https://www.dagstuhl.de/21351

August 30 – 31 , 2021, Dagstuhl Seminar 21351

Universals of Linguistic Idiosyncrasy in Multilingual Computational Linguistics

Organizers

Timothy Baldwin (The University of Melbourne, AU)
William Croft (University of New Mexico – Alburquerque, US)
Joakim Nivre (Uppsala University, SE)
Agata Savary (Université de Tours – Blois, FR)

For support, please contact

Dagstuhl Service Team

Documents

Dagstuhl Report, Volume 11, Issue 7 Dagstuhl Report
Aims & Scope
List of Participants
Shared Documents
Dagstuhl Seminar Schedule [pdf]

Summary

This Dagstuhl Seminar was initially planned as a 1-week event in June 2020 (with number 20261) with the following objectives:

  • Theoretical: To deepen the understanding of language universals, and of how they apply to linguistic idiosyncrasy, so as to further promote unified modelling while preserving diversity.
  • Practical: To improve the treatment of idiosyncrasy in treebanking frameworks, in computationally tractable ways and, thus, to foster high quality NLP tools for more languages with greater typological diversity.
  • Networking: To promote a higher degree of convergence across typology-driven initiatives, while focusing on three main aspects of language modelling: morphology, syntax, and semantics.

Due to the COVID-19 pandemic, the event was first rescheduled and finally reduced to a 2-day online event on 30-31 August 2021, with two 3-hour sessions, repeated for better inclusiveness of various time zones (which corresponds to about 20% of the initially planned duration).

Prior to the event, participants submitted discussion issues, based on which working groups and the program were formed, as described in our Wiki space.

More precisely, the program of the event followed the Dagstuhl model:

  • A list of recommended readings was published prior to the event
  • Introductory talks, given by the 4 organizers, ensured common understanding of the scope and challenges to address.
  • Personal introductions of all participants helped achieve a community building effect, despite the online setting.
  • Working groups (WGs) were built on the basis of the discussion issues submitted by the participants. Each WG had 4 co-leaders, at least one of which could attend repeated sessions, so as to ensure consistency between the 2 time-zone sub-groups. The following WGs were created:
    • WG1: What counts as a word?
    • WG2: What counts as a MWE and as a construction?
    • WG3: Syntax vs. semantics
  • Discussion issues were addressed in WGs by the proposers' short introductions followed by brainstorming.
  • Plenary reporting sessions from WGs took place twice for every time zone.

The event attracted 51 participants, who judged it successful and expressed the need for a full-size onsite follow-up event. All the organizational details and outcomes of the seminar are gathered in our Wiki space.

Despite its very reduced and fully online format, the seminar achieved part of its objectives, stressed the importance of some initially-defined research questions, gave rise to new questions, and showed the efficiency of some instruments.

  • On the networking side, the intended convergence effect was clearly apparent. While the initial proposal and invitee list was dominated by NLP-oriented members of the UD and PARSEME communities, strong contributions came notably from the less numerous typology and UniMorph experts. The four communities interacted actively, and reinforcing these interactions is intended for the near future. Notably, steps were taken towards:

    • integrating typology experts in the PARSEME core group
    • accompanying a seminal work in typology (Croft, to appear) with a "companion volume" about practical implementation of morphosyntactic concepts in UD.
  • On the theoretical side, the event showed:
    • The importance of the research question How to identify words across languages? (item I.A in the seminar proposal), to which the whole of Working Group 1 was dedicated. In particular, new insights from lesser-studied languages, brought by typology experts, allowed us to broaden the perspective on this issue.
    • The need for capturing the relationship between the two fundamental notions in this proposal: a multiword expression and a construction, studied by Working Group 2. From the linguistic and typology perspective, a MWE is a special case of a construction, which is rarely made explicit in current NLP models. But the notion of a construction needs a more formal definition to be implementable in NLP, notably as far as the type-token opposition is concerned (question II.B in the seminar proposal). Thus, the typology-NLP interactions are essential in the quest for an optimal model.
    • The scope of the syntax-semantics interface issues (question II in the proposal) addressed by Working Group 3. On the one hand, the interests of the community in this respect exceeded the scope intended by the event organizers. Namely corpus-lexicon interlinking for all language units, not only for MWEs, was targeted. On the other hand, MWEs are exemplars of condensed syntax-semantic interface issues, and as such provide good case studies in this domain.
  • On the practical side, some initial proposals emerged as to harmonizing UD treebank annotation guidelines with: (i) modelling morphological properties at the subword level (heavily studied by UniMorph), (ii) labelling MWEs (core activity of PARSEME).

Each multidisciplinary approach like ours bears heavy risks of intractability. This is because different communities often have different objectives and points of view on the same phenomena, and they may fail to agree on a unified approach, or even on the usefulness of working towards such a unification. In our case, there is a tension between:

  • diversity and descriptive detail required in linguistics,
  • necessary simplifications for the sake of robustness in NLP.

In other words, it is legitimate to question the usefulness of universality-driven initiatives (in NLP) if idiosyncrasy and diversity are basic properties of language data. Yet even typologists seek language universals which abstract away from the idiosyncrasy.

We feel that the event allowed us to mitigate this tension. Namely, even if a universality-based treebank fails to render the diversity of possible analyses of a language phenomenon, it is still useful not only for NLP applications but also for linguistic and typological analyses. This is because relevant examples are easy to extract (and to further re-interpret), as long as the annotation is consistent and well-documented.

Another barrier-lifting effect of the event concerned the relation between UD and PARSEME. It seems that the MWE categories defined by UD and PARSEME are less incompatible than initially expected, simply because the definition of an MWE in itself is different in UD and PARSEME. This could have been a source of major incompatibility but since a MWE does not really have a status in the UD annotation process, the discrepancies could (at least in some cases) be overcome relatively easily.

In conclusion, the event provided, in our opinion, a proof of concept for the framing objectives set up in the original Dagstuhl seminar proposal. However, since the effective framework and duration was severely reduced as compared to the initially intended setting, only part of these objectives could be achieved. Thus, we are currently putting efforts to ensure follow-up events. In particular, a new Dagstuhl seminar with roughly the same objectives has been submitted.

Summary text license
  Creative Commons BY 4.0
  Timothy Baldwin, William Croft, Joakim Nivre, and Agata Savary

Classification

  • Artificial Intelligence / Robotics

Keywords

  • Computational linguistics
  • Morphosyntax
  • Multiword expressions
  • Language universals
  • Idiosyncrasy

Documentation

In the series Dagstuhl Reports each Dagstuhl Seminar and Dagstuhl Perspectives Workshop is documented. The seminar organizers, in cooperation with the collector, prepare a report that includes contributions from the participants' talks together with a summary of the seminar.

 

Download overview leaflet (PDF).

Publications

Furthermore, a comprehensive peer-reviewed collection of research papers can be published in the series Dagstuhl Follow-Ups.

Dagstuhl's Impact

Please inform us when a publication was published as a result from your seminar. These publications are listed in the category Dagstuhl's Impact and are presented on a special shelf on the ground floor of the library.