Home

Awesome

MISO for Universal Decompositional Semantic Parsing

What is MISO?

MISO stands for Multimodal Inputs, Semantic Outputs. It is a deep learning framework with re-usable components for parsing a variety of semantic parsing formalisms. In various iterations, MISO has been used in the following publications:

If you use the code in a publication, please do cite these works.

What is Universal Decompositional Semantics?

Universal Decompositional Semantics (UDS) is a flexible semantic formalism built on English Web Treebank Universal Dependencies parses. UDS graphs are directed acyclic graphs on top of UD parses which encode the predicate-argument structure of an utterance. These graphs are annotated with rich, scalar-valued semantic inferences obtained from human annotators via crowdsourcing, encoding speaker intuitions about a variety of semantic phenomena including factuality, genericity, and semantic proto-roles. More details about the dataset can be found in the following paper: The Universal Decompositional Semantics Dataset and Decomp Toolkit, White et al., LREC 2020 and at decomp.io.

What is UDS Parsing?

UDS parsing is the task of transforming an utterance into a UDS graph, automatically. Using the existing dataset and the MISO framework, we can learn to parse into UDS. This is a particularly challenging parsing problem, as it involves three levels of parsing

  1. Syntactic parsing of the utterance into UD
  2. Parsing the utterance into the UDS graph structure
  3. Annotating the graph structure with UDS attributes

MISO overview

MISO builds heavily on AllenNLP, and so many of its core functionalities are the same.

Installation

Using conda, all required libraries can be installed by running:

Useful scripts

experiments/decomp_train.sh has several functions for training and evaluating UDS parsers via the command-line. This script is used for DecompParser models, trained and evaluated without UD parses. Specifically:

If training/evaluating a model with syntactic info, a similar script is used: syntax_experiments/decomp_train.sh. This script has the same functions as experiments/decomp_train, but also contains:

Configurations

For info on configurations, see the configs page

Models

For info on configurations, see the models page

Training

For info on configurations, see the training page

Testing

For info on configurations, see the testing page For predicting UDS graphs from arbitrary text, see the prediction page

Multilingual experiments

For info on configurations, see the multilingual page