Retrieval augmented generation (RAG) combines language models with external corpora to support knowledge-intensive tasks, such as open-domain question answering. Standard RAG systems typically employ a fixed top-k retrieval strategy, retrieving the same number of passages regardless of query needs. This can lead to either insufficient evidence, or the inclusion of irrelevant contexts that lead to a degradation of generation performance. In this work, we conduct an empirical study of how irrelevant retrieved passages affect downstream generation, analyzing their impact across multiple standard generator models. Building on these insights, we propose a lightweight, context-size classification module that dynamically predicts how much context is required based on query-specific needs. We integrate this approach into a full RAG pipeline and demonstrate improved performance over several baselines.

Abstract Term position feature is widely and successfully used in IR and Web search engines, to enhance the retrieval effectiveness. This feature is essentially used for two purposes: to capture query terms proximity or to boost the weight of terms appearing in some parts of a document. In this paper, we are interested in this second category. We propose two novel query‐independent techniques based on absolute term positions in a document, whose goal is to boost the weight of terms appearing in the beginning of a document. The first one considers only the earliest occurrence of a term in a document. The second one takes into account all term positions in a document. We formalize each of these two techniques as a document model based on term position, and then we incorporate it into a basic language model (LM). Two smoothing techniques, Dirichlet and Jelinek‐Mercer, are considered in the basic LM. Experiments conducted on three TREC test collections show that our model, especially the version based on all term positions, achieves significant improvements over the baseline LMs, and it also often performs better than two state‐of‐the‐art baseline models, the chronological term rank model and the Markov random field model.

The Information Retrieval (IR) community has witnessed a flourishing development of deep neural networks, however, only a few managed to beat strong baselines. Among them, models like DRMM and DUET were able to achieve better results thanks to the proper handling of exact match signals. Nowadays, the application of pre-trained language models to IR tasks has achieved impressive results exceeding all previous work. In this paper, we assume that established IR cues like exact term-matching, proven to be valuable for deep neural models, can be used to augment the direct supervision from labeled data for training these pre-trained models. To study the effectiveness of this assumption, we propose MarkedBERT a modified version of one of the most popular pre-trained models via language modeling tasks, BERT. MarkedBERT integrates exact match signals using a marking technique that locates and highlights Exact Matched query-document terms using marker tokens. Experiments on MS MARCO Passage Ranking task show that our rather simple approach is actually effective. We find that augmenting the input with marker tokens allows the model to focus on valuable text sequences for IR.

The overwhelming advances in mobile technologies allow recommender systems to be highly contextualized and able to deliver recommendation without an explicit request. However, it is no longer enough for a recommender system to determine what to recommend according to the users' needs, but it also has to deal with the risk of disturbing the user during recommendation. We believe that mobile technologies along with contextual information may help alleviate this issue. In this paper, we address intrusiveness as a probabilistic approach that makes use of the several embedded applications within the user's device and the user's contextual information in order to figure out intrusive recommendations that are subject to rejection. The experiments that we conducted have shown that the proposed approach yields promising results.

In this paper, we extensively study the impact of social signals (users' actions) obtained from several social networks on search ranking task. Social signals associated with web resources (documents) can be considered as an additional information that can play a vital role to estimate a priori importance of these resources. Particularly, we are interested in the freshness of signals and their diversity. We hypothesize that the moment (the date) when the user actions occur and the diversity of actions may impact the search performance. We propose to model these heterogeneous social features as document prior. We evaluate the effectiveness of our approach by carrying out extensive experiments on two different INEX datasets, namely SBS and IMDb, enriched with several social signals collected from social networks. Our experimental results consistently demonstrate the interest of integrating fresh and diverse signals in the retrieval process.

Social media platforms such as weblogs and social networking sites provide Internet users with an unprecedented means to express their opinions and debate on a wide range of issues. Concurrently with their growing importance in public communication, social media platforms may foster echo chambers and filter bubbles: homophily and content personalization lead users to be increasingly exposed to conforming opinions. There is therefore a need for unbiased systems able to identify and provide access to varied viewpoints. To address this task, we propose in this paper a novel unsupervised topic model, the Social Network Viewpoint Discovery Model (SNVDM). Given a specific issue (e.g., U.S. policy) as well as the text and social interactions from the users discussing this issue on a social networking site, SNVDM jointly identifies the issue's topics, the users' viewpoints, and the discourse pertaining to the different topics and viewpoints. In order to overcome the potential sparsity of the social network (i.e., some users interact with only a few other users), we propose an extension to SNVDM based on the Generalized Pólya Urn sampling scheme (SNVDM-GPU) to leverage "acquaintances of acquaintances" relationships. We benchmark the different proposed models against three baselines, namely TAM, SN-LDA, and VODUM, on a viewpoint clustering task using two real-world datasets. We thereby provide evidence that our model SNVDM and its extension SNVDM-GPU significantly outperform state-of-the-art baselines, and we show that utilizing social interactions greatly improves viewpoint clustering performance.

A large amount of social feedback expressed by social signals (e.g. like, +1, rating) are assigned to web resources. These signals are often exploited as additional sources of evidence in search engines. Our objective in this paper is to study the impact of the new social signals, called Facebook reactions (love, haha, angry, wow, sad) in the retrieval. These reactions allow users to express more nuanced emotions compared to classic signals (e.g. like, share). First, we analyze these reactions and show how users use these signals to interact with posts. Second, we evaluate the impact of each such reaction in the retrieval, by comparing them to both the textual model without social features and the first classical signal (like-based model). These social features are modeled as document prior and are integrated into a language model. We conducted a series of experiments on IMDb dataset. Our findings reveal that incorporating social features is a promising approach for improving the retrieval ranking performance.

Just-In-Time Recommender Systems involve all systems able to provide recommendations tailored to the preferences and needs of users in order to help them access useful and interesting resources within a large data space. The user does not need to formulate a query, this latter is implicit and corresponds to the resources that match the user's interests at the right time. In this paper, we propose a proactive context-aware recommendation approach for mobile devices that covers many domains. It aims at recommending relevant items that match users' personal interests at the right time without waiting for users to initiate any interaction.

Due to the diversity of alternative programs to watch and the change of viewers' contexts, real-time prediction of viewers' preferences in certain circumstances becomes increasingly hard. However, most existing TV recommender systems used only current time and location in a heuristic way and ignore other contextual information on which viewers' preferences may depend. This paper proposes a probabilistic approach that incorporates contextual information in order to predict the relevance of TV contents. We consider several viewer's current context elements and integrate them into a probabilistic model. We conduct a comprehensive effectiveness evaluation on a real dataset crawled from Pinhole platform. Experimental results demonstrate that our model outperforms the other context-aware models.

Real time summarization in microblog aims at providing new relevant and non redundant information about an event as soon as it occurs. In this paper, we introduce a new tweet summarization approach where the decision of selecting an incoming tweet is made immediately when a tweet is vailable. Unlike existing approaches where thresholds are redefined, the proposed method estimates thresholds for decision taking in real time as soon as the new tweet arrives. Tweet selection is based upon three criterion namely informativeness, novelty and relevance with regards of the user's interest which are combined as conjunctive condition. Only tweets having an informativeness and novelty scores above a parametric-free threshold are added to the summary. The evaluation of our approach was carried out on the TREC MB RTF 2015 data set and it was compared with well known baselines. The results have revealed that our approach produces the most precise summaries in comparison to all baselines and official runs of the TREC MB RTF 2015 task.

Traditionally, many information retrieval models assume that terms occur in documents independently. Although these models have already shown good performance, the word independency assumption seems to be unrealistic from a natural language point of view, which considers that terms are related to each other. Therefore, such an assumption leads to two well‐known problems in information retrieval (IR), namely, polysemy, or term mismatch, and synonymy. In language models, these issues have been addressed by considering dependencies such as bigrams, phrasal‐concepts, or word relationships, but such models are estimated using simple n‐grams or concept counting. In this paper, we address polysemy and synonymy mismatch with a concept‐based language modeling approach that combines ontological concepts from external resources with frequently found collocations from the document collection. In addition, the concept‐based model is enriched with subconcepts and semantic relationships through a semantic smoothing technique so as to perform semantic matching. Experiments carried out on TREC collections show that our model achieves significant improvements over a single word‐based model and the Markov Random Field model (using a Markov classifier).

Social signals (users' actions) associated with web resources (documents) can be considered as an additional information that can play a role to estimate a priori importance of the resource. In this paper, we are particularly interested in: first, showing the impact of signals diversity associated to a resource on information retrieval performance; second, studying the influence of their social networks origin on their quality. We propose to model these social features as prior that we integrate into language model. We evaluated the effectiveness of our approach on IMDb dataset containing 167438 resources and their social signals collected from several social networks. Our experimental results are statistically significant and show the interest of integrating signals diversity in the retrieval process.

This paper interests in social search over social networking services, typically in microblogging networks. We propose a new approach that integrates, within a Bayesian network model, new relevance factors such as the social importance of microbloggers and the temporal magnitude of tweets. In particular, the social importance of a microblogger is assimilated to his influence on the social network. This property is evaluated by applying PageRank algorithm on the social network of retweets and mentions. The temporal magnitude of microblogs is estimated based on temporal neighbors that present similar query terms. To validate our approach, we conducted a series of experiments on the TREC 2011 Microblog dataset. Results show that the integration of social and temporal features increases the retrieval effectiveness.

In this paper, we propose an attribute retrieval approach which extracts and ranks attributes from HTML tables. We distinguish between class attribute retrieval and instance attribute retrieval. On one hand, given an instance (e.g. University of Strathclyde) we retrieve from the Web its attributes (e.g. principal, location, number of students). On the other hand, given a class (e.g. universities) represented by a set of instances, we retrieve common attributes of its instances. Furthermore, we show we can reinforce instance attribute retrieval if similar instances are available. Our approach uses HTML tables which are probably the largest source for attribute retrieval. Three recall oriented filters are applied over tables to check the following three properties: (i) is the table relational, (ii) has the table a header, and (iii) the conformity of its attributes and values. Candidate attributes are extracted from tables and ranked with a combination of relevance features. Our approach is shown to have a high recall and a reasonable precision. Moreover, it outperforms state of the art techniques.

General Web search engines characterized by "onesize fits all" provide the same results for the same keyword queries even though these latter are submitted by different users with different intentions. In mobile Web search, the expected results for some queries could vary depending upon the user'slocation. We believe that identifying user's geographic intent in Web search can help to personalize search results by ranking local search results higher in the search results lists. Therefore, the objective of this paper is twofold: first to identify whether a mobile user query is location sensitive and second to personalize Web search results for these queries. In order to achieve these objectives, we propose to build a location language model for queries as a location query profile. Based on this latter, we compute two features issued from the domains of probability theory and Information theory, namely the Kurtosis and Kullback-Leibler Divergence measures in order to automatically classify location sensitive queries. The classification scheme is then integrated into a personalization process according to two approaches: refinement and re-ranking. Experimental evaluation using a sample of queries from AOL log and top documents returned by Google search, shows that the proposed model achieves high accuracy in identifying local sensitive queries and shows significant improvement on search relevance when integrated to a search engine.

In the past, most personalized retrieval models have been solely based on the computational behavior of the user to model the user profile. Personalized mobile search should however take the changing environment of the mobile user into account in order to better improve the search results quality. In this paper we propose an approach to personalize search results for mobile users by exploiting both cognitive and spatiotemporal context of the user. We propose to model the user on three semantic dimensions: time, location and interests. A case based reasoning approach is adopted to select the appropriate user profile for re-ranking the search results. In the absence of a standard evaluation framework for mobile search, we propose an evaluation scenario based on diary study entries. Our experiments undertaken in front of Yahoo boss search service shows that our retrieval approach is effective.

Most traditional information retrieval systems are based on single terms indexing. However, it is admitted that semantic content of a document (or a query) cannot be accurately captured by a simple set of independent keywords. Although, several works have incorporated phrases or other syntactic information in IR, such attempts have shown slight benefit, at best. Particularly in language modeling approaches this is achieved through the use of the big ram or n-gram models. However, in these models all big rams/n-grams are considered and weighted uniformly. In this paper we introduce a new approach to weight and consider only certain types of N-grams "compound terms". Experimental results on three test collections showed an improvement.

Major search engines perform what is known as Aggregated Search (AS). They integrate results coming from different vertical search engines (images, videos, news, etc.) with typical Web search results. Aggregated search is relatively new and its advantages need to be evaluated. Some existing works have already tried to evaluate the interest (usefulness) of aggregated search as well as the effectiveness of the existing approaches. However, most of evaluation methodologies were based (i) on what we call relevance by intent (i.e. search results were not shown to real users), and (ii) short text queries. In this paper, we conducted a user study which was designed to revisit and compare the interest of aggregated search, by exploiting both relevance by intent and content, and using both short text and fixed need queries. This user study allowed us to analyze the distribution of relevant results across different verticals, and to show that AS helps to identify complementary relevant sources for the same information need. Comparison between relevance by intent and relevance by content showed that relevance by intent introduces a bias in evaluation. Discussion about the results also allowed us to identify some useful thoughts concerning the evaluation of AS approaches.

In recent years, a lot of work has been done in the field of Opinion Detection in blogs but most of the research is based on machine learning or lexical based approaches. The objective of this paper is to focus on Social Network based evidences that can be exploited for the task of Opinion Detection. We propose a framework that makes use of the major elements of the blogosphere for extracting opinions from blogs. Besides this, we highlight the tasks of opinion prediction and multidimensional ranking. In addition, we also discuss the challenges that researchers might face while realizing the proposed framework. At the end, we demonstrate the importance of social networking evidences by performing experimentation.

In this poster, we describe a personalized search approach, which involves a graph based user profile issued from ontology and a session boundary recognition mechanism. The user profile refers to the short term user interest and is used for re-ranking the search results of queries in the same search session. The session boundary recognition is based on tracking changes in the dominant concepts held by the query and the user profile. Experimental evaluation was carried out using the HARD 2003 TREC collection and shows that our approach is effective.

The main objective in XML Retrieval is to select the relevant elements of XML document instead of the whole document. Many open issues appear when considering Relevance Feedback (RF) in XML documents. They are mainly related to the form of XML documents, which mix content and structure information and to the new information granularity. In this paper, a new flexible method of relevance feedback in XML retrieval using two sources of evidence is described. We propose to use the context criterion to select terms to extend the initial query and to use generative structures to express structural constraints. Both approaches are applied in different combined forms. Experiments are carried out with the INEX evaluation campaign and results show the effectiveness of our approach.

Relevance Feedback (RF) is a technique allowing to enrich an initial query according to the user feedback. The goal is to express more precisily the user's needs. Some open issues appear when considering semi-structured documents like XML documents. Most of the RF approaches proposed in XML retrieval are simple adaptations of traditional RF to the new granularity of information. They enrich queries by adding terms extracted from relevant elements instead of terms extracted from whole documents. In this paper we show how structural constraints can also be used in RF. We propose a new approach that is able to extend the initial query by adding one or more generative structures. This approach is applied to unstructured queries. Experiments are carried out on INEX collection and results show the interest of our method.

Abstract This article presents a genetic relevance optimization process performed in an information retrieval system. The process uses genetic techniques for solving multimodal problems (niching) and query reformulation techniques commonly used in information retrieval. The niching technique allows the process to reach different relevance regions of the document space. Query reformulation techniques represent domain knowledge integrated in the genetic operators structure to improve the convergence conditions of the algorithm. Experimental analysis performed using a TREC subcollection validates our approach.

No abstract available.

Compares two learning profile strategies in an information routing task: relevance backpropagation and gradient backpropagation. Gradient backpropagation is a learning approach used in multilayered neural networks in general. The convergence of the gradient backpropagation algorithm is still to be discussed, but we show its convergence in the majority of cases. Relevance backpropagation is a relevance feedback method used in our connectionist model called Mercure. Experiments carried out on Amaryllis documents showed the effectiveness of both methods, with a slight benefit for the relevance backpropagation strategy.