The Mind Reading Robot: A Fun Game of Word Prediction

Mind Reading Robot uses NLP algorithms to finish your phrases in the form of a “Human vs. Robot” game with a very simple scoring system

• You provide a phrase and the next word you would follow it with (“Crazy like a”, and “fox”)

• The Robot gets three guesses at the next word

• Play again and again, see yours and the Robot's wins over time

App Functionality

• Easy-step user input: Clear instruction provided in the input panel as well as the “How to Play”“ tab
• Diverse result presentation: "Result”“ and "Score Board” tabs show word predictions, scores, and charts

Prediction Algorithm

The app uses a 4-gram model with backoff smoothing method. It first searches the possible highest N-gram(ex. 4-gram if the search phrase consists 3 or more words), and returns the top 3 “next word” by frequency. If the evidence is insufficient, it then searches the next lower–order N-gram, so on and so forth (Jurafsky & Martin, 2014).

N-gram Data

60% random samples of english blogs, news, and tweets corpus from the Coursera Capstone project were preprocessed and generated into N-gram tables using R’s quanteda{} package. Features that occurred only once in each N-gram were removed as minimally advantageous to prediction, a decision which greatly enhanced app speed.

Anonymous Data Collection

Data is recorded to a private Google sheet. This allows the tracking of algorithm performance and will allow customized and improved accuracy by training the algorithm through past user inputs(see future enhancement).

Accuracy Improvement

- Enabling real-time algorithm training with user inputs, without compromising system processing time - Deploying sophisticated prediction methods such as Kneser-Ney smoothing and skip-grams - Looking at trends in tracked responses. For example, present historical data shows that users are unlikely to choose prepositions, conjunctions, or delimiters as their “next word”. A future algorithm will include weights to improve prediction according to how humans play this game. - Word association weighting

Potential applied usages include

- An educational tool: variants of the app could train children or second language learners. - User experience: the app could be used to elicit word associations in entertainment and social media spaces. For example, the Robot could provide the phrase, and the human the final word. - Learning a specific group: Millennials, skating enthusiasts, dreamers etc. (could result in improved accuracy). - Marketing research: “I drink coke when I”, “_____”