![](https://csdnimg.cn/release/download_crawler_static/8187609/bg4.jpg)
proximity based methods we experimented with for boosting web retrieval effec-
tiveness; we motivate them, and give examples. Next, in Section 4, we describe
our experimental framework, largely based on the TREC web track retrieval
evaluations. We follow with an account of our results, comparing them to the
performance of other techniques for web retrieval. In Sec tion 5 we provide a
deeper analysis for some topics, aiming to understand where our methods are
especially beneficial or detrimental to web retrieval effectiveness. Finally, our
conclusions and ongoing work come in Section 6.
2 Background and Re lated Work
Web Retrieval. In recent years, web retrieval tasks were divided into two cate-
gories: Named Page Finding and Topic Distillation. Named page finding targets
scenarios where a user searches for a specific page (which is known to exist, such
as a personal home page); this task is often evaluated with MRR or Success@N
for low values of N, since the user is known to be interested in only one result, and
prefers it to be as high on the ranked list as possible. Topic distillation, on the
other hand, involves finding key resources for a particular subject. Distillation
is normally evaluated with traditional MAP and precision@N scores [10, 11].
We focus on retrieval for topic distillation. Why? First, current performance
on the named page task is very high, making it almost a solved problem. In
the 2003 edition of the TREC web track, top performing systems achieved 90%
Success@10 and 0.7 MRR scores for this task [11], meaning that in most of
the cases the single relevant document is returned at rank 1. Furthermore, the
median sc ores over all participating systems are 80% for Success@10 and well
over 0.5 for MRR. In contrast, the topic distillation task has lots of room for
improvements: at TREC 2003, the best performing system scored les s than 0.13
on Precision@10 and less than 0.16 on MAP.
Secondly, the good results on named page retrieval are partly due to the
heavy usage of factors not directly related to the ranking formula (e.g., indegree
information); this makes the task highly sensitive to these external factors, thus
making it more complex to study the effects of changes in the ranking algorithm
or query processing on retrieval performance.
Finally, we focus on topic distillation because we want to determine the im-
pact of the use of phrases and proximity terms both in terms of the traditional
MAP scores and in terms of (very) high precision measures such as MRR, Pre-
cision@1/Precision@5, and Success@1/Success@5. Topic distillation is unique as
a task where both types of evaluation scores make sense.
Phrases and Proximity Terms. Intuitively, proximity and phrase operators are
factors which improve retrieval effectiveness; indeed, lots of research was directed
in this direction. The relative merits of statistical and syntactic phrases were
extensively investigated by Fagan [14], and again by Hull et al. [19]. Until the
late 1990s, usage of phrases and proximity operators—as well as a careful usage
of bo olean operators—did show varying degrees of improvements of retrieval
results [17, 12, 22], but rarely anything substantial.