As you may know, there are several open-source AMR parsers and our aligner improves these parsers. I would like to share some experiences with how to plugin in our alignments into the existing AMR parsers, although running most of these parser requires a certain amount of hacking.
"A Discriminative Graph-Based Parser for the Abstract Meaning Representation", Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, Chris Dyer, and Noah A. Smith.
The JAMR experiments are carried out with a pipeline of shell scripts. This made plugining our alignments very easy and saved a lot of my life. The hook for replacing the alignment is in the preprocessing script:
jamr/scripts/preprocessing/cmd.aligned
It takes an input AMR file with # ::tok header for each graph and adds
an additional # ::alignments header to each graph.
To replace the alignment, you can use the replace_comments.py scripts.
| JAMR parser | Smatch |
|---|---|
| +JAMR alignment | 65.9 |
| +Our alignment | 67.6 |
- JAMR uses the
cdectokenizer and our released alignments include the one preprocessed withcdec.
"A Transition-based Algorithm for AMR Parsing", Chuan Wang, Nianwen Xue, and Sameer Pradhan
The CAMR uses a single program entry amr_parsing.py in their project.
You can replace the JAMR aligner generated training file with ours,
using the same replace_comments.py script.
| CAMR parser | Smatch |
|---|---|
| +JAMR alignment | 64.6 |
| +Our alignment | 65.1 |
- CAMR uses StanfordCoreNLP as tokenizer. In our release,
we includes the alignment results using this tokenization (noted as
sd).
"Broad-coverage CCG Semantic Parsing with AMR", Yoav Artzi, Kenton Lee, and Luke Zettlemoyer.
[TBD]
[TBD]
[TBD]