Global FrameNet Tuesdays are a continuation of the successful International FrameNet Workshop 2020 (IFNW 2020) that went online due to COVID-19 and reached a wide audience within the worldwide community working on framenets, beyond those listed in GlobalFrameNet. We built on this experience and hosted regular meetings - work-in-progress reports, talks and discussion sessions - on the first Tuesday of a month. We started with a first series in 2021 from April to June and October to December, and will call for contributions from FrameNet projects all over the world for the year to follow. Reports and talks are recorded (without the ensuing plenary discussion) and made available on this website.

Notes

• Contribution submissions (see below for more info) can be sent to czulo (at) uni (minus) leipzig (dot) de

• Session links will only be sent to email addresses registered for the Global FrameNet news mailing list, you can:

Program

December 7th, 4pm UTC

The German FrameNet Constructicon: towards a fully integrated repository of words, constructions, conceptual metaphors and their meanings
(Fabian Barteld, Nina Böbel, Stefan Hartmann, Anastasia Neumann-Schneider, Alexander Willich, Alexander Ziem)

November 9th, 11am (PDT) - 3pm (BRT) - 7pm (CET)

Exploring Semantic Frame Relations across Languages
(Arthur Lorenzi and Collin Baker)

Separate research projects have developed Frame Semantic lexical databases for roughly a dozen languages (including Brazilian Portuguese, Chinese, Hebrew, Japanese, Korean, and Swedish, among others) based in varying degrees on the original ICSI FrameNet structure, methodology, and annotation practices (Ruppenhofer et al. 2016).  We will explore the extent to which semantic frames are similar across languages.  The underlying research questions include:
• To what extent are semantic frames the same cross-culturally and cross-linguistically? 
• Where framings differ, what are the reasons for these differences?
• Should we expect translations to be ``frame preserving''?

The FrameNet team has developed algorithms to find alignments across these FrameNets, using both symbolic and statistical approaches and utilizing the internal structure of the FrameNets, such as the number and type of frame elements and their definitions.  In order to take full advantage of annotation data in different FrameNet databases, BERT models were used to obtain contextual embeddings of LUs. First, using aligned corpora from EUROPARL, multilingual BERT (Devlin et al. 2019) was re-trained (fine-tuned) for each language with the objective of aligning word vectors across languages. This step is necessary because, although this version of BERT is multilingual, it doesn't force words with the same meaning  in different languages to be close in the embedding space. Then each annotated sentence was processed to create a vector for the LU in that specific sentence. The final embedding of each LU is the centroid (average) of all of its vectors in the annotated sentences. One type of alignment score between a pair of frames is obtained by measuring the distance between their LUs in this LU embedding space.

The team has also built an interactive tool called ViToXF for the visualization of alignments of frames (and LUs) across languages, which we will demonstrate during the talk.  The GUI allows the user to choose the frames to be compared, choose one of seven alignment techniques, and set parameters to control the number of alignments to be displayed. We will also present some preliminary results on evaluating the accuracy of the alignments. Both the database of alignment vectors and the visualization tool are available at https://github.com/icsi-berkeley/framenet-multilingual-alignment/releases/tag/1.0.3-2.

June 1st, 5pm - 6pm (UTC)

Breeding Fillmore’s Chickens and Hatching the Eggs: Recombining Frames and Roles in Frame-Semantic Parsing
(Gosse Minnema and Malvina Nissim)

Frame-semantic parsers traditionally predict predicates, frames, and semantic roles in a fixed order. This paper explores the ‘chicken-or-egg’ problem of interdependencies between these components theoretically and practically. We introduce a flexible BERT-based sequence labeling architecture that allows for predicting frames and roles independently from each other or combining them in several ways. Our results show that our setups can approximate more complex traditional models’ performance, while allowing for a clearer view of the interdependencies between the pipeline’s components, and of how frame and role prediction models make different use of BERT’s layers.

April 6th, 5pm - 6pm (UTC)

Lutma: a frame-maker tool
(Torrent, Matos, Gamonal, Belcavello, Viridiano and Almeida)

In this talk, we present Lutma, a tool for collaborative frame-making. We discuss the motivation for developing the tool and its implementation as a semi-constrained step-by-step process. Next, we present the structure envisioned for the database, discussing the importance of its multilingual nature. We then provide a demo of the main functionalities of the tool. Finally we highlight the centrality of the Global FrameNet community in fostering the development of new frames.

Organizing Committee

Collin Baker
FrameNet
International Computer Science Institute, USA

Oliver Czulo
Department of Translation Studies
Universität Leipzig, Germany

Tiago Timponi Torrent
FrameNet Brasil
Federal University of Juiz de Fora, Brazil