TREC Collections Research Articles

Conditional variational autoencoder for query expansion in ad-hoc information retrieval

Query expansion (QE) is commonly used to improve the performance of traditional information retrieval (IR) models. With the adoption of deep learning in IR research, neural QE models have emerged in recent years. Many of these models focus on learning embeddings by leveraging query-document relevance. These embedding models allow computing semantic similarities between queries and documents to generate expansion terms. However, existing models often ignore query-document interactions. This research aims to address that gap by proposing a QE model using a conditional variational autoencoder. It first maps a query-document pair into a latent space based on their interaction, then estimates an expansion model from that latent space. The proposed model is trained on relevance feedback data and generates expansions using pseudo-relevance feedback at test time. The proposed model is evaluated on three standard TREC collections for document ranking: AP and Robust 04 and GOV02, and the MS MARCO dataset for passage ranking. Results show the model outperforms state-of-the-art traditional and neural QE models. It also demonstrates higher additivity with neural matching than baselines.

Pseudo-Relevance Feedback Method Based on the Topic Relevance Model

In the field of information retrieval, most pseudo-relevance feedback models select candidate terms from the top k documents returned by the first-pass retrieval, but they cannot identify the reliability of these documents. This paper proposed a new approach to obtain feedback information more comprehensively by constructing four corresponding models. Firstly, the algorithm incorporated topic-based relevance information into the relevance model RM3 and constructed a topic-based relevance model, denoted as TopRM3, with two corresponding variants. TopRM3 estimated the reliability of a feedback document in language modeling when executing pseudo-relevance feedback from both term and topic-based perspectives. Secondly, the algorithm introduced topic-based relevance information into Rocchio’s model and constructed the corresponding model, denoted as TopRoc, with two corresponding variants. Experimental results on the five TREC collections show that the proposed TopRM3 and TopRoc are effective and generally superior to the state-of-the-art pseudo-relevance feedback models with optimal parameter settings in terms of mean average precision.

A retrieval model family based on the probability ranking principle for ad hoc retrieval

AbstractMany successful retrieval models are derived based on or conform to the probability ranking principle (PRP). We present a new derivation of a document ranking function given by the probability of relevance of a document, conforming to the PRP. Our formulation yields a family of retrieval models, called probabilistic binary relevance (PBR) models, with various instantiations obtained by different probability estimations. By extensive experiments on a range of TREC collections, improvement of the PBR models over some established baselines with statistical significance is observed, especially in the large Clueweb09 Cat‐B collection.

Learning deep relevance couplings for ad-hoc document retrieval

Modeling the relevance matching between a query and a document is challenging in ad-hoc document retrieval due to diverse semantic matching aspects. Both explicit and implicit semantic matching aspects exist, corresponding to the query term occurrences in a document and the term transformations between queries and documents respectively. While most of existing neural retrieval models involve both explicit and implicit matching aspects, they do not clearly and effectively distinguish and couple explicit and implicit matching signals, resulting in either only capturing the local knowledge of query topics or lacking of in-depth understanding of user query need. In this work, we propose a deep relevance coupling model (DRC) for document retrieval to capture the diverse semantic matching aspects. DRC effectively learns and couples three aspects: explicit semantic learning, implicit semantic learning, and relevance coupling. The explicit semantic learning employs a frequency layer and a probability layer to obtain the exact term-level relevance signals between queries and documents. The implicit semantic learning adopts a transformation layer and a dependency layer to automatically capture the implicit term-level relevance signals between queries and documents. The relevance coupling transforms the term-level signals to the query-document matching signals by using a term weight layer and then integrates all the query-document matching signals into the final relevance scores between queries and documents. Experiments conducted on the TREC collections demonstrate that the DRC model delivers better retrieval accuracy and robustness than the baseline methods.

Effectiveness evaluation without human relevance judgments: A systematic analysis of existing methods and of their combinations

In test collection based evaluation of retrieval effectiveness, it has been suggested to completely avoid using human relevance judgments. Although several methods have been proposed, their accuracy is still limited. In this paper we present two overall contributions. First, we provide a systematic comparison of all the most widely adopted previous approaches on a large set of 14 TREC collections. We aim at analyzing the methods in a homogeneous and complete way, in terms of the accuracy measures used as well as in terms of the datasets selected, showing that considerably different results may be achieved considering different methods, datasets, and measures. Second, we study the combination of such methods, which, to the best of our knowledge, has not been investigated so far. Our experimental results show that simple combination strategies based on data fusion techniques are usually not effective and even harmful. However, some more sophisticated solutions, based on machine learning, are indeed effective and often outperform all individual methods. Moreover, they are more stable, as they show a smaller variation across datasets. Our results have the practical implication that, when trying to automatically evaluate retrieval effectiveness, researchers should not use a single method, but a (machine-learning based) combination of them.

How do interval scales help us with better understanding IR evaluation measures?

Evaluation measures are the basis for quantifying the performance of IR systems and the way in which their values can be processed to perform statistical analyses depends on the scales on which these measures are defined. For example, mean and variance should be computed only when relying on interval scales. In our previous work we defined a theory of IR evaluation measures, based on the representational theory of measurement, which allowed us to determine whether and when IR measures are interval scales. We found that common set-based retrieval measures—namely precision, recall, and F-measure—always are interval scales in the case of binary relevance while this does not happen in the multi-graded relevance case. In the case of rank-based retrieval measures—namely AP, gRBP, DCG, and ERR—only gRBP is an interval scale when we choose a specific value of the parameter p and define a specific total order among systems while all the other IR measures are not interval scales. In this work, we build on our previous findings and we carry out an extensive evaluation, based on standard TREC collections, to study how our theoretical findings impact on the experimental ones. In particular, we conduct a correlation analysis to study the relationship among the above-mentioned state-of-the-art evaluation measures and their scales. We study how the scales of evaluation measures impact on non parametric and parametric statistical tests for multiple comparisons of IR system performance. Finally, we analyse how incomplete information and pool downsampling affect different scales and evaluation measures.

Binary Independence Language Model in a Relevance Feedback Environment

Model construction is a kind of knowledge engineering, and building retrieval models is critical to the success of search engines. This article proposes a new (retrieval) language model, called binary independence language model (BILM). It integrates two document-context based language models together into one by the log-odds ratio where these two are language models applied to describe document-contexts of query terms. One model is based on relevance information while the other is based on the non-relevance information. Each model incorporates link dependencies and multiple query term dependencies. The probabilities are interpolated between the relative frequency and the background probabilities. In a simulated relevance feedback environment of top 20 judged documents, our BILM performed statistically significantly better than the other highly effective retrieval models at 95% confidence level across four TREC collections using fixed parameter values for the mean average precision. For the less stable performance measure (i.e. precision at the top 10), no statistical significance is shown between the different models for the individual test collections although numerically our BILM is better than two other models with a confidence level of 95% based on a paired sign test across the test collections of both relevance feedback and retrospective experiments.

A novel unsupervised corpus-based stemming technique using lexicon and corpus statistics

Word Stemming is a widely used mechanism in the fields of Natural Language Processing, Information Retrieval, and Language Modeling. Language-independent stemmers discover classes of morphologically related words from the ambient corpus without using any language related rules. In this article, we proposed a fully unsupervised language-independent text stemming technique that clusters morphologically related words from the corpus of the language using both lexical and co-occurrence features such as lexical similarity, suffix knowledge, and co-occurrence similarity. The method applies to a wide range of inflectional languages as it identifies morphological variants formed through different linguistic processes such as affixation, compounding, conversion, etc.The proposed approach has been tested in Information Retrieval application for four languages (English, Marathi, Hungarian, and Bengali) using standard TREC, CLEF, and FIRE test collections. A significant improvement over word-based retrieval, five other corpus-based stemmers, and rule-based stemmers has been achieved in all the languages. Besides, information retrieval, the proposed approach has also been tested in text classification and inflection removal tasks. Our algorithm excelled over other baseline methods in all the test scenarios. Thus, we successfully achieved the objective of developing a multipurpose stemming algorithm that cannot only be used for information retrieval task but also for non-traditional tasks such as text classification, sentiment analysis, inflection removal, etc.

Fewer topics? A million topics? Both?! On topics subsets in test collections

When evaluating IR run effectiveness using a test collection, a key question is: What search topics should be used? We explore what happens to measurement accuracy when the number of topics in a test collection is reduced, using the Million Query 2007, TeraByte 2006, and Robust 2004 TREC collections, which all feature more than 50 topics, something that has not been examined in past work. Our analysis finds that a subset of topics can be found that is as accurate as the full topic set at ranking runs. Further, we show that the size of the subset, relative to the full topic set, can be substantially smaller than was shown in past work. We also study the topic subsets in the context of the power of statistical significance tests. We find that there is a trade off with using such sets in that significant results may be missed, but the loss of statistical significance is much smaller than when selecting random subsets. We also find topic subsets that can result in a low accuracy test collection, even when the number of queries in the subset is quite large. These negatively correlated subsets suggest we still lack good methodologies which provide stability guarantees on topic selection in new collections. Finally, we examine whether clustering of topics is an appropriate strategy to find and characterize good topic subsets. Our results contribute to the understanding of information retrieval effectiveness evaluation, and offer insights for the construction of test collections.

Modeling positive and negative feedback for improving document retrieval

Pseudo-relevance feedback (PRF) has evident potential for enriching the representation of short queries. Traditional PRF methods treat top-ranked documents as feedback, since they are assumed to be relevant to the query. However, some of these feedback documents may actually distract from the query topic for a range of reasons and accordingly downgrade PRF system performance. Such documents constitute negative examples (negative feedback) but could also be valuable in retrieval. In this paper, a novel framework of query language model construction is proposed in order to improve retrieval performance by integrating both positive and negative feedback. First, an improvement-based method is proposed to automatically identify the types of feedback documents (i.e. positive or negative) according to whether the document enhances the retrieval’s effectiveness. Subsequently, based on the learned positive and negative examples, the positive feedback models and the negative feedback models are estimated using an Expectation-Maximization algorithm with the assumptions: the positive term distribution is affected by the context term distribution and the negative term distribution is affected by both the positive term distribution and the context term distribution (such that the positive feedback model upgrades the rankings of relevant documents and the negative feedback model prunes the irrelevant documents from a query). Finally, a content-based representativeness criterion is proposed in order to obtain the representative negative feedback documents. Experiments conducted on the TREC collections demonstrate that our proposed approach results in better retrieval accuracy and robustness than baseline methods.

A survey of statistical approaches for query expansion

A major issue in effective information retrieval is the problem of vocabulary mismatches. The method called query expansion addresses this issue by reformulating each search query with additional terms that better define the information needs of the user. Many researchers have contributed to improving the accuracy of information retrieval systems, through different approaches to query expansion. In this article, we primarily discuss statistical query expansion approaches that include document analysis, search and browse log analyses, and web knowledge analyses. In addition to proposing a comprehensive classification for these approaches, we also briefly analyse the pros and cons of each technique. Finally, we evaluate these techniques using five functional features and experimental settings such as TREC collection and results of performance metrics. An in-depth survey of different statistical query expansion approaches suggests that the selection of the best approach depends on the type of search query, the nature and availability of data resources, and performance efficiency requirements.

Modeling Queries with Contextual Snippets for Information Retrieval

Query expansion under the pseudo-relevance feedback (PRF) framework has been extensively studied in information retrieval. However, most expansion methods are mainly based on the statistics of single terms, which can generate plenty of irrelevant query terms and decrease retrieval performance. To alleviate this problem, we propose an approach that adapts the PRF-based contextual snippets into a context-aware topic model to enhance query representations. Specifically, instead of selecting a series of independent terms, we make full use of the query contextual information and focus on the snippets with the length of n in the PRF documents. Furthermore, we propose a context-aware topic (CAT) model to mine the topic distributions of the query-relevant snippets, namely, fine contextual snippets. In contrast to the traditional topic models that infer the topics from the whole corpus, we establish a bridge between the snippets and the corresponding PRF documents, which can be used for modeling the topics more precisely and efficiently. Finally, the topic distributions of the fine snippets are used for context-aware and topic-sensitive query representations. To evaluate the performance of our approach, we integrate the obtained queries into a topic-based hybrid retrieval model and conduct extensive experiments on various TREC collections. The experimental results show that our query-modeling approach is more effective in boosting retrieval performance compared with the state-of-the-art methods.

Improving Arabic information retrieval using word embedding similarities

Term mismatch is a common limitation of traditional information retrieval (IR) models where relevance scores are estimated based on exact matching of documents and queries. Typically, good IR model should consider distinct but semantically similar words in the matching process. In this paper, we propose a method to incorporate word embedding (WE) semantic similarities into existing probabilistic IR models for Arabic in order to deal with term mismatch. Experiments are performed on the standard Arabic TREC collection using three neural word embedding models. The results show that extending the existing IR models improves significantly baseline bag-of-words models. Although the proposed extensions significantly outperform their baseline bag-of-words, the difference between the evaluated neural word embedding models is not statistically significant. Moreover, the overall comparison results show that our extensions significantly improve the Arabic WordNet based semantic indexing approach and three recent WE-based IR language models.

Improving Pseudo-Relevance Feedback With Neural Network-Based Word Representations

In information retrieval, query expansion methods, such as pseudo-relevance feedback, are designed to enrich users’ queries with relevant terms for comprehensively interpreting the desired information. One of the key issues for query expansion is how to obtain high-quality expansion terms to capture the information needs. Recent advances in neural network language models have demonstrated that these models can learn powerful distributed word representations, which have been successfully applied to solve various natural language processing tasks. In this paper, we propose a novel query expansion framework based on neural network-based word representations. Our framework first selects abundant candidate expansion terms using a modified term-dependency method and then generates term features for candidate terms based on word representations to encode relationships between given queries and corresponding candidate terms. Furthermore, we adopt learning-to-rank methods to train term-ranking models with the generated features for term refinement. We conduct extensive experiments to examine the performance of the learned term-ranking models and compare the effectiveness of the representation-based and context-based features for selecting relevant expansion terms. Experimental results using four TREC collections show that neural network-based word representations are effective in query expansion and can significantly improve retrieval performance.

EveTAR: building a large-scale multi-task test collection over Arabic tweets

This article introduces a new language-independent approach for creating a large-scale high-quality test collection of tweets that supports multiple information retrieval (IR) tasks without running a shared-task campaign. The adopted approach (demonstrated over Arabic tweets) designs the collection around significant (i.e., popular) events, which enables the development of topics that represent frequent information needs of Twitter users for which rich content exists. That inherently facilitates the support of multiple tasks that generally revolve around events, namely event detection, ad-hoc search, timeline generation, and real-time summarization. The key highlights of the approach include diversifying the judgment pool via interactive search and multiple manually-crafted queries per topic, collecting high-quality annotations via crowd-workers for relevancy and in-house annotators for novelty, filtering out low-agreement topics and inaccessible tweets, and providing multiple subsets of the collection for better availability. Applying our methodology on Arabic tweets resulted in EveTAR, the first freely-available tweet test collection for multiple IR tasks. EveTAR includes a crawl of 355M Arabic tweets and covers 50 significant events for which about 62K tweets were judged with substantial average inter-annotator agreement (Kappa value of 0.71). We demonstrate the usability of EveTAR by evaluating existing algorithms in the respective tasks. Results indicate that the new collection can support reliable ranking of IR systems that is comparable to similar TREC collections, while providing strong baseline results for future studies over Arabic tweets.

Verbosity normalized pseudo-relevance feedback in information retrieval

Document length normalization is one of the fundamental components in a retrieval model because term frequencies can readily be increased in long documents. The key hypotheses in literature regarding document length normalization are the verbosity and scope hypotheses, which imply that document length normalization should consider the distinguishing effects of verbosity and scope on term frequencies. In this article, we extend these hypotheses in a pseudo-relevance feedback setting by assuming the verbosity hypothesis on the feedback query model, which states that the verbosity of an expanded query should not be high. Furthermore, we postulate the following two effects of document verbosity on a feedback query model that easily and typically holds in modern pseudo-relevance feedback methods: 1) the verbosity-preserving effect: the query verbosity of a feedback query model is determined by feedback document verbosities; 2) the verbosity-sensitive effect: highly verbose documents more significantly and unfairly affect the resulting query model than normal documents do. By considering these effects, we propose verbosity normalized pseudo-relevance feedback, which is straightforwardly obtained by replacing original term frequencies with their verbosity-normalized term frequencies in the pseudo-relevance feedback method. The results of the experiments performed on three standard TREC collections show that the proposed verbosity normalized pseudo-relevance feedback consistently provides statistically significant improvements over conventional methods, under the settings of the relevance model and latent concept expansion.

Toward an anatomy of IR system component performances

Information retrieval (IR) systems are the prominent means for searching and accessing huge amounts of unstructured information on the web and elsewhere. They are complex systems, constituted by many different components interacting together, and evaluation is crucial to both tune and improve them. Nevertheless, in the current evaluation methodology, there is still no way to determine how much each component contributes to the overall performances and how the components interact together. This hampers the possibility of a deep understanding of IR system behavior and, in turn, prevents us from designing ahead which components are best suited to work together for a specific search task. In this paper, we move the evaluation methodology one step forward by overcoming these barriers and beginning to devise an “anatomy” of IR systems and their internals. In particular, we propose a methodology based on the General Linear Mixed Model (GLMM) and analysis of variance (ANOVA) to develop statistical models able to isolate system variance and component effects as well as their interaction, by relying on a grid of points (GoP) containing all the combinations of the analyzed components. We apply the proposed methodology to the analysis of two relevant search tasks—news search and web search—by using standard TREC collections. We analyze the basic set of components typically part of an IR system, namely, stop lists, stemmers, and n‐grams, and IR models. In this way, we derive insights about English text retrieval.

AWARE

We propose the Assessor-driven Weighted Averages for Retrieval Evaluation (AWARE) probabilistic framework, a novel methodology for dealing with multiple crowd assessors that may be contradictory and/or noisy. By modeling relevance judgements and crowd assessors as sources of uncertainty, AWARE takes the expectation of a generic performance measure, like Average Precision, composed with these random variables. In this way, it approaches the problem of aggregating different crowd assessors from a new perspective, that is, directly combining the performance measures computed on the ground truth generated by the crowd assessors instead of adopting some classification technique to merge the labels produced by them. We propose several unsupervised estimators that instantiate the AWARE framework and we compare them with state-of-the-art approaches, that is,Majoriity Vote and Expectation Maximization, on TREC collections. We found that AWARE approaches improve in terms of their capability of correctly ranking systems and predicting their actual performance scores.

Using Replicates in Information Retrieval Evaluation.

This article explores a method for more accurately estimating the main effect of the system in a typical test-collection-based evaluation of information retrieval systems, thus increasing the sensitivity of system comparisons. Randomly partitioning the test document collection allows for multiple tests of a given system and topic (replicates). Bootstrap ANOVA can use these replicates to extract system-topic interactions-something not possible without replicates-yielding a more precise value for the system effect and a narrower confidence interval around that value. Experiments using multiple TREC collections demonstrate that removing the topic-system interactions substantially reduces the confidence intervals around the system effect as well as increases the number of significant pairwise differences found. Further, the method is robust against small changes in the number of partitions used, against variability in the documents that constitute the partitions, and the measure of effectiveness used to quantify system effectiveness.

Document Language Models, Query Models, and Risk Minimization for Information Retrieval

We present a framework for information retrieval that combines document models and query models using a probabilistic ranking function based on Bayesian decision theory. The framework suggests an operational retrieval model that extends recent developments in the language modeling approach to information retrieval. A language model for each document is estimated, as well as a language model for each query, and the retrieval problem is cast in terms of risk minimization. The query language model can be exploited to model user preferences, the context of a query, synonomy and word senses. While recent work has incorporated word translation models for this purpose, we introduce a new method using Markov chains defined on a set of documents to estimate the query models. The Markov chain method has connections to algorithms from link analysis and social networks. The new approach is evaluated on TREC collections and compared to the basic language modeling approach and vector space models together with query expansion using Rocchio. Significant improvements are obtained over standard query expansion methods for strong baseline TF-IDF systems, with the greatest improvements attained for short queries on Web data.