nlp.js

• Fix Czech Stemmer

• Add spanish corpus for benchmark
• Improves in spanish stemmer
• logic in process actions
• Update docs
• Fix duplicated stem method in several stemmers

[3.10.0] - 2019-10-07

Added

• Traverse for handlebars, so answers now can be arrays or objects
• Automatic stemmer: is able to learn rules from languages without stemmer when the languages are inflected.
• Tests of the automatic stemmer in polish
• Spell checking: now users can write with small typos
• Changelog
• Portuguese sentiment analysis
• Contributor pictures to the readme
• Bengali sentiment analysis

Changed

• Faster Levenshtein implementation
• Now the browser version is generated with terser

Fixed

• Extended NER to support datetimerange
• Sort classifications in the NER manager
• Use performance.now instead of process.hrtime for browser compatibility

• Integration with duckling
• Ukrainian support
• Update dependencies

Support for emojis

Performance, performance, performance

• Actions
• Japanese
• Update dependencies
• Tweek hyperparameters for best performance

The version 3 comes with some important changes, mainly focused on improve performance:

• NlpClassifier no longer exists, in favor of NluManager as the manager of several NLU classes, and is able to manage several languages and several domains inside each language.
• Now by default, each domain of a language has it's own neural network classifier. When a language has more than 1 domain, a master neural network is trained that instead of classifying into the intent, classify into de domain. That way the models are faster to train and have a better score.
• The language guesser is now trained with the trigrams from the utterances used to train. That means that has a best guessing, and also that non-existing languages are guessed (example, klingon).
• Added Tagalog and Galician languages.
• The console-bot example training time in version 2.x in my laptop was 108 seconds, in the version 3.x the training time went down to 3 seconds, so the improvement in performance is notable.
• Also the size of the model.nlp files is decreased, the console-bot example went from 1614KB down to 928KB.
• The browser version has decreased from 5.08MB down to 2.3MB

• When using LRC and Neural, give some different ways to calculate the score, by default average.