TREC发展历程及现状分析

依据TREC会议集对历年参与团队与项目进行了统计,重点介绍了中国的TREC历程、TREC-16新推出的Million Query Track,指明了TREC三个未来关注焦点：非正式交流信息、特定学科领域以及用户交互。认为国内研究者应更加关注TREC以及中文语料库的建设。     

Threshold Setting and Performance Optimization in Adaptive Filtering

An experimental adaptive filtering system, built on the Okapi search engine, is described. In addition to the regular text retrieval functions, the system requires a complex set of procedures for setting score thresholds and adapting them following feedback. These procedures need to be closely related to the evaluation measures to be used. A mixture of quantitative methods relating a threshold to the number of documents expected to be retrieved in a time period, and qualitative methods relating to the probability of relevance, is defined. Experiments under the TREC-9 Adaptive Filtering Track rules are reported. The system is seen to perform reasonably well in comparison with other systems at TREC. Some of the variables that may affect performance are investigated.     

Comparing the Performance of Adaptive Filtering and Ranked Output Systems

Some insight into the behavior of adaptive filtering systems may be gained by comparing them with similar ranked-output retrieval systems. This is not easy; however, a new optimization measure, introduced for the TREC-9 filtering track, makes some such comparison possible. A series of experiments using the TREC-9 filtering data shows that filtering effectiveness is comparable to routing effectiveness, and demonstrates the gains to be made from adaptation.     

开源搜索引擎索引性能的比较研究

索引性能是衡量搜索引擎优劣的一个重要指标。文章通过实验的方法,用TREC-4和WT10g数据集合分别对12种主流开源搜索引擎做了两组索引测试,记录了它们在索引不同大小数据集合时的时间开销和资源消耗情况,并根据实验数据对这些搜索引擎的索引性能进行了比较和评估。     

Methods for automatically evaluating answers to complex questions

Evaluation is a major driving force in advancing the state of the art in language technologies. In particular, methods for automatically assessing the quality of machine output is the preferred method for measuring progress, provided that these metrics have been validated against human judgments. Following recent developments in the automatic evaluation of machine translation and document summarization, we present a similar approach, implemented in a measure called POURPRE, an automatic technique for evaluating answers to complex questions based on n-gram co-occurrences between machine output and a human-generated answer key. Until now, the only way to assess the correctness of answers to such questions involves manual determination of whether an information “nugget” appears in a system's response. The lack of automatic methods for scoring system output is an impediment to progress in the field, which we address with this work. Experiments with the TREC 2003, TREC 2004, and TREC 2005 QA tracks indicate that rankings produced by our metric correlate highly with official rankings, and that POURPRE outperforms direct application of existing metrics.

A goodness of fit test approach in information retrieval

In many probabilistic modeling approaches to Information Retrieval we are interested in estimating how well a document model “fits” the user’s information need (query model). On the other hand in statistics, goodness of fit tests are well established techniques for assessing the assumptions about the underlying distribution of a data set. Supposing that the query terms are randomly distributed in the various documents of the collection, we actually want to know whether the occurrences of the query terms are more frequently distributed by chance in a particular document. This can be quantified by the so-called goodness of fit tests. In this paper, we present a new document ranking technique based on Chi-square goodness of fit tests. Given the null hypothesis that there is no association between the query terms q and the document d irrespective of any chance occurrences, we perform a Chi-square goodness of fit test for assessing this hypothesis and calculate the corresponding Chi-square values. Our retrieval formula is based on ranking the documents in the collection according to these calculated Chi-square values. The method was evaluated over the entire test collection of TREC data, on disks 4 and 5, using the topics of TREC-7 and TREC-8 (50 topics each) conferences. It performs well, outperforming steadily the classical OKAPI term frequency weighting formula but below that of KL-Divergence from language modeling approach. Despite this, we believe that the technique is an important non-parametric way of thinking of retrieval, offering the possibility to try simple alternative retrieval formulas within goodness-of-fit statistical tests’ framework, modeling the data in various ways estimating or assigning any arbitrary theoretical distribution in terms.     

Scaling Up the TREC Collection

Due to the popularity of Web search engines, a large proportion of real text retrieval queries are now processed over collections measured in tens or hundreds of gigabytes. A new Very Large test Collection (VLC) has been created to support qualification, measurement and comparison of systems operating at this level and to permit the study of the properties of very large collections. The VLC is an extension of the well-known TREC collection and has been distributed under the same conditions. A simple set of efficiency and effectiveness measures have been defined to encourage comparability of reporting. The 20 gigabyte first-edition of the VLC and a representative 10% sample have been used in a special interest track of the 1997 Text Retrieval Conference (TREC-6). The unaffordable cost of obtaining complete relevance assessments over collections of this scale is avoided by concentrating on early precision and relying on the core TREC collection to support detailed effectiveness studies. Results obtained by TREC-6 VLC track participants are presented here. All groups observed a significant increase in early precision as collection size increased. Explanatory hypotheses are advanced for future empirical testing. A 100 gigabyte second edition VLC (VLC2) has recently been compiled and distributed for use in TREC-7 in 1998.     

Graph-based term weighting for information retrieval

A standard approach to Information Retrieval (IR) is to model text as a bag of words. Alternatively, text can be modelled as a graph, whose vertices represent words, and whose edges represent relations between the words, defined on the basis of any meaningful statistical or linguistic relation. Given such a text graph, graph theoretic computations can be applied to measure various properties of the graph, and hence of the text. This work explores the usefulness of such graph-based text representations for IR. Specifically, we propose a principled graph-theoretic approach of (1) computing term weights and (2) integrating discourse aspects into retrieval. Given a text graph, whose vertices denote terms linked by co-occurrence and grammatical modification, we use graph ranking computations (e.g. PageRank Page et al. in The pagerank citation ranking: Bringing order to the Web. Technical report, Stanford Digital Library Technologies Project, 1998) to derive weights for each vertex, i.e. term weights, which we use to rank documents against queries. We reason that our graph-based term weights do not necessarily need to be normalised by document length (unlike existing term weights) because they are already scaled by their graph-ranking computation. This is a departure from existing IR ranking functions, and we experimentally show that it performs comparably to a tuned ranking baseline, such as BM25 (Robertson et al. in NIST Special Publication 500-236: TREC-4, 1995). In addition, we integrate into ranking graph properties, such as the average path length, or clustering coefficient, which represent different aspects of the topology of the graph, and by extension of the document represented as a graph. Integrating such properties into ranking allows us to consider issues such as discourse coherence, flow and density during retrieval. We experimentally show that this type of ranking performs comparably to BM25, and can even outperform it, across different TREC (Voorhees and Harman in TREC: Experiment and evaluation in information retrieval, MIT Press, 2005) datasets and evaluation measures.     

Evolving local and global weighting schemes in information retrieval

This paper describes a method, using Genetic Programming, to automatically determine term weighting schemes for the vector space model. Based on a set of queries and their human determined relevant documents, weighting schemes are evolved which achieve a high average precision. In Information Retrieval (IR) systems, useful information for term weighting schemes is available from the query, individual documents and the collection as a whole. We evolve term weighting schemes in both local (within-document) and global (collection-wide) domains which interact with each other correctly to achieve a high average precision. These weighting schemes are tested on well-known test collections and are compared to the traditional tf-idf weighting scheme and to the BM25 weighting scheme using standard IR performance metrics. Furthermore, we show that the global weighting schemes evolved on small collections also increase average precision on larger TREC data. These global weighting schemes are shown to adhere to Luhn’s resolving power as both high and low frequency terms are assigned low weights. However, the local weightings evolved on small collections do not perform as well on large collections. We conclude that in order to evolve improved local (within-document) weighting schemes it is necessary to evolve these on large collections.     

The effect of pooling and evaluation depth on IR metrics

Batch IR evaluations are usually performed in a framework that consists of a document collection, a set of queries, a set of relevance judgments, and one or more effectiveness metrics. A large number of evaluation metrics have been proposed, with two primary families having emerged: recall-based metrics, and utility-based metrics. In both families, the pragmatics of forming judgments mean that it is usual to evaluate the metric to some chosen depth such as \(k=20\) or \(k=100\), without necessarily fully considering the ramifications associated with that choice. Our aim is this paper is to explore the relative risks arising with fixed-depth evaluation in the two families, and document the complex interplay between metric evaluation depth and judgment pooling depth. Using a range of TREC resources including NewsWire data and the ClueWeb collection, we: (1) examine the implications of finite pooling on the subsequent usefulness of different test collections, including specifying options for truncated evaluation; and (2) determine the extent to which various metrics correlate with themselves when computed to different evaluation depths using those judgments. We demonstrate that the judgment pools constructed for the ClueWeb collections lack resilience, and are suited primarily to the application of top-heavy utility-based metrics rather than recall-based metrics; and that on the majority of the established test collections, and across a range of evaluation depths, recall-based metrics tend to be more volatile in the system rankings they generate than are utility-based metrics. That is, experimentation using utility-based metrics is more robust to choices such as the evaluation depth employed than is experimentation using recall-based metrics. This distinction should be noted by researchers as they plan and execute system-versus-system retrieval experiments.     

Click-based evidence for decaying weight distributions in search effectiveness metrics

Search effectiveness metrics are used to evaluate the quality of the answer lists returned by search services, usually based on a set of relevance judgments. One plausible way of calculating an effectiveness score for a system run is to compute the inner-product of the run’s relevance vector and a “utility” vector, where the ith element in the utility vector represents the relative benefit obtained by the user of the system if they encounter a relevant document at depth i in the ranking. This paper uses such a framework to examine the user behavior patterns—and hence utility weightings—that can be inferred from a web query log. We describe a process for extrapolating user observations from query log clickthroughs, and employ this user model to measure the quality of effectiveness weighting distributions. Our results show that for measures with static distributions (that is, utility weighting schemes for which the weight vector is independent of the relevance vector), the geometric weighting model employed in the rank-biased precision effectiveness metric offers the closest fit to the user observation model. In addition, using past TREC data as to indicate likelihood of relevance, we also show that the distributions employed in the BPref and MRR metrics are the best fit out of the measures for which static distributions do not exist.     

Using Corpus-Based Approaches in a System for Multilingual Information Retrieval

We present a system for multilingual information retrieval that allows users to formulate queries in their preferred language and retrieve relevant information from a collection containing documents in multiple languages. The system is based on a process of document level alignments, where documents of different languages are paired according to their similarity. The resulting mapping allows us to produce a multilingual comparable corpus. Such a corpus has multiple interesting applications. It allows us to build a data structure for query translation in cross-language information retrieval (CLIR). Moreover, we also perform pseudo relevance feedback on the alignments to improve our retrieval results. And finally, multiple retrieval runs can be merged into one unified result list. The resulting system is inexpensive, adaptable to domain-specific collections and new languages and has performed very well at the TREC-7 conference CLIR system comparison.     

Universality of scholarly impact metrics

Given the growing use of impact metrics in the evaluation of scholars, journals, academic institutions, and even countries, there is a critical need for means to compare scientific impact across disciplinary boundaries. Unfortunately, citation-based metrics are strongly biased by diverse field sizes and publication and citation practices. As a result, we have witnessed an explosion in the number of newly proposed metrics that claim to be “universal.” However, there is currently no way to objectively assess whether a normalized metric can actually compensate for disciplinary bias. We introduce a new method to assess the universality of any scholarly impact metric, and apply it to evaluate a number of established metrics. We also define a very simple new metric h_s, which proves to be universal, thus allowing to compare the impact of scholars across scientific disciplines. These results move us closer to a formal methodology in the measure of scholarly impact.     

Construction of query concepts based on feature clustering of documents

In Information Retrieval, since it is hard to identify users’ information needs, many approaches have been tried to solve this problem by expanding initial queries and reweighting the terms in the expanded queries using users’ relevance judgments. Although relevance feedback is most effective when relevance information about retrieved documents is provided by users, it is not always available. Another solution is to use correlated terms for query expansion. The main problem with this approach is how to construct the term-term correlations that can be used effectively to improve retrieval performance. In this study, we try to construct query concepts that denote users’ information needs from a document space, rather than to reformulate initial queries using the term correlations and/or users’ relevance feedback. To form query concepts, we extract features from each document, and then cluster the features into primitive concepts that are then used to form query concepts. Experiments are performed on the Associated Press (AP) dataset taken from the TREC collection. The experimental evaluation shows that our proposed framework called QCM (Query Concept Method) outperforms baseline probabilistic retrieval model on TREC retrieval.     

Detection As Multi-Topic Tracking

The topic tracking task from TDT is a variant of information filtering tasks that focuses on event-based topics in streams of broadcast news. In this study, we compare tracking to another TDT task, detection, which has the goal of partitioning all arriving news into topics, regardless of whether the topics are of interest to anyone, and even when a new topic appears that had not been previous anticipated. There are clear relationships between the two tasks (under some assumptions, a perfect tracking system could solve the detection problem), but they are evaluated quite differently. We describe the two tasks and discuss their similarities. We show how viewing detection as a form of multi-topic parallel tracking can illuminate the performance tradeoffs of detection over tracking.     

Diversified search evaluation: lessons from the NTCIR-9 INTENT task

The evaluation of diversified web search results is a relatively new research topic and is not as well-understood as the time-honoured evaluation methodology of traditional IR based on precision and recall. In diversity evaluation, one topic may have more than one intent, and systems are expected to balance relevance and diversity. The recent NTCIR-9 evaluation workshop launched a new task called INTENT which included a diversified web search subtask that differs from the TREC web diversity task in several aspects: the choice of evaluation metrics, the use of intent popularity and per-intent graded relevance, and the use of topic sets that are twice as large as those of TREC. The objective of this study is to examine whether these differences are useful, using the actual data recently obtained from the NTCIR-9 INTENT task. Our main experimental findings are: (1) The $\hbox{D}\,\sharp$ evaluation framework used at NTCIR provides more “intuitive” and statistically reliable results than Intent-Aware Expected Reciprocal Rank; (2) Utilising both intent popularity and per-intent graded relevance as is done at NTCIR tends to improve discriminative power, particularly for $\hbox{D}\,\sharp$ -nDCG; and (3) Reducing the topic set size, even by just 10 topics, can affect not only significance testing but also the entire system ranking; when 50 topics are used (as in TREC) instead of 100 (as in NTCIR), the system ranking can be substantially different from the original ranking and the discriminative power can be halved. These results suggest that the directions being explored at NTCIR are valuable.     

基于TREC目标优化的过滤阈值调整算法

检出阈值的优化调整是自适应信息过滤的重点和难点之一。分析现有的阈值调整方法中普遍存在的问题,以TREC效用指标为目标函数,对阈值调整方法中的极大似然估计法和局部优化法进行比较分析,提出基于TREC目标优化的全局极大似然估计法与局部效用指标优化相结合的自适应过滤阈值调整算法。实验结果表明该方法能有效地提高信息过滤系统的性能。     

On information retrieval metrics designed for evaluation with incomplete relevance assessments

Modern information retrieval (IR) test collections have grown in size, but the available manpower for relevance assessments has more or less remained constant. Hence, how to reliably evaluate and compare IR systems using incomplete relevance data, where many documents exist that were never examined by the relevance assessors, is receiving a lot of attention. This article compares the robustness of IR metrics to incomplete relevance assessments, using four different sets of graded-relevance test collections with submitted runs—the TREC 2003 and 2004 robust track data and the NTCIR-6 Japanese and Chinese IR data from the crosslingual task. Following previous work, we artificially reduce the original relevance data to simulate IR evaluation environments with extremely incomplete relevance data. We then investigate the effect of this reduction on discriminative power, which we define as the proportion of system pairs with a statistically significant difference for a given probability of Type I Error, and on Kendall’s rank correlation, which reflects the overall resemblance of two system rankings according to two different metrics or two different relevance data sets. According to these experiments, Q′, nDCG′ and AP′ proposed by Sakai are superior to bpref proposed by Buckley and Voorhees and to Rank-Biased Precision proposed by Moffat and Zobel. We also point out some weaknesses of bpref and Rank-Biased Precision by examining their formal definitions.

The TREC-5 Confusion Track: Comparing Retrieval Methods for Scanned Text

A known-item search is a particular information retrieval task in which the system is asked to find a single target document in a large document set. The TREC-5 confusion track used a set of 49 known-item tasks to study the impact of data corruption on retrieval system performance. Two corrupted versions of a 55,600 document corpus whose true content was known were created by applying OCR techniques to page images. The first version of the corpus used the page images as scanned, resulting in an estimated character error rate of approximately 5%. The second version used page images that had been down-sampled, resulting in an estimated character error rate of approximately 20%. The true text and each of the corrupted versions were then searched using the same set of 49 questions. In general, retrieval methods that attempted a probabilistic reconstruction of the original clean text fared better than methods that simply accepted corrupted versions of the query text.