First, we take a look on the current issues in Natural Language Generation and then see how these have been taken care of in the paper. Analysis and generation are two sides of a coin in NLP applications such as Machine Translation. A wide range of complex problems which are considered to be specific for generation - content specification, sentence planning and surface realization - can be better handled if due consideration is given to the analysis techniques. The question arises whether it is possible to better chose and integrate analysis techniques which could be efficiently applied at different stages of generation. This issue has not been explicitly discussed with respect to LCS.
The next issue that arises is the reusability and adaptability of analysis techniques and tools for generation. The use of reversible grammar may lead to efficient and flexible natural language parsing and generation system. The questions are whether this reversibility is possible for close language pairs, like Indian languages and how the LCS is equipped to handle this reversibility.
Another important issue is the reusability of analysis knowledge and methodology of its acquisition for generation. It is worth discussing how to establish whether a resource built for analysis can be used for generation and at what price. The paper has not addressed this issue.
At present, there is a strong tendency towards engineering approach for natural language generation systems. Integrating the human factor (i.e., cognitive approach) into the engineering approach would greatly enhance the overall quality of the existing generation systems. It seems that LCS approach is rather cognitive, than engineering. It would be nice if LCS is augmented with an Cognitive Engineering approach.
Now, there are specific questions with respect to what have been described in the paper.
If LCS has been used as the interlingua of several projects such as UNITRAN and MILT, then how the present implementation differs from them?
How LCS captures the syntax of the input sentence? This is not clear from the paper. A generation interlingua must capture the syntax and semantics of the input.
Why the translation from LCS interlingua to LCS-AMR interlingua is necessary? Is it because LCS is not well-suited for generation or because the LCS-AMR can be fed to the Nitrogen Generation System?
What are meant by grammatical features of LCS nodes? The paper mentions the point but does not explain it.
The English sentence corresponding to the CLCS in (5) is not mentioned.
The target language lexicon is stored in a hash-table. Does it store surface forms of words or root words only? What about the source language lexicon? How the bilingual dictionary is accomplished is not clear from the paper.
The realization module is implemented in Nitrogen system. Is it possible to have a realization module based on LCS only?
In page 5, the interface between the Linearization module and the Statistical Extraction module is mentioned but the Statistical Extraction Module is not mentioned in Figure 1 or anywhere in the text. The heading of the subsubsection 3.2.2. should be either changed to 'bigram preferences' or the Figure 1 should include 'statistical preferences'.
Check the last sentence in Para 2 of page 7. There is some typos in 'E.g.,'.
The 'lcs-amr' in 'The lcs-amr creation' in the second para of section 7 should in all caps.
Copyright 2000 Computing Research Lab.