A literal reverse english dictionary. Predicts the word for a given definition.
We need to map a sequence of words (Our definition) to a target word. This can be done rather effectively using Recurrent Neural Networks, or its variants (Used cudNN-GRUs here).
We can build the RNN rather easily using Keras. The real nuisance is collecting and cleaning the data. The approach I chose was to gather a list of 3000 words from here, and use scrap_dict.py (which uses bs4) to scrap the definitions. For convenience, i've included dataset.txt, which has the words and its corresponding definitions. The rest of the data processing, and the RNN is implemented in ReverseDict.ipynb.
- Keras
- BeautifulSoup (Only if you want to try scrapping data yourself).
- TensorFlow (Only if you want to use cudNN-GRU. Use your favourite backend otherwise).
The code works as intended. Maps a definition to a word. But, the output usually doesn't make sense (Gives the wrong word for the definition). Need a smarter way to feed data.
The reason I added this repo is that, it serves as a general experiment on seq2seq models. It can be generalised to several use cases. For example, you can make a chatbot using the same code, by just feeding the right data.