NEWS.md

In text: Analyses of Text using Transformers Models from HuggingFace, Natural Language Processing and Machine Learning

text (development version)

text 1.0

Function

• multinomial regression in textTrainRegression()
• textPredictTest() can handle auc
• textEmbed() is faster (thanks to faster handling of aggregating layers)
• Added sort parameter in textEmbedRawLayers().

Bug/unexpected behaviour

• Tests using training with random forest was updated since outcomes changed when updating from R 4.2 to R 4.3.1. (see test_2_textTrain.R in tests/testthat folder)

text 0.9.99.9

Function

Possibility to use GPU for MacOS M1 and M2 chip using device = "mps" in textEmbed()

text 0.9.99.8

Function

textFineTune() as an experimental function is implemented max_length implemented in textTranslate()

text 0.9.99.7

Function

• textEmbedReduce() implemented

text 0.9.99.6

Bug Fix

• Fixing textEmbed error for many non-BERT models.

text 0.9.99.5

Bug Fix

• fixed textEmbed(decontextualize=TRUE), which gave error.

text 0.9.99.3

Functions changes

• Removing textSimialirtyTest() for version 1.0 because it needs more evaluations.

Bug Fix

• changed hard coded "bert-base-uncased" to model, so that layers = -2 works in textEmbed().
• Update logging level critical using integer 50 with set_verbosity.
• changed in sorting_xs_and_x_append from Dim to Dim0 when renaming x_appended variables.
• changed first to append_first and made it an option in textTrainRegression() and textTrainRandomForest().

text 0.9.99.2

DEFAULT CHANGES

• The default setting of textEmbed() is now providing token-level embeddings and text-level embeddings. Word_type embeddings are optional.
• In textEmbed() layers = 11:12 is now second_to_last.
• In textEmbedRawLayers default is now second_to_last.
• In textEmbedLayerAggregation() layers = 11:12 is now layers = "all".
• In textEmbed() and textEmbedRawLayers() x is now called texts.
• textEmbedLayerAggregation() now uses layers = "all", aggregation_from_layers_to_tokens, aggregation_from_tokens_to_texts.

New Function

• textZeroShot() is implemented.
• textDistanceNorm() and textDistanceMatrix()
• textDistance() can compute cosine distance.
• textModelLayers() provides N layers for a given model

New Setting

max_token_to_sentence in textEmbed()

Setting name changes

• aggregate_layers is now called aggregation_from_layers_to_tokens.
• aggregate_tokens is now called aggregation_from_tokens_to_texts. single_word_embeddings is now called word_types_embeddings

Function name changes

• textEmbedLayersOutput() is now called textEmbedRawLayers()

text 0.9.98

• adding textDimName()
• DEFAULT CHANGE in textEmbed(): dim_name = TRUE
• DEFAULT CHANGE in textEmbed(): single_context_embeddings = TRUE
• DEFAULT CHANGE in textEmbed(): device = "gpu"
• Adding specific layer aggregations for explore_words in textPlot()
• Adding x_append_target in textPredict() function

text 0.9.97

• updating textClassify(), textGeneration(), textNER(), textSum(), textQA(), and textTranslate().

text 0.9.96

New features

• harmonizing x_add to x_append across functions
• adding set_seed to language analysis tasks

Code changes

• abstracting function for sorting out x' in training and prediction
• textPredict does not take word_embeddings and x_append (not new_data)

text 0.9.95

New features

• textClassify() (under development)
• textGeneration() (under development)
• textNER() (under development)
• textSum() (under development)
• textQA() (under development)
• textTranslate() (under development)

text 0.9.93

New features

• New function: textSentiment(), from huggingface transformers models.
• add progression for time consuming functions including textEmbed(), textTrainRegression(), textTrainRandomForest() and textProjection().

text 0.9.92

New features

• Option dim_names to set unique dimension names in textEmbed() and textEmbedStatic().
• textPreictAll() function that can take several models, word embeddings, and variables as input to provide multiple outputs.
• option to add variables to the embeddings in textTrain() functions with x_append.

text 0.9.91

New features

• text version is printed from DESCRIPTION-file (rather than manual update)
• textPredict related functions are located in its own file
• textEmbed comment includes text_version number
• textEmbedLayersOutput and textEmbed can provide single_context_embeddings

Fixes

• Removed return_tokens option from textEmbed (since it is only relevant for textEmbedLayersOutput)
• removed the empty list $single_we when decontexts is FALSE.

text 0.9.90

New features

• Visualization of the download process of language models
• Can set error level from python
• Logistic regression is default for classification in textTrain.
• Megatron language model functionality

Bug Fixes

• When GPU is not found, CPU is used.

text 0.9.80

New features

• Option to set model_max_length in textEmbed().
• textModels() show downloaded models.
• textModelsRemove() deletes specified models.

Bug Fixes

• Fixed error for unpaired textSimilarityTest() when uneven number of cases are tested.

text 0.9.70

New Features

• Inclusion of textDistance() function with distance measures.
• Adding more measures to textSimilarity().
• Adding functionality from textSimilarity() in textSimilarityTest(), textProjection() and textCentrality() for plotting.
• Adding information about how textTrainRegression() concatenates word embeddings when provided with a list of several word embeddings.
• Adding two word embedding dimensions to example data of single word embeddings to match the 10 of the contextualized embeddings in word_embeddings_4$singlewords_we.

Bug Fixes

• In textCentrality(), words to be plotted are selected with word_data1_all$extremes_all_x >= 1 (rather than ==1).

text 0.9.60

• textSimilarityMatrix() computes semantic similarity among all combinations in a given word embedding.

text 0.9.54

• textDescriptives() gets options to remove NA and compute total scores.

text 0.9.53

• inclusion of textDescriptives()

text 0.9.52

• prompt option added to textrpp_initiate()

text 0.9.51

• tokenization is made with NLTK from python.

text 0.9.50

• Code has been cleaned up and prepared for CRAN

text 0.9.20

New Features

• New functions being tested: textWordPredictions() (which has a trial period/not fully developed and might be removed in future versions); p-values are not yet implemented.
• Possibility to use textPlot() for objects from both textProjection() and textWordPredictions()

Minor changes

• Changed wordembeddigs to word_embeddings through out the code/package.

Bug Fixes

• Warnings about seed when using multi-cores on Mac is addressed.

text 0.9.17

New Features

• textrpp_initiate() runs automatically in library(text) when default environment exits
• Python warnings a captured in embedding comments
• Option to print python options to console
• Updated the permutation test for plotting and textSimilarityTest().

Minor changes

• Changed from stringr to stringi (and removed tokenizer) as imported package

text 0.9.16

New Features

• textrpp_install() installs a conda environment with text required python packages.
• textrpp_install_virtualenv() install a virtual environment with text required python packages.
• textrpp_initialize() initializes installed environment.
• textrpp_uninstall() uninstalls conda environment.

text 0.9.13

New Features

• textEmbed() and textEmbedLayersOutput() support the use of GPU using the device setting.
• remove_words makes it possible to remove specific words from textProjectionPlot()

text 0.9.12

New Features

• In textProjetion() and textProjetionPlot() it now possible to add points of the aggregated word embeddings in the plot
• In textProjetion() it now possible to manually add words to the plot in order to explore them in the word embedding space.
• In textProjetion() it is possible to add color or remove words that are more frequent on the opposite "side" of its dot product projection.
• In textProjection() with split == quartile, the comparison distribution is now based on the quartile data (rather than the data for mean)

Bug Fixes

• If any of the tokens to remove is "[CLS]", subtract 1 on token_id so that it works with layer_aggregation_helper. 0.9.11
• Can now submit one word to textEmbed() with decontexts=TRUE.

text 0.9.11

• textSimilarityTest() is not giving error when using method = unpaired, with unequal number of participants in each group.

New Features

• textPredictTest() function to significance test correlations of different models. 0.9.11

Bug Fixes

• If any of the tokens to remove is "[CLS]", subtract 1 on token_id so that it works with layer_aggregation_helper. 0.9.11

text 0.9.10

This version is now on CRAN.

New Features

• Adding option to deselect the step_centre and step_scale in training.
• Cross-validation method in textTrainRegression() and textTrainRandomForrest() have two options cv_folds and validation_split. (0.9.02)
• Better handling of NA in step_naomit in training.
• DistilBert model works (0.9.03)

Major changes

Bug Fixes

• textProjectionPlot() plots words extreme in more than just one feature (i.e., words are now plotted that satisfy, for example, both plot_n_word_extreme and plot_n_word_frequency). (0.9.01)
• textTrainRegression() and textTrainRandomForest() also have function that select the max evaluation measure results (before only minimum was selected all the time, which, e.g., was correct for rmse but not for r) (0.9.02)
• removed id_nr in training and predict by using workflows (0.9.02).

Minor changes

text documentation built on Aug. 9, 2023, 5:08 p.m.