“…For our purposes the existing approaches can be characterised by their application scenarios and assumptions about available datasets. Interesting work has been done within the neural, supervised and semi-supervised frameworks, e.g., (Ahlberg et al, 2015), (Ahlberg et al, 2014), (Koskenniemi et al, 2018), (Silfverberg et al, 2018), (Wolf-Sonkin et al, 2018), (Kirov and Cotterell, 2018), (Faruqui et al, 2016), (Faruqui et al, 2015), (Aharoni and Goldberg, 2016), (Cotterell et al, 2017). Much of this work assumes availability of partially labelled data, such as word paradigms and/or clean datasets, such as lists of 'headwords' (lemmas) from which paradigms are generated.…”