Universals of Linguistic Idiosyncrasy in Multilingual Computational Linguistics

Two specific objections to the current (UDv2) relations were discussed. First, there is a top-level indirect object relation (iobj), but it has been a source of confusion, owing in part to the diverse range of uses of the term “indirect object” in different linguistic traditions. We concluded that iobj should be considered for removal in UDv3, with multiple obj for double object constructions (different phrases meeting the UD criteria for obj could optionally distinguished via subtypes). Second, UD’s broad interpretation of the adverbial clause relation (advcl) lumps together adjunct clauses and complement clauses with oblique-like marking.¹⁰¹⁰10Consider the sentence “As it was raining, we worked on improving annotation.” The second subordinate clause, “on improving annotation”, is not adverbial in the ordinary use of the term, but rather is a complement clause with prepositional marking. “As it was raining” is a true adverbial clause adjunct, and should remain advcl. The group agreed to endorse the idea of a new cobl relation in UDv3 that would target oblique subordinate clauses [1].

Various specialized patterns in names, numbers, dates, and measurements are known to be challenging for syntactic annotation [2, 3]. We discussed cases like “Eminent linguist Mr. Bill Croft”, where the phrase “eminent linguist” and the title “Mr.” are arguably modifiers as they can be freely omitted. We agreed that the UD guidelines should be changed to treat these as modifiers, adopting the suggested label nmod:desc (for “descriptor”) to distinguish these from other kinds of nmod [2].

More controversial were date expressions (“July 30, 1980” – should this be considered headed?), numbered entities (“Room S108”), and other names with an entity type (“Michigan Street”, “Lake Michigan”) or suffix (“Richard III”, “BMW Inc.”). Some of these cases are currently listed as flat in the guidelines. It emerged from the discussion that there were two different interpretations of flat held by members of the group. One interpretation is that flat is an “escape hatch” for tricky cases related to names (etc.) which do not easily fit general-purpose syntactic constructions. This is the expansive view of flat. Another interpretation is that flat should be reserved for expressions that are truly headless, defying all possible attempts to identify one part as the syntactically most important element. We leave these issues to further discussion.

Negation is a grammatical category in all human languages. However, there is no explicit uniform encoding of negation in treebanks or in UD. Negation is also often expressed through a combination of different morphosyntactic elements, but, being a grammatical category, it is usually not considered a multiword expression or a phraseme. The purpose of the discussion point was to raise awareness for this phenomenon and to explore where to locate the representation of negation in annotated corpora. The following data can serve to illustrate the phenomenon.

Examples of morphosyntactic strategies for clausal negation include:

• $\mathrm{■}$

  In German, a clause can be negated by simply adding the negative adverb nicht “not”.

• $\mathrm{■}$

  In standard French, clausal negation is expressed by a combination of a pre-verbal particle, ne, and post-verbal pas, as in Il ne pleut pas. “It isn’t raining.” In this example, negation is expressed through two elements, both of which only express the category of negation.

• $\mathrm{■}$

  We also find so-called neg-words such as English nothing, i.e., indefinites that fulfill another dependency in a clause but, in addition, also mark negation, as in English Alex said nothing.

It is common that languages can use more than one of these elements to express a single negation, as in the French example Personne ne fait confiance à personne. (gloss: nobody NE makes confidence to nobody) ‘Nobody trusts anybody.’

Negation itself is not a uniform phenomenon on the meaning and usage side either, i.e., morphosyntactic negation marking can encode clausal negation, but also constituent negation, meta-linguistic negation, or expletive negation. A negative meaning can be explicit, or can be inferred. In addition, we find idiomatic combinations that express negation of various types: The German expression einen Dreck (lit.: “a dirt”) “nothing” marks a sentence as morphosyntactically negative. The English expression I’ll be damned if … has the effect that the if clause is semantically, though not morphosyntactically, negative. Finally, we find items that mark only pragmatically inferable negativity like sarcasm or irony. An example is German clause final, intonationally separated …– also nicht! (lit.: “dots thus not”) or …und ich bin der Kaiser von China “…and I am the emperor of China,” as in Alex ist echt schlau – also nicht/und ich bin der Kaiser von China! “Alex is really clever – certainly not!”

The various types of semantic negativity are also relevant for the treatment of MWEs, as we find lexical expressions that are restricted to occur in clauses with a particular type of negation, so-called negative polarity items. In the simplest case, there is a fixed expression with a particular negator, such as the German bound word Unterlass “stop”, which only occurs in the combination ohne Unterlass “without stop”. The German modal verb brauchen “need” requires a semantically negative clause, but the choice of the negation strategy is irrelevant. From a collocational or MWE perspective, brauchen would not be a single word, but rather form a collocation or MWE together with the semantic category of negation. Finally, NPIs such as English lift a finger or fine in cases with an explicit semantic negation, but also in some cases of (conventionalized) pragmatic negation, such as in denial (But I DID lift a finger). However, sarcasm and irony (i.e., only conversational pragmatic negation) do not seem to license any known NPIs.

Turning back to the seminar theme: Multiple exponence of negation is not marked as a dependency in UD, nor as a MWE in PARSEME. If Negation is a morphosyntactic category, shouldn’t it be marked? Yes, but probably as a grammatical category at the clausal level. As for the other phenomena, negation and negation-related phenomena might be a good motivation and testing ground for adding constructional information (as proposed by WG7) in addition to syntactic dependencies and classical MWEs to corpus annotation.

The working group was formed to address three main discussion topics:

• $\mathrm{■}$

  Discovering linguistic idiosyncrasy (Nurit Melnik)

• $\mathrm{■}$

  Identifying non-compositional MWEs in text (Francis Bond)

• $\mathrm{■}$

  NLP-based study of universals of linguistic idiosyncrasy (Agata Savary)

Over the course of six sessions, the group engaged in intense discussions which included but were not limited to these issues. This document aims to provide a comprehensive overview of our discussions and suggest potential directions for future research.

The central theme of our discussions revolved around the concept of idiosyncrasy. We were fortunate to have a diverse group of participants with expertise in various fields such as NLP, computational linguistics, theoretical linguistics (particularly Head-driven Phrase Structure Grammar and Construction Grammar), typology, grammar engineering, and more. This diverse range of perspectives greatly enriched our discussions.

Our conversations were structured around three interrelated topics.

Our initial strategy was not to spend time trying to come up with an precise definition of idiosyncrasy. Instead, we began by looking for phenomena which we would intuitively identify as idiosyncratic. We realized that such phenomena generally stood in opposition to what is “regular”.

One type of idiosyncratic phenomena are syntactic structures which are not part of a regular core. These structures are often discussed in the Construction Grammar literature. For example, the English the X-er the Y-er construction has a very unique morphosyntactic pattern.

• (1)

  The more the merrier. [en]

Another example is the do-be construction, which is subject to various idiosyncratic constraints that are not derived from general properties of the language.

• (2)

  What you have to do is *(to) get ready. [en]

Some constructions have a regular syntactic structure, but they can host lexical items which are not expected to appear in them.

• (3)

  I sneezed the foam off my cappuccino. [en]

Regular syntactic structures may also have idiosyncratic meanings. Thus, for example, in the following exchange, the meaning of the coordination of the two identical Swedish adjectives meaning ‘happy’ is ‘not so good’.

• (4)

  Are you happy? [en] Happy and happy [sw]

Another type of idiosyncrasy involves exception to rules. For example, although adjectives in English precede the nouns that they modify, the adjective enough can only appear post-nominally.

• (5)
  • a.

    • That sounds good enough. [en]

  • b.
    • *

      That sounds enough good.

We also considered the notion of idiosyncrasy from a typological cross-linguistic perspective and asked whether there are particular domains which are more susceptible to idiosyncrasy. One phenomenon that we focused on was what is referred to as pro-drop, namely the ability to “drop” pronominal subjects, which is found in Japanese but not in English.

• (6)
  • a.

    着いた tuita “$\varphi$ arrived” [ja]

  • b.

    I arrived. [en]

The common wisdom is that pro-drop depends on the richness of the morphology. However as the Japanese example indicates, this is not necessarily the case and this property is more idiosyncratic than is believed.

Other domains which were mentioned as potentially exhibiting idiosyncrasy were: temporals, kinship terms, negation, existentials, possessives, inherently reflexive verbs. This topic was later developed in the discussions of WG7, where they proposed to compile a typologically informed collection of “meta-constructions” (or domains) and the “morphosyntactic strategies” which languages employ to realize them.

One general type of idiosyncrasy that we discussed is mismatches between levels (phonology, morphology, semantics, syntax), particularly in the domain of MWEs. Following are some more specific cases that were presented as particularly challenging for NLP.

MWEs can encode single predications. In other words, MWEs can appear in the sense hierarchy in the same way as a single word. For example, the meaning of the English MWE look up is similar to phseek, yet it is subject to idiosyncratic syntactic constraints.

• (7)
  • a.

    • I looked up the word. [en]

  • b.

    • I looked the word up.

  • c.
    • *

      I looked up it.

  • d.

    • I looked it up.

Moreover, some MWEs may look like “regular” phrases but as MWEs their parts do not exhibit “regular” syntactic behavior. For example, the relationship between the noun and adjective in the MWE hot dog is not intersective modification; hot dog is not a dog_n:1 that is hot_a:1, but a kind of sausage (hot_dog_n:1). This effects the syntax:

• (8)
  • a.
    • #

      I ate a very hot dog. [en]

  • b.

    • I ate a very hot pizza.

However, even if we think of a MWE as a single predicate, bits of it can still be accessed and modified. In some cases, internal syntactic modification becomes semantically external. For example, although Texan is modifying the dust part of the MWE in (9a), it is interpreted as modifying the entire meaning of the MWE (9b).

• (9)
  • a.

    He bit the Texan dust. [en]

  • b.

    He died in Texas.

It should be noted that internal modification being interpreted externally is not only a feature of MWEs. Consider the following example.

• (10)
  • a.

    I have an occasional drink. [en]

  • b.

    I drink occasionally.

Syntactically, the adjective occasional is modifying the noun drink, but semantically, what is occasional is not the drink but rather the entire drinking event. Accounting for mismatches between levels was also discussed in WG1, WG3, WG9.

Following our WG discussions several of the participants joined forces with members of WG2 (Annotation of particular kinds of constructions) to discuss the relation between Construction Grammar and Universal Dependencies, thus forming WG7, initially named “CxG meets UD”. Some of the topics which we identified as ideas for future work were discussed in WG7 sessions.

• $\mathrm{■}$

  Create a cross-linguistic idiosyncracycon/constructicons
  even the ’rare’ constructions are often cross-linguistic

• $\mathrm{■}$

  Sense-tag corpora

Other ideas for future work were related to the challenge of automatic discovery of idiosyncrasy.

• $\mathrm{■}$

  Investigate the feasibility of detecting suprisal in automatic parsing

• $\mathrm{■}$

  Do 10-fold cross-validation and mine the errors
  (But how do we distinguish errors, creativity and idiosyncrasy?)

Finally, one issue that came up and prompted ideas for future research addressed the notion of idiosyncrasy with respect to large language models (LLMs): How do LLM-generated texts compare to natural texts in terms of the frequency and distribution of various idiosyncratic phenomena such as idioms, MWEs and that-less relative clauses?

In the current state of affairs, both UD and UniMorph contain morphological data but they tag it according to different although generally compatible annotation schemas. Therefore we suggest an effort to align these two resources. This effort will probably not include any changes needed to be done in UD.

As a first step we propose to map the differences between UD and UniMorph, both in terms of the features used and the structure of the features. The mapping will be done as part of an effort to extract morphological data from UD to UniMorph. Once it’ll be done, we could produce a list of the changes we believe are required and approve through the relevant committees in the management of both datasets.

Having both projects using the same annotation schema for morphological data will make it the cooperation between the project organizers smoother.

Currently, morphological information in UD is confined to the “feats” column where features are attributed to the entire node, and it makes languages seem different depending on the frequency of white space usage. Decomposing words into morphemes and marking the relations between morphemes with dependency arcs as done between words, will equate the structure of different languages and will make it easier to typologically compare languages.

The group proposed segmenting words into morphemes, mostly using the segmentation files from UniMorph, currently existing for about a dozen of languages. The lemma of each morpheme will be its canonical form and the features be associated only with the relevant morpheme rather than with the entire word.

There were 2 main implementation options discussed: one where the content of each “morphological node” is a morpheme, and one where it is a truncated word. For example, a word like “industrialization” will be decomposed into “industry”, “-al”, “-ize” and “-ation” according to the first option, and to “industry”, “industrial”, “industrialize” and “industrialization” according to the second. The benefits of the latter is that it does not require the annotators to decide on the “canonical form” and whether it even exists and it is more closely aligned with the UniMorph derivational morphology data, but on the other hand this option is less intuitive and diverges to some extent from the structure of UD for words.

This discussion should be revisited after the completion of the first phase.

This is a summary of the discussions that took place in Working Group 7 during the Dagstuhl Seminar 23191 Universals of Linguistic Idiosyncrasy in Multilingual Computational Linguistics.

WG7 was formed from the initial working groups 2 (Annotation of particular kinds of constructions) and 4 (Finding idiosyncrasy in corpora), who independently of each other realised that they wanted a group that was focused on the relation between Construction Grammar and Universal Dependencies. So for the second half of the seminar we formed WG7 with the initial name “CxG meets UD”, but we changed the name of the group to the more catchy UniCoDeX (Universal Construction Dependency Xrammar).

The group met for four sessions and had intense discussions which resulted in the formation of three interconnected tasks. This document tries to reflect the results of our discussions and the way forward.

The overall question of WG7 was how to connect Construction Grammar (CxG) with Universal Dependencies (UD) in a way that both Construction Grammar projects and Computational Linguistics projects can benefit from. Our discussions were organized around three interrelated topics:

The general goal of this topic is to verify the typological coverage of UD, increase its consistency, and advise users on how to analyze constructions in their languages. There are two important concepts: meta-constructions framed in comparative terms based on [1]; and morphosyntactic strategies that specific languages may employ. The group working on this topic will create a collection of annotation examples and guidelines for different languages and meta-constructions:¹⁵¹⁵15[1] calls them “constructions”, but we refer to these as meta-constructions to differentiate them from language-specific constructions in the Construction Grammar sense.

• $\mathrm{■}$

  This collection corresponds to the “Swadesh list” for morphosyntax from WG2 – we dub it the Nivre list since it was proposed by Joakim Nivre. Using this checklist, treebankers can determine which mophosyntactic strategies are used in their languages for each meta-construction and how to annotate them.

• $\mathrm{■}$

  One component of this collection is a spreadsheet with languages in the rows and meta-constructions in the columns, where the cells contain links to queries yielding annotation examples and notes indicating which morphosyntactic strategy is illustrated in each example.

• $\mathrm{■}$

  Another component is a table of meta-constructions and morpho-syntactic strategies, with examples from different languages and language families.

This collection will be used to promote consistent and typologically informed coverage of morphosyntactic strategies in UD and update the general UD annotation guidelines from a typological perspective, including morphosyntactic strategies as defined by [1].

• $\mathrm{■}$

  This will help UD to find areas where we can improve current explanations and analyses to be more typologically oriented.

• $\mathrm{■}$

  It will also be of help when extending the guidelines to improve the coverage of explanations.

• $\mathrm{■}$

  It will ensure that it is possible to represent all (or at least most) known typological diversity in UD.

• $\mathrm{■}$

  It will facilitate using UD for research in language typology.

The general goal of this topic is to develop recommendations for how to annotate UD treebanks with constructions. They should be useful for several different use cases, workflows and granularities, such as:

• $\mathrm{■}$

  Use cases: we could be building a new constructicon from the ground up and want to come up with good definitions of constructions, or we might already have an existing constructicon which we want to use for annotation or extend with new constructions

• $\mathrm{■}$

  Different annotation granularities: from the coarsest level (to just annotate the head of a construction with its name), to the most fine-grained (to also annotate all the construction elements with their names and spans within the sentence)

• $\mathrm{■}$

  Different workflows: people might want to annotate one construction at a time, or several at once – or they might want to use an iterative approach where they start with coarse-level annotation and then refine them

The general goal of this topic is how to formulate search queries that can locate interesting examples of a given construction. This is very closely related to the previous two topics, as they all depend on being able to search for constructions in treebanks.

The group discussed some issues that arise when it comes to formulating search queries, such as:

• $\mathrm{■}$

  we want guidelines that help people with writing and refining queries

• $\mathrm{■}$

  we want (semi-)automatic techniques for extracting relevant search queries from an existing constructicon entry

• $\mathrm{■}$

  precision/recall tradeoff: it is probably more important to have a good recall than good precision, but it is usually easier to improve the precision by modifying a query

• $\mathrm{■}$

  possible strategies to increase the recall can be to use approximate tree matching or to loosen some constraints in the query

The following are the existing constructicons that we are aware of:

• $\mathrm{■}$

  English: Berkeley FrameNet Constructicon: http://sato.fm.senshu-u.ac.jp/frameSQL/cxn/CxNeng/cxn00/21colorTag/

• $\mathrm{■}$

  English: Birmingham English Constructicon: https://englishconstructicon.bham.ac.uk/

• $\mathrm{■}$

  English: CASA (FAU Erlangen-Nürnberg): https://constructicon.de/

• $\mathrm{■}$

  German: FrameNet-Konstruktikon (HHU Düsseldorf): http://framenet-constructicon.hhu.de/

• $\mathrm{■}$

  Swedish: Svenskt konstruktikon (Univ. of Gothenburg): https://spraakbanken.gu.se/karp/#?mode=konstruktikon

• $\mathrm{■}$

  Brazilian Portuguese: FrameNet Brasil (FU Juiz de Fora): https://www2.ufjf.br/framenetbr-en/

• $\mathrm{■}$

  Japanese: Japanese FrameNet (Keio University): https://jfn.st.hc.keio.ac.jp/

• $\mathrm{■}$

  Russian: Russian Constructicon (UiT Arctic University of Norway): https://constructicon.github.io/russian/

• $\mathrm{■}$

  Most of the different constructions were presented at the Constructicon Alignment Workshop (CAW, December 2022), and video recordings are available here: https://www.globalframenet.org/caw2022

Croft’s “Morphosyntax: Constructions of the World’s Languages” [1] contains a glossary of different comparative concepts (meta-constructions, strategies, information packaging, etc.), and this glossary is available online here:

• $\mathrm{■}$

  Interactive interface: https://spraakbanken.github.io/ComparativeConcepts/

• $\mathrm{■}$

  GitHub repo: https://github.com/spraakbanken/ComparativeConcepts

Finally, here is a list of different search engines for corpora, tools and treebanks, that can be used to find constructions:

• $\mathrm{■}$

  Grew-match: https://match.grew.fr/

• $\mathrm{■}$

  SPIKE (query-by-example): https://spike.apps.allenai.org/

• $\mathrm{■}$

  DepEdit: https://gucorpling.org/depedit/

• $\mathrm{■}$

  UDAPI: https://udapi.github.io

• $\mathrm{■}$

  Korap (IDS-Mannheim: https://korap.ids-mannheim.de/ and https://github.com/KorAP/

• $\mathrm{■}$

  Corpus workbench (CWB, useful for larger corpora) – several sites use CWB, such as:

  • –

    Språkbanken Korp (Univ. of Gothenburg): https://spraakbanken.gu.se/korp/

  • –

    CQPWeb (Lancaster Univ.): https://cqpweb.lancs.ac.uk/

  • –

    CWB source code can be found here: https://cwb.sourceforge.io/

We propose a new layer for selectively annotating constructions on top of UD trees. This is intended for constructions (in the sense of Construction Grammar) whose form and meaning/function is not already captured well by the UD tree. Construction instances receive a type name (possibly from a constructicon resource) and may contain relations to construction elements. The elements of the construction are not constrained by the UD tree: e.g., a construction element may cut across multiple UD subtrees. For now, we envision that they would be marked in the MISC column of .conllu files, though in principle they could be moved to a separate extension column.

The annotation layer does not have the goal of directly indicating the elements of form or meaning that are characteristic of or required by the construction, beyond indicating the construction evoker and spans of construction elements. Aspects of the UD analysis (tags, deprels, morphological features) that are characteristic of a construction’s form should be described as such in a type-level constructicon entry. The precise contents of such an entry are not part of this proposal, but constructicons incorporating UD information in some way already exist (e.g., the Russian Constructicon).

Somewhat continuing the discussion started in WG3, Working Group 8 (WG8) To Semantics and Beyond at the Dagstuhl Seminar 23191 on Universals of Linguistic Idiosyncracy in Multilingual Computational Linguistics identified topics related to semantics that have presented to be problematic from various angles, such as semantic representations, inconsistencies or differences in representations/handling of notions across frameworks, cross-lingual handling of constructions or semantic notions, and coverage, e.g., both for arguments and modifiers. Specifically, the following four topics were identified, and we focused on the first two topics during the WG8 discussion sessions:

1. 1.

   Serial verb constructions and light verb constructions from a cross-lingual perspective in terms of representation of their semantics, validity of statements such as “Predicates describe one event”, representations across different frameworks, leveraging other annotations or resources (e.g., the notion of scope tree in MRS), The topic was also discussed in WG3.

2. 2.

   Inventory of semantic relations to describe states, processes and events cross-linguistically and for both arguments and adjuncts (also discussed in WG3 and WG5)

3. 3.

   Inconsistent handling of entities and coreference across resources

4. 4.

   Aspects of lexical semantics such as linking with wordnets, internal semantic structure, relation between internal semantic structure and compositionality, representation in existing resources/corpora

In the next two sections, we summarize the discussions corresponding to the two focused topics.¹⁷¹⁷17The document with more detailed notes for the discussions can be found at: https://docs.google.com/document/d/13-JOkaCKAshDShRFE9NN8lYysc7jmLMmlswmrfneRRo

Verbal constructions can present interesting phenomena involving a wide range of semantic considerations, e.g., in terms of determining whether a single event is involved or multiple events are involved, while at the same time each of these constructions themselves may also present a broad continuum to make it harder to demarcate the boundaries of the construction. This can present issues for describing such constructions or developing proper theoretical accounts of them. We discussed two verbal constructions, the serial verb constructions (SVCs) and light verb constructions (LVCs), that illustrate this issue.

SVCs, e.g., persuade X to take a hike, or try to run, are argued to involve control structures where one of the verbs is considered to be incorporated into the other verb. In such constructions, multiple independent lexical verbs are combined to indicate a single event (or two sub-events of a single event connected temporally indicating either simultaneous or consecutive temporality). LVCs, e.g., take a bath, give a bath, or give a speech, involve a verb which does not indicate the event itself but provides some aspectual or causal information about the event and the semantic core of the event is expressed by the nominal element.¹⁸¹⁸18In some languages, e.g., in Hindi, as the Hindi PARSEME annotated data indicate, the event may also be determined based on a combination of other elements such as an adjective with a verb..

These constructions raise important questions about how one can determine semantic uniqueness of an event, whether the event is expressed by a verb, and also what an event is. We discussed tests for semantic uniqueness of an event and noted that neither of these tests were sufficient by themselves in indicating whether a single event was expressed or multiple events were expressed, but they might indicate tendencies that multiple tests together could help confirm to determine if a single event was involved. These included tests such as modification, argument sharing, and coordination.

In regards to modification, we identified examples in Thai SVCs indicating that negation (as modification) could be used but either in only one place (i.e., with one of the verbs) or it may mean the same irrespective of the position of negation. Cross-lingual examples such as these and the English, I persuaded Francis not to take a hike are interesting. Note that the use of negation in this position could be used to indicate persuading Francis to not take a particular hike or any hike, but in this position, it does not negate the act of persuading Francis to take a particular/any hike. Quantification (as a modification strategy) can also be useful. In terms of semantic representation, “one event” may be represented semantically through, for example, the same position in the (Minimal Recursion Semantics/MRS representation) quantifier scope tree which can also entail a test of semantic uniqueness. However, quantifiers based test also does not always work. It is not always clear how to create the quantifier examples.

The argument sharing test is also not diagnostic. A vast amount of literature involving language specific approaches to SVCs discusses argument sharing as a criterion for a single event. While it may be a necessary criterion for some languages, it is not sufficient in determining semantic uniqueness of event. Also, while English SVC I tried to run involves subject sharing, a Thai construction equivalent to “pound X flat”, does not involve “subject sharing”¹⁹¹⁹19although an argument is shared but such resultative constructions in Thai are also considered to be SVCs. To take another example, the argument sharing test does not help resolve whether the Japanese construction equivalent to “jump-rise the stairs” involves one event with manner information or two events. In such verb-verb constructions, one verb (“jump”) is a hyponym of the other (“rise”).

Coordination can also be considered as a test for semantic uniqueness but is again found to not be a strong diagnostic. For example, in the case of LVCs, since the N+V combination expresses an event, one can expect coordination of nouns in such constructions to not be possible. But PARSEME corpus shows there are quite a few of such cases. For example, it is possible to say He took a bath and a shower. Coordination also seems possible where the nominal part of the LVC is modified and is coordinated with a non-LVC noun, e.g., the bath that I took and the bathtub were wonderful.

The group found such examples with LVCs involving these diagnostics, and their flexibility when testing their boundaries particularly interesting and ended up focusing a large part of the WG8 sessions on discussing LVCs further. We are continuing our discussions involving LVCs, their characteristics and behavior across languages to arrive at a more general and typologically informed semantic representation of these constructions beyond Dagstuhl with many participants from WG8 and other groups (the current participants’ list continuing these discussions includes: Emily M. Bender, Archna Bhatia, Kilian Evang, Dan Flickinger, Jan Hajič, Dag Haug, Laura Kallmeyer, Carlos Ramisch, Nianwen Xue). We plan to continue these discussions to also include other MWEs to develop their semantic representations while taking into account the observed cross-lingual patterns as well as the idiosyncratic behaviors they demonstrate.

There is currently no commonly accepted inventory of semantic roles and relations that can be applied across languages, domains – something like Universal Dependencies, with similarly low barriers to use in applications, but for semantic roles and relations. Kilian Evang started the discussion by formulating some desiderata for such a scheme: it should be usable without reference to a lexicon, it should cover at minimum all dependencies between content words (thus, both arguments and modifiers) with a unified vocabulary, and handle states and events in a unified way. Semantic relations should not be too fine-grained, He illustrated what such a scheme might look like using a possession/control relation, labeled pss. This label could be used directly for modifier relations:

For arguments of predicates denoting states, the same relation is split over two dependency edges, dubbed the “domain” and the “range” of the (possession) relation:

(a) (b)

For arguments of predicates denoting events, there is the additional possibility of having an argument denoting an initial range (ir):

(a) (b)
(c)

Differences to existing schemes were discussed: MRS first assigns semantically unspecified argument labels (ARG0 = event, then ARG1, 2, 3 in order of decreasing obliqueness). Separate, language-specific lexicons à la FrameNet assign specific roles to these argument labels, per predicate (e.g., own, belong may be in the same frame, but for belong ARG2 is the owner, and for own, ARG1). In PropBank, all arguments are verb-specific. SynSemClass is an event type ontology under development that defines coarser, but still hierarchically ordered classes of events (and states), with e.g. the “ownership” class containing verbs like own, possess, hold. “Buying” would be a separate class, e.g., it is not clear that the seller is also the initial owner in all cases. This links to the general question of how much annotators should be allowed to infer that is not strictly entailed by the predicate, e.g. I visited my mother at 4 o’clock yesterday – should the annotator include the information that it was in the afternoon? Verbs in the Prague Dependency Treebank are linked to FrameNet, VerbNet, OntoNotes, PropBank, English WordNet, and SynSemClass. In the TRIPS ontology, there are events of change, events of state, and within those further classes.

Potential issues and next steps with the proposed schemes were discussed. Compared to UD, additional semantic edges need to be annotated compared to UD, e.g., in control constructions. It was therefore suggested that annotation start atop Enhanced UD, even if automatically predicted and imperfect. Similarly, existing PARSEME data could be used to automatically pre-annotate semantic edges in VMWEs, e.g. kick $\to$ bucket with special labels. The next steps should focus on annotating a bunch of data to evaluate the proposal, and also on lexical mapping to other resources like SynSemClass, PropBank, Universal Proposition Banks, VerbAtlas, VerbNet, FrameNet, SNACS, UMR to study synergies and enable comparisons.