transition-neural-parser 0.5.2.dev0

The package for transition based nueral AMR parser

Transition-based Neural Parser

transition-neural-parser

transition-neural-parser is a powerful and easy-to-use Python package that provides a state-of-the-art neural transition-based parser for Abstract Meaning Representation (AMR). AMR is a semantic formalism used to represent the meaning of natural language sentences in a structured and machine-readable format. The package is designed to enable users to perform AMR parsing with high accuracy and generate reliable token-to-node alignments, which are crucial for various natural language understanding and generation tasks.

Pip Installation Instructions

Step 1: Create and activate a new conda environment To ensure compatibility and prevent potential conflicts, create a new conda environment with Python 3.8:

conda create -n amr-parser python=3.8

Activate the newly created environment:

conda activate amr-parser

Step 2: Install the package and dependencies Install the transition-neural-parser package using pip:

pip install transition-neural-parser

Next, install the torch-scatter dependency. For Linux servers, use the following command:

pip install torch-scatter -f https://data.pyg.org/whl/torch-1.13.1+cu117.html

For other operating systems, please visit the official torch-scatter repository to find the appropriate installation instructions.

Step 3: Download a pretrained AMR parser and run inference Here is an example of how to download and use a pretrained AMR parser:

from transition_amr_parser.parse import AMRParser

# Download and save a model named AMR3.0 to cache
parser = AMRParser.from_pretrained('AMR3-structbart-L')
tokens, positions = parser.tokenize('The girl travels and visits places')

# Use parse_sentence() for single sentences or parse_sentences() for a batch
annotations, machines = parser.parse_sentence(tokens)

# Print Penman notation
print(annotations)

# Print Penman notation without JAMR, with ISI
amr = machines.get_amr()
print(amr.to_penman(jamr=False, isi=True))

# Plot the graph (requires matplotlib)
amr.plot()

This example demonstrates how to tokenize a sentence, parse it using the pretrained AMR parser, and print the resulting AMR graph in Penman notation. If you have matplotlib installed, you can also visualize the graph.

Available Pretrained Models

The models downloaded using from_pretrained() method will be stored to the pytorch cache folder as follows:

cache_dir = torch.hub._get_torch_home()

This table shows you available pretrained model names to download;

pretrained model name	corresponding file name
AMR3-structbart-L-smpl	amr3.0-structured-bart-large-neur-al-sampling5-seed42.zip
AMR3-structbart-L	amr3.0-structured-bart-large-neur-al-seed42.zip
AMR2-structbart-L	amr2.0-structured-bart-large-neur-al-seed42.zip
AMR2-joint-ontowiki-seed42	amr2joint_ontowiki2_g2g-structured-bart-large-seed42.zip
AMR2-joint-ontowiki-seed43	amr2joint_ontowiki2_g2g-structured-bart-large-seed43.zip
AMR2-joint-ontowiki-seed44	amr2joint_ontowiki2_g2g-structured-bart-large-seed44.zip
AMR3-joint-ontowiki-seed42	amr3joint_ontowiki2_g2g-structured-bart-large-seed42.zip
AMR3-joint-ontowiki-seed43	amr3joint_ontowiki2_g2g-structured-bart-large-seed43.zip
AMR3-joint-ontowiki-seed44	amr3joint_ontowiki2_g2g-structured-bart-large-seed44.zip

Upcoming Features

The current release primarily supports model inference using Python scripts. In future versions, we plan to expand the capabilities of this package by:

• Adding training and evaluation scripts for a more comprehensive user experience. Interested users can refer to the IBM/transition-amr-parser repository for training and evaluation in the meantime.
• Broadening platform support to include MacOS and higher versions of Python, in addition to the current support for the Linux operating system and Python 3.8.

Release History

• v1.0.0: This release corresponds to v0.5.2 in the IBM/transition-amr-parser repository.

Research and Evaluation Results

Structured-BART

Current version (0.5.2). Structured-BART (Zhou et al 2021b) encodes the parser state using specialized cross and self-attention heads and leverages BART's language model to replace the use of subgraph actions and lemmatizer, thus enabling a much simpler oracle with 100% coverage. It yields 84.2 Smatch (84.7 with silver data and 84.9 with ensemble). Version 0.5.2 introduces the ibm-neural-aligner (Drozdov et al 2022) yielding a base AMR3.0 performance of 82.7 (83.1 with latent alignment training). Structured-BART is also used for (Lee et al 2022) which yields a new single model SoTA of 85.7 for AMR2.0 and 84.1 for AMR3.0 by introducing Smatch-based ensemble distillation.

Action Pointer

Checkout the action-pointer branch (derived from version 0.4.2) for the Action Pointer Transformer model (Zhou et al 2021) from NAACL2021. As the stack-Transformer, APT encodes the parser state in dedicated attention heads. APT uses however actions creating nodes to represent them. This decouples token and node representations yielding much shorter sequences than previous oracles with higher coverage. APT achieves 81.8 Smatch (83.4 with silver data and partial ensemble) on AMR2.0 test using RoBERTa embeddings and has an efficient shallow decoder. Due to aligner implementation improvements this code reaches 82.1 on AMR2.0 test, better that what is reported in the paper.

Stack-Transformer

Checkout the stack-transformer branch (derived from version 0.3.4) for the stack-Transformer model (Fernandez Astudillo et al 2020) from EMNLP findings 2020. The stack-Transformer masks dedicated cross attention heads to encode the parser state represented by stack and buffer. It yields 80.2 Smatch (81.3 with self-learning) on AMR2.0 test (this code reaches 80.5 due to the aligner implementation). Stack-Transformer can be used to reproduce our works on self-learning and cycle consistency in AMR parsing (Lee et al 2020) from EMNLP findings 2020, alignment-based multi-lingual AMR parsing (Sheth et al 2021) from EACL 2021 and Knowledge Base Question Answering (Kapanipathi et al 2021) from ACL findings 2021.

The code also contains an implementation of the AMR aligner from (Naseem et al 2019) with the forced-alignment introduced in (Fernandez Astudillo et al 2020).

Aside from listed contributors, the initial commit was developed by Miguel Ballesteros and Austin Blodgett while at IBM.

Evaluating Trained checkpoints

We offer some trained checkpoints on demand, and their evalution score measured in Smatch is below:

paper	config(.zip)	beam	Smatch
(Drozdov et al 2022) MAP	amr2.0-structured-bart-large-neur-al-seed42	10	84.0
(Drozdov et al 2022) MAP	amr3.0-structured-bart-large-neur-al-seed42	10	82.6
(Drozdov et al 2022) PR	amr3.0-structured-bart-large-neur-al-sampling5-seed42	1	82.9
(Lee et al 2022) (ensemble)	amr2joint_ontowiki2_g2g-structured-bart-large	10	85.9
(Lee et al 2022) (ensemble)	amr3joint_ontowiki2_g2g-structured-bart-large	10	84.4

we also provide the trained ibm-neural-aligner under names AMR2.0_ibm_neural_aligner.zip and AMR3.0_ibm_neural_aligner.zip. For the ensemble we provide the three seeds. Following fairseq conventions, to run the ensemble just give the three checkpoint paths joined by : to the normal checkpoint argument -c. Note that the checkpoints were trained with the v0.5.1 tokenizer, this reduces performance by 0.1 on v0.5.2 tokenized data.

Note that we allways report average of three seeds in papers while these are individual models. A fast way to test models standalone is

bash tests/standalone.sh configs/<config>.sh