📄 Generation of noun-noun compounds in the Spanish-English machine translation system SPANAM®

Summary

In Spanish, the relation between two nouns is usually expressed by the "noun + preposition + noun" (NPN) construction. In English, a noun compound (NN), such as water supply, sounds more natural. If the NPN is preserved literally, the translation comes out awkward. If, however, compounds are formed too eagerly, ambiguity arises that may mislead the post-editor. The author describes how this problem was solved in SPANAM, the machine translation system of the Pan American Health Organization (PAHO). It is a transfer system with an ATN parser. The rules were derived empirically from a corpus containing thousands of sentences with NPN constructions. The strategy has four components:

• dictionary coding of English nouns and of special analysis units (AU) that allow or block compound formation;
• a cautious algorithm that operates only when a full syntactic parse is available and filters out risky cases with a series of syntactic and lexical filters;
• a simplified mode for partial parses, in which a compound is formed only when it is specified in the dictionary;
• macros for post-editing.

Semantic coding was not needed for this. The proportion of compounds formed increased markedly, and most of the new ones made the translation more fluent. The work shows how a rule-based system can become more natural without losing accuracy.

Key points

• SPANAM has been in use at PAHO since 1980. Its dictionaries contain more than 95,000 source entries. In the ARPA tests of 1992–1994 it received the highest scores among 8 systems, but its fluency scores were somewhat lower than its scores on the other criteria.
• The target dictionary has 214 English nouns coded, and 93 % of them block compound formation. There are also about 4300 NPN AUs (76 % require a compound, 24 % forbid one) and 73 PN AUs such as de calidad. About 2000 NPN AUs were obtained by converting old fixed units (SU).
• Among the algorithm's filters are the following: a noun has no more than two target words, a phrase has no more than four nouns and modifiers in total, there are no proper names, and there are no subordinate clauses.
• Over 20,000 words of text, the proportion of NN compounds formed rose from 4.6 % (18 compounds in the 1994 version) to 23.7 % (93 in the 2001 version). Of the new compounds, 73 % improved fluency, 24 % were erroneous, and 3 % were questionable. No desirable compound was lost in the process.
• There are two macros for MS Word: one forms an NN compound from an NPN construction, the other decomposes a compound back. The authors consider adding semantic information for doubtful cases a possible next step.