Parser combinators for Tigrinya and Oromo morphology

Parser combinators for Tigrinya and Oromo morphology Patrick Littell1, Tom McCoy2, Na-Rae Han3, Shruti Rijhwani4 Zaid Sheikh4, David Mortensen4, Teruko Mitamura4, Lori Levin4 1National Research Council of Canada, 2Johns Hopkins University, 3University of Pittsburgh, 4Carnegie Mellon University Digital Technologies, Dept. of Cognitive Science, Dept. of Linguistics, Language Technologies Institute 1200 Montreal Road, 3400 N. Charles Street, 2816 Cathedral of Learning, 5000 Forbes Avenue, Ottawa, ON K1A 0R6, Baltimore, MD 21218, Pittsburgh, PA 15260, Pittsburgh, PA 15213 patrick.littell@nrc.gc.ca, tom.mccoy@jhu.edu, naraehan@pitt.edu, {srijhwan, zsheikh, dmortens, teruko, lsl}@cs.cmu.edu Abstract We present rule-based morphological parsers in the Tigrinya and Oromo languages, based on a parser-combinator rather than finite-state paradigm. This paradigm allows rapid development and ease of integration with other systems, although at the cost of non-optimal theoretical efficiency. These parsers produce multiple output representations simultaneously, including lemmatization, morphological segmentation, and an English word-for-word gloss, and we evaluate these representations as input for entity detection and linking and humanitarian need detection. Keywords: morphology, parsing, Tigrinya, Oromo 1. Introduction In this paper, we experiment with using parser combi- nators (Hutton and Meijer, 1988; Frost and Launchbury, 1989) for the rapid development of practical morphological parsers, as an alternative or supplement to the typical finite- state transducers (Karttunen and Beesley, 1992; Karttunen, 1993). This paradigm offered some practical advantages over a finite-state system, allowing parsers to be written very rapidly in a familiar programming language, although at a cost of runtime efficiency. We present morphological parsers for two Afroasiatic lan- guages, the Tigrinya language of Eritrea and Ethiopia (§4.1.), and the Oromo language of Ethiopia and Kenya (§4.2.).1 These parsers were designed during the LoReHLT17 “surprise-language” evaluation (Strassel and Tracey, 2016) (§3.) to support machine translation, en- tity detection and linking, and humanitarian need detection (Strassel et al., 2017). These parsers were operable within about 36 hours of learning the identity of the languages, al- though they underwent further development during the next two weeks of evaluation. 2. Parser combinators 2.1. Introduction The “parser combinator” paradigm (Burge, 1975; Wadler, 1985; Hutton and Meijer, 1988; Frost and Launchbury, 1989) is a kind of declarative programming that simultane- ously defines the grammar being parsed and the executable code that parses it. This paradigm involves defining parser functions with a particular semantics, using a general- purpose programming language (in our case, Python), as well as defining higher-order “combinators” that take one or more component parsing functions as arguments and return a complex parsing function (for example, the composition of the two functions).2 The resulting grammar is itself an 1github.com/littell/ethi_morph 2The usage of “combinator” here comes not from combina- toric logic but from the “combinator pattern” (wiki.haskell. org/Combinator_pattern) of software design. executable function, that parses text as a recursive-descent parser, but one in which every component has access to all the capabilities and libraries of the general-purpose pro- gramming language (e.g. regular expression libraries, file I/O, etc.). The basic building block of a parser-combinator grammar is the atomic parser. An atomic parser can do something as simple as recognize a single …

Content truncated to 500 words