Deeppavlov

- Downloads resuming after interruptions
- Small documentation tweaks
- [Multilingual SQUAD configuration](https://github.com/deepmipt/DeepPavlov/blob/0.5.1/deeppavlov/configs/squad/squad_bert_multilingual_freezed_emb.json) for inference

- Conversational BERT model in English (trained on informal lexicon data)
- Updated requirements
- Python 3.7 support
- New DSL for building rule-based skills
- New SOTA insult-detection model based on the conversational BERT
- Updated documentation
- Refactored MorphoTagger classes to better conform DeepPavlov ideology
- Many smaller changes and fixes

* [New rule-based obscenity classifier for Russian language](https://github.com/deepmipt/DeepPavlov/blob/0.4.0/deeppavlov/configs/classifiers/ru_obscenity_classifier.json)
* [Lemmatization option in Morphotagger pipeline for Russian language](http://docs.deeppavlov.ai/en/0.4.0/components/morphotagger.html#python)
* Unification of default configuration variables as per 844
* Fixes in knowledge base dataset readers
* other minor fixes

* Fixes in documentation
* Compressed BERT-based models
* Minor fixes

* Full spectrum of [BERT-based models](http://docs.deeppavlov.ai/en/master/components/bert.html)
* [Multilingual BERT based NER](http://docs.deeppavlov.ai/en/master/components/ner.html#ner-multi-bert) for 100 languages and 19 classes
* Pre-trained [BERT-base model for Russian](http://docs.deeppavlov.ai/en/master/components/bert.html) Language
* Updated [rusentiment model](http://docs.deeppavlov.ai/en/master/components/bert.html#bert-for-classification) trained on twitter
* New [AIML](http://docs.deeppavlov.ai/en/master/skills/aiml_skill.html) wrapper component
* New [SMN, DAM, DAM-USE-T ranking models](http://docs.deeppavlov.ai/en/master/intro/features.html#ranking-component)
* Minor fixes

* [Multilingual and English BERT-based models for classification and SQuAD](http://docs.deeppavlov.ai/en/master/components/squad.html)
* [Knowledge base based Question Answering model](http://docs.deeppavlov.ai/en/master/components/kbqa.html)
* [Refactor train script to class to simplify its modification and extension](http://docs.deeppavlov.ai/en/master/apiref/core/trainers.html)
* [Super-convergence Keras model class]()
* +3 F1 for speller correction with new custom `vocab_penalty` property for `searcher_component`
* Refactored similarity matching skill
* Add universal method `destroy` in `Component` class
* `--no-default-skill` optional flag if your component implements an interface of DeepPavlov skill
* `--stateful` and `--multiinstance` flags for Amazon Alexa integration
* Move `utils` module from project root into `deeppavlov` module
* Many minor improvements and bug fixes, including logging, ELMo save/load and many other