Special Features

One unique feature in our approach is the manner in which the MT system handles Restrictive Relative clauses (such as #1 below).

Our Deep Parser has been specifically designed to identify Gaps and their Fillers. Using this information, complex sentences such as #1 can be split into a corresponding set of simpler sentences (see #1a, 1b below).

Now, each of these simpler sentences can be translated reasonably well. The syntactic complexity of the resultant set of sentences in Hindi is significantly lower, making it more 'accessible'. It must also be kept in mind that long Restrictive Relative clauses are even more complex in Hindi, due in part to its Subject-Object-Verb (SOV) word order.

Another feature is the high-quality information that the Deep Parser provides to the MT system to enable it correctly handle the complexities of Case, Number, Gender, and Aspect in Hindi.

As mentioned previously, if the syntactic parsing of the input sentence (English) is wrong, the resulting translation will usually be unacceptable. However, the highly scalable Statistical Deep Parsers available in the public domain today are designed to accept conversational fragments (i.e. 'ungrammatical' sentences); they do not check for conformance with the rules of grammar. This approach has a significant impact on a number of downstream applications (such as Machine Translation and Text-To-Speech) that are crucially dependent on the parse-tree generated by the Deep Parser. Our approach is different; we follow Chomsky's Government & Binding Theory[1][2][3][4][5][6][7][8] to identify analyses that may be flawed. We do not attempt to translate sentences whose parse-trees fail this test, thereby avoiding egregious errors in translation. As a result, several sentences will not be translated (including sentences that are 'ungrammatical', as well as highly complex sentences that could not be parsed), but that is in line with our design philosophy that values correctness far higher than speed or scalability. However, note that even this test is not sufficient for ambiguous structures. For example, in sentence #1, it is not immediately clear whether 'the lawyer advised the man in the park' or 'the dog bit the man in the park'. Hence, we could have two well-formed parse-trees that correspond to the above two analyses. The Deep Parser needs more information to figure out whether to attach the Preposition Phrase 'in the park' to the higher clause or the lower clause. This information must be provided by a 'reasoning-about-the-world' module i.e. by reasoning that lawyers do not usually advise clients in parks, while dogs occasionally bite people in parks. Our Deep Parser does not include such a 'reasoning-about-the-world' module, and therefore cannot always handle such ambiguous structures correctly.

References