Change Log¶

Fixed¶

• removed unnecessary ClassVar import

Fixed¶

• removed obsolete --output parameter of train.py. use --path instead. fixes #473

Added¶

• increased test coverage to avoid regressions (ongoing)
• added regex featurization to support intent classification and entity extraction (intent_entity_featurizer_regex)

Changed¶

• replaced existing CRF library (python-crfsuite) with sklearn-crfsuite (due to better windows support)
• updated to spacy 1.8.2
• logging format of logged request now includes model name and timestamp
• use module specific loggers instead of default python root logger
• output format of the duckling extractor changed. the value field now includes the complete value from duckling instead of just text (so this is an property is an object now instead of just text). includes granularity information now.
• deprecated intent_examples and entity_examples sections in training data. all examples should go into the common_examples section
• weight training samples based on class distribution during ner_crf cross validation and sklearn intent classification training
• large refactoring of the internal training data structure and pipeline architecture
• numpy is now a required dependency

Removed¶

• luis data tokenizer configuration value (not used anymore, luis exports char offsets now)

Fixed¶

• properly update coveralls coverage report from travis
• persistence of duckling dimensions
• changed default response of untrained intent_classifier_sklearn from "intent": None to "intent": {"name": None, "confidence": 0.0}
• /status endpoint showing all available models instead of only those whose name starts with model
• properly return training process ids #391

Fixed¶

• fixed missing argument attribute error

Fixed¶

• updated mitie installation documentation

Fixed¶

• fixed documentation about training data format

Fixed¶

• /status endpoint showing all available models instead of only those whose name starts with model

Fixed¶

• Fixed range calculation for crf #355

Fixed¶

• Fixed duckling dimension persistence. fixes #358

Fixed¶

• Fixed pypi installation dependencies (e.g. flask). fixes #354

Fixed¶

• Fixed CRF model training without entities. fixes #345

Fixed¶

• Fixed Luis emulation and added test to catch regression. Fixes #353

Fixed¶

• deepcopy of context #343

Fixed¶

• NER training reuses context inbetween requests

Added¶

• ngram character featurizer (allows better handling of out-of-vocab words)
• replaced pre-wired backends with more flexible pipeline definitions
• return top 10 intents with sklearn classifier #199
• python type annotations for nearly all public functions
• added alternative method of defining entity synonyms
• support for arbitrary spacy language model names
• duckling components to provide normalized output for structured entities
• Conditional random field entity extraction (Markov model for entity tagging, better named entity recognition with low and medium data and similarly well at big data level)
• allow naming of trained models instead of generated model names
• dynamic check of requirements for the different components & error messages on missing dependencies
• support for using multiple entity extractors and combining results downstream

Changed¶

• unified tokenizers, classifiers and feature extractors to implement common component interface

• src directory renamed to rasa_nlu

• when loading data in a foreign format (api.ai, luis, wit) the data gets properly split into intent & entity examples

• Configuration:

  • added max_number_of_ngrams
  • removed backend and added pipeline as a replacement
  • added luis_data_tokenizer
  • added duckling_dimensions

• parser output format changed

  from {"intent": "greeting", "confidence": 0.9, "entities": []}

  to {"intent": {"name": "greeting", "confidence": 0.9}, "entities": []}

• entities output format changed

  from {"start": 15, "end": 28, "value": "New York City", "entity": "GPE"}

  to {"extractor": "ner_mitie", "processors": ["ner_synonyms"], "start": 15, "end": 28, "value": "New York City", "entity": "GPE"}

  where extractor denotes the entity extractor that originally found an entity, and processor denotes components that alter entities, such as the synonym component.

• camel cased MITIE classes (e.g. MITIETokenizer → MitieTokenizer)

• model metadata changed, see migration guide

• updated to spacy 1.7 and dropped training and loading capabilities for the spacy component (breaks existing spacy models!)

• introduced compatibility with both Python 2 and 3

Removed¶

Fixed¶

• properly parse str additionally to unicode #210
• support entity only training #181
• resolved conflicts between metadata and configuration values #219
• removed tokenization when reading Luis.ai data (they changed their format) #241

Fixed¶

• fixed failed loading of example data after renaming attributes, i.e. “KeyError: ‘entities’”

Fixed¶

• fixed regression in mitie entity extraction on special characters
• fixed spacy fine tuning and entity recognition on passed language instance

Fixed¶

• python documentation about calling rasa NLU from python

Fixed¶

• mitie tokenization value generation #207, thanks @cristinacaputo
• changed log file extension from .json to .log, since the contained text is not proper json

Added¶

• Changelog ;)
• option to use multi-threading during classifier training
• entity synonym support
• proper temporary file creation during tests
• mitie_sklearn backend using mitie tokenization and sklearn classification
• option to fine-tune spacy NER models
• multithreading support of build in REST server (e.g. using gunicorn)
• multitenancy implementation to allow loading multiple models which share the same backend

Fixed¶

• error propagation on failed vector model loading (spacy)
• escaping of special characters during mitie tokenization