Building a Multi-label Text Classifier using BERT and TensorFlow
source link: https://www.tuicool.com/articles/hit/VFjaeyj
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.
In a multi-label classification problem, the training set is composed of instances each can be assigned with multiple categories represented as a set of target labels and the task is to predict the label set of test data. e.g.,
- A text might be about any of religion, politics, finance or education at the same time or none of these.
- A movie can be categorized into action, comedy and romance genre based on its summary content. There is possibility that a movie falls into multiple genres like romcoms [romance & comedy].
How is it different from multi-class classification problem?
In Multi-class classification each sample is assigned to one and only one label: a fruit can be either an apple or a pear but not both at the same time. Let us consider an example of three classes C= ["Sun, "Moon, Cloud”]. In multi-class each sample can belong to only one of C classes. In multi-label case each sample can belong to one or more than one class.
For our discussion we will use Kaggle’s Toxic Comment Classification Challenge dataset consisting of a large number of Wikipedia comments which have been labeled by human raters for toxic behavior. The types of toxicity are:
toxic, severe_toxic, obscene, threat, insult, identity_hate
“Hi! I am back again! Last warning! Stop undoing my edits or die!”
is labelled as [1,0[1,0,0,1,0,0]aning it is both
Please refer here for detailed EDA of the dataset.
Let us briefly discuss BERT
In Oct 2018, Google released a new language representation model called BERT , which stands for Bidirectional Encoder Representations from Transformers. BERT builds upon recent work in pre-training contextual representations — including Semi-supervised Sequence Learning , Generative Pre-Training , ELMo , and ULMFit . However, unlike these previous models, BERT is the first deeply bidirectional , unsupervised language representation, pre-trained using only a plain text corpus ( Wikipedia ).
Pre-trained representations can either be context-free or contextual
- Context-free models such as word2vec or GloVe generate a single word embedding representation for each word in the vocabulary. For example, the word “ bank ” would have the same context-free representation in “ bank account ” and “ bank of the river. ”
- Contextual models instead generate a representation of each word that is based on the other words in the sentence. Contextual representations can further be unidirectional or bidirectional . For example, in the sentence “ I accessed the bank account ,” a unidirectional contextual model would represent “ bank ” based on “ I accessed the ” but not “ account .” However, BERT represents “ bank ” using both its previous and next context — “ I accessed the … account ” — starting from the very bottom of a deep neural network, making it deeply bidirectional.
Bidirectional LSTM based language models train a standard left-to-right language model and also train a right-to-left (reverse) language model that predicts previous words from subsequent words like in ELMO. In ELMo, there is a single LSTM for the forward language model and backward language model each. The crucial difference is that neither LSTM takes both the previous and subsequent tokens into account at the same time.
Why BERT is superior to other Bidirectional models?
Intuitively, a deep bidirectional model is strictly more powerful than either a left-to-right model or the concatenation of a left-to-right and right-to left model. Unfortunately, standard conditional language models can only be trained left-to-right or right-to-left, since bidirectional conditioning would allow each word to indirectly “see itself” in a multi-layered context.
To solve this problem, BERT uses “MASKING” technique to mask out some of the words in the input and then condition each word bidirectionally to predict the masked words. For example:
BERT also learns to model relationships between sentences by pre-training on a very simple task that can be generated from any text corpus: Given two sentences A and B, is B the actual next sentence that comes after A in the corpus, or just a random sentence? For example:
This is just a very basic overview of what BERT is. For details please refer to the original paper and some references,d .>
Good News:Google has uploaded BERT to TensorFlow Hub which means we can directly use the pre-trained models for our NLP problems be it text classification or sentence similarity etc.
The example of predicting movie review , a binary classification problem is provided as an example code in the repository. In this article, we will focus on application of BERT to the problem of multi-label text classification . So we will be basically modifying the example code and applying changes necessary to make it work for multi-label scenario.
Install the BERT using
!pip install bert-tensorflow
Downloading pre-trained BERT models: These are the weights and other necessary files to represent the information BERT learned in pre-training. You’ll need to pick which BERT pre-trained weights you want. There are two ways to download and use the pre-trained BERT model:
- Directly using from the tensorflow-hub:
Following pre-trained models are available to choose from.
BERT-Base, Uncased BERT-Large, Uncased BERT-Base, Cased BERT-Large, Cased BERT-Base, Multilingual Case BERT-Base, Chinese
We will use basic model: ‘uncased_L-12_H-768_A-12’
BERT_MODEL_HUB = “https://tfhub.dev/google/bert_uncased_L-12_H-768_A-12/1"
2. Manually Download the BERT model files : Download and save into a directory and unzip it. Here are links to the files for English:
BERT-Base, Uncased, BERT-Base, Cased, BERT-Large, Cased, BERT-Large, Uncased
You can use either way, but let us see what are the files actually in the pre-trained models. When I download
BERT-Base, Uncased, these are 3 important files as follows:
BERT_INIT_CHKPNT = ‘uncased-l12-h768-a12/bert_model.ckpt’
BERT_CONFIG = ‘uncased-l12-h768-a12/bert_config.json’
BERT_VOCAB : Contains model vocabulary [ wo[ words to indexes mapping]
BERT_INIT_CHKPNT : Contains weights of the pre-trained model
BERT_CONFIG : Contains BERT model architecture.
Tokenization involves breaking up of input text into its individual words. In order to do so, the first step is to create the tokenizer object. Two ways we can do that:
- Directly from the tensorflow-hub
2. From manually downloaded files:
BERT_INIT_CHKPNT & BERT_VOCAB files
After you have created the tokenizer, it is time to use it. Let us tokenize sentence:
“This here’s an example of using the BERT tokenizer”
Size of the vocabulary: ~30K
Let us first read the data set provided :
We need to convert our data into a format that BERT understands. Some utility functions are provided to do that.
create_examples() , reads data-frame and loads input text and corresponding target labels into
Using tokenizer, we’ll call
convert_examples_to_features method on our examples to convert them into features BERT understands. This method adds the special “CLS” and “SEP” tokens used by BERT to identify sentence start and end. It also appends “index” and “segment” tokens to each input. So all the job of formatting input as per the BERT is done by this function.
Here we use the pre-trained BERT model and fine-tune it for our classification task. Basically we load the pre-trained model and then train the last layer for classification task.
In multi-label classification instead of softmax(), we use sigmoid() to get the probabilities.
- In simple binary classification, there’s no big difference between the two, however in case of multinational classification, sigmoid allows to deal with non-exclusive labels (a.k.a. multi-labels ), while softmax deals with exclusive classes.
- A logit (also called a score) is a raw unscaled value associated with a class , before computing the probability. In terms of neural network architecture, this means that a logit is an output of a dense (fully-connected) layer .i>
So, to compute probabilities, we make the following change:
### multi-class case: probabilities = tf.nn.softmax(logits) ### multi-label case: probabilities = tf.nn.sigmoid(logits)
To compute per example loss, tensorflow provides another method:
tf.nn.sigmoid_cross_entropy_with_logits Measures the probability error in discrete classification tasks in which each class is independent and not mutually exclusive. This is suitable for multi-label classification problems.>
Rest of the code is mostly from the BERT reference.e complete code is available at github .
Kaggle Submission Score:
Just by running 1 epoch, got very good results. This is the power of transfer learning : using pre-trained model which has been trained on a huge dataset and then fine-tuning it for a specific task.
So try it out on some other dataset and run for few epochs[3�[3–4] see the results.
Thanks for reading.
 href="https://ai.googleblog.com/2018/11/open-sourcing-bert-state-of-art-pre.html" rel="nofollow noopener noreferrer" target="_blank">https://ai.googleblog.com/2018/11/open-sourcing-bert-state-of-art-pre.html
 href="https://mlexplained.com/2019/01/07/paper-dissected-bert-pre-training-of-deep-bidirectional-transformers-for-language-understanding-explained/" rel="nofollow noopener noreferrer" target="_blank">https://mlexplained.com/2019/01/07/paper-dissected-bert-pre-training-of-deep-bidirectional-transformers-for-language-understanding-explained/
 href="https://www.depends-on-the-definition.com/guide-to-multi-label-classification-with-neural-networks/" rel="nofollow noopener noreferrer" target="_blank">https://www.depends-on-the-definition.com/guide-to-multi-label-classification-with-neural-networks/
 href="https://github.com/google-research/bert/blob/master/run_classifier.py" rel="nofollow noopener noreferrer" target="_blank">https://github.com/google-research/bert/blob/master/run_classifier.py
 href="https://stackoverflow.com/questions/47034888/how-to-choose-cross-entropy-loss-in-tensorflow" rel="nofollow noopener noreferrer" target="_blank">https://stackoverflow.com/questions/47034888/how-to-choose-cross-entropy-loss-in-tensorflow
<ref="https://www.tensorflow.org/api_docs/python/tf/nn/sigmoid_cross_entropy_with_logits" rel="nofollow noopener noreferrer" target="_blank">https://www.tensorflow.org/api_docs/python/tf/nn/sigmoid_cross_entropy_with_logits
 href="https://towardsdatascience.com/journey-to-the-center-of-multi-label-classification-384c40229bff" rel="nofollow noopener noreferrer" target="_blank">https://towardsdatascience.com/journey-to-the-center-of-multi-label-classification-384c40229bff
 href="https://gombru.github.io/2018/05/23/cross_entropy_loss/" rel="nofollow noopener noreferrer" target="_blank">https://gombru.github.io/2018/05/23/cross_entropy_loss/
Aggregate valuable and interesting links.
Joyk means Joy of geeK