# Effective Phrase Prediction VLDB 2007 Arnab Nandi Dept. of EECS University of Michigan, Ann Arbor H. V. Jagadish Dept. of EECS University.

## Presentation on theme: "Effective Phrase Prediction VLDB 2007 Arnab Nandi Dept. of EECS University of Michigan, Ann Arbor H. V. Jagadish Dept. of EECS University."— Presentation transcript:

Effective Phrase Prediction VLDB 2007 Arnab Nandi Dept. of EECS University of Michigan, Ann Arbor arnab@umich.edu H. V. Jagadish Dept. of EECS University of Michigan, Ann Arbor jag@umich.edu

Outline INTRODUCTION Motivation Motivation Effective suggestions for autocompletion Simple FussyTree Construction algorithm& Significance FussyTree EVALUATION METRICS& Total Profit Metric(TPM) EXPERIMENTS

INTRODUCTION Autocompletion is a widely deployed facility in systems that require user input.

Motivation Ex: Hello.f 1. Hello.foo 2. Hello.freeze 3. Hello.frozen? - Decrease the number of keystrokes typed by up to 20% for email

Effective suggestions for autocompletion τ = 2 z = 2 y = 3

Effective suggestions for autocompletion “please call” meets all three conditions of co-occurrence, comparability “please call me” fails to meet the uniqueness requirement, since “please call me asap” has the same frequency. τ = 2 z = 2 y = 3

Simple FussyTree Construction algorithm our tree using a sliding window of 4 The first phrase to be added is (please, call, me, asap) (please, call, me), (please, call)

Simple FussyTree Construction algorithm Occurs with a Threshold frequency τ=2

Significance FussyTree the branch point C is considered for flag promotion

EVALUATION METRICS& Total Profit Metric(TPM)  n: number of accepted completions

EVALUATION METRICS& Total Profit Metric(TPM) d : distraction parameter TPM metric measures the effectiveness of our suggestion mechanism while the precision and recall metrics refer to the quality of the suggestions themselves TPM(0): the fraction of keystrokes saved as a result of the autocompletion TPM(1):is an extreme case where we consider every suggestion(right or wrong) to be a blocking factor that costs us one keystroke

EXPERIMENTS

Training size 8 Prefix length

Download ppt "Effective Phrase Prediction VLDB 2007 Arnab Nandi Dept. of EECS University of Michigan, Ann Arbor H. V. Jagadish Dept. of EECS University."

Similar presentations