We introduce basic notation and then describe several retention functions serving as temporal document priors. We say that document D in document collection D has time timepDq and text textpDq. A query q has time timepqq and text textpqq. We write dgpq; Dq as the time difference between timepqq and timepDq with the granularity g. We introduce a series of retention functions. The memory chain models ((1) and (2)) build on the assumptions that there are different memories. The Weibull functions ((3) and (4)) are of interest to psychologists because they fit human retention behavior well. In contrast, the retention functions linear and hyperbolic ((6) and (7))
Every moment of our life we retrieve information from our brain:
we remember. We remember items to a certain degree: for a
mentally healthy human being retrieving very recent memories is
virtually effortless, while retrieving untraumatic memories from the past
is more difficult [
Inspired by the cognitive psychology literature on human memory and on retention functions in particular, we consider seven temporal document priors. We propose a framework for assessing them, building on four key notions: performance, parameter sensitivity, efficiency, and cognitive plausibility, and then use this framework to assess those seven document priors. We show that on several data sets (newspaper and microblog), with different retrieval models, the exponential function as a document prior should not be the first choice. Overall, other functions, like the Weibull function, score better within our proposed framework.
The full version of this paper appeared in ECIR 2013 [
Memory Chain Model. The memory chain model [
One can also consider the Weibull function
μ1 e a1dgpq;Dq a2μ2a1 pe a2dgpq;Dq e a1dgpq;Dqq ; and its extension fBWpD; q; gq e adgpD;qq d d ; fEWpD; q; gq b p1 bqμe adgpD;qq d d b p1 bqμpdgpD; qq 1qa; where a, b, and μ are the decay, an asymptote, and the initial learning performance.
A very intuitive baseline is given by the linear function, fLpD; q; gq pa dgpq; Dq b bq ; where a is the gradient and b is dgpq; argmaxD1PD dgpq; D1qq. Its range is between 0 and 1 for all documents in D.
The hyperbolic discounting functionhas been used to model how humans value rewards: the later the reward the less they consider (1) (3) (4) (5) (6) the reward worth. Here, MCM-1 k dgpq; Dqq (7) where k is the discounting factor.
We propose a set of three criteria for assessing temporal document priors and we determine whether the priors meet the criteria. A framework for assessing temporal document priors.
Performance. A document prior should improve the performance on a set of test queries for a collection of time-aware documents. A well-performing document prior improves on the standard evaluation measures across different collections and across different query sets. We use the number of improved queries as well as the stability of effectiveness with respect to different evaluation measures as an assessment for performance, where stability refers to that improved or non-decreasing performance over several test collections.
Sensitivity of parameters. A well-performing document prior is not overly sensitive with respect to parameter selection: the best parameter values for a prior are in a region of the parameter space and not a single value.
Efficiency. Query runtime efficiency is of little importance when it comes to distinguishing between document priors: if the parameters are known, all document priors boil down to simple look-ups. We use the number of parameters as a way of assessing the efficiency of a prior.
Cognitive plausibility. We define the cognitive plausibility
of a document prior (derived from a retention function) with the
goodness of fit in large scale human experiments [
Discussion. To ensure comparability with previous work, we use
different models for different datasets: TREC-2 and TREC-{6,7,8}
for news and Tweets2011 for social media. On the news data set,
we analyse the effect of different temporal priors on the
performance of the baseline, query likelihood with Dirichlet
smoothing [
Table 1 gives an overview of the assessment of different document priors. We find that all but BW, AP, and L are stable in the parameter optimisation. Of those functions, BW and L have only few parameters, and BW performs best. 4.
We have proposed a new perspective on functions used for temporal document priors used for retrieving recent documents. We showed how functions with a cognitive moti- vation yield similar, if not significantly better results than others on news and microblog datasets. In particular, the Weibull function is stable, easy to optimize, and motivated by psychological experiments.
Acknowledgments. We thank Jessika Reissland for her inspiration. This research was partially supported by the European Union’s ICT Policy Support Programme as part of the Competitiveness and Innovation Framework Programme, CIP ICT-PSP under grant agreement nr 250430, the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreements nr 258191 (PROMISE Network of Excellence) and 288024 (LiMoSINe project), the Netherlands Organisation for Scientific Research (NWO) under project nrs 612.061.814, 612.061.815, 640.004.802, 727.011.005, 612.001.116, HOR-11-10, the Center for Creation, Content and Technology (CCCT), the BILAND project funded by the CLARINnl program, the Dutch national program COMMIT, by the ESF Research Network Program ELIAS, and the Elite Network Shifts project funded by the Royal Dutch Academy of Sciences.