SaarlandPcfgDependency

PCFG and dependency conversion+parsing with DeepBank(s) SIG

Notes provided by AngelinaIvanova

---

• 1) Head-to-head comparison of Stanford Basic, CoNLL and DELPH-IN Syntactic Derivation Tree-derived formats

As a follow-up on the talk about parsing Stanford Basic, CoNLL and DELPH-IN Syntactic Derivation Tree-derived formats: it could be interesting to look closely at the structural differences and test if combination of different annotations could lead to improvement. There is an ongoing work in Oslo about mapping DELPH-IN Syntactic Derivation Tree-derived format to Stanford Basic and CoNLL. First, the formats are aligned and then it is analyzed how parallel the formats are . Finally, the decision about the approach to mapping is going to be taken: rule-based or learning-based.

• 2) Parse-ranking results

In order to conduct fair comparison of parsers on out-of-domain data it is necessary to freeze the data so that the knowledge about the new domains is not present in the grammar.

• 3) Parsing ERG derivation trees with Berkeley

Hypothesis in Berkeley parser: human cannot annotate a big corpus manually with fine-grained annotations. Therefore the main problem for this parser is the fine-grained set of supertags. Berkeley has an option for separate learning of syntactic categories and lexical categories. Other parsers: Charniak and Johnson, Stanford

• 4) DELPH-IN MRS-derived format

Possible future developments:

• participation in data-driven dependency parsing competition
• using DELPH-IN MRS-derived format in graph-structure prediction task
• linking DM to semantic roles
• using DM in social media analysis

Motivation to use prepositions in the names of dependency labels instead of including them as nodes in the dependency tree is that prepositions cannot be joined and modified. Prepositions are not predicates by themselves.

What are non-lexical elements of MRS good for? The main motivation to exclude them during the conversion to DELPH-IN Derivation Tree- and MRS-derived formats was to apply off the shelf dependency parsers on the resulting dependencies. Perhaps we loose something we treasure when we get rid of non-lexical elements of MRS.

There is no way to go back from bilexical dependencies to MRS-like structures.