The purpose of this project is to categorize poems into different types.

The dataset used for training was compiled by PoetryFoundation.org which is available on Kaggle and is also included in the repo. The dataset contains the 574 poems along with their poets and their genre (type). We only require the poems and their type for our training needs.

The code was built on a Python 3.6 environment with the following external libraries:

• keras
• pandas
• numpy
• scikit-learn
• nltk
• textblob

1. poems.csv - Dataset used for training the model (also available on Kaggle)
2. poems.py - Main trainer program that pre-processes the data, builds and trains the model

I had never worked with text data for categorisation before this point so I had to research a bit in order to train a model that could categorize poems in the form of text. Cleaning of the dataset comes first. First of all, not all the words in the poem necessarily contribute in determining the category of the poem. Words like a, you, my, is, etc. have least impact so they were removed with the help of stopwords python module. The most frequent and least frequent words were also removed. Suffixes such as ing, ly, s, etc. were removed for maintaining consistency across all poems. At this point all the basic pre-processing has been done.

Now in order to work with text in neural networks we need to convert them to certain numbers or hashes so that they can be fed into the neural networks and trained. For this task the function for one hot coding in the keras pre-processing library was used. The function creates hashes for each word in the dataset vocabulary.

I have built a fairly simple model that uses three layers namely Embedding Layer, Flatten Layer, and Dense Layer for training. Embedding is a layer that’s frequently used while dealing with text classification. Flatten is used to reduce the dimensionality of Embedding layer’s output since the Dense layer takes in one dimensional data as input.

• A snippet showing last few Epochs and the Accuracy

The model gives an accuracy that varies from 71.55% to 72.3% which I think is fair enough for a model this simple and for a dataset this limited. But is this accuracy good enough? Probably not, but the model can definitely be improved using various other layers and a better dataset.

There are three categories of poems in the dataset. So, even if a person randomly guesses the categories his probability of being accurate will be 33.33% and the model gives an accuracy of 72% so we can say that model is performing a significantly good categorization.

I haven't deployed this model but I probably will in the near future.