TALP Research Center

On the one side, we have studied the problem of lexical selection in statistical MT.  For that purpose, we have constructed a Spanish-to-English baseline phrase-based statistical MT system and iterated across its development cycle, analyzing how to ameliorate its performance through the incorporation of linguistic knowledge.  First, we have extended the system by combining shallow-syntactic translation models based on linguistic data views. A significant improvement is reported.  This system is further enhanced using dedicated discriminative phrase translation models. These models allow for a better representation of the translation context in which phrases occur, effectively yielding an improved lexical choice. However, we have found that improvements in lexical selection do not necessarily imply an improved overall syntactic or semantic structure. The incorporation of dedicated predictions into the statistical framework requires, therefore, further study. As a side question, we have studied one of the main criticisms against empirical MT systems, i.e., their strong domain dependence, and how its negative effects may be mitigated by properly combining outer knowledge sources when porting a system into a new domain.  We have successfully ported an English-to-Spanish phrase-based statistical MT system trained on the political domain to the domain of dictionary definitions.

On the other side, we have studied the problem of automatic MT evaluation.  We have analyzed the main deficiencies of current evaluation methods, which arise, in our opinion, from the shallow quality principles upon which they are based. Instead of relying on the lexical dimension alone, we suggest a novel path towards heterogeneous evaluations.  Our approach is based on the design of a rich set of automatic metrics devoted to capture a wide variety of translation quality aspects at different linguistic levels (lexical, syntactic and semantic). These metrics have been evaluated over different scenarios. The most notable finding is that metrics based on deeper linguistic information (syntactic/semantic) are able to produce more reliable system rankings than metrics which limit their scope to the lexical dimension, specially when the systems under evaluation are different in nature. However, at the sentence level, some of these metrics suffer a significant decrease, which is mainly attributable to parsing errors. In order to improve sentence-level evaluation, apart from backing off to lexical similarity in the absence of parsing, we have also studied the possibility of combining the scores conferred by metrics at different linguistic levels into a single measure of quality.  Two valid non-parametric strategies for metric combination have been presented. These offer the important advantage of not having to adjust the relative contribution of each metric to the overall score. In spite of their simplicity, both schemes lead to a significantly improved evaluation quality.  As a complementary issue, we show how to use the heterogeneous set of metrics to obtain automatic and detailed linguistic error analysis reports.

The two parts of this thesis are tightly connected, since the hands-on development of an actual MT system has allowed us to experience in first person the role of the evaluation methodology in the development cycle of MT systems.