NEWS.md

In sbo: Text Prediction via Stupid Back-Off N-Gram Models

sbo 0.5.0

API and UI changes

• Former kgram_freqs class is now called sbo_kgram_freqs. The constructor kgram_freqs() is still available as an alias to sbo_kgram_freqs().

• Former sbo_preds class is now substituted by two classes:

  - `sbo_predictor`: for interactive use
  - `sbo_predtable`: for storing text predictors out of memory (e.g. 
  `save()` to file)
  

• sbo_predictor and sbo_predtable objects are obtained by the homonym constructors, which are now S3 generics accepting character input, as well as sbo_kgram_freqs and sbo_predtable (for the sbo_predictor() constructor) class objects. In particular, these allow to directly train a text predictor without storing the intermediate sbo_dictionary, and kgram_freqs objects.

• The behaviour of the dict argument in kgram_freqs() and kgram_freqs_fast() has changed, now accepting either a sbo_dictionary, a character or a formula (see also 'New features').
• The sbo_predictor implementation dramatically improves the speed of predict() (by a factor of x10). A single call to predict() now allocates a few kBs of RAM (whereas it previously allocated few MBs, c.f. issue #10).
• Metadata of sbo_kgram_freqs and sbo_pred* objects is now stored via attributes (#11).

New features

• New S3 class sbo_dictionary.
• New S3 class word_coverage with generic constructors and a preconfigured plot() method.
• Dictionaries in kgram_freqs() and sbo_pred*() can now be built also with a fixed target coverage fraction of training corpus.
• Added prune() generic function for reducing -gram order of kgram_freqs and sbo_predtable's.
• Added summary() methods for sbo_kgram_freqs and sbo_pred* objects; correspondingly, the output of print() has been simplified considerably (#5).
• The object of class sbo_kgram_freqs, sbo_dictionary, sbo_predictor and sbo_predtable can be constructed either through the homonymous constructors, or through the aliases kgram_freqs(), dictionary(), predictor(), predtable().

Other improvements and patches

• sbo now has SystemRequirements: C++11, for correct integration with C++11 code (in particular std::unordered_map).
• Model training (with sbo_predictor()) is now considerably faster, due to optimizations in the algorithm for building Stupid Back-Off prediction tables.
• The Stupid Back-Off algorithm is now thoroughly tested, and small inconsistencies between the predict.kgram_freqs() and predict.sbo_predictor() methods have been fixed, including:

  - Proper handling of unknown words
  - Consistent handling of ties in prediction probabilities.
  

  • Model evaluation in eval_sbo_predictor() is now carried out by sampling a single sentence from each document in test corpus.
  • Removed unnecessary dependencies from Depends and Imports package fields.

sbo 0.3.2

• Patch addressing unexpected behaviour of erase argument in preprocess() and kgram_freqs_fast(), c.f. issue #17.

sbo 0.3.1

• Changed leading to trailing underscore in private variables definition of C++ kgramFreqs class, as per §1.6.4 of the "Writing R extensions" guide.
• Removed Catch tests infrastructure for C++ code.

sbo 0.3.0

• Added kgram_freqs_fast() for fast and memory efficient kgram tokenization using the default text preprocessing utility.

sbo 0.2.0

• The infrastructure of kgram_freqs(), get_word_freqs(), preprocess(), and predict.sbo_preds() has been entirely rewritten in C++.
• Added tokenize_sentences() function for sentence level tokenization.
• kgram_freqs() now accepts any user defined single character EOS token, through the EOS argument.

sbo 0.1.2

• Added preproc argument to kgram_freqs() and get_word_freqs(), for custom training corpus preprocessing.
• The dict argument of kgram_freqs() now also accepts numeric values, allowing to build a dictionary directly from the training corpus.

sbo 0.1.1

• Added predict method for sbo_kgram_freqs class.

sbo documentation built on Dec. 6, 2020, 1:06 a.m.