UniMorph Seminars: Starting Aug,16th (w/Reut Tsarfaty)
19 views
Skip to first unread message
Ekaterina Vylomova
Aug 15, 2021, 7:19:56 AM8/15/21
to unimorph, Reut Tsarfaty, Ryan Cotterell, David Yarowsky, Mans Hulden, khuyagb...@gmail.com
Dear UniMorph colleagues,
Tomorrow (Aug, 16th) we are starting a series of meetings (most of which will be in September) during which we will be focusing on various aspects of the UniMorph database that can be improved.
Each meeting will have a shared theme and a leading speaker, the expected duration is ~2-3 hours. Everyone who would like to contribute or is interested in a specific theme is welcome to attend. A preliminary list of topics is provided here: https://unimorph.github.io/unimorph2021/
UMSeminar-1: Aug, 16th 1pm–3pm UTC
(9am–11am EDT; 3pm–5pm Zurich; 4pm–6pm Haifa; 9pm–11pm Ulaanbaatar; 11pm–01am, Aug17th Melbourne).
Theme: Word forms, phrases, or clauses: what level of granularity do we need in UniMorph?
Leading Speaker: Reut Tsarfaty
Zoom: https://unimelb.zoom.us/j/81174130031?pwd=ei83SWIxSnRXVjlLSGgxUHhEdTBVdz09 PASS: 728750
Title: More Than Morphs: Getting More Out of UniMorph
Abstract: Morphological processes such as inflection and reinflection are studied and evaluated in NLP nowadays with the help of UniMorph (UM), a large collection of labeled inflection tables of over a hundred typologically different languages. In this talk we look closely at the current version of UniMorph and assess its design and content. Specifically, we ask whether UM is a necessary component of morphological reinflection (or would minimal supervision be enough), whether the current version of UM is sufficient for morphological reinflection (or whether there are some aspects missing), and, importantly, whether the word forms in UM provide the right level of granularity for annotating morphology (as opposed to, for instance, phrase-level or clause-level). We derive answers to these questions from both theoretical arguments and empirical evidence, and conclude with concrete suggestions on steps that may be taken to push UM to the next level of studying computational morphology, in accord with contextualized embeddings and downstream tasks.
Related:
- Annotation of clitics. How should we store the paradigms with clitics (they typically(?) copy those that are without; maybe allow rules to be stored and paradigms for them generated on the fly?)
- Annotation of MWEs that require dependencies. Samples in some language require syntactic information, e.g. “legkaja_ADJ promyshlennost_N” vs. genitive constructions “sindrom_N Aspergera_N”.
Abstract: Morphological processes such as inflection and reinflection are studied and evaluated in NLP nowadays with the help of UniMorph (UM), a large collection of labeled inflection tables of over a hundred typologically different languages. In this talk we look closely at the current version of UniMorph and assess its design and content. Specifically, we ask whether UM is a necessary component of morphological reinflection (or would minimal supervision be enough), whether the current version of UM is sufficient for morphological reinflection (or whether there are some aspects missing), and, importantly, whether the word forms in UM provide the right level of granularity for annotating morphology (as opposed to, for instance, phrase-level or clause-level). We derive answers to these questions from both theoretical arguments and empirical evidence, and conclude with concrete suggestions on steps that may be taken to push UM to the next level of studying computational morphology, in accord with contextualized embeddings and downstream tasks.
Related:
- Annotation of clitics. How should we store the paradigms with clitics (they typically(?) copy those that are without; maybe allow rules to be stored and paradigms for them generated on the fly?)
- Annotation of MWEs that require dependencies. Samples in some language require syntactic information, e.g. “legkaja_ADJ promyshlennost_N” vs. genitive constructions “sindrom_N Aspergera_N”.
Warm regards,
Ekaterina
Reply all
Reply to author
Forward
0 new messages