Martin Kroon & Sjef Barbiers. (2018, January 26). A filter for syntactically comparable parallel sentences [Poster session]. Computational Linguistics in The Netherlands 28 (CLIN28), Radboud University, Nijmegen.
Massive automatic comparison of languages by using parallel corpora will greatly enhance and speed up research in the field of comparative syntax. The goal of this research is to develop a theoretical model of the syntactic properties that all languages have in common and the range and limits of syntactic variation.
It only makes sense to compare parallel translated sentences that are sufficiently similar syntactically. A method and measure is needed to filter out parallel sentences that are syntactically too different. In this paper I compare three possible filters with each other and with a baseline, (relative) Levenshtein (Levenshtein, 1966) distance based on POS-tags.
The first filter uses sentence-length ratio between aligned sentences as a measure for syntactic comparability. Alignments in which one of the two sentences is twice as long as the other are most probably noise, and should not be used for syntactic comparison. Sentence-length ratio as a measure will be computationally cheap, though it might not be fine-grained enough to adequately grasp syntactic differences caused by differences in syntactic hierarchies.
The second filter uses dependency parses. Hierarchical input, such as parses, should give better results than the baseline, which takes a linear string as input. I evaluate the usefulness of Oommen et al.’s (1996) or Zhang and Shasha’s (1989) algorithms to measure distance between trees. A weakness of using dependency parses is that a parser must be available for all languages involved.
The third filter uses sentence vectors. We will see whether Mikolov et al.’s (2013) method to translate between two vector spaces determines whether two aligned sentences are syntactically comparable.
The three filters will be evaluated with a manually labeled set of syntactically comparable parallel sentences in English, Swedish and Dutch.
We will conclude by showing which method or combination of methods provides the best filter.
References
Levenshtein, V. I. (1966). Binary codes capable of correcting deletions, insertions, and reversals. In Soviet physics doklady (Vol. 10, 8, pp. 707–710).
Mikolov, T., Le, Q. V., & Sutskever, I. (2013). Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168.
Oommen, B. J., Zhang, K., & Lee, W. (1996). Numerical similarity and dissimilarity measures between two trees. IEEE Transactions on Computers, 45(12), 1426–1434.
Zhang, K. & Shasha, D. (1989). Simple fast algorithms for the editing distance between trees and related problems. SIAM journal on computing, 18(6), 1245–1262.