BerlinProcessing

Discussion on HPSG Processing Strategies

Moderator: StephanOepen

Scribe: YiZhang

This discussion was initiated by the request of contributing a chapter on HPSG processing to the language and linguistics compass online volume. The general discussion of HPSG processing strategies dates back to a workshop in Berlin in 1998. People from SB, Tokyo, Stanford came together to discuss the strategies of overcoming obstacles in HPSG processing: FS manipulation, efficient unification and copying, etc. Stochastic disambiguation models joined the game around 2000. More recently, the development has been focusing on improving robustness of the HPSG parsing. The essential goal nowadays is to achieve the balance (sweet-point) between coverage accuracy and efficiency. While the on-going effort to achieve better efficiency in Tokyo presents a competitive response to other deep parsing communities (i.e. CCG), there is also a growing interest in exploring cognitive apparatus in HPSG parsing (deterministic parsing, incremental parsing, more general psycholinguistic human sentence processing, etc) from others (StephanOepen).

The general open question for discussion is:

How to achieve the sweet-point of balance between efficiency, accuracy and coverage?

MiyaoYusuke: it should be emphasized that the three aspects of the parsing (coverage/robustness, accuracy, efficiency) must been considered together when attempting to improve the parser. It is also closely related to grammar engineering. And from some aspects, the generation is similar the parsing. A good disambiguation model can help to overcome the problem with an overgenerating grammar.

JohnCarroll: Traditional grammar developers do not rely on statistical disambiguation model. What does grammar developers think of this problem?

DanFlickinger: For instance, when treebanking the grammar outputs, the results are sometimes somehow skewed by the disambiguation model.

AnnCopestake: To a degree, the disambiguation model makes the grammar domain specific. For large scale parsing it skews the search space. We should avoid be less different from treebank-induced grammars. One should not be too much affected by practical issues.

DanFlickinger: What's the workload distribution between parser and grammar? Many parsing edges are created for the need to handle modifier attachment. To which degree the syntactic structure are isomorphic?

It is mentioned in the discussion that there are ongoing work on generation with treebank-induced grammar (i.e. at Tokyo).

StephanOepen: Incremental processing strategies, inter-dependencies between the grammar and processing

BertholdCrysmann: Interlinguality need to be considered in processing, as well. A close cooperation between grammar engineering and processing people is needed.