Summer Workshop 415037. 2014 project. [37] J. Panevová, Valency Frames and the Meaning of [47] O. Bojar, S. Cinková, O. Dušek, T. O’Gorman, the Sentence, The Prague School of Structural and M. Popel, R. Sudarikov, Z. Urešová, Tecto to AMR Functional Linguistics 41 (1994) 223–243. and translation, unpublished, 2014. URL: https:// [38] M. Lopatková, V. Kettnerová, E. Bejček, www.clsp.jhu.edu/workshops/14-workshop/, JHU A. Vernerová, Z. Žabokrtský, Valenční slovník Summer Workshop 2014 project. českých sloves VALLEX (The Valency Dictionary of Czech Verbs VALLEX), Karolinum, Praha, 2016. [39] M. Lopatková, V. Kettnerová, J. Mírovský,