Helsinki-NLP at the EACL workshops: VarDial and SlavNLP

The Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial) celebrated its tenth edition in 2023 with a record number of submissions. Various members of the Helsinki-NLP group have contributed to this success, most notably Jörg Tiedemann as a co-founder of the workshop series, and Yves Scherrer and Tommi Jauhiainen as current main organizers. This year, three research papers were co-authored by researchers from Helsinki (Olli Kuparinen, Aleksandra Miletic, Janine Siewert, Yves Scherrer). Tommi and Yves also contributed to the three shared tasks proposed in the evaluation campaign this year. After two fully online and one hybrid edition, the 2023 edition was characterized by a growing on-site participation, with about two thirds of talks and posters presented in person.

One of the highlights of VarDial was the panel discussion (initially called “round table”, but we were assigned a fairly small room without tables, let alone round ones) that reflected on the past of VarDial and its future in the era of large language models. Our experts (Antonis Anastasopoulos, Gabriel Bernier-Colborne, Preslav Nakov, Tanja Samardzic, Ivan Vulic) agreed that even though the methods changed drastically over the last ten years, the VarDial themes were more relevant than ever. VarDial has also been known for its evaluation campaign primarily focusing on “hard” language identification problems such as those between closely related language varieties. The panelists were happy to see the continued interest in this campaign, but also wished for more varied downstream tasks to be included in the evaluation campaign. As organizers, we have always been strived to propose a wide variety of tasks, but struggled to attract sufficient numbers of participants. Furthermore, it was highlighted that dialects are first and foremost spoken varieties and that we should therefore focus more on spoken data in the future. The panelists viewed the ongoing consolidation of methods and the dominance of Transformer-based paradigms as a potential silver bullet that would hopefully make it easier for everybody to participate in future shared tasks.

This year Slavic NLP (formerly Balto-Slavic NLP) included two papers from the members of our unit (Roman Yangarber and Anna Dmitrieva) who also contributed to the organization of the workshop and the shared task. The Slav-NER shared task is a multilingual named entity recognition challenge for Czech, Polish, and Russian, which also includes name normalization and entity linking. The top system’s performance on NER and normalization this year reached an F1 score of 90. Entity linking, a more challenging task, had an F1 score of 72-80, while cross-lingual entity linking had an F1 score of ~67, which is a great improvement compared to the previous challenge. The workshop’s best paper, “Resources and Few-shot Learners for In-context Learning in Slavic Languages” (Štefánik et al.), also touched on Polish, Czech, and Russian NER. The authors created an evaluation benchmark for in-context learning for these languages, supported tasks being NER, Classification, QA, and NLI.

Notes from EACL 2023

Hello! I’m Timothee Mickus, one of the postdocs in the FoTran project. I was at this year’s EACL to present a paper we submitted a year ago to TACL. The piece was about throwing linear algebra at Transformers and seeing what comes out. It was my first live conference since the pandemic, so that was a nice change of pace from Zoom-based and Underline-based conferences. It’s always easier to get feedback with live audiences.

I left the conference with a few tentative opportunities to collaborate with people in other labs, some new ideas for experiments, and, of course, numerous papers to read, with topics ranging from biases in annotation practices to the interpretability of contextualization in Transformer embeddings and from unsupervised machine translation evaluation to knowing when to expect failure and success in multitask scenarios.

The keynotes ranged through a wide variety of topics, and special attention had been paid so that they would also come from different backgrounds. Ed Grefenstette talked about large-language models and why instruction-based LMs seemed especially promising from a very NLP-centric standpoint. Kevin Munger provided a more media-studies oriented take on the same topic, and questioned how we should change our habits (from how we teach to how we write about AI to what new social policies we need) given the rise of the chatbots in today and tomorrow’s information landscape. Joyce Chai brought an overview of how advancements in NLP impact other AI research fields, and in particular embodied AI.

How do we keep ACL events affordable?

Coming into EACL, I expected that most of this blog post would be about the science. It turned out that EACL 2023 was also out of the ordinary in terms of organization.

One contentious point that was thoroughly discussed during the conference breaks was the price. It turned out to be one of the major points addressed during the business meeting (which was the only session where questions could be directed to the conference organizers). I tend to agree that $800 for a five day event (without counting the ACL membership dues) is a hefty sum; it contributed to the decisions of some of my colleagues and co-authors to not attend the conference. Next year’s venue has yet to be decided, since next year’s organizers want to explore more affordable options than what they initially had in mind. I would personally appreciate some transparency as to how the conference is funded (how much comes from sponsors? from the ACL dues? from the conference attendees?) and where that funding goes (how much are we paying for the conference venue and coffee break caterings? for the conference handbooks? for our EACL-branded tote-bags?).

The challenge of organizing EACL 2023

But maybe the most important thing to mention is the numerous rescheduling. The conference was initially supposed to happen in Kiev, and did not happen in Kiev for reasons. The backup venue in Dubrovnik was something of a last minute change of plans: the original plan once Kiev was ruled out was that the conference would just go online. I suppose this rushed relocation also explains some of the mishaps and communications issues during the final weeks before the conference. To take some concrete examples: the handbook we were provided contained duplicate entries for some of the papers presented at the conference; conversely, some poster sessions were not included. On a more personal note, I had no clear indication as to whether my talk would be presented in an oral or poster session: instead, I was sent a link to the underline page while it was still under construction, and directed to search for my name in this semi-functional website.

Presentations, posters, findings and all the confusion

This leads me to another unconventional organizational decision: every paper that was presented orally also had a poster presentation slot. This happened to be my case as well. I’m still not fully certain what I think of it: I’m not complaining about the extra opportunity to advertise my work, but it is extra work (both ahead of and during the conference). It’s also somewhat unfair that some of the presenters didn’t get the opportunity–and all the more jarring when it comes to Findings papers: Officially, Findings are works that are good enough but for which there’s not extra space in the conference itself, and yet organizers did manage to double-book quite a number of papers in the main conference. This means we have a tiered acceptance system: some of us got two presentation slots, some of us only got a poster, and some of us (Findings authors) only got virtual presentations. In and of itself, that’s not necessarily a bad thing, but I’m uncomfortable with this tiered system being left as implicit. Let’s hope that future editions of EACL and *ACL conferences will be more transparent on that front.

Where is NLP as a field and EACL as a conference?

The last point I’d like to mention comes from the opening session. As it turns out, most submissions to the European chapter of the association for computational linguistics did not come from Europe (with a good third of submissions stemming from the USA), and very few actually focused on linguistics. Most works submitted focused more on engineering, processing and dataset description rather than syntax, semantics or phonology modeling. While this state of affairs begs the question of what is EACL precisely, it also highlights how things have changed across the last few years: NLP has become a more international and engineering focused field of research. Despite these changes, it’s good to see that the community remains very directly involved in deciding where it is going–be it local volunteers in Dubrovnik picking up the ball at the last minute to ensure that EACL would not be an online-only event, or attendees openly discussing whether the current conference practices we have are a good fit for the community.