An exploration in using the pre-trained BERT model to perform Named Entity Recognition (NER) where labelled training data is limited but there is a considerable amount of unlabelled data. Two different regularisation terms using Kullback–Leibler (KL) divergence are proposed that aim to leverage the unlabelled data to help the model generalise to unseen data.
- NER Baseline: A simple approach to predict named entity labels through a MLP.
- BERT NER: Words are encoded using a static pre-trained BERT layer, a MLP on top predicts the named entity label.
- BERT NER Data Distribution KL: Same architecture as
BERT NERbut in the final epochs of training a KL term is introduced to encourage predicted labels for the unlabelled data to match the expected probability distribution of the data. - BERT NER Confidence KL: Same architecture as
BERT NERbut in the final epochs of training a KL term is introduced to encourage the model to have high confidence when predicting the unlabelled training examples.
- Motivation
- Solution Components
- Data
- Implementation Details
- Results
- Future Work
- Install
- Run
- Library Structure
- References
Identifying named entities in a sentence is a common task in NLP pipelines. There are an extensive set of datasets available online with marked entities such as famous person or geographic location but often more bespoke categories are needed for particular application. An example of this could be a chatbot application which may begin processing a message from a user by labelling problem specific relevant entities in their message. Often the process of labelling this data must be done laboriously internally especially in a PoC phase. This results in there being only a small amount of labelled data with the potential addition of some unlabelled data.
The aim behind this project is to design a solution to learn as much as possible from the small amount of labelled data without over-fitting as well as leveraging unlabelled data to improve generalisation to unseen data.
Learning representation for words on large corpora which are applicable across many NLP application has been an active area of research in the last decade. A large success came from Mikolov et al. [1] and Pennington et al. [2] in using deep neural networks to produce the pre-train word embeddings Word2Vec and GloVe respectively. These representation were hugely popular across the NLP field and gave a considerable boost in performance of the naive one-hot encoding approach especially when training data is limited.
More recently, there was a huge breakthrough in learned representation from Devlin et al. with the design of the BERT model [3]. The model comprises of 12 transformer layers and learns a representation for the context of a word in a given sentence. For different downstream tasks, minimal additional parameter are added and the whole model is fine tuned to the data. This differs from the use of the pre-trained word embeddings such as GloVe which are kept static during optimisation in a downstream task.
The pre-trained BERT model has achieved state-of-the-art performance on a number of NLP tasks and seems like the most appropriate architecture for the NER problem especially when data is limited. However, fine tuning hundreds of millions of parameters requires considerable computing power that was not available for this project. Lan et al. developed a lighter version of BERT called ALBERT by utilising factorisation of embedding parameters and cross-layer parameter sharing [4] but this was still considered to be too heavy. As a work around, the pre-trained BERT layers were kept fixed with an additional MLP added to fine tune the BERT output embeddings to solve the named entity recognition task. The precise architecture can be found in section 4.1.
The Kullback–Leibler (KL) divergence is a measure of how one probability distribution is different from another. Often in our data we can estimate a prior distribution for the categorical labels by observing our labelled data or from knowledge of an industry e.g. roughly knowing the percentage of credit transactions which are fraudulent. The proposed Data Distribution KL Regularizer is designed to leverage this prior knowledge in combination with the predictions on the unlabelled data to improve the generalisation of the model.
We aim for the distribution of our model assigned labels, ,
to replicate our prior distribution,
,
on the unlabelled data. We estimate
by:
where is the index of
the word in the flattened batch of 128 sentences. The KL loss is defined by:
This loss is optimised on batches of the unlabelled data on alternating steps with the optimisation for the cross entropy loss on the labelled data in the later stages of training.
The second KL regularizer explored was designed to reward a model that had high confidence in the predicted labels made on the unlabelled data. The model will use the unlabelled data to produce better representations of the words that are more generalisable. Xie et al. proposed the following prior to encourage confidence in unsupervised cluster assignment [5]:
where is the index of
the word in the flattened batch of 128 sentences. Xie devised calculated
from
distance metrics to cluster centroids whereas we continue to use the probabilities produced by
the softmax layer of the network. The KL loss is defined by:
This loss is optimised on batches of the unlabelled data on alternating steps with the optimisation for the cross entropy loss on the labelled data in the later stages of training.
The data consists of 48,000 sentences from newspaper articles with the following labelled named entities:
| Tag | Description | % in data |
|---|---|---|
| O | Other | 84.68% |
| geo | Geographical Entity | 4.3% |
| gpe | Geopolitical Entity | 1.53% |
| per | Person | 3.27% |
| org | Organization | 3.52% |
| tim | Time indicator | 2.56% |
| art | Artifact | 0.07% |
| nat | Natural Phenomenon | 0.02% |
| eve | Event | 0.05% |
The labelled training dataset consists of 2,560 random sentences, there are 9,600 sentences in the test set. All the data is labelled but to simulate a case where we have unlabelled data, we ignore the labels on the remaining 36,000 sentences. The data is placed into batches of size 128.
The 12 layer trained BERT model was downloaded from tf_hub at the following URL: https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12. The multi-layered perceptron on top has 3 layers of size 256, 128 and 64 with relu activation. There is then a dense layer of size 32 with no activation which represents the latent space to visualise the data clusters that are forming. There is then a final dense layer with a softmax activation to assign the probabilities of the labels.
The BERT NER is trained over 20 epochs on the training dataset (400 batches); after this point the model began to over-fit. For the models with KL divergence, the BERT NER is then fine tuned by doing a gradient descent on the KL divergence using an unlabelled batch followed by a gradient descent on the cross entropy loss on a labelled batch. This process is repeated for 40 labelled and unlabelled batches for the data distribution KL model and 80 labelled and unlabelled batches for the confidence KL model respectively.
The tokenisation of the sentences is done using the tokenizer from the bert-for-tf2 library.
All words are lower cased.
Sentences are padded such that they all have a length of 50. These padded token are ignored when optimising both cross entropy loses and KL divergence terms. Very small epsilons are added to denominators of calculations on the probability because the probability on labels of masked tokens are set to zero.
A baseline model that uses an embedding layer of size 784 instead of the BERT layer was also trained for comparison. The MLP layers in this model are the same as described in Section 4.1. This is a simple baseline model where each word is seen to be independent in a sentence.
| Metric | Description |
|---|---|
| Validation Accuracy | Overall accuracy on the full test set |
| Validation Accuracy no Other | Overall accuracy on the full test set when words with ground truth tag of O are removed |
| Validation Mean F1 | The mean F1 score across all categories |
| Model Name | Validation Accuracy | Validation Accuracy no Other | Validation Mean F1 |
|---|---|---|---|
| NER Baseline | 93.65% | 66.73% | 0.5457 |
| BERT NER | 95.90% | 78.41% | 0.6099 |
| BERT NER Data Distribution KL | 94.44% | 83.50% | 0.6320 |
| BERT NER Confidence KL | 96.05% | 80.82% | 0.6514 |
The addition of the KL optimisation steps has improved the overall performance of the model when the hugely dominant O category is removed from consideration and when assessing all categories equally with a mean F1. The overall accuracy in the model using the data distribution KL is lower than the BERT NER but it has a significant boost in the accuracy of the other categories; this is because the KL term is encouraging the model to categorise less words as O.
Further results files can be found in the results folder of the repo.
The images below are 10 batches of test sentences encoded using the model to a dimension of 32 (the layer before the dense softmax layer). These representations are reduced to 2D using PCA such that they can be visualised.
| Model | True Labels | Predicted Labels |
|---|---|---|
| NER Baseline |  |
 |
| BERT NER |  |
 |
| BERT NER Data Distribution KL |  |
 |
| BERT NER Confidence KL |  |
 |
These provide thoughtful insight into how the KL optimisation steps are affecting the representations the model is learning. The representation for the NER baseline is very compact and the large gaps within the latent space highlight that it is not learning a very strong representation. This is because the model is encoding each work independantly and so there is no information encoded about the overall sentence.
The model with the data distribution KL can be seen to be classifying more samples into the tim class opposed to O to match the prior distribution. To improve the precision of this approach, it has slightly shifted away the centroid of the tim class further from O. Furthermore, the addition of the unlabelled seems to have results in much better clusters forming for the geo, gpe and per categories.
The model with the confidence KL has a very similar encoding space to the BERT NER model but to increase the confidence in predictions, it has dispersed the cluster centroids. This has resulted in encodings that are more spread over the latent space.
- Compare with bias labelled data: The labelled data was just a random sample of the set of training data available. It would be interesting to compare the approaches if the labelled data was sampled in some bias way e.g. picking sentences with less O tags and seeing if the KL divergance term would better leverage the unlabelled data.
- Use both KL terms: Experiment using both KL regularizers at the same time.
- Different optimisation techniques: Explored different ways to optimise the network to use the KL term more effectively.
- Compared traditional regularizers: Compared how effective these KL terms were in comparison to more traditiomal methods such as L2 norm or dropout.
Create Environment: conda create --name nerbert python=3.6
Activate Environment: conda activate nerbert
Make Install Executable: chmod +x install.sh
Install Requirements: ./install.sh
Save All Models To saved_models Directory From: https://drive.google.com/drive/folders/1HgHJtuW1fOuO8bWxSAxTZZQL48FW-rRI?usp=sharing
Train: python -m examples.example_train_<model_name>
Evaluate: python -m examples.example_evaluate_model
- config: Yaml files with configurations for models and input data
- data: CSV file of raw data and preprocessed
- examples: Examples of how to effectively use the library
- ner: Library
- models: Tensorflow model code, instances of tf.keras.Model
- trainers: Wrappers around the models to train them and save the resulting weights
- kl_loss: Functions to calculate the proposed KL losses
- model_evaluator: Functions to evaluate the performance of the model and produce visualisations
- preprocessor: Preprocesses training and test data
- utils: Extra utility functions
- readme_equations: Images of equations for readme
- results: Saved results and visualisations from experiments
- saved_models: Saved model weights
[1] Tomas Mikolov and Kai Chen and Greg Corrado and Jeffrey Dean. Efficient Estimation of Word Representations in Vector Space. 2013, https://arxiv.org/pdf/1301.3781.pdf
[2] Jeffrey Pennington and Richard Socher and Christopher D. Manning. Glove: Global vectors for word representation. 2014, https://nlp.stanford.edu/pubs/glove.pdf
[3] Jacob Devlin and Ming-Wei Chang and Kenton Lee and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 2018, https://arxiv.org/pdf/1810.04805.pdf
[4] Zhenzhong Lan and Mingda Chen and Sebastian Goodman and Kevin Gimpel and Piyush Sharma and Radu Soricut. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. 2019, https://arxiv.org/pdf/1909.11942.pdf
[5] Junyuan Xie and Ross Girshick and Ali Farhadi. Unsupervised Deep Embedding for Clustering Analysis. 2015, https://arxiv.org/pdf/1511.06335.pdf