zensols.calamr 0.1.0

CALAMR: Component ALignment for AMR

CALAMR: Component ALignment for Abstract Meaning Representation

This repository contains code for the paper CALAMR: Component ALignment for Abstract Meaning Representation and aligns the components of a bipartite source and summary AMR graph. To reproduce the results of the paper, see the paper repository.

The results are useful as a semantic graph similarity score (like SMATCH) or to find the summarized portion (as AMR nodes, edges and subgraphs) of a document or the portion of the source that represents the summary. If you use this library or the PropBank API/curated database, please cite our paper.

Features:

• Align source/summary AMR graphs.
• Scores for extent to which AMRs are summarized or represented in their source text.
• Rendering of the alignments.
• Support for four AMR corpora.

Table of Contents

• Documentation
• Obtaining
• Corpora
• Usage
  • Command Line
    • Aligning Corpus Documents
    • Ad hoc Corpora
  • AMR Release 3.0 Corpus (LDC2020T02)
  • API
  • Docker
• Example Graphs
  • GraphViz
    • The Nascent Graph (with flow data)
    • The Source Graph
  • Plotly
• Attribution
• Citation
• Changelog
• License

Documentation

The recommended reading order for this project:

1. The abstract and introduction of the paper CALAMR: Component ALignment for Abstract Meaning Representation
2. Overview and implementation guide
3. Full documentation
4. API reference

Obtaining

The library can be installed with pip from the pypi repository:

pip3 install zensols.calamr

Corpora

This repository contains code to support the following corpora with source/summary AMR for alignment:

• LDC2020T02 (AMR Release 3.0) Proxy Corpus
• ISI Little Prince
• ISI Bio AMR
• A micro corpus (toy corpus) used in the paper examples and usage.

Usage

The command-line tool and API does not depend on the repository. However, it has a template configuration file that both the CLI and the API use. The examples also use data in the repository. Do the following to get started:

1. Clone this repository and change the working directory to it:

   git clone https://github.com/plandes/calamr && cd calamr
   

2. Copy the resource file:

   cp src/config/dot-calamrrc ~/.calamrrc
   

Command Line

The steps below show how to use the command-line tool. First set up the application environment:

1. Edit the ~/.calamrrc file to choose the corpus and visualization. Keep the calamr_corpus set to adhoc for these examples. (Note that you can also set the the CALAMRRC environment variable to a file in a different location if you prefer.)
2. Create the micro corpus:

   calamr mkadhoc --corpusfile corpus/micro/source.json
   

3. Print the document keys of the corpus:

   calamr keys
   

Aligning Corpus Documents

AMR corpora that distinguish between source and summary documents are needed so the API knows what data to align. The following examples utilize preexisting corpora (including the last section's micro corpus):

1. Generate the Liu et al. graph for the micro corpus in directory example:

   calamr aligncorp liu-example -f txt -d example
   

2. Force the Little Prince AMR corpus download and confirm success with the single document key 1943:

   calamr keys --override=calamr_corpus.name=little-prince
   

3. Use the default AMR parser to extract sentence text from the Little Prince AMR corpus using the SPRING (Bevilacqua et al. 2021) parser:

   calamr penman -d lp.txt --limit 5 \
   	--override amr_default.parse_model=spring \
   	~/.cache/calamr/corpus/amr-rel/amr-bank-struct-v3.0.txt
   

4. Score the parsed sentences using CALAMR, SMATCH and WLK:

   calamr score --parsed lp.txt \
   	--methods calamr,smatch,wlk \
   	~/.cache/calamr/corpus/amr-rel/amr-bank-struct-v3.0.txt
   

Ad hoc Corpora

The micro corpus can be edited and rebuilt to add your own data to be aligned. However, there's an easier way to align ad hoc documents.

1. Align a summarized document not included in any corpus. First create the annotated documents as files short-story.json.

   [
       {
    	   "id": "intro",
    	   "body": "The Dow Jones Industrial Average and other major indexes pared losses.",
    	   "summary": "Dow Jones and other major indexes reduced losses."
       },
       {
    	   "id": "dow-stats",
    	   "body": "The Dow ended 0.5% lower on Friday while the S&P 500 fell 0.7%. Among the S&P sectors, energy and utilities gained while technology and communication services lagged.",
    	   "summary": "Dow sank 0.5%, S&P 500 lost 0.7% and energy, utilities up, tech, comms came down."
       }
   ]
   

   Now align the documents using the XFM Bart Base AMR parser, rendering with the maximum number of steps (-r 10), and save results to example:

   calamr align short-story.json --override amr_default.parse_model=xfm_bart_base -r 10 -d example -f txt
   

The -r option controls how many intermediate graphs generated to show the iteration of the algorithm over all the steps (see the paper for details).

AMR Release 3.0 Corpus (LDC2020T02)

If you are using the AMR 3.0 corpus, there is a preprocessing step that needs executing before it can be used.

The Proxy Report corpus from the AMR 3.0 does not have both the alignments (text-to-graph alignments) and snt-type (indicates if a sentence is part of the source or the summary) metadata. By default, this API expects both. To merge them into one dataset do the following:

1. Obtain or purchase the corpus.
2. Move the file where the software can find it:

   mkdir ~/.cache/calamr/download
   cp /path/to/amr_annotation_3.0_LDC2020T02.tgz ~/.cache/calamr/download
   

3. Merge the alignments and sentence descriptors:

   ./src/bin/merge-proxy-anons.py
   

4. Confirm the merge was successful by printing the document keys and align a report:

   calamr keys --override=calamr_corpus.name=proxy-report
   calamr aligncorp 20041010_0024 -f txt -d example \
       --override calamr_corpus.name=proxy-report
   

API

To use the package programmatically:

1. Get the resource bundle:

   from pathlib import Path
   from zensols.amr import AmrFeatureDocument
   from zensols.calamr import DocumentGraph, Resource, ApplicationFactory
   
   # get the resource bundle
   res: Resource = ApplicationFactory.get_resource()
   

2. Get the Liu et al. AMR feature document example and print it.

   doc: AmrFeatureDocument = res.get_corpus_document('liu-example')
   doc.write()
   

   output:

   [T]: Joe's dog was chasing a cat in the garden. I saw Joe's dog, which was running in the garden. The dog was chasing a cat.
   sentences:
       [N]: Joe's dog was chasing a cat in the garden.
           (c0 / chase-01~e.4
                 :location (g0 / garden~e.9)
                 :ARG0 (d0 / dog~e.2
                       :poss (p0 / person
                             :name (n0 / name
                                   :op1 "Joe"~e.0)))
                 :ARG1 (c1 / cat~e.6))
   .
   .
   .
   amr:
       summary:
           Joe's dog was chasing a cat in the garden.
       sections:
           no section sentences
               I saw Joe's dog, which was running in the garden.
               The dog was chasing a cat.
   

3. Align (if not already and cached) and get the flow results of the example:

   flow = res.align_corpus_document('liu-example')
   flow.write()
   

   output:

   summary:
       Joe's dog was chasing a cat in the garden.
   sections:
       no section sentences
           I saw Joe's dog, which was running in the garden.
           The dog was chasing a cat.
   statistics:
       agg:
           aligned_portion_hmean: 0.8695652173913044
           mean_flow: 0.7131309357900468
           tot_alignable: 21
           tot_aligned: 18
           aligned_portion: 0.8571428571428571
           reentrancies: 0
   

4. Parse the first document from the ad hoc JSON file align it, and give its statistics:

   doc: AmrFeatureDocument = next(iter(res.parse_documents(Path('short-story.json'))))
   graph: DocumentGraph = res.create_graph(doc)
   flow = res.align(graph)
   flow.write()
   

   output:

   summary:
       Dow Jones and other major indexes reduced losses.
   sections:
       no section sentences
           The Dow Jones Industrial Average and other major indexes pared losses.
   statistics:
       agg:
           aligned_portion_hmean: 1.0
           mean_flow: 0.9269955839429582
           tot_alignable: 24
           tot_aligned: 24
           aligned_portion: 1.0
           reentrancies: 0
    ...
   

5. Render the results of a flow:

   flow = res.align_corpus_document('liu-example')
   flow.render()
   

6. Render all graphs of the flow results of the flow to directory example:

   flow.render(
       contexts=flow.get_render_contexts(include_nascent=True),
       directory=Path('example'),
       display=False)
   

Docker

A stand-alone docker image is also available (see CALAMR Docker image). This docker image provides stand-alone container with all models, configuration and the adhoc micro corpus installed.

Example Graphs

The Liu et al. example graphs were created from the last step of the API examples, which is equivalent the first step of the command line example.

GraphViz

To create these graphs, set your ~/.calamrrc configuration to:

[calamr_default]
renderer = graphviz

The Nascent Graph (with flow data)

The Source Graph

Plotly

To create these graphs, set your ~/.calamrrc configuration to:

[calamr_default]
renderer = plotly

See the interactive version.

Attribution

This project, or reference model code, uses:

• Python 3.11
• amrlib for AMR parsing.
• amr_coref for AMR co-reference
• zensols.amr for AMR features and summarization data structures.
• Sentence-BERT embeddings
• zensols.propbankdb and zensols.deepnlp for PropBank embeddings
• zensols.nlparse for natural language features and NLP scoring
• Smatch (Cai and Knight. 2013) and WLK (Opitz et. al. 2021) for scoring.

Citation

If you use this project in your research please use the following BibTeX entry:

@inproceedings{landes-di-eugenio-2024-calamr-component,
    title = "{CALAMR}: Component {AL}ignment for {A}bstract {M}eaning {R}epresentation",
    author = "Landes, Paul  and
      Di Eugenio, Barbara",
    editor = "Calzolari, Nicoletta  and
      Kan, Min-Yen  and
      Hoste, Veronique  and
      Lenci, Alessandro  and
      Sakti, Sakriani  and
      Xue, Nianwen",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italy",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.236",
    pages = "2622--2637"
}

Changelog

An extensive changelog is available here.

License

MIT License

Copyright (c) 2023 - 2024 Paul Landes