Augmenting Input Method Language Model with user Location Type Information

TL;DR

This study investigates whether incorporating user location type information from geo-tags can improve input method predictions, finding a modest 2% accuracy gain using an LSTM model with location data.

Contribution

It introduces a novel approach of integrating location type data into input prediction models and evaluates its impact on accuracy.

Findings

Location type information provides a small accuracy boost in word prediction.

Weak correlation found between location types and word distribution.

LSTM models benefit slightly from additional location context.

Abstract

Geo-tags from micro-blog posts have been shown to be useful in many data mining applications. This work seeks to find out if the location type derived from these geo-tags can benefit input methods, which attempts to predict the next word a user will input during typing. If a correlation between different location types and a change in word distribution can be found, the location type information can be used to make the input method more accurate. This work queried micro-blog posts from Twitter API and location type of these posts from Google Place API, forming a dataset of around 500k samples. A statistical study on the word distribution found weak support for the assumption. An LSTM based prediction experiment found a 2% edge in the accuracy from language models leveraging location type information when compared to a baseline without that information.