Abstract
This Chapter examines the extent to which machine translation output is constrained by the structure of source language sentences, i.e., whether MT can enjoy the freedom of the human translator to produce translations that depart in significant ways from the structure of a source sentence. We focus in particular on the issue of clause shifts in the target translation of a source sentence. We describe experiments with Logos Model that both succeeded and failed in this effort to escape from this structure-preserving tendency in MT.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
There is of course the parallel tendency for translators to adhere more closely than necessary to source language constructions, a tendency called “translationese” (Koppel and Odan 2011).
- 2.
- 3.
Gdaniec and Schmid (1995). Schmid subsequently further described the experiment in her doctoral dissertation (Schmid 1996) where she relates Logos Model to Langacker’s Cognitive Grammar (1991). She observes that the two grammars share basic assumptions about the nature of language as “general cognitive activity.” She writes (p. 281) that Langacker sees language as “a psychological and ultimately neurological phenomenon which is acquired through experience.” She notes that “Both theories [Cognitive Grammar and Logos Model] consider neural networks as a valid model of human mental activity….”
- 4.
The E-G Semantic Table contains roughly 15,000 pattern-rules that for the most part serve specific semantic purposes, as the name implies. If Logos Model had been allowed to continue development after 2000, the volume of semantically oriented pattern-rules could by now easily exceed double or triple that number. Even as it now stands, it is not unusual for a certain verb to have forty or more SEMTAB rules covering its multiple meanings and the various generalized SAL contexts that trigger these meanings.
- 5.
Ideally, given time and motivation, the offending pattern-rules presumably could have been constrained sufficiently to avoid the unintended effects.
- 6.
Tests designed to compare systems are of limited value if they do not correlate testing with the kinds of texts used to train each system. There will probably never be such a thing as a perfect, all-purpose system, no more than there are perfect, all-purpose translators, although clearly some (in either case) approach that ideal more so than others.
- 7.
Xing Shi et al. (2016, web): “As the [neural] model first encodes the source sentence into a high-dimensional vector, then decodes into target sentence, it is hard to understand and interpret what is going on inside such a procedure.”
References
Gdaniec C, Schmid P (1995) Constituent shifts in the logos English-German system. In: Proceedings of the sixth international conference on theoretical and methodological issues in machine translation. Centre for Computational Linguistics, Catholieke Universiteit Leuven, Leuven, pp 311–318
Koppel M, Odan N (2011) Translationese and its dialects. In: Proceedings of the 49th annual meeting of the association for computational linguistics. Portland, Oregon, pp 1318–1326
Schmid P (1996) Clausal constituent shifts: a study in cognitive Grammar and machine translation. Ph.D. dissertation. Georgetown University. UMI Dissertation Services
Somers HL (1992/3) Current research in machine translation. Machs Trans 7:231–246
Shi X, Padhi I, Knight K (2016) Does string-based neural MT learn syntax? In: Proceedings of the 2016 conference on empirical methods in natural language processing. Austin, pp 1526–1534. http://xingshi.me/data/pdf/EMNLP2016long.pdf. Accessed 18 Mar 2017
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 2018 Springer International Publishing AG, part of Springer Nature
About this chapter
Cite this chapter
Scott, B. (2018). Some Limits on Translation Quality. In: Translation, Brains and the Computer. Machine Translation: Technologies and Applications, vol 2. Springer, Cham. https://doi.org/10.1007/978-3-319-76629-4_7
Download citation
DOI: https://doi.org/10.1007/978-3-319-76629-4_7
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-76628-7
Online ISBN: 978-3-319-76629-4
eBook Packages: Computer ScienceComputer Science (R0)