“…this study estimates that TREC’s existence was responsible for approximately one-third of an improvement of more than 200% in web search products that was observed between 1999 and 2009.”

More here.

The goal of this project is to create a database of web search activity that will be provided to the information retrieval research community to use on current and future information retrieval research projects.

To accomplish this, the Lemur Project developed a toolbar to be voluntarily installed by users. After a year of data collection, the project has been aborted,

Given that we have gathered the equivalent of less than 6 seconds of Google traffic (assuming 500 million queries per day) in one year, we have decided to terminate the project.

This is pretty depressing news. Admittedly, part of this depression originates from my guilt over not having contributed to the project myself. However, a more substantial part stems from the potential this data set had to be groundbreaking, perhaps similar to the release of the first Tipster collections. Although this was way before my time, I imagine the sudden release of a large, public corpus resulted in a tremendous amount of activity and excitement.

Information retrieval research has had large collections of documents for a few decades now. We evaluate on a few hundred queries and publish results. With some exceptions, the majority of interest in the field has focused on scaling up corpora. As a result, we have rich set of tools to analyze and retrieve documents from large corpora.

There are two things missing from this model: a rich stream of queries coming into the system and a rich stream of interactions between users and documents. Our friends in the CHI and information science communities have been doing a great job with understanding the important factors involved in user behavior on laboratory scale. However, I’m going to draw an analogy here between small scale user studies for IR and document-level NLP analysis for IR that may raise a few eyebrows. I believe that many IR researchers would argue that, given the choice between a corpus-driven approaches and NLP approaches to IR, they would opt for more data. This is despite the rich analysis NLP can provide. Similarly, I believe that the fine-grained analysis provided by laboratory studies may be less important than very large scale analysis of user behavior. Of course, both the results about NLP for IR and the claim about laboratory experiments are based on relatively limited experiments (e.g. small sets of queries). We should, as a community, continue research in all of these directions.

Having said this, let’s consider some motivations for web query logs and IR research,

Claim 1. Web query logs will help with the contribution to web search research.

There is no doubt that query logs are important for any search engine, web or otherwise. However, query logs are only one of the many sources of interaction data available in production. There are many, many other signals which can be effectively exploited for query understanding and document ranking. In my opinion, outside of starting its own web search engine, academia will always be scurrying to catchup to industry’s data sources.

I convinced myself a few years ago that the resources required to build and maintain a web search engine may never exist in academia. This is not to say that academic IR researchers should give up on having impact on web search engines. IR research several decades old continues to impact modern search engine design. What needs to be determined is how the current academic IR researchers can more directly address the problems confronted by web search companies. I personally believe that a tight coupling between academic and industrial research labs needs to exist. This could be accomplished in a number of ways.

1. add value to an existing search engine’s interface. If search engines provide ranker APIs, academics can develop new interfaces which may attract users and, as a result, interaction data.
2. teach the IR fundamentals during the academic year/perform intense interaction during the summer during internships or other collaborations. I am most familiar with Yahoo’s Key Scientific Challenges Fellowships and Faculty Engagement Grants. Similar programs exist at other web search engines.
3. develop high-quality, public web search engine simulators which provide students/researchers with the ability to test algorithms in silico. Our SIGIR 2009 paper made extensive use of simulation whose parameters were grounded in real world data. Systems research in computer architecture or computer networking have adopted this approach for a while. SIGIR 2010 will be hosting a workshop on simulated interaction.

No doubt there are many, many other alternatives.

Claim 2. Web query logs will help with the contribution to production search research.

As stated earlier, IR research has looked at the document side for many, many problems. This research has benefited web search as well as search in other domains such as legal, news, and enterprise search.

User behavior data improved production web search engines; user behavior data will no doubt improve production non-web search engines. Just as with web search though, this data does not exist in academia.

I believe, though, that the barrier to entry for non-web/vertical search engines is somewhat lower. The collections are smaller and manageable. At the same time, document representations can be richer for verticals, interaction is less constrained, and, as a result, the potential for attracting users may be higher than with portal web search engines.

If an academic institution maintained a domain-specific production search engine, academic research could become more relevant to industrial search engines. For example, academic institutions would easily be able to publish about query logs, interaction, large scale adaptation, and online learning with large scale real world data.  One important, unresolved question is how to come to terms with experimental reproducibility and production data which is often closed due to privacy reasons.

Academic IR research will continue to contribute to general IR research. Students trained in IR fundamentals will continue to be strong candidates for research and development in production search companies. I believe that there is room for greater impact. How that happens remains to be seen.

There’s no existing IR test collection for such a task, although some similar datasets do exist. For various reasons we weren’t able to create a traditional test collection, with user-issued queries and deep pools of relevance judgements. But, this particular dataset and possibly other online dialog archives can be mined to produce a ready-made IR test collection.

The users of the online forum we’ve been looking at frequently include links in their forum posts — often to previous messages and threads in the same forum. These links are sometimes in response to a new user’s question, and refer the user to a previous instance of the same (or similar) question and an answer contributed by another user. Here’s a few examples to illustrate my point. This interaction among forum users can be used as a form of query/relevance judgement pair. See the paper for a few more details on how we characterize the presence of a question-post/answer-link pair.

This type of test collection creation does have some distinct advantages over the typical retrieval test collections used at TREC. First, the queries represent real information needs of real users of the online forum. Many TREC queries are pulled from search engine logs, but frequently (as in the Blog Track’s Feed Distillation task) the queries are invented by participants or assessors. The information needs present in the online forum posts are much more verbose than typical keyword queries on a web search engine, providing a retrieval system more evidence with which to use in relevance scoring. The “relevance judgement”, provided by another forum user linking to a previous thread, also presents in-situ relevance information — sensitive not only to the original question, but also to the overall nature of the forum and the time when the question was asked.

There are several drawbacks inherent in this type of corpus creation, most importantly with regard to the exhaustiveness of the relevance assessment. Typically in TREC-style collection development, ranked results from several retrieval systems are pooled and those pooled documents are assessed for relevance. When the systems’ output is sufficiently diverse and relevance assessment is sufficiently deep, this produces a reasonably complete relevance assessment for each query — if a relevant document is in the collection, it would most likely be retrieved by one of the systems and be judged by being admitted into the pool. The method of collecting relevance judgements we use in our SIGIR poster, on the other hand, will not produce anything close to an exhaustive set of relevant threads. In the great majority of cases, only a single thread is linked to in a subsequent reply message. There is no guarantee that this thread is the best or only relevant thread in the collection. For this reason, we must take care not to assume non-judged threads are necessarily irrelevant.

There are plenty of datasets that seem to be ready-made for classification or regression tasks, without any need for annotation — for example the classic 20 newsgroups for text classification and Yahoo! Answers for a number of prediction tasks. For relevance ranking, however, I haven’t seen any ready-made datasets with real relevance judgements, as opposed to noisy interaction indicators such as click-through statistics. Conversation archives like the one we use offer one way to mine behavioral data for relevance judgements, offering ground-truth preferable in many ways to post-hoc relevance assessment.

This problem appeared in a relatively early paper by Ian Soboroff and others and it has continued to be problematic since then (Aslam and Savell discuss it, as well).

The mechanics at work here are easy enough to understand. Most judgment-free ranking is based on an analysis of the documents that are commonly retrieved by many systems. Systems that perform fairly well tend to return many of the same documents as other ‘pretty good’ systems. Poor performers tend to miss these documents.  But what about the best performers?  Aslam and Savell argue that most judgment-free evaluation leads to a “tyranny of the masses,” punishing systems that do anything really different from the norm.  Wu and Crestani suggest that the best performers “are somewhat peculiar”; they do something qualitatively different from average performers.  Simply by deviating from the norm, then, the best systems look bad under the common judgment-free lenses.

If the best systems are doing something qualitatively different from the great unwashed, what is that difference?  Can we model it in order to improve our ranking of systems in the absence of relevance judgments?

Most of the literature on this topic focuses on TREC data.  In this context it is often the case that the best retrievals result from complex manual runs, as opposed to automatic, title-only runs. When I started looking into this a bit, I assumed that high-performance runs, by virtue of resulting from detailed statements of information need, would retrieve relevant documents that were missed by most other systems (e.g. documents that were relevant but that lacked terms from the topic title).

Pursuing this hypothesis will take some real work, but I was surprised by this figure:

Average recall for TREC-8 systems

The plot shows #rel_returned/#rel averaged over the 50 topics used in TREC-8 (ranking is by MAP).  Now it certainly could be true that the best systems are finding relevant documents that other systems are not.  But the best performers don’t appear to be finding more relevant documents than others.  

To me the mystery here is why these high-performing runs appear so bad using most judgment-free evaluation measures.  Retrieving “hidden” relevant documents would indeed lead to apparent bad performance under a tyranny of the masses.  But these systems don’t have especially high recall (quite the contrary, in fact).  Are they retrieving hidden relevant documents and failing to return obvious ones?  That seems unlikely.

What are the best performers doing that sets them apart from the crowd?  Can we account for this difference in judgment-free evaluation?  Until we can I can only be skeptical: what are we really measuring when we estimate performance using Cranfield-type methods without relevance judgments?