Slides from Tutorial Now Available
The slides from this tutorial are now linked from the LREC website at https://bit.ly/3AvgQzM
Cross-linguistic Frame Alignment
The goal of this tutorial was to introduce participants to Multilingual FrameNet, algorithms for cross-linguistic frame alignments, and the alignment visualization tool, ViToXF, and demonstrate its features and functionality. We also discussed ways of evaluating the alignments and algorithms combining resource-based and distributional approaches to alignment. This tutorial was intended to be of interest to theoreticians creating curated resources, advocates of vector-based statistical approaches, and developers of practical applications, such as virtual assistants.
Background
The proliferation of ontologies and lexica based on human categories (e.g. Palmer, et al. 2005, Ehrlinger et al. 2016) leaves an abundance of possibilities for researchers and engineers who need these resources. This tutorial addresses the alignment of resources in a theoretically grounded way, confirming semantic identity and similarity, and provides the tools to consider one such set of semantic resources.
The building of the FrameNet lexical resource for English (Ruppenhofer et al. 2016) has encouraged the development of frame semantic resources for more than ten other languages, both Indo-European (e.g. Spanish, German, Portuguese, Swedish) and non-IE (e.g. Japanese, Chinese, Korean, Hebrew). This has led to increased interest in the similarities and differences between semantic frames across languages and cultures.
To study this phenomenon, Baker and Lorenzi (2020) produced a database of vectors that represent alignments between frames in different FrameNet resources (e.g., English-Spanish, English-Japanese, etc.). The alignments were created using 4 resource-based and 7 vector-based alignment methods. The database reveals both consistent differences between certain language pairs and also different results from resource-based and vector-based algorithms. An interactive tool for visualizing all of these alignments, called ViToXF, was produced at the same time, showing comparisons between English and one of 7 other languages: Chinese, French, German, Japanese, Portuguese, Spanish, and Swedish. Interested researchers will find the latest release of the database of vectors at https://github.com/icsi-berkeley/framenet-multilingual-alignment/releases/tag/1.0.3-2, and ViToXF is freely available at https://github.com/icsi-berkeley/framenet-multilingual-alignment.
Take-homes
Participants learned to:
- Understand the challenge of semantic alignment across languages,
- Explore cross-linguistic alignment with ViToXF (symbolic and distributional techniques),
- Identify strengths and weaknesses of each approach,
- Improve the resources based on the discovered weaknesses, and
- Exploit relations within resources to improve mappings between them