Arekit

Release

Fixed:
* Using custom check of duplicated opinions during `OpinionCollection` initialization.
Changes:
* Speed-up and engine optimizations:
* Optionally loading neutral annotator.
* Multi-Instance networks: now we consider that the next appered context always continues the prior.
(check out multi-instance bags creation for details)
* Now shuffling in models performed for bags, not for bag groups.
* Networks: added `allow_growth=True` flag for tensorflow based neural networks.
Memory fraction parameter has been removed.

Collection of parsed news become dispatched from text opinions collection.
* **News parsing now is assumed to be performed using `TextParser.parse(news, options)` call. Related refactoring.**
* Stemmer application from `RuAttitiudes` parser has been removed.
* Removed dependency from `RelatedParsedNewCollection` in TextOpinionCollection.
* Labeling now separated from LinkedTextOpinion collection.
* `ParsedText` class has been refactored, removed unused methods. Keep tokens has been discarded.
* BERT tsv-format-encoders are now in a Factory (at contrib directory).
* Fixed: `RuSentRelTextOpinion` replaced with `TextOpinion`, and independent from `OpinionRef`.
* `Single`/`Multi` models now are not exist, as the latter prefixes affects only onto batch types selection. Refactoring.

Release Notes

* Labels conversion `to_str` and `from_str` now a part of external formatters (unique for each source, experiment, etc.).
* Added labels-scaler, and labels casing (to int or uint) now depends on scaler;
* Added bert exporter in contribution folder: with related formatters according to the [[paper]](https://www.aclweb.org/anthology/N19-1035.pdf):
* **NLI** -- (Natural language inference) format, assumes to provide an additional sentence, which describes
attitude should be extracted
* **QA** -- (Question answering) provides an additional question onto attitude sentiment.
With Label encoding in following format:
* **Multiple** -- all the supported sentiment labels (positive, negative, neutral)
* **Binary** -- (YES, NO) according to mention (additional sentence), provided by **NLI** and **QA** formatters.
* Refactoring experiments in order to apply the latter also for classifiers (models from scikit-learn)
* Updated nn-engine API
* Refactoring tf-based neural network implementation.
* Bert now moved into separated folder from `contrib` directory.
* frame_variants moved to `frames` directory.
* Frame variants labeling in news now performed during `parse` operation.
* `DataType` now enumeration. List of Supported data-types now a part of experiment
The latter were moved onto sample level.
* `Service` folder removed as the latter assumes to be apart of this repository.

v0.20.3-wims-rc
Release Notes
* Experiments now supports two-scale and three-scale task representation with the related evaluation formats.