Warning: This document is for an old version of Rasa NLU. The latest version is 0.15.1.

Change Log

All notable changes to this project will be documented in this file. This project adheres to Semantic Versioning starting with version 0.7.0.

[Unreleased 0.13.0.aX] - master

Note

This version is not yet released and is under active development.

Added

Changed

Removed

Fixed

[0.13.6] - 2018-10-04

Changed

  • boto3 is now loaded lazily in AWSPersistor and is not included in requirements_bare.txt anymore

Fixed

  • Allow training of pipelines containing EmbeddingIntentClassifier in a separate thread on python 3. This makes http server calls to /train non-blocking
  • require scikit-learn<0.20 in setup py to avoid corrupted installations with the most recent scikit learn

[0.13.5] - 2018-09-28

Changed

  • Training data is now validated after loading from files in loading.py instead of on initialisation of TrainingData object

Fixed

  • Project set up to pull models from a remote server only use the pulled model instead of searching for models locally

[0.13.4] - 2018-09-19

Fixed

  • pinned matplotlib to 2.x (not ready for 3.0 yet)
  • pytest-services since it wasn’t used and caused issues on Windows

[0.13.3] - 2018-08-28

Added

  • EndpointConfig class that handles authenticated requests (ported from Rasa Core)
  • DataRouter() class supports a model_server EndpointConfig, which it regularly queries to fetch NLU models
  • this can be used with rasa_nlu.server with the --endpoint option (the key for this the model server config is model)
  • docs on model fetching from a URL
  • ability to specify lookup tables in training data

Changed

  • loading training data from a URL requires an instance of EndpointConfig
  • Changed evaluate behaviour to plot two histogram bars per bin. Plotting confidence of right predictions in a wine-ish colour and wrong ones in a blue-ish colour.

Removed

Fixed

  • re-added support for entity names with special characters in markdown format

[0.13.2] - 2018-08-28

Changed

  • added information about migrating the CRF component from 0.12 to 0.13

Fixed

  • pipelines containing the EmbeddingIntentClassifier are not trained in a

separate thread, as this may lead to freezing during training

[0.13.1] - 2018-08-07

Added

  • documentation example for creating a custom component

Fixed

  • correctly pass reference time in miliseconds to duckling_http

[0.13.0] - 2018-08-02

Warning

This is a release breaking backwards compatibility. Unfortunately, it is not possible to load previously trained models as the parameters for the tensorflow and CRF models changed.

Added

  • support for tokenizer_jieba load custom dictionary from config
  • allow pure json including pipeline configuration on train endpoint
  • doc link to a community contribution for Rasa NLU in Chinese
  • support for component count_vectors_featurizer use tokens feature provide by tokenizer
  • 2-character and a 5-character prefix features to ner_crf
  • ner_crf with whitespaced tokens to tensorflow_embedding pipeline
  • predict empty string instead of None for intent name
  • update default parameters for tensorflow embedding classifier
  • do not predict anything if feature vector contains only zeros in tensorflow embedding classifier
  • change persistence keywords in tensorflow embedding classifier (make previously trained models impossible to load)
  • intent_featurizer_count_vectors adds features to text_features instead of overwriting them
  • add basic OOV support to intent_featurizer_count_vectors (make previously trained models impossible to load)
  • add a feature for each regex in the training set for crf_entity_extractor
  • Current training processes count for server and projects.
  • the /version endpoint returns a new field minimum_compatible_version
  • added logging of intent prediction errors to evaluation script
  • added histogram of confidence scores to evaluation script
  • documentation for the ner_duckling_http component

Changed

  • renamed CRF features wordX to suffixX and preX to suffixX
  • L1 and L2 regularisation defaults in ner_crf both set to 0.1
  • whitespace_tokenizer ignores punctuation .,!? before whitespace or end of string
  • Allow multiple training processes per project
  • Changed AlreadyTrainingError to MaxTrainingError. The first one was used to indicate that the project was already training. The latest will show an error when the server isn’t able to training more models.
  • Interpreter.ensure_model_compatibility takes a new parameters for the version to compare the model version against
  • confusion matrix plot gets saved to file automatically during evaluation

Removed

  • dependence on spaCy when training ner_crf without POS features
  • documentation for the ner_duckling component - facebook doesn’t maintain the underlying clojure version of duckling anymore. component will be removed in the next release.

Fixed

  • Fixed Luis emulation output to add start, end position and confidence for each entity.
  • Fixed byte encoding issue where training data could not be loaded by URL in python 3.

[0.12.3] - 2018-05-02

Added

  • Returning used model name and project name in the response of GET /parse and POST /parse as model and project respectively.

Fixed

  • readded possibility to set fixed model name from http train endpoint

[0.12.2] - 2018-04-20

Fixed

  • fixed duckling text extraction for ner_duckling_http

[0.12.1] - 2018-04-18

Added

  • support for retrieving training data from a URL

Fixed

  • properly set duckling http url through environment setting
  • improvements and fixes to the configuration and pipeline documentation

[0.12.0] - 2018-04-17

Added

  • support for inline entity synonyms in markdown training format
  • support for regex features in markdown training format
  • support for splitting and training data into multiple and mixing formats
  • support for markdown files containing regex-features or synonyms only
  • added ability to list projects in cloud storage services for model loading
  • server evaluation endpoint at POST /evaluate
  • server endpoint at DELETE /models to unload models from server memory
  • CRF entity recognizer now returns a confidence score when extracting entities
  • added count vector featurizer to create bag of words representation
  • added embedding intent classifier implemented in tensorflow
  • added tensorflow requirements
  • added docs blurb on handling contextual dialogue
  • distribute package as wheel file in addition to source distribution (faster install)
  • allow a component to specify which languages it supports
  • support for persisting models to Azure Storage
  • added tokenizer for CHINESE (zh) as well as instructions on how to load MITIE model

Changed

  • model configuration is separated from server / train configuration. This is a breaking change and models need to be retrained. See migrations guide.
  • Regex features are now sorted internally. retrain your model if you use regex features
  • The keyword intent classifier now returns null instead of "None" as intent name in the json result if there’s no match
  • in teh evaluation results, replaced O with the string no_entity for better understanding
  • The CRFEntityExtractor now only trains entity examples that have "extractor": "ner_crf" or no extractor at all
  • Ignore hidden files when listing projects or models
  • Docker Images now run on python 3.6 for better non-latin character set support
  • changed key name for a file in ngram featurizer
  • changed jsonObserver to generate logs without a record seperator
  • Improve jsonschema validation: text attribute of training data samples can not be empty
  • made the NLU server’s /evaluate endpoint asynchronous

Fixed

  • fixed certain command line arguments not getting passed into the data_router

[0.11.4] - 2018-03-19

Fixed

  • google analytics docs survey code

[0.11.3] - 2018-02-13

Fixed

  • capitalization issues during spacy named entity recognition

[0.11.2] - 2018-02-06

Fixed

  • Formatting of tokens without assigned entities in evaluation

[0.11.1] - 2018-02-02

Fixed

  • Changelog doc formatting
  • fixed project loading for newly added projects to a running server
  • fixed certain command line arguments not getting passed into the data_router

[0.11.0] - 2018-01-30

Added

  • non ascii character support for anything that gets json dumped (e.g. training data received over HTTP endpoint)
  • evaluation of entity extraction performance in evaluation.py
  • support for spacy 2.0
  • evaluation of intent classification with crossvalidation in evaluation.py
  • support for splitting training data into multiple files (markdown and JSON only)

Changed

  • removed -e . from requirements files - if you want to install the app use pip install -e .
  • fixed http duckling parsing for non en languages
  • fixed parsing of entities from markdown training data files

[0.10.6] - 2018-01-02

Added

  • support asterisk style annotation of examples in markdown format

Fixed

  • Preventing capitalized entities from becoming synonyms of the form lower-cased -> capitalized

[0.10.5] - 2017-12-01

Fixed

  • read token in server from config instead of data router
  • fixed reading of models with none date name prefix in server

[0.10.4] - 2017-10-27

Fixed

  • docker image build

[0.10.3] - 2017-10-26

Added

  • support for new dialogflow data format (previously api.ai)
  • improved support for custom components (components are stored by class name in stored metadata to allow for components that are not mentioned in the Rasa NLU registry)
  • language option to convert script

Fixed

  • Fixed loading of default model from S3. Fixes #633
  • fixed permanent training status when training fails #652
  • quick fix for None “_formatter_parser” bug

[0.10.1] - 2017-10-06

Fixed

  • readme issues
  • improved setup py welcome message

[0.10.0] - 2017-09-27

Added

  • Support for training data in Markdown format
  • Cors support. You can now specify allowed cors origins within your configuration file.
  • The HTTP server is now backed by Klein (Twisted) instead of Flask. The server is now asynchronous but is no more WSGI compatible
  • Improved Docker automated builds
  • Rasa NLU now works with projects instead of models. A project can be the basis for a restaurant search bot in German or a customer service bot in English. A model can be seen as a snapshot of a project.

Changed

  • Root project directories have been slightly rearranged to clean up new docker support
  • use Interpreter.create(metadata, ...) to create interpreter from dict and Interpreter.load(file_name, ...) to create interpreter with metadata from a file
  • Renamed name parameter to project
  • Docs hosted on GitHub pages now: Documentation
  • Adapted remote cloud storages to support projects (backwards incompatible!)

Fixed

  • Fixed training data persistence. Fixes #510
  • Fixed UTF-8 character handling when training through HTTP interface
  • Invalid handling of numbers extracted from duckling during synonym handling. Fixes #517
  • Only log a warning (instead of throwing an exception) on misaligned entities during mitie NER

[0.9.2] - 2017-08-16

Fixed

  • removed unnecessary ClassVar import

[0.9.1] - 2017-07-11

Fixed

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

[0.9.0] - 2017-07-07

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

[0.8.12] - 2017-06-29

Fixed

  • fixed missing argument attribute error

[0.8.11] - 2017-06-07

Fixed

  • updated mitie installation documentation

[0.8.10] - 2017-05-31

Fixed

  • fixed documentation about training data format

[0.8.9] - 2017-05-26

Fixed

  • properly handle response_log configuration variable being set to null

[0.8.8] - 2017-05-26

Fixed

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

[0.8.7] - 2017-05-24

Fixed

  • Fixed range calculation for crf #355

[0.8.6] - 2017-05-15

Fixed

  • Fixed duckling dimension persistence. fixes #358

[0.8.5] - 2017-05-10

Fixed

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

[0.8.4] - 2017-05-10

Fixed

  • Fixed CRF model training without entities. fixes #345

[0.8.3] - 2017-05-10

Fixed

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

[0.8.2] - 2017-05-08

Fixed

  • deepcopy of context #343

[0.8.1] - 2017-05-08

Fixed

  • NER training reuses context inbetween requests

[0.8.0] - 2017-05-08

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. MITIETokenizerMitieTokenizer)

  • 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

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

[0.7.4] - 2017-03-27

Fixed

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

[0.7.3] - 2017-03-15

Fixed

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

[0.7.2] - 2017-03-13

Fixed

  • python documentation about calling rasa NLU from python

[0.7.1] - 2017-03-10

Fixed

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

[0.7.0] - 2017-03-10

This is a major version update. Please also have a look at the Migration Guide.

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

[0.6-beta] - 2017-01-31