The premise of entity retrieval is to better answer search queries by returning specific entities instead of documents. Many queries mention particular entities; recognizing and linking them to the corresponding entry in a knowledge base is known as the task of entity linking in queries. In this paper we make a first attempt at bringing together these two, i.e., leveraging entity annotations of queries in the entity retrieval model. We introduce a new probabilistic component and show how it can be applied on top of any term-based entity retrieval model that can be emulated in the Markov Random Field framework, including language models, sequential dependence models, as well as their fielded variations. Using a standard entity retrieval test collection, we show that our extension brings consistent improvements over all baseline methods, including the current state-of-the-art. We further show that our extension is robust against parameter settings.
{"title":"Exploiting Entity Linking in Queries for Entity Retrieval","authors":"Faegheh Hasibi, K. Balog, Svein Erik Bratsberg","doi":"10.1145/2970398.2970406","DOIUrl":"https://doi.org/10.1145/2970398.2970406","url":null,"abstract":"The premise of entity retrieval is to better answer search queries by returning specific entities instead of documents. Many queries mention particular entities; recognizing and linking them to the corresponding entry in a knowledge base is known as the task of entity linking in queries. In this paper we make a first attempt at bringing together these two, i.e., leveraging entity annotations of queries in the entity retrieval model. We introduce a new probabilistic component and show how it can be applied on top of any term-based entity retrieval model that can be emulated in the Markov Random Field framework, including language models, sequential dependence models, as well as their fielded variations. Using a standard entity retrieval test collection, we show that our extension brings consistent improvements over all baseline methods, including the current state-of-the-art. We further show that our extension is robust against parameter settings.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129450631","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Developing effective information retrieval models has been a long standing challenge in Information Retrieval (IR), and significant progresses have been made over the years. With the increasing number of developed retrieval functions and the release of new data collections, it becomes more difficult, if not impossible, to compare a new retrieval function with all existing retrieval functions over all available data collections. To tackle thisproblem, this paper describes our efforts on constructing a platform that aims to improve the reproducibility of IR researchand facilitate the evaluation and comparison of retrieval functions. With the developed platform, more than 20 state of the art retrieval functions have been implemented and systematically evaluated over 16 standard TREC collections (including the newly released ClueWeb datasets). Our reproducibility study leads to several interesting observations. First, the performance difference between the reproduced results and those reported in the original papers is small for most retrieval functions. Second, the optimal performance of a few representative retrieval functions is still comparable over the new TREC ClueWeb collections. Finally, the developed platform (i.e., RISE) is made publicly available so that any IR researchers would be able to utilize it to evaluate other retrieval functions.
{"title":"A Reproducibility Study of Information Retrieval Models","authors":"Peilin Yang, Hui Fang","doi":"10.1145/2970398.2970415","DOIUrl":"https://doi.org/10.1145/2970398.2970415","url":null,"abstract":"Developing effective information retrieval models has been a long standing challenge in Information Retrieval (IR), and significant progresses have been made over the years. With the increasing number of developed retrieval functions and the release of new data collections, it becomes more difficult, if not impossible, to compare a new retrieval function with all existing retrieval functions over all available data collections. To tackle thisproblem, this paper describes our efforts on constructing a platform that aims to improve the reproducibility of IR researchand facilitate the evaluation and comparison of retrieval functions. With the developed platform, more than 20 state of the art retrieval functions have been implemented and systematically evaluated over 16 standard TREC collections (including the newly released ClueWeb datasets). Our reproducibility study leads to several interesting observations. First, the performance difference between the reproduced results and those reported in the original papers is small for most retrieval functions. Second, the optimal performance of a few representative retrieval functions is still comparable over the new TREC ClueWeb collections. Finally, the developed platform (i.e., RISE) is made publicly available so that any IR researchers would be able to utilize it to evaluate other retrieval functions.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"9 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126569781","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Over a decade of research on document expansion methods resulted in several independent avenues, including smoothing methods, translation models, and dimensionality reduction techniques, such as matrix decompositions and topic models. Although these research avenues have been individually explored in many previous studies, there is still a lack of understanding of how state-of-the-art methods for each of these directions compare with each other in terms of retrieval accuracy. This paper eliminates this gap by reporting the results of an empirical comparison of document expansion methods using translation models estimated based on word co-occurrence and cosine similarity between low-dimensional word embeddings, Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization (NMF), on standard TREC collections. Experimental results indicate that LDA-based document expansion consistently outperforms both types of translation models and NMF according to all evaluation metrics for all and difficult queries, which is closely followed by translation model using word embeddings.
{"title":"A Study of Document Expansion using Translation Models and Dimensionality Reduction Methods","authors":"Saeid Balaneshinkordan, Alexander Kotov","doi":"10.1145/2970398.2970439","DOIUrl":"https://doi.org/10.1145/2970398.2970439","url":null,"abstract":"Over a decade of research on document expansion methods resulted in several independent avenues, including smoothing methods, translation models, and dimensionality reduction techniques, such as matrix decompositions and topic models. Although these research avenues have been individually explored in many previous studies, there is still a lack of understanding of how state-of-the-art methods for each of these directions compare with each other in terms of retrieval accuracy. This paper eliminates this gap by reporting the results of an empirical comparison of document expansion methods using translation models estimated based on word co-occurrence and cosine similarity between low-dimensional word embeddings, Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization (NMF), on standard TREC collections. Experimental results indicate that LDA-based document expansion consistently outperforms both types of translation models and NMF according to all evaluation metrics for all and difficult queries, which is closely followed by translation model using word embeddings.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116858507","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
significant increase in the number of questions in question answering forums has led to the interest in text categorization methods for classifying a newly posted question as good (suitable) or bad (otherwise) for the forum. Standard text categorization approaches, e.g. multinomial Naive Bayes, are likely to be unsuitable for this classification task because of: i) the lack of sufficient informative content in the questions due to their relatively short length; and ii) considerable vocabulary overlap between the classes. To increase the robustness of this classification task, we propose to use the neighbourhood of existing questions which are similar to the newly asked question. Instead of learning the classification boundary from the questions alone, we transform each question vector into a different one in the feature space. We explore two different neighbourhood functions using: the discrete term space, the continuous vector space of real numbers obtained from vector embeddings of documents. Experiments conducted on StackOverflow data show that our approach of using the neighborhood transformation can improve classification accuracy by up to about 8%.
{"title":"Nearest Neighbour based Transformation Functions for Text Classification: A Case Study with StackOverflow","authors":"Piyush Arora, Debasis Ganguly, G. Jones","doi":"10.1145/2970398.2970426","DOIUrl":"https://doi.org/10.1145/2970398.2970426","url":null,"abstract":"significant increase in the number of questions in question answering forums has led to the interest in text categorization methods for classifying a newly posted question as good (suitable) or bad (otherwise) for the forum. Standard text categorization approaches, e.g. multinomial Naive Bayes, are likely to be unsuitable for this classification task because of: i) the lack of sufficient informative content in the questions due to their relatively short length; and ii) considerable vocabulary overlap between the classes. To increase the robustness of this classification task, we propose to use the neighbourhood of existing questions which are similar to the newly asked question. Instead of learning the classification boundary from the questions alone, we transform each question vector into a different one in the feature space. We explore two different neighbourhood functions using: the discrete term space, the continuous vector space of real numbers obtained from vector embeddings of documents. Experiments conducted on StackOverflow data show that our approach of using the neighborhood transformation can improve classification accuracy by up to about 8%.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"98 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128383893","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The prodigious amount of user-generated content continues to grow at an enormous rate. While it greatly facilitates the flow of information and ideas among people and communities, it may pose great threat to our individual privacy. In this paper, we demonstrate that the private traits of individuals can be inferred from user-generated content by using text classification techniques. Specifically, we study three private attributes on Twitter users: religion, political leaning, and marital status. The ground truth labels of the private traits can be readily collected from the Twitter bio field. Based on the tweets posted by the users and their corresponding bios, we show that text classification yields a high accuracy of identification of these personal attributes, which poses a great privacy risk on user-generated content. We further propose a constrained utility maximization framework for preserving user privacy. The goal is to maximize the utility of data when modifying the user-generated content, while degrading the prediction performance of the adversary. The KL divergence is minimized between the prior knowledge about the private attribute and the posterior probability after seeing the user-generated data. Based on this proposed framework, we investigate several specific data sanitization operations for privacy preservation: add, delete, or replace words in the tweets. We derive the exact transformation of the data under each operation. The experiments demonstrate the effectiveness of the proposed framework.
{"title":"A Utility Maximization Framework for Privacy Preservation of User Generated Content","authors":"Yi Fang, Archana Godavarthy, Haibing Lu","doi":"10.1145/2970398.2970417","DOIUrl":"https://doi.org/10.1145/2970398.2970417","url":null,"abstract":"The prodigious amount of user-generated content continues to grow at an enormous rate. While it greatly facilitates the flow of information and ideas among people and communities, it may pose great threat to our individual privacy. In this paper, we demonstrate that the private traits of individuals can be inferred from user-generated content by using text classification techniques. Specifically, we study three private attributes on Twitter users: religion, political leaning, and marital status. The ground truth labels of the private traits can be readily collected from the Twitter bio field. Based on the tweets posted by the users and their corresponding bios, we show that text classification yields a high accuracy of identification of these personal attributes, which poses a great privacy risk on user-generated content. We further propose a constrained utility maximization framework for preserving user privacy. The goal is to maximize the utility of data when modifying the user-generated content, while degrading the prediction performance of the adversary. The KL divergence is minimized between the prior knowledge about the private attribute and the posterior probability after seeing the user-generated data. Based on this proposed framework, we investigate several specific data sanitization operations for privacy preservation: add, delete, or replace words in the tweets. We derive the exact transformation of the data under each operation. The experiments demonstrate the effectiveness of the proposed framework.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"45 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134371509","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
General-purpose knowledge bases are increasingly growing in terms of depth (content) and width (coverage). Moreover, algorithms for entity linking and entity retrieval have improved tremendously in the past years. These developments give rise to a new line of research that exploits and combines these developments for the purposes of text-centric information retrieval applications. This tutorial focuses on a) how to retrieve a set of entities for an ad-hoc query, or more broadly, assessing relevance of KB elements for the information need, b) how to annotate text with such elements, and c) how to use this information to assess the relevance of text. We discuss different kinds of information available in a knowledge graph and how to leverage each most effectively. We start the tutorial with a brief overview of different types of knowledge bases, their structure and information contained in popular general-purpose and domain-specific knowledge bases. In particular, we focus on the representation of entity-centric information in the knowledge base through names, terms, relations, and type taxonomies. Next, we will provide a recap on ad-hoc object retrieval from knowledge graphs as well as entity linking and retrieval. This is essential technology, which the remainder of the tutorial builds on. Next we will cover essential components within successful entity linking systems, including the collection of entity name information and techniques for disambiguation with contextual entity mentions. We will present the details of four previously proposed systems that successfully leverage knowledge bases to improve ad-hoc document retrieval. These systems combine the notion of entity retrieval and semantic search on one hand, with text retrieval models and entity linking on the other. Finally, we also touch on entity aspects and links in the knowledge graph as it can help to understand the entities' context. This tutorial is the first to compile, summarize, and disseminate progress in this emerging area and we provide both an overview of state-of-the-art methods and outline open research problems to encourage new contributions.
{"title":"Utilizing Knowledge Bases in Text-centric Information Retrieval","authors":"Laura Dietz, Alexander Kotov, E. Meij","doi":"10.1145/2970398.2970441","DOIUrl":"https://doi.org/10.1145/2970398.2970441","url":null,"abstract":"General-purpose knowledge bases are increasingly growing in terms of depth (content) and width (coverage). Moreover, algorithms for entity linking and entity retrieval have improved tremendously in the past years. These developments give rise to a new line of research that exploits and combines these developments for the purposes of text-centric information retrieval applications. This tutorial focuses on a) how to retrieve a set of entities for an ad-hoc query, or more broadly, assessing relevance of KB elements for the information need, b) how to annotate text with such elements, and c) how to use this information to assess the relevance of text. We discuss different kinds of information available in a knowledge graph and how to leverage each most effectively. We start the tutorial with a brief overview of different types of knowledge bases, their structure and information contained in popular general-purpose and domain-specific knowledge bases. In particular, we focus on the representation of entity-centric information in the knowledge base through names, terms, relations, and type taxonomies. Next, we will provide a recap on ad-hoc object retrieval from knowledge graphs as well as entity linking and retrieval. This is essential technology, which the remainder of the tutorial builds on. Next we will cover essential components within successful entity linking systems, including the collection of entity name information and techniques for disambiguation with contextual entity mentions. We will present the details of four previously proposed systems that successfully leverage knowledge bases to improve ad-hoc document retrieval. These systems combine the notion of entity retrieval and semantic search on one hand, with text retrieval models and entity linking on the other. Finally, we also touch on entity aspects and links in the knowledge graph as it can help to understand the entities' context. This tutorial is the first to compile, summarize, and disseminate progress in this emerging area and we provide both an overview of state-of-the-art methods and outline open research problems to encourage new contributions.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"36 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116122292","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Classic learning to rank algorithms are trained using a set of labeled documents, pairs of documents, or rankings of documents. Unfortunately, in many situations, gathering such labels requires significant overhead in terms of time and money. We present an algorithm for training a learning to rank model using a set of labeled features elicited from system designers or domain experts. Labeled features incorporate a system designer's belief about the correlation between certain features and relative relevance. We demonstrate the efficacy of our model on a public learning to rank dataset. Our results show that we outperform our baselines even when using as little as a single feature label.
{"title":"Learning to Rank with Labeled Features","authors":"Fernando Diaz","doi":"10.1145/2970398.2970435","DOIUrl":"https://doi.org/10.1145/2970398.2970435","url":null,"abstract":"Classic learning to rank algorithms are trained using a set of labeled documents, pairs of documents, or rankings of documents. Unfortunately, in many situations, gathering such labels requires significant overhead in terms of time and money. We present an algorithm for training a learning to rank model using a set of labeled features elicited from system designers or domain experts. Labeled features incorporate a system designer's belief about the correlation between certain features and relative relevance. We demonstrate the efficacy of our model on a public learning to rank dataset. Our results show that we outperform our baselines even when using as little as a single feature label.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"90 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121754700","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Events are central in human history and thus also in Web queries, in particular if they relate to history or news. However, ambiguity issues arise as queries may refer to ambiguous events differing in time, geography, or participating entities. Thus, users would greatly benefit if search results were presented along different events. In this paper, we present EventMiner, an algorithm that mines events from top-k pseudo-relevant documents for a given query. It is a probabilistic framework that leverages semantic annotations in the form of temporal expressions, geographic locations, and named entities to analyze natural language text and determine important events. Using a large news corpus, we show that using semantic annotations, EventMiner detects important events and presents documents covering the identified events in the order of their importance.
{"title":"EventMiner: Mining Events from Annotated Documents","authors":"Dhruv Gupta, Jannik Strotgen, K. Berberich","doi":"10.1145/2970398.2970411","DOIUrl":"https://doi.org/10.1145/2970398.2970411","url":null,"abstract":"Events are central in human history and thus also in Web queries, in particular if they relate to history or news. However, ambiguity issues arise as queries may refer to ambiguous events differing in time, geography, or participating entities. Thus, users would greatly benefit if search results were presented along different events. In this paper, we present EventMiner, an algorithm that mines events from top-k pseudo-relevant documents for a given query. It is a probabilistic framework that leverages semantic annotations in the form of temporal expressions, geographic locations, and named entities to analyze natural language text and determine important events. Using a large news corpus, we show that using semantic annotations, EventMiner detects important events and presents documents covering the identified events in the order of their importance.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"94 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127068090","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Previous studies have shown that semantically meaningful representations of words and text can be acquired through neural embedding models. In particular, paragraph vector (PV) models have shown impressive performance in some natural language processing tasks by estimating a document (topic) level language model. Integrating the PV models with traditional language model approaches to retrieval, however, produces unstable performance and limited improvements. In this paper, we formally discuss three intrinsic problems of the original PV model that restrict its performance in retrieval tasks. We also describe modifications to the model that make it more suitable for the IR task, and show their impact through experiments and case studies. The three issues we address are (1) the unregulated training process of PV is vulnerable to short document over-fitting that produces length bias in the final retrieval model; (2) the corpus-based negative sampling of PV leads to a weighting scheme for words that overly suppresses the importance of frequent words; and (3) the lack of word-context information makes PV unable to capture word substitution relationships.
{"title":"Analysis of the Paragraph Vector Model for Information Retrieval","authors":"Qingyao Ai, Liu Yang, Jiafeng Guo, W. Bruce Croft","doi":"10.1145/2970398.2970409","DOIUrl":"https://doi.org/10.1145/2970398.2970409","url":null,"abstract":"Previous studies have shown that semantically meaningful representations of words and text can be acquired through neural embedding models. In particular, paragraph vector (PV) models have shown impressive performance in some natural language processing tasks by estimating a document (topic) level language model. Integrating the PV models with traditional language model approaches to retrieval, however, produces unstable performance and limited improvements. In this paper, we formally discuss three intrinsic problems of the original PV model that restrict its performance in retrieval tasks. We also describe modifications to the model that make it more suitable for the IR task, and show their impact through experiments and case studies. The three issues we address are (1) the unregulated training process of PV is vulnerable to short document over-fitting that produces length bias in the final retrieval model; (2) the corpus-based negative sampling of PV leads to a weighting scheme for words that overly suppresses the importance of frequent words; and (3) the lack of word-context information makes PV unable to capture word substitution relationships.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"43 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123629949","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Word embeddings, which are low-dimensional vector representations of vocabulary terms that capture the semantic similarity between them, have recently been shown to achieve impressive performance in many natural language processing tasks. The use of word embeddings in information retrieval, however, has only begun to be studied. In this paper, we explore the use of word embeddings to enhance the accuracy of query language models in the ad-hoc retrieval task. To this end, we propose to use word embeddings to incorporate and weight terms that do not occur in the query, but are semantically related to the query terms. We describe two embedding-based query expansion models with different assumptions. Since pseudo-relevance feedback methods that use the top retrieved documents to update the original query model are well-known to be effective, we also develop an embedding-based relevance model, an extension of the effective and robust relevance model approach. In these models, we transform the similarity values obtained by the widely-used cosine similarity with a sigmoid function to have more discriminative semantic similarity values. We evaluate our proposed methods using three TREC newswire and web collections. The experimental results demonstrate that the embedding-based methods significantly outperform competitive baselines in most cases. The embedding-based methods are also shown to be more robust than the baselines.
{"title":"Embedding-based Query Language Models","authors":"Hamed Zamani, W. Bruce Croft","doi":"10.1145/2970398.2970405","DOIUrl":"https://doi.org/10.1145/2970398.2970405","url":null,"abstract":"Word embeddings, which are low-dimensional vector representations of vocabulary terms that capture the semantic similarity between them, have recently been shown to achieve impressive performance in many natural language processing tasks. The use of word embeddings in information retrieval, however, has only begun to be studied. In this paper, we explore the use of word embeddings to enhance the accuracy of query language models in the ad-hoc retrieval task. To this end, we propose to use word embeddings to incorporate and weight terms that do not occur in the query, but are semantically related to the query terms. We describe two embedding-based query expansion models with different assumptions. Since pseudo-relevance feedback methods that use the top retrieved documents to update the original query model are well-known to be effective, we also develop an embedding-based relevance model, an extension of the effective and robust relevance model approach. In these models, we transform the similarity values obtained by the widely-used cosine similarity with a sigmoid function to have more discriminative semantic similarity values. We evaluate our proposed methods using three TREC newswire and web collections. The experimental results demonstrate that the embedding-based methods significantly outperform competitive baselines in most cases. The embedding-based methods are also shown to be more robust than the baselines.","PeriodicalId":443715,"journal":{"name":"Proceedings of the 2016 ACM International Conference on the Theory of Information Retrieval","volume":"14 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2016-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115770565","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：