-
Notifications
You must be signed in to change notification settings - Fork 10
Home
Welcome to the am-parser wiki!
This documentation is a work in progress. If anything is unclear, please open an issue and we will get back to you as quickly as possible.
A characteristic feature of the AM parser is that it learns to parse sentences into AM dependency trees, which then evaluate to graphs in the AM algebra. We have defined our own file format, AM-CoNLL, to store AM dependency trees; it is a mild extension of the well-known CoNLL format for storing dependency trees.
Internal note: There is an UdS specific documentation on how to set up the environment and where to find the files.
The AM parser consists of four steps:
- Preprocessing and decomposition: convert the graphbank to AM dependency trees
- Training the neural network to predict AM dependency trees
- Use the model to parse and evaluate the test data.
If you want to use the AM parser for a graphbank based on one of the graph formalisms supported here, you can use the above pipeline.
If however you want to use the AM parser on a new graph formalism, we recommend that you use our unsupervised training regiment that learns AM dependency trees from sentence-graph pairs. This makes it easy to use our compositional, interpretable parser on novel graphbanks! You can find a guide here.
You can use our pretrained models to directly parse sentences into graphs (this is also explained in the main README).
Alternatively, you can just let the am-parser compute the scores, and use some other parser to produce trees in the AM-CoNLL format, e.g. the A* parser. In this case, you will need to evaluate the AM-CoNLL files yourself.
- There are multiple options to visualize AM dependency trees (by sentence or for the whole corpus) for further error analysis.
- You can visualize graphbanks, creating a pdf file for every graph in the graph bank.
- UCCA preprocessing
An overview of the third party packages we used can be found here.