Frugal Paradigm Completion

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8248–8273 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8248 Frugal Paradigm Completion Alexander Erdmann,† Tom Kenter, Markus Becker and Christian Schallhart ∗ Google, UK †Ohio State University, USA erdmann.6@osu.edu, {tomkenter,mabecker,schallhart}@google.com Abstract Lexica distinguishing all morphologically re- lated forms of each lexeme are crucial to many language technologies, yet building them is expensive. We propose Frugal Paradigm Completion, an approach that predicts all re- lated forms in a morphological paradigm from as few manually provided forms as possi- ble. It induces typological information dur- ing training which it uses to determine the best sources at test time. We evaluate our language-agnostic approach on 7 diverse lan- guages. Compared to popular alternative ap- proaches, our Frugal Paradigm Completion approach reduces manual labor by 16-63% and is the most robust to typological variation. 1 Introduction From syntactic parsing (Seeker and Kuhn, 2013) to text-to-speech (Zen et al., 2016; Wan et al., 2019), many linguistic technologies rely on accu- rate lexica decorated with morphological informa- tion. Yet, building such lexica requires much hu- man effort (Buckwalter, 2002; Tadi´c and Fulgosi, 2003; Forsberg et al., 2006; Sagot, 2010; Eskan- der et al., 2013). We present a language-agnostic method for minimizing the manual labor required to add new paradigms to an existing lexicon. Formally, let each lexicon entry, or realization, be a triple (P, C, f). P marks membership in some paradigm P of morphologically related words, C defines a cell in P as a bundle of morphosyn- tactic features, and f is the form realizing C in P. Hence, paradigm SING can be expressed (in the UniMorph schema (Kirov et al., 2018)) as a set of realizations: {(SING, NFIN, sing), (SING, 3.SG.PRES, sings), . . . }. For each paradigm to be added to the lexicon, e.g., FLY, we aim to select as few sources as pos- ∗This work was carried out during the first author’s intern- ship at Google UK in 2019. sible to be manually realized, e.g., {(FLY, NFIN, fly), (FLY, PST, flew)} such that the forms realiz- ing the remaining cells can be predicted, i.e., flies, flying, flown. Here, sources are manually provided realizations. Targets are realizations whose forms must be predicted from sources. Our work dif- fers from traditional paradigm completion (Durrett and DeNero, 2013) in that sources are not given blindly, but the system must strategically select which sources it wants to be given at test time. Paradigm completion from one source is typ- ically non-deterministic due to multiple inflec- tion classes realizing different exponents in some cells, e.g., suffixing +ed generates the past tense for WALK, but not for SING or FLY which are members of different classes. Hence, many works discuss paradigm completion in the con- text of (implicit) inflection class disambiguation (Ackerman et al., 2009; Montermini and Bonami, 2013; Beniamine et al., 2018). Finkel and Stump (2007) propose three approaches to select the fewest sources required to deterministically iden- tify class. …

Content truncated to 500 words