AMR World 🌎: Public AMR tools and resources 🎊

An overview of available tools and resources for AMR graphs. Requirements to be included are:

• Publicly available (main requirement)
• If it's a tool: easy to use (ideally)

Please file an issue or make a pull request if you want to make a suggestion for inclusion. There should be a link to a github-repository and (if available) a link to a related publication (📜). Note that work/papers on single parsers or generators should be listed separately, except if they are easy to use off-the-shelf.

Table of Contents

• AMR Data
  • Standard English Sembanks
  • Other Languages
  • Pair-wise annotations
  • Miscelanea
• AMR Tools
  • Parsing and Generation
  • Annotation and Visualization
  • Graph reading, writing, standardization
  • Metrics and Evaluation
  • Graph-to-text alignment

Elements in each categorty are sorted alphabetically.

AMR Data

Standard English Sembanks

Please find information about the "traditional/long-standing" English public manual AMR annotations in this sub-directory.

Other Languages

Non-English Sembanks contain Non-English texts and their AMR graphs.

• Brazilian AMR, 📜: 1561 Brazilian AMRs for Little Prince. Manually created.
• Chinese AMR 📜: 1562 Chinese AMRs for Little Prince. Manually crafted. Double annotated.
• German AMR, 📜: 400 German AMRs. Manually created.
• Spanish AMR, 📜: 486 AMRs (source sentences available through Linguistic Data Consortium). Manually created.
• Korean AMR, 📜: 1,253 sentences. Manually created.
• Turkish AMR, 📜: 700 gold sentences.

English Sembanks that relate to other languages.

• MASSIVE AMR, 📜: A dataset with more than 84,000 text-to-graph Abstract Meaning Representation (AMR) annotations for 1,685 information-seeking utterances mapped to 50+ typologically diverse languages. Manually created.

Pair-wise annotations

Pairs of AMRs with annotation, e.g., entailment, or similarity.

• BAMBOO benchmark, 📜: several thousand silver AMRs for testing and evaluating AMR metrics.
• Chinese AMR Double Annotation 📜: 1562 Chinese AMRs for Little Prince. Manually crafted. Double annotated.
• NLI AMR, 📜: >1 mio Silver AMR pairs of five NLI data sets.
• Spanish-English, 📜: 50 pairs of parallel AMRs with annotated divergence type.
• ParsEval, 📜: 800 Parsed AMRs with human quality annoations (domain: little Prince, AMR3).
• Textual Similarity, 📜: Silver AMRs of text similarity benchmarks (e.g., STS) that are annoated with human textual similarity ratings.

Miscelanea

• AMR Bibliography, 📜: Up-to-date and very large collection of publications on AMR.
• AMR dictionaries and lexica
• AMR Guidelines, 📜: The backbone of everything. AMR examples, and how to create AMRs.
• GrAPES, 📜: AMR parser challenge.
• UMR: Uniform Meaning Representation is a successor to AMR that aims to be more flexible across languages and to cover more semantic phenomena.

AMR Tools

If not mentioned otherwise, tools are all in Python.

Parsing and Generation

These are tools for generating AMRs from text (parsing), or generating text from AMR (generation).

• amrlib: Many pre-trained (English) models for parsing and generation.

Annotation and Visualization

(Usually) GUI-based tools for creating, saving, and showing AMRs

• metamorphosed, 📜: graphical editor to edit Abstract Meaning Representations graphs

Graph reading, writing, standardization

These are tools for reading, writing and handling AMR graphs.

• amr-utils: read, write, visualize graphs
• penman, 📜: read, write, modify graphs
• Smatch++, 📜: read, write, modify graphs, standardize graphs

Metrics and Evaluation

Structure matching

The following are graph metrics that compare graphs structurally (i.e., they don't use embeddings etc.).

• Ancast, 📜: Approximates Smatch, more efficient than Smatch.
• Sema, 📜: triple match heuristic, fast.
• SemBleu, 📜: path extraction heuristic, fast.
• Smatch, 📜: triple match, heuristic solver.
• Smatch++, 📜: triple match, optimal solver, standardized scoring. Fine-grained scoring.

Structure and content matching

These are "advanced" metrics that use embeddings or other features, for finer graph meaning similarity (e.g., structures like kitten vs. cat vs. possibly cat :mod young, etc.).

• AMRSim, 📜: Self supervised learning of graph matching witn neural network
• CALAMR, 📜: Flow-based alignment of graph components
• Rematch, 📜: Efficient graph feature matching with node label generalization
• S2match, 📜: Smatch with word embeddings
• SXmatch, 📜: Smatch with cross-lingual embeddings
• WWLK, 📜: Contextualized matching with Wasserstein distance Weisfeiler Leman Algorithm

Graph-to-text alignment

These are methods for finding out which parts of two AMRs relate to which parts of a sentence.

• GPLA-aligner, 📜, alignment as parsing step
• JAMR-aligner, 📜, alignment as parsing step
• Leamr, 📜: Learned statistical amr-text aligmemnt
• WWLK-aligner, 📜: Wasserstein amr-text alignment