Kick Starting an NLP Project
We can divide an NLP project into several sub-projects or phases. These phases are completed in a particular sequence. This tends to increase the overall efficiency of the process, as memory usage changes from one phase to the next. An NLP project has to go through six major phases, which are outlined in the following figure:
Figure 1.4: Phases of an NLP project
Suppose you are working on a project in which you need to classify emails as important and unimportant. We will explain how this is carried out by discussing each phase in detail.
Data Collection
This is the initial phase of any NLP project. Our sole purpose is to collect data as per our requirements. For this, we may either use existing data, collect data from various online repositories, or create our own dataset by crawling the web. In our case, we will collect different email data. We can even get this data from our personal emails as well, to start with.
Data Preprocessing
Once the data is collected, we need to clean it. For the process of cleaning, we will make use of the different preprocessing steps that we have learned about in this chapter. It is necessary to clean the collected data to ensure effectiveness and accuracy. In our case, we will follow these preprocessing steps:
- Converting all the text data to lowercase
- Stop word removal
- Text normalization, which will include replacing all numbers with some common term and replacing punctuation with empty strings
- Stemming and lemmatization
Feature Extraction
Computers understand only binary digits: 0 and 1. As such, every instruction we feed into a computer gets transformed into binary digits. Similarly, machine learning models tend to understand only numeric data. Therefore, it becomes necessary to convert text data into its equivalent numerical form.
To convert every email into its equivalent numerical form, we will create a dictionary of all the unique words in our data and assign a unique index to each word. Then, we will represent every email with a list having a length equal to the number of unique words in the data. The list will have 1 at the indices of words that are present in the email and 0 at the other indices. This is called one-hot encoding. We will learn more about this in coming chapters.
Model Development
Once the feature set is ready, we need to develop a suitable model that can be trained to gain knowledge from the data. These models are generally statistical, machine learning-based, deep learning-based, or reinforcement learning-based. In our case, we will build a model that is capable of differentiating between important and unimportant emails.
Model Assessment
After developing a model, it is essential to benchmark it. This process of benchmarking is known as model assessment. In this step, we will evaluate the performance of our model by comparing it to others. This can be done by using different parameters or metrics. These parameters include precision, recall, and accuracy. In our case, we will evaluate the newly created model by seeing how well it performs at classifying emails as important and unimportant.
Model Deployment
This is the final stage for most industrial NLP projects. In this stage, the models are put into production. They are either integrated into an existing system or new products are created by keeping this model as a base. In our case, we will deploy our model to production, so that it can classify emails as important and unimportant in real time.