Final Result

The final result of applying the steps shown previously is the following translation for sentence #1:

It is important to note that, if sentence #1 is not split into two sentences as above, then its Hindi translation will be very complex (even for someone with a High Working Memory span). This is because, in an SOV (Subject-Object-Verb) language such as Hindi, the entire Relative clause that restricts the scope of the Object Noun Phrase (i.e. 'the man bitten by the dog') precedes the Verb of the Main Clause (i.e. 'advised'), making it very difficult to recollect the Subject of the Main Clause (i.e. 'the lawyer'). Keep in mind that even the Infinitive clause (i.e. 'not to sue the city') precedes the Verb of the Main Clause. It would require a radical restructuring of the sentence to reduce the complexity of the Hindi translation; this is beyond the capabilities of most MT systems.

The above approach (of splitting sentences that contain Restrictive Relative clauses into multiple sentences) is one of the unique features of our MT system. However, this approach suffers from the limitation that, if Gaps and their Fillers are not correctly identified by the Deep Parser, the resulting translation will be wrong. As mentioned earlier, even the best Deep Parsers cannot always do this reliably because of the incredible complexity of natural languages. This takes us back to the design question of the extent to which Deep Parsers can detect errors in the parse-trees that they produce. A statistical parser has no way to determine whether a previously unseen sentence has been parsed correctly. A hybrid parser, on the other hand, has a deep understanding of the rules of grammar.