Word Prediction Model for Mobile Devices

1. The Problem
2. Approach to the Problem
3. The Solution
4. Conclusion

Texting on mobile devices represents an important activity in our everyday lifes, why there exist tons of different text messaging apps on Android and iOS. Nevertheless,these apps tend to waste our time, as:

• Time wasted for typing recurrent words
• Time wasted for typing recurrent expressions
• Mobile device's capacity not used appropriately

Therefore, we will use the mobile device's memory and processing power to predict the next words as well as expressions and save typing. Thus, the mobile device will save its user's time and allow for faster answers to messages.

1. Downloaded a collection of texts from Blogs, News and Twitter and explored it for its content
2. Analyzed the set of solutions for this problem like heuristic models, statistical models and machine learning.
3. Decision for statistical model using Stupid Backoff Model a simplification of the Katz's Backoff Model
4. Cleaning of the training data
5. Creation of uni-, bi- and trigram
6. Creation of the prediction database from the three ngrams, by splitting its content into a “base”“ (without last word), "prediction”“ (last word) and its given "frequency”
7. Manual testing for given test sentences from an external entity

The prediction model is based on the Stupid Backoff Model with three data sets, namely a unigram (0.2 MB), bigram (2.4 MB) and trigram (11.4 MB) with a total of 441894 entries.

When using the application, the user simply types text in the appropriate textbox and a prediction will automatically be displayed in the output section in real time.

If the most recent entries in the input are known as “base” to the database, the model calculates a score for each possible prediction. The predcition with the highest score will be returned. If not known to the model, the model suggests a random word in the unigram as prediction.