Language Modeling Approach Research Articles

Integrating multiple document features in language models for expert finding

We argue that expert finding is sensitive to multiple document features in an organizational intranet. These document features include multiple levels of associations between experts and a query topic from sentence, paragraph, up to document levels, document authority information such as the PageRank, indegree, and URL length of documents, and internal document structures that indicate the experts’ relationship with the content of documents. Our assumption is that expert finding can largely benefit from the incorporation of these document features. However, existing language modeling approaches for expert finding have not sufficiently taken into account these document features. We propose a novel language modeling approach, which integrates multiple document features, for expert finding. Our experiments on two large scale TREC Enterprise Track datasets, i.e., the W3C and CSIRO datasets, demonstrate that the natures of the two organizational intranets and two types of expert finding tasks, i.e., key contact finding for CSIRO and knowledgeable person finding for W3C, influence the effectiveness of different document features. Our work provides insights into which document features work for certain types of expert finding tasks, and helps design expert finding strategies that are effective for different scenarios. Our main contribution is to develop an effective formal method for modeling multiple document features in expert finding, and conduct a systematic investigation of their effects. It is worth noting that our novel approach achieves better results in terms of MAP than previous language model based approaches and the best automatic runs in both the TREC2006 and TREC2007 expert search tasks, respectively.

Context-Aware Person Identification in Personal Photo Collections

Identifying the people in photos is an important need for users of photo management systems. We present MediAssist, one such system which facilitates browsing, searching and semi-automatic annotation of personal photos, using analysis of both image content and the context in which the photo is captured. This semi-automatic annotation includes annotation of the identity of people in photos. In this paper, we focus on such person annotation, and propose person identification techniques based on a combination of context and content. We propose language modelling and nearest neighbor approaches to context-based person identification, in addition to novel face color and image color content-based features (used alongside face recognition and body patch features). We conduct a comprehensive empirical study of these techniques using the real private photo collections of a number of users, and show that combining context- and content-based analysis improves performance over content or context alone.

An empirical study of gene synonym query expansion in biomedical information retrieval

Due to the heavy use of gene synonyms in biomedical text, people have tried many query expansion techniques using synonyms in order to improve performance in biomedical information retrieval. However, mixed results have been reported. The main challenge is that it is not trivial to assign appropriate weights to the added gene synonyms in the expanded query; under-weighting of synonyms would not bring much benefit, while overweighting some unreliable synonyms can hurt performance significantly. So far, there has been no systematic evaluation of various synonym query expansion strategies for biomedical text. In this work, we propose two different strategies to extend a standard language modeling approach for gene synonym query expansion and conduct a systematic evaluation of these methods on all the available TREC biomedical text collections for ad hoc document retrieval. Our experiment results show that synonym expansion can significantly improve the retrieval accuracy. However, different query types require different synonym expansion methods, and appropriate weighting of gene names and synonym terms is critical for improving performance.

A language modeling framework for expert finding

Statistical language models have been successfully applied to many information retrieval tasks, including expert finding: the process of identifying experts given a particular topic. In this paper, we introduce and detail language modeling approaches that integrate the representation, association and search of experts using various textual data sources into a generative probabilistic framework. This provides a simple, intuitive, and extensible theoretical framework to underpin research into expertise search. To demonstrate the flexibility of the framework, two search strategies to find experts are modeled that incorporate different types of evidence extracted from the data, before being extended to also incorporate co-occurrence information. The models proposed are evaluated in the context of enterprise search systems within an intranet environment, where it is reasonable to assume that the list of experts is known, and that data to be mined is publicly accessible. Our experiments show that excellent performance can be achieved by using these models in such environments, and that this theoretical and empirical work paves the way for future principled extensions.

Fast exact maximum likelihood estimation for mixture of language model

Language modeling is an effective and theoretically attractive probabilistic framework for text information retrieval. The basic idea of this approach is to estimate a language model of a given document (or document set), and then do retrieval or classification based on this model. A common language modeling approach assumes the data D is generated from a mixture of several language models. The core problem is to find the maximum likelihood estimation of one language model mixture, given the fixed mixture weights and the other language model mixture. The EM algorithm is usually used to find the solution. In this paper, we proof that an exact maximum likelihood estimation of the unknown mixture component exists and can be calculated using the new algorithm we proposed. We further improve the algorithm and provide an efficient algorithm of O ( k ) complexity to find the exact solution, where k is the number of words occurring at least once in data D. Furthermore, we proof the probabilities of many words are exactly zeros, and the MLE estimation is implemented as a feature selection technique explicitly.

Smoothing document language models with probabilistic term count propagation

Smoothing of document language models is critical in language modeling approaches to information retrieval. In this paper, we present a novel way of smoothing document language models based on propagating term counts probabilistically in a graph of documents. A key difference between our approach and previous approaches is that our smoothing algorithm can iteratively propagate counts and achieve smoothing with remotely related documents. Evaluation results on several TREC data sets show that the proposed method significantly outperforms the simple collection-based smoothing method. Compared with those other smoothing methods that also exploit local corpus structures, our method is especially effective in improving precision in top-ranked documents through "filling in" missing query terms in relevant documents, which is attractive since most users only pay attention to the top-ranked documents in search engine applications.

Ordinal Regression for Information Retrieval

This letter presents a new discriminative model for Information Retrieval (IR), referred to as Ordinal Regression Model (ORM). ORM is different from most existing models in that it views IR as ordinal regression problem (i.e. ranking problem) instead of binary classification. It is noted that the task of IR is to rank documents according to the user information needed, so IR can be viewed as ordinal regression problem. Two parameter learning algorithms for ORM are presented. One is a perceptron-based algorithm. The other is the ranking Support Vector Machine (SVM). The effectiveness of the proposed approach has been evaluated on the task of ad hoc retrieval using three English Text REtrieval Conference (TREC) sets and two Chinese TREC sets. Results show that ORM significantly outperforms the state-of-the-art language model approaches and OKAPI system in all test sets; and it is more appropriate to view IR as ordinal regression other than binary classification.

Language‐modeling kernel based approach for information retrieval

AbstractIn this presentation, we propose a novel integrated information retrieval approach that provides a unified solution for two challenging problems in the field of information retrieval. The first problem is how to build an optimal vector space corresponding to users' different information needs when applying the vector space model. The second one is how to smoothly incorporate the advantages of machine learning techniques into the language modeling approach. To solve these problems, we designed the language‐modeling kernel function, which has all the modeling powers provided by language modeling techniques. In addition, for each information need, this kernel function automatically determines an optimal vector space, for which a discriminative learning machine, such as the support vector machine, can be applied to find an optimal decision boundary between relevant and nonrelevant documents. Large‐scale experiments on standard test‐beds show that our approach makes significant improvements over other state‐of‐the‐art information retrieval methods.

In this issue

In this issue

Machine learning for Asian language text classification

PurposeThe purpose of this research is to compare several machine learning techniques on the task of Asian language text classification, such as Chinese and Japanese where no word boundary information is available in written text. The paper advocates a simple language modeling based approach for this task.Design/methodology/approachNaïve Bayes, maximum entropy model, support vector machines, and language modeling approaches were implemented and were applied to Chinese and Japanese text classification. To investigate the influence of word segmentation, different word segmentation approaches were investigated and applied to Chinese text. A segmentation‐based approach was compared with the non‐segmentation‐based approach.FindingsThere were two findings: the experiments show that statistical language modeling can significantly outperform standard techniques, given the same set of features; and it was found that classification with word level features normally yields improved classification performance, but that classification performance is not monotonically related to segmentation accuracy. In particular, classification performance may initially improve with increased segmentation accuracy, but eventually classification performance stops improving, and can in fact even decrease, after a certain level of segmentation accuracy.Practical implicationsApply the findings to real web text classification is ongoing work.Originality/valueThe paper is very relevant to Chinese and Japanese information processing, e.g. webpage classification, web search.

A novel dependency language model for information retrieval

This paper explores the application of term dependency in information retrieval (IR) and proposes a novel dependency retrieval model. This retrieval model suggests an extension to the existing language modeling (LM) approach to IR by introducing dependency models for both query and document. Relevance between document and query is then evaluated by reference to the Kullback-Leibler divergence between their dependency models. This paper introduces a novel hybrid dependency structure, which allows integration of various forms of dependency within a single framework. A pseudo relevance feedback based method is also introduced for constructing query dependency model. The basic idea is to use query-relevant top-ranking sentences extracted from the top documents at retrieval time as the augmented representation of query, from which the relationships between query terms are identified. A Markov Random Field (MRF) based approach is presented to ensure the relevance of the extracted sentences, which utilizes the association features between query terms within a sentence to evaluate the relevance of each sentence. This dependency retrieval model was compared with other traditional retrieval models. Experiments indicated that it produces significant improvements in retrieval effectiveness.

Generating gene summaries from biomedical literature: A study of semi-structured summarization

Most knowledge accumulated through scientific discoveries in genomics and related biomedical disciplines is buried in the vast amount of biomedical literature. Since understanding gene regulations is fundamental to biomedical research, summarizing all the existing knowledge about a gene based on literature is highly desirable to help biologists digest the literature. In this paper, we present a study of methods for automatically generating gene summaries from biomedical literature. Unlike most existing work on automatic text summarization, in which the generated summary is often a list of extracted sentences, we propose to generate a semi-structured summary which consists of sentences covering specific semantic aspects of a gene. Such a semi-structured summary is more appropriate for describing genes and poses special challenges for automatic text summarization. We propose a two-stage approach to generate such a summary for a given gene – first retrieving articles about a gene and then extracting sentences for each specified semantic aspect. We address the issue of gene name variation in the first stage and propose several different methods for sentence extraction in the second stage. We evaluate the proposed methods using a test set with 20 genes. Experiment results show that the proposed methods can generate useful semi-structured gene summaries automatically from biomedical literature, and our proposed methods outperform general purpose summarization methods. Among all the proposed methods for sentence extraction, a probabilistic language modeling approach that models gene context performs the best.

An empirical study of query expansion and cluster-based retrieval in language modeling approach

The term mismatch problem in information retrieval is a critical problem, and several techniques have been developed, such as query expansion, cluster-based retrieval and dimensionality reduction to resolve this issue. Of these techniques, this paper performs an empirical study on query expansion and cluster-based retrieval. We examine the effect of using parsimony in query expansion and the effect of clustering algorithms in cluster-based retrieval. In addition, query expansion and cluster-based retrieval are compared, and their combinations are evaluated in terms of retrieval performance by performing experimentations on seven test collections of NTCIR and TREC.

Table extraction for answer retrieval

The ability to find tables and extract information from them is a necessary component of many information retrieval tasks. Documents often contain tables in order to communicate densely packed, multi-dimensional information. Tables do this by employing layout patterns to efficiently indicate fields and records in two-dimensional form. Their rich combination of formatting and content presents difficulties for traditional retrieval techniques. This paper describes techniques for extracting tables from text and retrieving answers from the extracted information. We compare machine learning (especially, Conditional Random Fields) and heuristic methods for table extraction. To retrieve answers, our approach creates a cell document, which contains the cell and its metadata (headers, titles) for each table cell, and the retrieval model ranks the cells of the extracted tables using a language-modeling approach. Performance is tested using government statistical Web sites and news articles, and errors are analyzed in order to improve the system.

N-gram Adaptation with Dynamic Interpolation Coefficient Using Information Retrieval Technique

This study presents an N-gram adaptation technique when additional text data for the adaptation do not exist. We use a language modeling approach to the information retrieval (IR) technique to collect the appropriate adaptation corpus from baseline text data. We propose to use a dynamic interpolation coefficient to merge the N-gram, where the interpolation coefficient is estimated from the word hypotheses obtained by segmenting the input speech. Experimental results show that the proposed adapted N-gram always has better performance than the background N-gram.

Parsimonious translation models for information retrieval

In the KL divergence framework, the extended language modeling approach has a critical problem of estimating a query model, which is the probabilistic model that encodes the user’s information need. For query expansion in initial retrieval, the translation model had been proposed to involve term co-occurrence statistics. However, the translation model was difficult to apply, because the term co-occurrence statistics must be constructed in the offline time. Especially in a large collection, constructing such a large matrix of term co-occurrences statistics prohibitively increases time and space complexity. In addition, reliable retrieval performance cannot be guaranteed because the translation model may comprise noisy non-topical terms in documents. To resolve these problems, this paper investigates an effective method to construct co-occurrence statistics and eliminate noisy terms by employing a parsimonious translation model. The parsimonious translation model is a compact version of a translation model that can reduce the number of terms containing non-zero probabilities by eliminating non-topical terms in documents. Through experimentation on seven different test collections, we show that the query model estimated from the parsimonious translation model significantly outperforms not only the baseline language modeling, but also the non-parsimonious models.

A goodness of fit test approach in information retrieval

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

Incorporating context within the language modeling approach for ad hoc information retrieval

Using context to improve retrieval performance is a current challenge within the discipline and presents a major challenge to the research community. In this thesis, the context of documents formed via semantic associations is utilized within the Language Modeling (LM) approach for ad hoc Information Retrieval.

Backoff hierarchical class n-gram language models: effectiveness to model unseen events in speech recognition

In this paper, we introduce the backoff hierarchical class n-gram language models to better estimate the likelihood of unseen n-gram events. This multi-level class hierarchy language modeling approach generalizes the well-known backoff n-gram language modeling technique. It uses a class hierarchy to define word contexts. Each node in the hierarchy is a class that contains all the words of its descendant nodes. The closer a node to the root, the more general the class (and context) is. We investigate the effectiveness of the approach to model unseen events in speech recognition. Our results illustrate that the proposed technique outperforms backoff n-gram language models. We also study the effect of the vocabulary size and the depth of the class hierarchy on the performance of the approach. Results are presented on Wall Street Journal (WSJ) corpus using two vocabulary set: 5000 words and 20,000 words. Experiments with 5000 word vocabulary, which contain a small numbers of unseen events in the test set, show up to 10% improvement of the unseen event perplexity when using the hierarchical class n-gram language models. With a vocabulary of 20,000 words, characterized by a larger number of unseen events, the perplexity of unseen events decreases by 26%, while the word error rate (WER) decreases by 12% when using the hierarchical approach. Our results suggest that the largest gains in performance are obtained when the test set contains a large number of unseen events.

Two-stage statistical language models for text database selection

As the number and diversity of distributed Web databases on the Internet exponentially increase, it is difficult for user to know which databases are appropriate to search. Given database language models that describe the content of each database, database selection services can provide assistance in locating databases relevant to the information needs of users. In this paper, we propose a database selection approach based on statistical language modeling. The basic idea behind the approach is that, for databases that are categorized into a topic hierarchy, individual language models are estimated at different search stages, and then the databases are ranked by the similarity to the query according to the estimated language model. Two-stage smoothed language models are presented to circumvent inaccuracy due to word sparseness. Experimental results demonstrate that such a language modeling approach is competitive with current state-of-the-art database selection approaches.