Syntactic Quality Measurement in Machine Translation with Interlinguas

The measurement of machine translation (MT) performances is an unsolved issue in NLP. This task can be done by a human but the time cost and the need for skilled workers to do it rise the necessity for automatic ways to measure the quality of a translation. In this work, we aim to develop a new methodology for measuring the quality of MT results from a syntactic point of view. The idea takes as a theoretical framework the work of Harris about the decomposition of sentences in elementary units called kernels. Our model parses Spanish sentences and UNL (Universal Networking Language) Graphs with a rule-based methodology and divides them into units of information. Comparing those units the model measures the quality of the translation. Our results show that decomposing the sentences in minimal syntactic units could mprove the evaluation performances also without a lexical/semantic analysis.