10/1/2015
Axiomatic Analysis of Smoothing Methods in Language Models for Pseudo-Relevance Feedback
HUSSEIN HAZIMEH AND CHENGXIANG ZHAI
UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN
1
Pseudo Relevance Feedback
Judgments:
d1 +
d2 -
d3 +
…
dk -
...
Query Retrieval
Engine
Results:
d1 3.5
d2 2.4
…
dk 0.5
...
User
Document
collection
Judgments:
d1 +
d2 +
d3 +
…
dk -
...
top 10
Pseudo feedback
Assume top 10 docsare relevant
Relevance feedback User judges documents
New
Query
FeedbackLearn from
Examples
2
Pseudo-Relevance Feedback
It’s blind!
Good for high recall information needs
A Blind Superhero. Courtesy of iStock
3
Collection-based Smoothing
Collection-based smoothing is generally used for LM-based retrieval functions and for PRF models
A commonly used collection-based smoothing scheme is Dirichletprior smoothing:
Dirichlet Prior (Smoothing Parameter)
Document Length
Count of Word in Document
4
Study of Smoothing Methods in PRF
We will establish both analytically and empirically that collection-based smoothing is not a good choice for PRF:◦ It forces PRF models to select very common words
Additive smoothing will be shown to outperform the collection-based counterpart
5
How Do LM PRF Models Work?
D1
Dn
… Averaging
Function: 𝑨
Scoring
Function: 𝒇
𝑃(𝑤|𝜃1)
𝑃(𝑤|𝜃𝑛)
𝑃(𝑤|𝜃𝐶)
𝑃(𝑤|𝜃𝐹)
6
How Do LM PRF Models Work?
The feedback LM, 𝜃𝐹, would generally have the following form:
𝐴:ℝ𝑛 → ℝ is an averaging function, e.g. geometric mean
𝑓:ℝ2 → ℝ is a function increasing in the first argument and decreasing in the second
Rewards common words in feedback set
Penalizes common words in collection
7
Problem!
The first argument rewards common words in the collection while the second penalizes them. The analysis shows that the first argument usually “wins”!
Rewards common words in feedback set
and collection
Penalizes common words in collection
Proportional to 𝑃(𝑤|Θ𝐶)
8
Overview of the Analysis
We considered three PRF models in the study:◦ Divergence Minimization Model
◦ Relevance Model
◦ Geometric Relevance Model
Next, we will briefly discuss how the DMM and GRM work and then give an overview of the axiomatic analysis.
The analysis of the RM is very similar to the GRM and the same results apply
9
Divergence Minimization Model (Zhai and Lafferty, 2001)
The DMM solves the following optimization problem:
The solution has a closed form and is given by:
10
Geometric Relevance Model (Seo and Croft, 2010)
An enhanced form of the Relevance Model (RM) that replaces the arithmetic mean used in RM by the geometric mean:
Note that the function above is not is not affected by 𝑃(𝑤|𝜃𝑐), i.e., the model is not designed to penalize common words.
11
Main Axiom: IDF Effect (Clinchant and Gaussier, 2013)
Rationale: A PRF model is expected to penalize common words in the collection in order to select high quality discriminative terms.
Given any two words 𝑤1and 𝑤2 from the feedback set 𝐷1, 𝐷2,
12
DMM with Collection-based smoothing: IDF Effect
Study the sign of:
Not straightforward. Strategy:◦ Find an attainable lower bound on the expression above
◦ Study the sign of the lower bound
◦ If the lower bound is strictly positive, then DMM supports the IDF effect
13
DMM: Results of Analysis
Conclusion: Using collection-based smoothing the DMM will be either consistently reward common terms or will select only one feedback term
14
GRM with Collection-based smoothing: IDF Effect
The GRM cannot support the IDF effect:
It consistently rewards favors common words in the collection
15
Proposed Solution: Additive Smoothing
Words get additional pseudo-counts:
Next, we show how additive smoothing prevents the models from rewarding common terms
16
DMM with Additive Smoothing: IDF Effect
The DMM unconditionally supports the IDF Effect:
Now it is performing the intended objective!
17
DMM: Empirical Validation
Query: “Computer”
18
GRM with Additive Smoothing: IDF Effect
Although the IDF effect is still not supported:
However, common terms are no longer being rewarded!
19
GRM: Empirical Validation
Query: “Computer”
20
Empirical Evaluation: Retrieval Measures
21
Empirical Evaluation: Robustness of Additive Smoothing
22
Measuring the Discrimination of PRF Models
In previous studies, the average of the IDF of the top terms was used as an indicator of how discriminative the terms selected by a PRF method are
Such a measure might not work well in some cases
We propose the Discrimination Measure (DM):
≈ Expected Document Frequency
Constant
23
Empirical Evaluation: Discrimination Measure
A several-fold decrease in the expected document frequency
24
Conclusion
Collection-based smoothing forces PRF models to select very common terms◦ The same problem might exist in other applications where LMs are aggregated
Additive smoothing prevents PRF models from rewarding common terms and increases the retrieval performance significantly
A new measure for quantifying PRF Discrimination
25
Future Work
Should PRF models penalize common words?
Analysis of other smoothing methods such as topic-based smoothing
Inspect areas, other than PRF, where collection-based smoothing is used in aggregating language models
26
Thanks to SIGIR for the Student Travel Grant!
Thank you for Listening!
27