News

• Tutorial proposal for GluonNLP is accepted at EMNLP 2019, Hong Kong, and KDD 2019, Anchorage.

Models and Scripts

• BERT pre-training on BooksCorpus and English Wikipedia with mixed precision and gradient accumulation on GPUs. We achieved the following fine-tuning results based on the produced checkpoint on validation sets(#482, #505, #489). Thank you @haven-jeon

  • Dataset	MRPC	SQuAD 1.1	SST-2	MNLI-mm
    Score	87.99%	80.99/88.60	93%	83.6%

• BERT fine-tuning on various sentence classification datasets with checkpoints converted from the official repository(#600, #571, #481). Thank you @kenjewu @haven-jeon

  • Dataset	MRPC	RTE	SST-2	MNLI-m/mm
    Score	88.7%	70.8%	93%	84.55%, 84.66%

• BERT fine-tuning on question answering datasets with checkpoints converted from the official repository(#493). Thank you @fiercex

  • Dataset	SQuAD 1.1	SQuAD 1.1	SQuAD 2.0
    Model	bert_12_768_12	bert_24_1024_16	bert_24_1024_16
    F1/EM	88.53/80.98	90.97/84.05	77.96/81.02

• BERT model convertion scripts for checkpoints from the original tensorflow repository, and more converted models(#456, #461, #449). Thank you @fiercex:

  • Multilingual Wikipedia (cased, BERT Base)
  • Chinese Wikipedia (cased, BERT Base)
  • Books Corpus & English Wikipedia (uncased, BERT Large)

• Scripts and command line interface for BERT embedding of raw sentences(#587, #618). Thank you @imgarylai

• Scripts for exporting BERT model for deployment (#624)

New Features

• [API] Add BERTVocab (#509) thanks @kenjewu
• [API] Add Transforms for BERT (#526) thanks @kenjewu
• [API] add data parallel for transformer (#387)
• [FEATURE] Add squad2.0 Dataset (#551) thanks @fiercex
• [FEATURE] Add NumpyDataset (#498)
• [FEATURE] Add TruncNorm initializer for BERT (#548) thanks @Ishitori
• [FEATURE] Add split sampler for distributed training (#494)
• [FEATURE] Custom metric for masked accuracy (#503)
• [FEATURE] Support custom sampler in SimpleDatasetStream (#507)
• [FEATURE] clip gradient norm by parameter (#470)

Bug Fixes

• [BUGFIX] Fix Data Preprocessing for Translation Data (#568)
• [FIX] fix parameter clip (#527)
• [FIX] Fix divergence of the training of transformer (#543)
• [FIX] Fix documentation and a bug in NCE Block (#558)
• [FIX] Fix hashing single ngrams in NGramHashes (#450)
• [FIX] Fix weight dying in BERTModel.decoder for BERT pre-training (#500)
• [BUGFIX] Modifying the FastText Classification training for accurate mean pooling (#529) thanks @sravanbabuiitm

API Changes

• [API] BERT return intermediate encodings per layer (#606) thanks @Ishitori
• [API] Better handle case when backoff is not possible in TokenEmbedding (#459)
• [FIX] Rename wiki_cn/wiki_multilingual to wiki_cn_cased/wiki_multilingual_uncased (#594) thanks @kenjewu
• [FIX] Update default value of BERTAdam epsilon to 1e-6 (#601)
• [FIX] Fix BERT decoder API for masked language model prediction (#501)
• [FIX] Remove bias correction term in BERTAdam (#499)

Enhancements

• [BUGFIX] use glove.840B.300d for NLI experiments (#567)
• [API] Add debug option for parallel (#584)
• [FEATURE] Skip dropout layer in Transformer when rate=0 (#597) thanks @TaoLv
• [FEATURE] update sharded loader (#468)
• [FIX] Update BERTLayerNorm Implementation (#485)
• [TUTORIAL] Use FixedBucketSampler in BERT tutorial for better performance (#506) thanks @Ishitori
• [API] Add Bert tokenizer to transforms.py (#464) thanks @fiercex
• [FEATURE] Add data parallel to big rnn lm script (#564)

Minor Fixes

• Various documentation fixes: #484, #613, #614, #438, #448, #550, #563, #611, #605, #440, #554, #445, #556, #603, #483, #576, #610, #547, #458, #574, #510, #447, #465, #436, #622, #583 thanks @anuragsarkar97 @brettkoonce
• [FIX] fix repeated unzipping in squad dataset (#553)
• [FIX] web fixes (#453)
• [FIX] Remove unused argument in fasttext_word_ngram.py (#486) thanks @kurtjanssensai
• [FIX] Remove unused code (#528)
• [FIX] Remove unused code in text_classification script (#442)
• [MISC] Bump up version (#454)
• [BUGFIX] fix pylint error (#549)
• [FIX] Simplify the data preprocessing code for the sentiment analysis script (#462)
• [FEATURE] BERT doc fixes and script usability enhancements (#444)
• [FIX] Fix Py2 compatibility of machine_translation/dataprocessor.py (#541) thanks @ymjiang
• [BUGFIX] Fix GluonNLP MXNet dependency (#555)
• [BUGFIX] Fix Weight Drop and Test (#546)
• [CI] Add version upper bound to doc.yml (#467)
• [CI] speed up tests (#582)
• [CI] upgrade mxnet to 1.4.0 (#617)
• [FIX] Revert an unintended change (#525)
• [BUGFIX] update paths and imports in bert scripts (#634)

Highlights

• Featured in AWS re:invent 2018

Models

• BERT

  • The Bidirectional Encoder Representations from Transformers model as introduced by Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." arXiv preprint arXiv:1810.04805 (2018) (#409). Model parameters are converted from the original model checkpoints from Google research, including:
    • BERT BASE model trained on
      • Book Corpus & English Wikipedia (cased)
      • Book Corpus & English Wikipedia (uncased)
      • multilingual Wikipedia (uncased)
    • BERT LARGE model trained on Book Corpus & English Wikipedia (uncased)

• ELMo

  • The Embeddings from Language Models as introduced by Peters, Matthew E., et al. "Deep contextualized word representations." arXiv preprint arXiv:1802.05365 (2018) (#227, #428). Model parameters are converted from the original model checkpoints in AllenNLP, including the small, medium, original models trained on 1 billion words dataset, and the original model trained on 5.5B tokens consisting of Wikipedia & monolingual news crawl data from WMT 2008-2012.

• Word Embedding

  • The GloVe model as introduced by Pennington, Jeffrey, Richard Socher, and Christopher Manning. "Glove: Global vectors for word representation." Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). 2014 (#359).

• Natural Language Inference 

  • The Decomposable Attention Model as introduced by Parikh, Ankur P., et al. "A decomposable attention model for natural language inference." arXiv preprint arXiv:1606.01933 (2016). (#404). On the SNLI test set, it achieves 84.6% accuracy (without intra-sentence attention) and 84.4% accuracy (with intra-sentence attention). Thank you @linmx0130 @hhexiy!

• Dependency Parsing

  • The Deep Biaffine Attention Dependency Parser as introduced by Dozat, Timothy, and Christopher D. Manning. "Deep biaffine attention for neural dependency parsing." arXiv preprint arXiv:1611.01734 (2016). (#408). It achieved 96% UAS on the Penn Treebank dataset. Thank you @hankcs!

• Text Classification

  • The Text CNN model as introduced by Kim, Yoon. "Convolutional neural networks for sentence classification." arXiv preprint arXiv:1408.5882 (2014). (#391) Thank you @xiaotinghe!

New Tutorials

• ELMo
  • A tutorial on generating contextualized representation with the pre-trained ELMo model, as introduced by Peters, Matthew E., et al. "Deep contextualized word representations." arXiv preprint arXiv:1802.05365 (2018) (#227, #428).
• BERT
  • A tutorial on fine-tuning the BERT model for sentence pair classification, as introduced by Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." arXiv preprint arXiv:1810.04805 (2018) (#437)

New Datasets

• Sentiment Analysis
  • MR, a movie-review data set of 10,662 sentences labeled with respect to their overall sentiment polarity (positive or negative). (#391)
  • SST_1, an extension of the MR data set with fine-grained labels (#391)
  • SST_2, an extension of the MR data set with binary sentiment polarity labels (#391)
  • SUBJ, a subjectivity data set for sentiment analysis (#391)
  • TREC, a movie-review data set of 10,000 sentences labeled with respect to their subjectivity status (subjective or objective). (#391)

API Updates

• Changed Vocab constructor from staticmethod to classmethod to handle inheritance (#386)
• Added Transformer Encoder APIs (#409)
• Added pre-trained ELMo model to model.get_model API (#227)
• Added pre-trained BERT model to model.get_model API (#409)
• Added unknown_lookup setter to TokenEmbedding (#429)
• Added dtype support to EmbeddingCenterContextBatchify (#416)
• Propagated exceptions from PrefetchingStream (#406)
• Added sentencepiece tokenizer detokenizer (#380)
• Added CSR format for variable length data in embedding training (#384)

Fixes & Small Changes

• Included output of nlp.embedding.list_sources() in API docs (#421)
• Supported symlinks in examples and scripts (#403)
• Fixed weight tying in GNMT and Transformer (#413)
• Simplified transformer notebook (#400)
• Fixed LazyTransformDataStream prefetching (#397)
• Adopted src/gluonnlp folder layout (#390)
• Fixed text8 archive file name for downloads from S3 (#388) Thanks @bkktimber!
• Fixed ppl reporting for training on multi gpu in the language model notebook (#365). Thanks @ThomasDelteil!
• Fixed a spelling mistake in QA script. (#379) Thanks @qyhfbqz!

Highlights

• Hands-on Tutorial at KDD 2018 (website) (code)

Models

• Language Model
  • The Large Scale Word Language Model as introduced by Jozefowicz, Rafal, et al. “Exploring the limits of language modeling”. arXiv preprint arXiv:1602.02410 (2016) achieved test PPL 43.62 on GBW dataset (#179 #270 #277 #278 #286 #294)
  • The NT-ASGD based Language Model as introduced by Merity, S., et al. “Regularizing and optimizing LSTM language models”. ICLR 2018 achieved test PPL 65.62 on WikiText-2 dataset (#170)
• Document Classification
  • The Classification Model as introduced by Joulin, Armand, et al. “Bag of tricks for efficient text classification” achieved validation accuracy validation accuracy 98 on Yelp review dataset (#258 #297)
• Question Answering
  • The QANet as introduced by Jozefowicz, Rafal, et al. “ QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension”. ICLR 2018 achieved F1 score 79.5 on SQuAD 1.1 dataset (#339) (coming soon to master branch)

New Tutorials

• Machine Translation
  • The Google NMT as introduced by Wu, Yonghui, et al. “Google's neural machine translation system: Bridging the gap between human and machine translation”. arXiv preprint arXiv:1609.08144 (2016) is introduced as part of the gluonnlp tutorial (#261)
  • The Transformer based Machine Translation by Vaswani, Ashish, et al. “Attention is all you need.” Advances in Neural Information Processing Systems. 2017 is introduced as part of the gluonnlp tutorial (#279)
• Sentence Embedding
  • A Structured Self-attentive Sentence Embedding (#366) by Z. Lin, M. Feng, C. Santos, M. Yu, B. Xiang, B. Zhou, Y. Bengio, "A Structured Self-attentive Sentence Embedding" ICLR 2017 is introduced in gluonnlp tutorial (#366)

New Datasets

• Word Embedding
  • Wikipedia (#218)
  • Fil9 dataset(#363)
  • FastText crawl-300d-2M-subword(#336), wiki-news-300d-1M-subword(#368), cc.en.300(#373)

API updates

• Added dataloader that allows multi-shard sampling (#237 #280 #285)
• Simplified DataStream, added DatasetStream, refactored and extended PrefetchingStream (#235)
• Unified BPTT batchify for dataset and stream (#246)
• Added symbolic beam search (#233)
• Added SequenceSampler (#272)
• Refactored Transform APIs (#282)
• Reorganized index of the repo and model zoo page (#357)

Fixes & Small Changes

• Fixed module name in batchify.py example (#239)
• Improved imports structure (#248)
• Added test for nmt scripts (#234)
• Speeded up batchify.Pad (#249)
• Fixed LanguageModelDataset.bptt_batchify (#243)
• Fixed weight drop and add tests (#268)
• Fixed relative links that pypi doesn't handle (#293)
• Updated notebook build logic (#309)
• Added community link (#313)
• Enabled run tests in parallel (#317)
• Enabled word embedding scripts tests (#321)

See all commits

GluonNLP v0.3 contains many exciting new features. (depends on MXNet 1.3.0b20180725)

Models

• Language Models
  • The Cache Language Model as introduced by Grave, E., et al. “Improving neural language models with a continuous cache”. ICLR 2017 is introduced as part of gluonnlp.model.train (#110)
  • The Activation Regularizer and Temporal Activation Regularizer as introduced by Merity, S., et al. "Regularizing and optimizing LSTM language models". ICLR 2018 is introduced as part of gluonnlp.loss (#110)
• Machine Translation
  • The Transformer Model as introduced by Vaswani, Ashish, et al. "Attention is all you need." Advances in Neural Information Processing Systems. 2017* is introduced as part of the gluonnlp nmt scripts (#133)
• Word embeddings
  • Trainable word embedding models are introduced as part of gluonnlp.model.train (#136)
    • Word2Vec by Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems (pp. 3111-3119).
    • FastText models by Bojanowski, P., Grave, E., Joulin, A., & Mikolov, T. (2017). Enriching Word Vectors with Subword Information. Transactions of the Association for Computational Linguistics, 5, 135-146.

New Datasets

• Machine Translation
  • WMT2014BPE (#135) (#177) (#180)
• Question Answering
  • Stanford Question Answering Dataset (SQuAD) Rajpurkar, P., Zhang, J., Lopyrev, K., & Liang, P. (2016). SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (pp. 2383-2392). (#113)
• Word Embeddings
  • Text8 (#136)

API changes

• The download directory for datasets and other artifacts can now be specified via the MXNET_HOME environment variable. (#106)
• TokenEmbedding class now exposes the Inverse Vocab as well (#123)
• SortedSampler now supports use_average_length option (#135)
• Add more strategies for bucket creation (#145)
• Add tokenizer to bleu (#154)
• Add Convolutional Encoder and Highway Layer (#129) (#186)
• Add plain text of translation data. (#158)
• Use Sherlock Holmes dataset instead of PTB for language model notebook (#174)
• Add classes JiebaToknizer and NLTKStanfordSegmenter for Chinese Word Segmentation (#164)
• Allow toggling output and prompt in documentation website (#184)
• Add shape assertion statements for better user experience to some attention cells (#201)
• Add support for computation of word embeddings for unknown words in TokenEmbedding class (#185)
• Distribute subword vectors for pretrained fastText embeddings enabling embeddings for unknown words (#185)

Fixes & Small Changes

• fixed bptt_batchify sometimes returned an invalid last batch (#120)
• Fixed wrong PPL calculation in word language model script for multi-GPU (#150)
• Fix split compound words and wmt16 results (#151)
• Adapt pretrained word embeddings example notebook for nd.topk change in mxnet 1.3 (#153)
• Fix beam search script (#175)
• Fix small bugs in parser (#183)
• TokenEmbedding: Skip lines with invalid bytes instead of crashing (#188)
• Fix overly large memory use in TokenEmbedding serialization/deserialization if some tokens are overly large (eg. 50k characters) (#187)
• Remove duplicates in WordSim353 when combining segments (#192)

See all commits

Features

GluonNLP provides its users with easy access to

• State of the art models
• Pre-trained word embeddings
• Many public datasets for different tasks
• Examples friendly to users that are new to the task
• Reproducible training scripts

Models

Gluon NLP Toolkit supplies model definitions for common NLP tasks. These can be adapted for the users requirements or taken as blueprint for new developments. All of these are implemented using Gluon Blocks allowing easy reuse as plug-and-play neural network building blocks.

• Language Models
  • Standard RNN language model
  • AWD language model by salesforce
• Attention Cells
  • Multi-head attention cell
  • MLP attention cell
  • Dot-product attention cell
• Beam Search
  • Beam Search Sampler
  • Beam Search Scorer

Data

Gluon NLP Toolkit provides tools for building efficient data pipelines for NLP tasks by defining a Dataset class interface and utilities for transforming them. Several datasets are included by default and will be automatically downloaded when used.

• Language modeling with WikiText
  • WikiText is a popular language modeling dataset from Salesforce. It is a collection of over 100 million tokens extracted from the set of verified Good and Featured articles on Wikipedia.
• Sentiment Analysis with IMDB
  • IMDB: IMDB is a popular dataset for binary sentiment classification. It provides a set of 25,000 highly polar movie reviews for training, 25,000 for testing, and additional unlabeled data.
• CoNLL datasets
  • These datasets include data for the shared tasks, such as part-of-speech (POS) tagging, chunking, named entity recognition (NER), semantic role labeling (SRL), etc.
  • We provide built in support for CoNLL 2000 – 2002, 2004, as well as the Universal Dependencies dataset which is used in the 2017 and 2018 competitions.
• Word embedding evaluation datasets
  • There are a number of commonly used datasets for intrinsic evaluation for word embeddings. We provide commonly used datasets for the similarity and analogy evaluation tasks.

Gluon NLP further ships with common datasets data transformation functions, dataset samplers to determine how to iterate through datasets as well as functions to generate data batches.

A complete and up-to-date list of supplied datasets and utilities is available in the API documentation.

Other features

• Vocabulary
• Pre-trained embeddings

Examples and scripts

The Gluon NLP toolkit also provides scripts that use the functionality of the toolkit for various tasks

• Word Embedding Evaluation
• Beam Search Generator
• Word language modeling
• Sentiment Analysis through Fine-tuning, w/ Bucketing
• Machine Translation