A bias–variance evaluation framework for information retrieval systems

In information retrieval (IR), the improvement of the effectiveness often sacrifices the stability of an IR system. To evaluate the stability, many risk-sensitive metrics have been proposed. Since the theoretical limitations, the current works study the effectiveness and stability separately, and have not explored the effectiveness–stability tradeoff. In this paper, we propose a Bias–Variance Tradeoff Evaluation (BV-Test) framework, based on the bias–variance decomposition of the mean squared error, to measure the overall performance (considering both effectiveness and stability) and the tradeoff between effectiveness and stability of a system. In this framework, we define generalized bias–variance metrics, based on the Cranfield-style experiment set-up where the document collection is fixed (across topics) or the set-up where document collection is a sample (per-topic). Compared with risk-sensitive evaluation methods, our work not only measures the effectiveness–stability tradeoff of a system, but also effectively tracks the source of system instability. Experiments on TREC Ad-hoc track (1993–1999) and Web track (2010–2014) show a clear effectiveness–stability tradeoff across topics and per-topic, and topic grouping and max–min normalization can effectively reduce the bias–variance tradeoff. Experimental results on TREC Session track (2010–2012) also show that the query reformulation and increase of user data are beneficial to both effectiveness and stability simultaneously.     

Multilingual sentence categorization and novelty mining

A challenge for sentence categorization and novelty mining is to detect not only when text is relevant to the user’s information need, but also when it contains something new which the user has not seen before. It involves two tasks that need to be solved. The first is identifying relevant sentences (categorization) and the second is identifying new information from those relevant sentences (novelty mining). Many previous studies of relevant sentence retrieval and novelty mining have been conducted on the English language, but few papers have addressed the problem of multilingual sentence categorization and novelty mining. This is an important issue in global business environments, where mining knowledge from text in a single language is not sufficient. In this paper, we perform the first task by categorizing Malay and Chinese sentences, then comparing their performances with that of English. Thereafter, we conduct novelty mining to identify the sentences with new information. Experimental results on TREC 2004 Novelty Track data show similar categorization performance on Malay and English sentences, which greatly outperform Chinese. In the second task, it is observed that we can achieve similar novelty mining results for all three languages, which indicates that our algorithm is suitable for novelty mining of multilingual sentences. In addition, after benchmarking our results with novelty mining without categorization, it is learnt that categorization is necessary for the successful performance of novelty mining.     

Using language models to improve opinion detection

Opinion mining is one of the most important research tasks in the information retrieval research community. With the huge volume of opinionated data available on the Web, approaches must be developed to differentiate opinion from fact. In this paper, we present a lexicon-based approach for opinion retrieval. Generally, opinion retrieval consists of two stages: relevance to the query and opinion detection. In our work, we focus on the second state which itself focusses on detecting opinionated documents . We compare the document to be analyzed with opinionated sources that contain subjective information. We hypothesize that a document with a strong similarity to opinionated sources is more likely to be opinionated itself. Typical lexicon-based approaches treat and choose their opinion sources according to their test collection, then calculate the opinion score based on the frequency of subjective terms in the document. In our work, we use different open opinion collections without any specific treatment and consider them as a reference collection. We then use language models to determine opinion scores. The analysis document and reference collection are represented by different language models (i.e., Dirichlet, Jelinek-Mercer and two-stage models). These language models are generally used in information retrieval to represent the relationship between documents and queries. However, in our study, we modify these language models to represent opinionated documents. We carry out several experiments using Text REtrieval Conference (TREC) Blogs 06 as our analysis collection and Internet Movie Data Bases (IMDB), Multi-Perspective Question Answering (MPQA) and CHESLY as our reference collection. To improve opinion detection, we study the impact of using different language models to represent the document and reference collection alongside different combinations of opinion and retrieval scores. We then use this data to deduce the best opinion detection models. Using the best models, our approach improves on the best baseline of TREC Blog (baseline4) by 30%.     

Indices of novelty for emerging topic detection

Emerging topic detection is a vital research area for researchers and scholars interested in searching for and tracking new research trends and topics. The current methods of text mining and data mining used for this purpose focus only on the frequency of which subjects are mentioned, and ignore the novelty of the subject which is also critical, but beyond the scope of a frequency study. This work tackles this inadequacy to propose a new set of indices for emerging topic detection. They are the novelty index (NI) and the published volume index (PVI). This new set of indices is created based on time, volume, frequency and represents a resolution to provide a more precise set of prediction indices. They are then utilized to determine the detection point (DP) of new emerging topics. Following the detection point, the intersection decides the worth of a new topic. The algorithms presented in this paper can be used to decide the novelty and life span of an emerging topic in a specific field. The entire comprehensive collection of the ACM Digital Library is examined in the experiments. The application of the NI and PVI gives a promising indication of emerging topics in conferences and journals.     

Improving the retrieval effectiveness of very short queries

This paper describes an automatic approach designed to improve the retrieval effectiveness of very short queries such as those used in web searching. The method is based on the observation that stemming, which is designed to maximize recall, often results in depressed precision. Our approach is based on pseudo-feedback and attempts to increase the number of relevant documents in the pseudo-relevant set by reranking those documents based on the presence of unstemmed query terms in the document text. The original experiments underlying this work were carried out using Smart 11.0 and the lnc.ltc weighting scheme on three sets of documents from the TREC collection with corresponding TREC (title only) topics as queries. (The average length of these queries after stoplisting ranges from 2.4 to 4.5 terms.) Results, evaluated in terms of P@20 and non-interpolated average precision, showed clearly that pseudo-feedback (PF) based on this approach was effective in increasing the number of relevant documents in the top ranks. Subsequent experiments, performed on the same data sets using Smart 13.0 and the improved Lnu.ltu weighting scheme, indicate that these results hold up even over the much higher baseline provided by the new weights. Query drift analysis presents a more detailed picture of the improvements produced by this process.     

Binary and graded relevance in IR evaluations—Comparison of the effects on ranking of IR systems

In this study the rankings of IR systems based on binary and graded relevance in TREC 7 and 8 data are compared. Relevance of a sample TREC results is reassessed using a relevance scale with four levels: non-relevant, marginally relevant, fairly relevant, highly relevant. Twenty-one topics and 90 systems from TREC 7 and 20 topics and 121 systems from TREC 8 form the data. Binary precision, and cumulated gain, discounted cumulated gain and normalised discounted cumulated gain are the measures compared. Different weighting schemes for relevance levels are tested with cumulated gain measures. Kendall’s rank correlations are computed to determine to what extent the rankings produced by different measures are similar. Weighting schemes from binary to emphasising highly relevant documents form a continuum, where the measures correlate strongly in the binary end, and less in the heavily weighted end. The results show the different character of the measures.     

Query reformulation using automatically generated query concepts from a document space

We propose a new query reformulation approach, using a set of query concepts that are introduced to precisely denote the user’s information need. Since a document collection is considered to be a domain which includes latent primitive concepts, we identify those concepts through a local pattern discovery and a global modeling using data mining techniques. For a new query, we select its most associated primitive concepts and choose the most probable interpretations as query concepts. We discuss the issue of constructing the primitive concepts from either the whole corpus or from the retrieved set of documents. Our experiments are performed on the TREC8 collection. The experimental evaluation shows that our approach is as good as current query reformulation approaches, while being particularly effective for poorly performing queries. Moreover, we find that the approach using the primitive concepts generated from the set of retrieved documents leads to the most effective performance.     

A novel emerging topic detection method: A knowledge ecology perspective

Emerging topic detection has attracted considerable attention in recent times. While various detection approaches have been proposed in this field, designing a method for accurately detecting emerging topics remains challenging. This paper introduces the perspective of knowledge ecology to the detection of emerging topics and utilizes author-keywords to represent research topics. More precisely, we first improve the novelty metric and recalculate emergence capabilities based on the “ecostate” and “ecorole” attributes of ecological niches. Then, we take the perspective that keywords are analogous to living bodies and map them to the knowledge ecosystem to construct an emerging topics detection method based on ecological niches (ETDEN). Finally, we conduct in-depth comparative experiments to verify the effectiveness and feasibility of ETDEN using data extracted from scientific literature in the ACM Digital Library database. The results demonstrate that the improved novelty indicator helps to differentiate the novelty values of keywords in the same interval. More importantly, ETDEN performs significantly better performance on three terms: the emergence time point and the growth rate of pre-and post-emergence.     

Web document summarization by exploiting social context with matrix co-factorization

In the context of social media, users usually post relevant information corresponding to the contents of events mentioned in a Web document. This information posses two important values in that (i) it reflects the content of an event and (ii) it shares hidden topics with sentences in the main document. In this paper, we present a novel model to capture the nature of relationships between document sentences and post information (comments or tweets) in sharing hidden topics for summarization of Web documents by utilizing relevant post information. Unlike previous methods which are usually based on hand-crafted features, our approach ranks document sentences and user posts based on their importance to the topics. The sentence-user-post relation is formulated in a share topic matrix, which presents their mutual reinforcement support. Our proposed matrix co-factorization algorithm computes the score of each document sentence and user post and extracts the top ranked document sentences and comments (or tweets) as a summary. We apply the model to the task of summarization on three datasets in two languages, English and Vietnamese, of social context summarization and also on DUC 2004 (a standard corpus of the traditional summarization task). According to the experimental results, our model significantly outperforms the basic matrix factorization and achieves competitive ROUGE-scores with state-of-the-art methods.     

Lexical cohesion and term proximity in document ranking

We demonstrate effective new methods of document ranking based on lexical cohesive relationships between query terms. The proposed methods rely solely on the lexical relationships between original query terms, and do not involve query expansion or relevance feedback. Two types of lexical cohesive relationship information between query terms are used in document ranking: short-distance collocation relationship between query terms, and long-distance relationship, determined by the collocation of query terms with other words. The methods are evaluated on TREC corpora, and show improvements over baseline systems.     

A probabilistic framework for integrating sentence-level semantics via BERT into pseudo-relevance feedback

Existing pseudo-relevance feedback (PRF) methods often divide an original query into individual terms for processing and select expansion terms based on the term frequency, proximity, position, etc. This process may lose some contextual semantic information from the original query. In this work, based on the classic Rocchio model, we propose a probabilistic framework that incorporates sentence-level semantics via Bidirectional Encoder Representations from Transformers (BERT) into PRF. First, we obtain the importance of terms at the term level. Then, we use BERT to interactively encode the query and sentences in the feedback document to acquire the semantic similarity score of a sentence and the query. Next, the semantic scores of different sentences are summed as the term score at the sentence level. Finally, we balance the term-level and sentence-level weights by adjusting factors and combine the terms with the top-k scores to form a new query for the next-round processing. We apply this method to three Rocchio-based models (Rocchio, PRoc2, and KRoc). A series of experiments are conducted based on six official TREC data sets. Various evaluation indicators suggest that the improved models achieve a significant improvement over the corresponding baseline models. Our proposed models provide a promising avenue for incorporating sentence-level semantics into PRF, which is feasible and robust. Through comparison and analysis of a case study, expansion terms obtained from the proposed models are shown to be more semantically consistent with the query.     

TREC概况及其最新发展研究

介绍了文本检索会议的基本情况，包括其研究目标、主要项目和研究内容，然后介绍了TREC-2003新增加的基因学项目的主要内容，最后介绍了TREC主要项目的最新变动情况以及它对国内信息检索评价的启示。     

The effect of user characteristics on search effectiveness in information retrieval

This paper investigates the influence of user characteristics (e.g. search experience and cognitive skills) on user effectiveness. A user study was conducted to investigate this effect, 56 participants completed searches for 56 topics using the TREC test collection. Results indicated that participants with search experience and high cognitive skills were more effective than those with less experience and slower perceptual abilities. However, all users rated themselves with the same level of satisfaction with the search results despite the fact they varied substantially in their effectiveness. Therefore, information retrieval evaluators should take these factors into consideration when investigating the impact of system effectiveness on user effectiveness.     

A diary study of information encountering triggered by visual stimuli on micro-blogging services

Despite various models established to present the process of information encountering (IE), little research has been done on the stimulus that plays an essential role in attracting users’ attention and eliciting the subsequent behavioral responses during an IE process. This study was particularly interested in visual stimuli which are superior to textual ones in enhancing information processing and sensory experience. A diary study of IE was conducted in the context of micro-blogging services. They demonstrate the environmental characteristics conducive to IE and are especially abundant in visual stimuli. A total of 189 valid IE incidents triggered by visual stimuli on a representative micro-blogging service were collected with an online questionnaire created based on the critical incident technique (CIT) and analyzed both quantitatively and qualitatively. As found in this study, most of the visual stimuli triggering IE excluded text or motion, and their comprehensibility and novelty were both perceived to be high while humorousness much lower. The encountered micro-posts covered a wide range of topics and were published by different types of micro-bloggers. When interacting with the posts, the participants sometimes just examined their visual or textual content, but sometimes further captured the posts by liking, reposting, commenting, and following, etc. The significant results indicate that the visual stimuli excluding text and those with higher comprehensibility or humorousness were more likely to induce intense approach to the micro-posts. These findings inform micro-bloggers of the means of engaging the audience in intense interactions with their posts to gain persistent profit or reputation. The combination of diaries and the CIT is effective for data collection in IE research.     

The impact of text browsing on text retrieval performance

The results from a series of three experiments that used Text Retrieval Conference (TREC) data and TREC search topics are compared. These experiments each involved three novel user interfaces (one per experiment). User interfaces that made it easier for users to view text were found to improve recall in all three experiments. A distinction was found between a cluster of subjects (a majority of whom were search experts) who tended to read fewer documents more carefully (readers, or exclusives) and subjects who skimmed through more documents without reading them as carefully (skimmers, or inclusives). Skimmers were found to have significantly better recall overall. A major outcome from our experiments at TREC and with the TREC data, is that hypertext interfaces to information retrieval (IR) tasks tend to increase recall. Our interpretation of this pattern of results across the three experiments is that increased interaction with the text (more pages viewed) generally improves recall. Findings from one of the experiments indicated that viewing a greater diversity of text on a single screen (i.e., not just more text per se, but more articles available at once) may also improve recall. In an experiment where a traditional (type-in) query interface was contrasted with a condition where queries were marked up on the text, the improvement in recall due to viewing more text was more pronounced with search novices. Our results demonstrate that markup and hypertext interfaces to text retrieval systems can benefit recall and can also benefit novices. The challenge now will be to find modified versions of hypertext interfaces that can improve precision, as well as recall and that can work with users who prefer to use different types of search strategy or have different types of training and experience.     

Automatic ranking of information retrieval systems using data fusion

Measuring effectiveness of information retrieval (IR) systems is essential for research and development and for monitoring search quality in dynamic environments. In this study, we employ new methods for automatic ranking of retrieval systems. In these methods, we merge the retrieval results of multiple systems using various data fusion algorithms, use the top-ranked documents in the merged result as the “(pseudo) relevant documents,” and employ these documents to evaluate and rank the systems. Experiments using Text REtrieval Conference (TREC) data provide statistically significant strong correlations with human-based assessments of the same systems. We hypothesize that the selection of systems that would return documents different from the majority could eliminate the ordinary systems from data fusion and provide better discrimination among the documents and systems. This could improve the effectiveness of automatic ranking. Based on this intuition, we introduce a new method for the selection of systems to be used for data fusion. For this purpose, we use the bias concept that measures the deviation of a system from the norm or majority and employ the systems with higher bias in the data fusion process. This approach provides even higher correlations with the human-based results. We demonstrate that our approach outperforms the previously proposed automatic ranking methods.     

Bias–variance analysis in estimating true query model for information retrieval

The estimation of query model is an important task in language modeling (LM) approaches to information retrieval (IR). The ideal estimation is expected to be not only effective in terms of high mean retrieval performance over all queries, but also stable in terms of low variance of retrieval performance across different queries. In practice, however, improving effectiveness can sacrifice stability, and vice versa. In this paper, we propose to study this tradeoff from a new perspective, i.e., the bias–variance tradeoff, which is a fundamental theory in statistics. We formulate the notion of bias–variance regarding retrieval performance and estimation quality of query models. We then investigate several estimated query models, by analyzing when and why the bias–variance tradeoff will occur, and how the bias and variance can be reduced simultaneously. A series of experiments on four TREC collections have been conducted to systematically evaluate our bias–variance analysis. Our approach and results will potentially form an analysis framework and a novel evaluation strategy for query language modeling.     

Probabilistic co-relevance for query-sensitive similarity measurement in information retrieval

Interdocument similarities are the fundamental information source required in cluster-based retrieval, which is an advanced retrieval approach that significantly improves performance during information retrieval (IR). An effective similarity metric is query-sensitive similarity, which was introduced by Tombros and Rijsbergen as method to more directly satisfy the cluster hypothesis that forms the basis of cluster-based retrieval. Although this method is reported to be effective, existing applications of query-specific similarity are still limited to vector space models wherein there is no connection to probabilistic approaches. We suggest a probabilistic framework that defines query-sensitive similarity based on probabilistic co-relevance, where the similarity between two documents is proportional to the probability that they are both co-relevant to a specific given query. We further simplify the proposed co-relevance-based similarity by decomposing it into two separate relevance models. We then formulate all the requisite components for the proposed similarity metric in terms of scoring functions used by language modeling methods. Experimental results obtained using standard TREC test collections consistently showed that the proposed query-sensitive similarity measure performs better than term-based similarity and existing query-sensitive similarity in the context of Voorhees’ nearest neighbor test (NNT).