Positional Relevance Model for Pseudo–Relevance Feedback Yuanhua Lv & ChengXiang Zhai Department of Computer Science, UIUC Presented by Bo Man 2014/11/18

Positional Relevance Model for Pseudo–Relevance Feedback Yuanhua Lv & ChengXiang Zhai Department of Computer Science, UIUC Presented by Bo Man 2014/11/18

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Review on Feedback (1)Explicit Feedback Easy for training But need user’s interaction

Review on Feedback (2)Implicit Feedback Not need interaction But work more on mining

(3)Pseudo- Relevance Feedback No need user’s interaction No Mining

Problems? Traditional Pseudo-Relevance Feedback assumes that the contents of a document are incoherent (sharing the same topic). What if a document shares different topics ？ Term-based? Or document-based?

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Motivation How to effectively select from feedback documents the words that are focused on the query topic based on positions of terms in feedback documents?

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Positional Relevance Model Relevance Model (one of the most robust) Θ rep- resent the set of smoothed document models for the pseudo feedback documents. p(θD) is a prior on documents and is often assumed to be uniform without any additional prior knowledge about document D. After the relevance model is estimated, the estimated P (w|Q) can then be interpolated with the original query model θQ to improve performance. α is a parameter to control the amount of feedback.

Positional Relevance Model(PRM) i indicates a position in document D F is the set of feedback documents (assumed to be relevant) Challenge? How to estimate joint probability? Positional Relevance Model

Methods (1) i.i.d. Sampling (2) conditional Sampling estimating P (w, Q, D, i) Positional Relevance Model

i.i.d. sampling Positional Relevance Model

i.i.d. sampling derivation(1) P(D) can be interpreted as a document prior and set to a uniform distribution with no prior knowledge about document D. assume that every position is equally likely but it is possible to estimate P(i|D) based on document structures assume that the generation of word w and that of query Q are independent Positional Relevance Model

i.i.d. sampling derivation(2) In the above equation, P (w|D, i) is the probability of sampling word w at position i in document D. To improve the efficiency of PRM, we simplify P (w|D, i) as: QUESTION: HOW to estimate ? The query likelihood at position i of document D. Positional Relevance Model

conditional sampling …… QUESTION: HOW to estimate ? The query likelihood at position i of document D. Positional Relevance Model

estimate the query likelihood at position i of document D (1)Use Positional Language Model (2)Use Gaussian kernel function (3)Approximate (4)Set parameters Positional Relevance Model

estimate the query likelihood at position i of document D (5)Use JM Smoothing (6)Compute The computation of positional query likelihood is the most time-consuming part in estimating PRM. Positional Relevance Model

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Experiments and Results Evaluation methods. (1) The basic retrieval model is the KL-divergence retrieval model, and we chose the Dirichlet smoothing method [33] for smoothing document language models, where the smoothing parameter μ was set empirically to This method was labeled as “NoFB”. (2) The baseline pseudo feedback method is the relevance model “RM3”, which is one of the most effective and robust pseudo feedback methods un- der language modeling framework. (3) Another baseline pseudo feedback method is a standard passage- based feed- back model, labeled as “RM3-p”, which estimates the RM3 relevance model based on the best matching passage of each feedback document. (4) We have two variations of PRM, i.e., “PRM1” and “PRM2”, which are based on the two estimation methods described in Section 3.2, respectively. (5) In addition, we also used PRM1 and PRM2 for passage feed- back in a way as RM3-p does. Specifically, we first computed a PLM for each position of the document, and then we estimate a PRM based on a passage of size 2σ centered at the position with the maximum positional query likelihood score

Results Experiments and Results

More results Robustness Analysis Experiments and Results

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Presentation Guideline Review on Feedback Motivation Positional Relevance Model Experiments, Results and Analyses Conclusions

Propose a novel positional relevance model(PRM) PRM exploits term position and proximity to assign more weights to words closer to query words, based on the intuition—words closer to query words are more likely to be consistent with the query topic. Experiments results show that PRM is quite effective and performs significantly better than others based on document or passage. Conclusions

Questions?