Login
Congrats in choosing to up-skill for your bright career! Please share correct details.
Home / Blog / Artificial Intelligence / Transformers – The New Breed of NLP
Bharani Kumar Depuru is a well known IT personality from Hyderabad. He is the Founder and Director of AiSPRY and 360DigiTMG. Bharani Kumar is an IIT and ISB alumni with more than 17 years of experience, he held prominent positions in the IT elites like HSBC, ITC Infotech, Infosys, and Deloitte. He is a prevalent IT consultant specializing in Industrial Revolution 4.0 implementation, Data Analytics practice setup, Artificial Intelligence, Big Data Analytics, Industrial IoT, Business Intelligence and Business Management. Bharani Kumar is also the chief trainer at 360DigiTMG with more than Ten years of experience and has been making the IT transition journey easy for his students. 360DigiTMG is at the forefront of delivering quality education, thereby bridging the gap between academia and industry.
Table of Content
Our capacity to communicate information through language is one of our most remarkable accomplishments. Textual data in many different languages is one of the numerous types of information that are created all around us. Machine learning models that can generate speech and words with astounding fluidity have been created using neural networks, which has been a great success. The programme "Transformers," which has brought representational learning back to life, is at the centre of this revolution. Researchers have created neural network models that are capable of human speaking, making them clever enough to grasp what to say and in what order, in order to produce high-quality human-level word sequences and to accomplish both effective and controlled text production. Natural Language Processing (NLP) is the name given to the branch of artificial intelligence (AI) that employs computer methods to represent and analyse human languages. Allowing robots to comprehend and generate language has emerged as one of AI's revolutionary talents.
Figure 1 Natural Language Processing (Source: bold360)
The Transformer Library's wealth of deep learning architectures and pre-trained models enable us to complete a wide range of text-related tasks with great ease, including language translation, question answering, text summarization, and many sequences to sequence tasks in more than 100 different languages. These pre-trained models are also capable of handling long-range dependencies successfully. A job that uses input in sequence order and output in another sequence order, which may or may not be the same length as the input, is known as a "sequence to sequence task." They are substantially pre-trained on data-rich tasks like language modelling and are based on transfer learning. The models can be adjusted to carry out certain tasks connected to our unique datasets. Transformers offers APIs that make sharing with the community and enabling smooth library integration across libraries simple, both of which are supported by well-known deep learning libraries, PyTorch and TensorFlow.
The Transformer Library contains many deep learning architectures and pre-trained models allowing us to solve text-related tasks with great ease such as language translation, question answering, text summarization and many sequences to sequence tasks in 100+ languages, these pre-trained models are capable of managing long-range dependencies effectively. Sequence to sequence task can also be defined as a task where input is in sequence order and output is another sequence order that may or may not be the same length as the input. They are based on transfer learning and are pre-trained heavily on data-rich tasks like language modeling. The models can be fine-tuned to perform specific tasks related to our specific datasets. Supported by popular deep learning libraries, PyTorch and TensorFlow, Transformers provide APIs that enable seamless integration between libraries and sharing with the community easy for further research experiments. Click here to learn Artificial Intelligence in Hyderabad
A Transformer has an encoder-decoder architecture in which an encoder turns the input text into numerical tensors and a decoder turns the tensors back into text. To complete an underlying job, it creates or extracts meaningful text data from the input representation. The 'Attention' model depicts the interdependencies between the different input and output components. Text input is divided into many tokens by tokenization as part of an encapsulation process. Each token is converted into a useful representation before being delivered to a decoder, where the representation is retrieved or produced as the output.The encoder-decoder architecture of a Transformer is where the encoder converts the input text into numeric tensors and a decoder converts the tensors into output text. It generates or extracts meaningful text data from the input representation to solve an underlying task. The ‘Attention’ models the dependencies between the various parts of the input and parts of the output. The input goes through an encapsulation pipeline involving tokenization, where the text input is split into multiple tokens. Each token is mapped into a meaningful representation and passed on to a Decoder where the representation is extracted or generated into a final output. These pre-trained models can be applied to solve a variety of tasks such as:
The illustration below provides a high-level view of the Transformer’s architecture and its building blocks. This article briefly introduces key elements of the architecture to understand what makes Transformer network a revolutionary breakthrough in Natural Language processing.
Figure 2 The Transformer Architecture (Source: Transfer Learning for NLP @Paul Azunre)
To summarize the illustration, a Transformer Architecture contains the following elements: encoder-decoder stack, positional embeddings, multi-head attention mechanism, feedforward, linear and SoftMax functions of neural network.
Figure 2's Transformer design can be seen with the encoder stack on the left and the decoder stack on the right. In Transformers, sentences are processed as a whole. As a result, the input consists of a long string of words, each of which is transformed into a tensor and then represented numerically. The input is divided into vectors known as tokens in order for the algorithm to comprehend, interpret, and store the input phrases. Embedding methods like word2vec or one-hot encoded vectors are used for this process.
Figure 3 Tokenizing Process (Source: NLP Tutorial Simplilearn)
To examine the word order of a phrase, tokens and word embedding must first be acquired. This is carried out in order to preserve sequential awareness and understand how input and output interdependence. Positional encodings are used in order to comprehend the relative order of words.
Figure 4 Input Embedding added with a position index (Source: The transformer walkthrough, @Matthew Barnett)
Each word in the input phrase is changed into an index number during position encoding. From 1, 2, 3, 4, on up to n, the index number stores the value that reflects the absolute location of each word in a phrase. To create a positional encoding, the word "stars"' location and its vector values are determined. The notion of positional encoding in Transformers was briefly presented in this section. The model may learn and use position encoding in a variety of ways. In the Transformer design, an encoder stack and a decoder stack typically have six identical levels. (Reference: Attention Is All You Need). You may dive down into each layer of the encoded and decoded stack to have a better understanding of the operation taking place there.
Figure 5 Simplified decomposition of Encoder-Decoder (Source: Attention is All You Need)
Multi-Head (Self) attention layer and a feed-forward neural network make up each encoder layer. Similar to the encoder layer, the decoder layer also consists of the same components, but between the feed-forward neural network and the self-attention layer is an encoder-decoder attention layer. As seen in the earlier illustration, the encoder stack receives input words, and the top of the stack outputs the results. The encoder stack's output vectors have the same size and form as the input vectors. Similar characteristics exist between the encoder and decoder stacks. The encoder generates its output and sends the data to the layers of the decoder. The encoder-decoder attention module receives this information next. In the Decoder stack, encoder-decoder attention functions somewhat differently than an encoder's own attention. Instead of the layers below in the stack, some of its inputs originate from the encoder stack.
In Neural Networks, Attention refers to the key details of input that are of value and focus and that must be attended to solve a specific task. The Attention mechanism allows the network to learn and locate what part of input should be of focus. i.e. apply weights on that input to impact output generation. The input vectors that are passed into the decoder stack hold the key and value vectors from the top of the encoder stack. The query vectors are passed from the layers below in the decoder stack. This is done to compute attention between every output and input token. In the Transformer model, Key, Value, and query vectors are constructed by computing the embedded vectors by a matrix of weights. These vectors are needed to establish links between the words and to represent the relationship between the words of a given input sentence.
Figure 6 Weights are assigned to input words at each step of translation (Source: Attention Mechanism @floydhub)
The feed-forward neural network is the following step after an output from the attention mechanism module has been created. To give an example, suppose the intended output is "How was your day?" Based on this target, loss functions are computed, backpropagation is carried out, gradients are derived, and weights are updated to produce an accurate translation of the input, which is then forwarded to the following decoder in the stack. The output of a decoder stack is processed via a linear and SoftMax layer at the top to produce the network's output probabilities. In NLP tasks, the attention mechanism is crucial since it enables the model to recall input words and pay attention to certain phrases while formulating a response.
Simplified terminology was used to explain some of the breakthroughs made by Transformers. Transformers' primary features are its capacity to support parallel processing, shorten calculation times, and enhance performance for input with lengthy dependencies. The computation of the link between words relies heavily on the positional encodings and attention mechanism. Transformers, as opposed to RNN and LSTM, perform non-sequentially, meaning the entire phrase is used as the input. This prevents recursion. It has the capacity to learn dependencies and minimise information loss. BERT, its variations, GPT, and XLNet are examples of NLP models that have developed from the Transformers design and have become widely used today. Transformers is the next generation of NLP due to its state-of-the-art performance on language processing tasks.
Click here to learn Data Science Course, Data Science Course in Hyderabad, Data Science Course in Bangalore
360DigiTMG - Data Science, Data Scientist Course Training in Bangalore
No 23, 2nd Floor, 9th Main Rd, 22nd Cross Rd, 7th Sector, HSR Layout, Bengaluru, Karnataka 560102
1800-212-654-321
Didn’t receive OTP? Resend
Let's Connect! Please share your details here
