A Noisy-Channel Approach to Question Answering

Authors: Echihabi and MarcuDate: 2003

Presenters: Omer PercinEmin Yigit Koksal

Ustun Ozgur

IR 2013

Question Answering Approaches

Matching same words Does not always work

Who is the leader of France? Hadjenberg, who is the leader of France Jewish

Community...

Solution: Some Kind of Transformation Needed

Noisy Channel Approach

IDEA: Question is a noisy version of the Answer.

ANSWER QUESTIONNOISY CHANNEL

MODEL

QA System

Two Modules IR Engine:

Get relevant documents and their sentences Answer Identification Module

From these sentences, identify substrings S(A) as candidate answers

For each substring (answer candidate) and sentence, calculate probability of obtaining the Question text using a Generative Model

Answer Identification Module

HORIZONTAL CUTTER PERMUTERANSWER

MARKER FERTILIZER REPLACER

CANDIDATE SENTENCE

QUESTION

p1 p2 p3 p4 p5

P(Q|A) = p1 x p2 x p3 x p4 x p5

GENERATIVE MODEL

FERTILIZER

REPLACER

PERMUTER

CUTTER

ANSWER MARKER

Training and Testing

Use the generative model Training:

Deterministic cutter Answer known

Testing: Given a Question Exhaustive cutter (try each cut for each sentence) Exhaustive answer marker (try each word as possible

answer) Generate probabilities: P(Q| Sentence+Cut+Answer) Select Sentence+Cut+Answer combination with Max. P

Performance

Not bad compared to systems with 10s modules, only 2 modules

Beats QA-base in MRR on TREC datasets (0.325-0.354 vs 0.291)

QA-base was state-of-art and ranked 2-7 in TREC in last 3 years (in 2003)