Ask2Transformers v0.4

New tasks, datasets and pre-trained entailment models

This updates introduces the tasks of Event Extraction (as TextClassificationTask) and EventArgumentClassificationTask tasks. Also, it adds dataset loaders for ACE and WikiEvents datasets. On this update we provide the necessary code to reproduce the results of the Textual Entailment for Event Argument Extraction: Zero- and Few-Shot with Multi-Source learning paper. We have also uploaded pre-trained entailment models to the HuggingFace model Hub that can be used directly by this framework:

• (TACRED) HiTZ/A2T_RoBERTa_SMFA_TACRED-re
• (ACE) HiTZ/A2T_RoBERTa_SMFA_ACE-arg
• (WikiEvents --> ACE) HiTZ/A2T_RoBERTa_SMFA_WikiEvents-arg_ACE-arg
• (WikiEvents) HiTZ/A2T_RoBERTa_SMFA_WikiEvents-arg
• (ACE --> WikiEvents) HiTZ/A2T_RoBERTa_SMFA_ACE-arg_WikiEvents-arg

Some minor corrections and new unit-tests and documentation have also been added.

What's Changed

• New tasks, datasets and pre-trained models by @osainz59 in #11

Full Changelog: v0.3...v0.4

Ask2Transformers v0.3

Repository complete redesign

This update changes drastically all the code of the repository. Before this change all the code was mostly done ad hoc to run the experiments. Now, with a framework perspective in mind, we refactored the code to offer a proper framework to build textual entailment based systems and run (or replicate) experiments on different tasks.

The repository now has three main modules: a2t.data, a2t.tasks and a2t.base. Each of these represents the main components of the framework: the code data loading, the task definition and the inference. Also, now we distinguish tasks based on the number of "spans" to classify: ZeroaryTask for text classificaiton, UnaryTask for span classification like NER and BinaryTask for span pairs classification like Relation Extraction. This decision was made with Information Extraction perspective in mind.

On this release we also added the documentation to the repository (finally!). It is available at https://osainz59.github.io/Ask2Transformers.

We also started doing Unit testing to try avoid most of the bugs that could be generated.

Roadmap

We plan to develop new Tasks with the entailment approach like: NER, event-extraction or event argument extraction.

We also plan to develop a pipeline feature, to run a predefined sequence of tasks with a single model.

We still did not finished to write all the documentation and tests, so we plan to continue working on it.

Ask2Transformers v0.2.1

Relation Extraction update

Add the new task: Relation Classification
Add support for TACREDClassifier

Ask2Transformers v0.1.2

Integrate WNDomains

• Add WNDomainsClassifier
• Update README

Ask2Transformers v0.1.1

Minor bugs correction

• FP16 error when CUDA is not available

Initial release

A2T 0.1

Initial release:

• Adding topic_classification package