Home

Welcome to the alignment-models wiki!

Here you can find information about how to navigate our alignment-models repository and reproduce our analysis on automatic alignment model from the paper Aligning Actions Across Recipe Graphs

We compare the performance of our two alignment models (base and extended) with a baseline and a naive model:

Model Name	Saves To
Alignment Model (extended)	./results1
Alignment Model (base)	./results2
Cosine Similarity Baseline	./results3
Naive Model	./results4

We test the alignment models on a recipes corpus that we extract from Lin et al. 2020.

You can look up how each model performs in our analysis by checking out the corresponding accuracy rates reported in the Results section of the README file.

================================================================================

How to run the models

Step 1:

• Download the corpus from here. Create directories to store the results (i.e., ./results1, ./results2, ./results3, ./results4). Optionally, change hyperparameters and path names in the file constants.py here.

Step 2:

• Run the following command from this directory:

python main.py [model_name] --embedding_name [embedding_name]

where [model_name] could be one of the following:

Sequence : Sequential Ordering of Alignments Cosine_similarity : Cosine model (Baseline) Naive : Common Action Pair Heuristics mode (Naive Model) Alignment-no-feature : Base Alignment model (w/o parent+child nodes) Alignment-with-feature : Extended Alignment model (with parent+child nodes)

and [embedding_name] could be one of the following:

bert : BERT embeddings (default) elmo : ELMO embeddings

Other files in the repo:

• property_testing.py: Testing for reverse and transitivity property
• sequence_model.py: Sequence ordering of alignments
• training_testing.py: Training and testing functions
• utils.py: Utility functions