The paper I have been asked to critique describes further developments within a well-established theoretical framework, the interlingual approach to Machine Translation (MT) that uses Lexical Conceptual Structures (LCSs) as its interlingua building blocks. I feel obliged to state first that I come from a different background, having mostly worked in the lexicalist framework. My familiarity with the interlingua approach comes more from readings than from working experience. Therefore I feel I lack the competence to attempt an evaluation of how the paper contributes to the further development of the interlingua framework. Rather, I will analyze the paper from two other perspectives, quite far apart from each other: (i) the first perspective is that of an external observer working in a different environment. From that viewpoint I will remark on the architecture being proposed, and discuss some similarities and differences with the framework I am most familiar with; (ii) the second perspective I will try to take is that of the authors themselves. From that perspective I will discuss how the described system accomplishes the goals that the authors set to themselves (or rather my understanding of such goals, from what I could gather from reading the paper).
The LCS interlingua and the lexicalist approach bear several resemblances: they are both lexically-based, in mapping from one language to another they both tend to abstract away from language-specific idiosyncrasies (e.g. word order), and they both regard generation as a problem of linearizing target lexical items according to a target grammar and a set of constraints coming from the source language. One obvious difference is that in the lexicalist approach there is an explicit bilingual lexical mapping, which is absent from the LCS interlingua approach.
In LCS interlingua, LCS for verbs are the backbone of the semantic, language-independent representations of sentences. These are described in terms of two kinds of information: semantic structure (the shape of the graph, plus its structural primitives and fields) and semantic content (the constants which are found at the leaves of an LCS graph). The paper describes a Chinese-English MT system, therefore the LCSs input to generation are LCSs for Chinese verbs. A pair of widely divergent languages like Chinese and English looks like a challenging test-bed for a pre-existing interlingua being newly ported to the language pair at hand. The paper focuses on the generation component, therefore little discussion is provided about the challenges of analyzing Chinese using LCSs. However, the description of the LCSs input to generation gives some hints about what LCSs created from Chinese source sentences look like. The semantic structure of a verb is characterized as "something the verb inherits from its Levin verb class". In other words, the semantic structure of an LCS is a detailed and structured representation of a semantic class as described in Levin's classification. This sounds like good news, as it gives some evidence that Levin classes, although originally devised for English, are sufficiently abstract and numerous to provide a good coverage to other languages as well, Chinese in this case. Also, if Levin classes can be mapped onto corresponding LCSs, then the latter can be automatically acquired form the former.
If LCSs are structured rendition of Levin classes, the drawback of their extra structure seems to be the amount of memory their storage requires, particularly in their long-hand text representation. The authors state at some point that the processing of input files can be problematic for some system components when the size of the files exceeds some given threshold.
However, the correspondence between Levin classes and LCSs exploited in generating the latter, could perhaps lend itself to be exploited more directly to alleviate the problem of LCS complexity. If the most common LCS semantic structures (i.e. LCSs minus constants) are in a one-to-one correspondence with Levin classes, and the latter are cross-linguistically adequate to cover different languages, would it be possible to use Levin classes directly as building blocks of an interlingua, instead of semantic structures? They would have a comparable expressive power and they would make for a leaner interlingua. Such Levin class constants, equipped with indices to express their relationships with other interlingua building blocks, would be part of the lexical representation of verbs. What would be lost in reducing semantic structures to verb class constants would be the relationships among different verb classes. However, these could perhaps be explicitly stipulated, given that the number of Levin classes is not only finite, but also manageable. I would be interested to know whether the authors consider this a viable path, as the resulting picture would bring LCS interlingua and lexicalist MT a little bit closer, and perhaps could be the basis of a fruitful interchange between the two approaches.
My second set of comments is about the goals of the work described in the paper and the way they are accomplished. One of the main goals of the described generation component seems to be the integration of an external generation package, Nitrogen, which performs linearization and statistical extraction of the most likely linear sequences. Nitrogen comes equipped with its own interlingua (AMR), from which it generates word lattices passed on to the statistical extractor. In order to harmonize the two interlinguas (LCR and AMR), an intermediate representation language was built, called LCS-AMR. In an initial stage of the project, LCS-AMR representations were recast as standard AMR representations and fed to Nitrogen. However, the mapping from LCS-AMR to AMR raised serious problems, eventually leading to replacing the Nitrogen grammar with a newly implemented English grammar. The output of such grammar are word lattices, fed into the Nitrogen statistical component. Given this background, one point I would be interested in is the current role of LCS-AMR. Since it seems to have been initially devised as a bridge language between LCS and AMR, what is its current role, once AMR has been removed? Would it be conceivable to have the English grammar work directly with LCS, instead of LCS-AMR?
Copyright 2000 Computing Research Lab.