Search

The parallel Hierarchical Dirichlet Process (pHDP) is an efficient topic model which explores the equivalence of the generation process between Hierarchical Dirichlet Process (HDP) and Gamma-Gamma-Poisson Process (G2PP), in order to achieve parallelism at the topic level. Unfortunately, pHDP loses the non-parametric feature of HDP, i.e., the number of topics in pHDP is predetermined and fixed. Furthermore, under the bootstrap structure of pHDP, the topic-indiscriminate words are of high probabilities to be assigned to different topics, resulting in poor qualities of the extracted topics. To achieve parallelism without sacrificing the non-parametric feature of HDP, in addition to improve the quality of extracted topics, we propose a parallel dynamic topic model by developing an adjustment mechanism of evolving topics and reducing the sampling probabilities of topic-indiscriminate words. Both supervised and unsupervised experiments on benchmark datasets show the competitive performance of our model.

An innovative model-based approach to coupling text clustering and topic modeling is introduced, in which the two tasks take advantage of each other. Specifically, the integration is enabled by a new generative model of text corpora. This explains topics, clusters and document content via a Bayesian generative process. In this process, documents include word vectors, to capture the (syntactic and semantic) regularities among words. Topics are multivariate Gaussian distributions on word vectors. Clusters are assigned corresponding topic distributions as their semantics. Content generation is ruled by text clusters and topics, which act as interacting latent factors. Documents are at first placed into respective clusters, then the semantics of these clusters is then repeatedly sampled to draw document topics, which are in turn sampled for word-vector generation. Under the proposed model, collapsed Gibbs sampling is derived mathematically and implemented algorithmically with parameter estimation for the simultaneous inference of text clusters and topics. A comparative assessment on real-world benchmark corpora demonstrates the effectiveness of this approach in clustering texts and uncovering their semantics. Intrinsic and extrinsic criteria are adopted to investigate its topic modeling performance, whose results are shown through a case study. Time efficiency and scalability are also studied.

Current research on subset selection for opinion analysis assumes that their methods can retrieve the opinions expressed in documents from general text features. However, such relaxed conditions can hardly maintain the performance of the analysis in opinion mining , especially when given strict limitations on the subset size. In this paper, we propose a framework for opinion subset selection. This framework can select a small set of instances from original data to convey a subjective representation for opinion classification and regression. Compared with our framework, the conventional submodular based subset selection approach cannot capture the fine-grained opinion features expressed in the corpus. Specifically, we propose a monotone non-decreasing score function and a framework based on topic modeling and submodular maximization for filtering irrelevant information and selecting the subsets. Our work further introduces an opinion-sensitive algorithm for optimizing the proposed function for opinion subset construction . We perform extensive experiments and comparative analysis of different subset selection methods in this work. The experimental result shows that the proposed opinion subset selection framework can compress the original text training set and preserve the test set’s classification and regression metric performance at the same time.

In recent years, XACML (eXtensible Access Control Markup Language) has been widely used in the development of various applications, especially Web services. The evaluation time of a PDP (Policy Decision Point) grows significantly when the PDP loads a large-scale policy set coded in XACML. In order to improve the PDP evaluation performance, we propose an optimized policy evaluation engine, namely XDLEngine, and make the following contributions. First, XDLEngine has an advantage in the process of handling a large-scale policy set, and innovatively adopts the LDA (Latent Dirichlet Allocation) topic model to cluster policies. Second, according to the clustering results of the LDA model, we digitize and vectorize all rules in policy sets, which facilitates the rule matching. Third, the cosine similarity is introduced to classify the rules under each topic, which greatly reduces the number of comparisons in the process of rule matching and improves the matching efficiency of XDLEngine. Finally, due to the independence between different topics, we use a multi-threaded parallel search in the process of rule matching, which significantly lowers the evaluation time of XDLEngine. The experimental results show that when the number of requests reaches 20,000, the evaluation time of XDLEngine for a practical large-scale policy set with 120,000 rules is approximately 2.48%, 3.47% and 3.68% of that of the Sun PDP, XEngine and HPEngine, respectively.

Question retrieval is an extremely important research field in Community Question Answering (CQA). Most existing question retrieval methods depend on semantic analysis of questions, whose effectiveness suffers from the short texts of the noise words in the question corpus. In order to recommend the questions with more advanced knowledge to users, the influence of the questions’ popularity should be considered during retrieving questions. To make retrieved questions with both similar semantics and high popularity, we propose an Integrated Retrieval Framework for Similar Questions named Word-semantic Embedded Label Clustering – LDA with Question Life Cycle (WELQLC-QR), consisting of Word-semantic Embedded Label Clustering – LDA (WEL) and Question Life Cycle Optimization Similar Question List Strategy (QLC). Firstly, WEL is proposed for question retrieval from the perspective of semantic matching. It not only overcomes the problem of over-generalization of the semantic information extracted by topic models when facing short questions with multi-levels labels, but also avoids the influence of noise vocabularies during semantic extracting of the questions. Then, based on the internal factors (i.e., the number of comments and answers to the question) and external factors (i.e., programming language ranking information) of questions, QLC constructs a popularity-predicted model to optimize the similar question set searched by WEL, making the final retrieval results both semantically similar and popular. Finally, experiments are conducted on CQADupStack dataset, and results show that the MRR@N of WELQLC-QR model has an average increase of 8.99%, 8.3%, 4.74% and 3.56% compared with that of L-LDA, LC-LDA, BM25 and Word2vec, respectively.

Hierarchical classification (HC) is effective when categories are organized hierarchically. However, the blocking problem makes the effect of hierarchical classification greatly reduced. Blocking means that samples are easily getting misclassified in high-level classifiers so that the samples are blocked at the high-level of the hierarchy. This issue is caused by the inconsistency between the artificially defined hierarchy and the actual hierarchy of the raw data. Another issue is that it is flippant to strictly process data following the hierarchy. Therefore, special treatment is required for some uncertain data. To address the first issue, we learn category relationships and modify the hierarchy. To address the second issue, we introduce three-way decisions (3WD) to targetedly deal with the ambiguous data. We extend original studies and propose two HC models based on 3WD, collectively referred to as TriHC, for carefully modifying the hierarchy to alleviate the blocking problem. The proposed TriHC model learns new category hierarchies by the following three steps: (1) mining category relations; (2) modifying category hierarchies according to the latent category relations; and (3) using 3WD to divide observed objects into three regions: positive region, boundary region, and negative region, and making decisions based on different strategies. Specifically, based on different category relation mining methods, there are two versions of TriHC, cross-level blocking priori knowledge based TriHC (CLPK-TriHC) and expert classifier based TriHC (EC-TriHC). The CLPK-TriHC model defines a cross-level blocking distribution matrix to mine the category relations between the higher and lower levels. To better exploit category hierarchical relations, the EC-TriHC model builds expert classifiers using topic model to learn latent category topics. Experimental results validate that the proposed methods can simultaneously reduce the blocking and improve the classification accuracy.

The Latent Dirichlet Allocation (LDA) model, which is a document-level probabilistic model, has been widely used in topic modeling. However, an essential issue of the LDA is its shortage in identifying co-occurrence relationships (e.g., aspect-aspect, aspect-opinion, etc.) in sentences. To address the problem, we propose an association constrained LDA (AC-LDA) for effectively capturing the co-occurrence relationships. Specifically, based on the basic features of the syntactic structure in product reviews, we formalize three major types of word association combinations and then carefully design corresponding identifications. For reducing the influence of global aspect words on the local distribution, we apply an important constraint on global aspects. Finally, the constraint and related association combinations are merged into the LDA to guide the topic-words allocation in the learning process. Based on the experiments on real-world product review data, we demonstrate that our model can effectively capture the relationships hidden in local sentences and further increase the extraction rate of fine-grained aspects and opinion words. Our results confirm the superiority of the AC-LDA over the state-of-the-art methods in terms of the extraction accuracy. We also verify the strength of our method in identifying irregularly appeared terms, such as non-aspect opinions, low-frequency words, and secondary aspects.

Disk-based algorithms have the ability to process large-scale data which do not fit into the memory, so they provide good scalability to a mobile device with limited memory resources. In general, the speed of disk I/O is much slower than that of memory access, the total amount of disk I/O is the most crucial factor which determines the efficiency of disk-based algorithms. This paper proposes BlockLDA, an efficient disk-based Latent Dirichlet Allocation (LDA) inference algorithm which can efficiently infer an LDA model when both of the data and model do not fit into the memory. BlockLDA manages the data and model as a set of small blocks so that it can support efficient disk I/O as well as process the LDA inference in a block-wise manner. In addition, it utilizes advanced techniques which help to minimize the amount of disk I/O, including 1) a space reduction algorithm to dynamically manage the block-wise model considering its changing sparsity and 2) a local scheduling algorithm to carefully select the next data blocks so that the number of page faults is minimized. Our experimental results demonstrate that BlockLDA shows better scalability and efficiency than its disk-based and in-memory competitors under the memory-limited environment.

Relational topic models (RTM) have been widely used to discover hidden topics in a collection of networked documents. In this paper, we introduce the class of Constrained Relational Topic Models (CRTM), a semi-supervised extension of RTM that, apart from modeling the structure of the document network, explicitly models some available domain knowledge. We propose two instances of CRTM that incorporate prior knowledge in the form of document constraints. The models smooth the probability distribution of topics such that two constrained documents can either share the same topics or denote distinct themes. Experimental results on benchmark relational datasets show significant performances of CRTM on a semi-supervised document classification task.

Hashtags have become a crucial social media tool. The categorization of posts in a simple and informal way helps to spread the content through the web. At the same time, it enables users to easily find messages within a specific topic. However, the flexibility provided to use and create a hashtag carries some problems. Equivalent expressions, like synonyms, are handled like entirely different words. On the other hand, the same hashtag may refer to different topics. In this paper, we present TORHID (Topic Relevant Hashtag Identification), a method that employs topic modeling with the purpose of retrieving and identifying hashtags relevant to a specific topic in Twitter streams, starting from a seed hashtag and resorting to a classifier to remove non relevant hashtags. The result is a network of hashtags related to the seed, that we can use to deepen the initial search.

Internet has become the most popular platform for people to exchange opinions and express stances. The stances implied in web texts indicate people's fundamental beliefs and viewpoints. Understanding the stances people take is beneficial and critical for many security and business related applications, such as policy design, emergency response and marketing management. Most previous work on stance detection focuses on identifying the supportive or unsupportive attitudes towards a specific target. However, another important type of stance detection, i.e. multiple standpoint detection, has been largely ignored. Multiple standpoint detection aims to identify the distinct standpoints people hold among multiple parties, which reflects people's intrinsic values and judgments. When expressing standpoints, people tend to discuss diverse topics, and the word usage in the topics of different standpoints often varies a lot. As topics can provide latent information for identifying various standpoints, in this paper, we propose a topic-based approach to detecting multiple standpoints in Web texts, by enhancing generative classification model as well as feature representation of texts. In addition, we develop an adaptive process to determine parameter values in our approach automatically. Experimental studies on several real-world datasets verify the effectiveness of our proposed approach in detecting multiple standpoints. (C) 2019 Elsevier Inc. All rights reserved.

Existing graph-based methods for extractive document summarization represent sentences of a corpus as the nodes of a graph or a hypergraph in which edges depict relationships of lexical similarity between sentences. Such approaches fail to capture semantic similarities between sentences when they express a similar information but have few words in common and are thus lexically dissimilar. To overcome this issue, we propose to extract semantic similarities based on topical representations of sentences. Inspired by the Hierarchical Dirichlet Process, we propose a probabilistic topic model in order to infer topic distributions of sentences. As each topic defines a semantic connection among a group of sentences with a certain degree of membership for each sentence, we propose a fuzzy hypergraph model in which nodes are sentences and fuzzy hyperedges are topics. To produce an informative summary, we extract a set of sentences from the corpus by simultaneously maximizing their relevance to a user-defined query, their centrality in the fuzzy hypergraph and their coverage of topics present in the corpus. We formulate a polynomial time algorithm building on the theory of submodular functions to solve the associated optimization problem. A thorough comparative analysis with other graph-based summarization systems is included in the paper. Our obtained results show the superiority of our method in terms of content coverage of the summaries., 8 figures

Fine-grained sentiment analysis for online reviews plays more and more important role in many applications. The key techniques here are how to efficiently extract multi-grained aspects, identify associated opinions, and classify sentiment polarity. Although various topic models have been proposed to process some of these tasks in recent years, there was little work available for effective sentiment analysis. In this paper, we propose a joint aspect based sentiment topic (JABST) model that jointly extracts multi-grained aspects and opinions through modeling aspects, opinions, sentiment polarities and granularities simultaneously. Moreover, by means of the supervised learning, we then propose a maximum entropy based JABST model (MaxEnt–JABST) to improve accuracy and performance in extracting opinions and aspects. Comprehensive evaluation results on real-world reviews for electronic devices and restaurants demonstrate that our JABST and MaxEnt–JABST models significantly outperform related proposals.

Modeling the interests of authors over time from documents has important applications in broad applications such as recommendation systems, authorship identification and opinion extraction. In this paper, we propose an Ordering-sensitive and Semantic-aware Dynamic Author Topic Model (OSDATM), which monitors the evolution of author interest in time-stamped documents. The model further uses the discovered author interest information to discover better topics. Unlike traditional topic models, OSDATM is sensitive to the ordering of words, thus it extracts more information from the semantic meaning of the context. The experimental results show that OSDATM learns better topics than state-of-the-art topic models. In addition, the dynamic interests of authors that the OSDATM model discovers are interpretable and consistent with the truth.

Tag-based recommendation has become increasingly important in recent years owing to the popularization of social tagging systems. Related studies on this subject can be categorized into item recommendation and tag recommendation based on the objective of the recommendation. In both categories, most existing recommendation approaches focus on improving prediction accuracy. But they ignore that the explanations of recommendations also greatly affect the decision-making of users. In social tagging systems, tags not only behave as auxiliary information of items but also show the implicit preferences of users. Therefore, they can be used to improve prediction accuracy and provide explanations to recommended items. Items and tags have interrelation and mutual effects. By focusing only on either item recommendation or tag recommendation, users may miss some information and can only achieve marginal gains. Fusing the two types of recommendations can improve the performance of both approaches. On the basis of the above ideas, in this study, we propose an EXPLainable item-tag CO-REcommendation (EXPLORE) framework that jointly recommends items and the corresponding tags. Different from conventional recommendations that utilize a single source of content, EXPLORE takes advantage of users’ interests, item contents, and item tags. The experiments conducted on three real-world datasets demonstrate that EXPLORE outperforms the state-of-the-art methods. Importantly, the recommended tags can provide explanations to the recommended items, thus making the recommendation results explainable.

Nowadays, massive short texts, such as social media posts and newspaper titles, are available on the Internet. Analyzing these short texts is very significant for many content analysis tasks. However, the commonly used text analysis tools, i.e., topic models, lose effectiveness on short texts because of the sparsity and noise problems. Recent topic models mainly attempt to solve the sparsity problem, but neglect the noise issue. To address this, we propose a common semantics topic model (CSTM) in this paper. The key idea is to introduce a new type of topic, namely common topic, to gather the noise words. The experimental results on real-world datasets indicate that our CSTM outperforms the existing short text topic models on the traditional tasks.

Sentiment topic models are used as unsupervised methods to solve the specific problems of the general aspect-based sentiment analysis (ABSA) problem. One of the main problems of the technique is its substandard aspect term extraction, which leads to difficulties in aspect label determination. This paper is focused on improving the aspect term extraction of topic models by incorporating product descriptions to the current state-of-the-art sentiment topic model, Aspect Sentiment Unification Model (ASUM). We present two models that extend from ASUM differently to leverage on the information found in the product description: Seller-aided Aspect-based Sentiment Model (SA-ASM) and Seller-aided Product-based Sentiment Model (SA-PSM). SA-ASM has its topic distribution inside the review while SA-PSM has its topic distribution inside the product description. Based on experiments conducted to reviews of laptops and mobile phones, results show that SA-ASM performs better in micro-level problems such as sentiment classification and aspect assignment and SA-PSM performs better in macro-level problems like aspect category detection. Both models achieve better performances compared to current topic modeling methods for the ABSA problem.

Despite impressive achievements in image processing and artificial intelligence in the past decade, understanding video-based action remains a challenge. However, the intensive development of 3D computer vision in recent years has brought more potential research opportunities in pose-based action detection and recognition. Thanks to the advantages of depth camera devices like the Microsoft Kinect sensor, we developed an effective approach to in-depth analysis of indoor actions using skeleton information, in which skeleton-based feature extraction and topic model-based learning are two major contributions. Geometric features, i.e. joint distance, joint angle, and joint-plane distance are calculated in the spatio-temporal dimension. These features are merged into two types, called pose and transition features, and then are provided to codebook construction to convert sparse features into visual words by k-means clustering. An efficient hierarchical model is developed to describe the full correlation of feature - poselet - action based on Pachinko Allocation Model. This model has the potential to uncover more hidden poselets, which have been recognized as the valuable information and help to differentiate pose-sharing actions. The experimental results on several well-known datasets, such as MSR Action 3D, MSR Daily Activity 3D, Florence 3D Action, UTKinect-Action 3D, and NTU RGB+D Action Recognition, demonstrate the high recognition accuracy of the proposed method. Our method outperforms state-of-the-art methods in the field in most dataset benchmarks.

After a news event, many different websites publish coverage of that event, each expressing their own unique commentary, perspectives, and viewpoints. Websites form around a specific set of interests to cater to different audiences, and discovering these interests can help audiences C especially people and organizations that are interested in news C select the most appropriate websites to use as their sources of information. This paper presents three methods for formally defining and mining a websites interests, each of which is explicitly or implicitly based on a hierarchial structure: website-webpage-keyword. The first, and most straightforward, method explicitly uses keyword-layer network communities and the mapping relations between websites and keywords. The second method expands upon the first method with an iterative algorithm that combines both the mapping relations and the network relations from the website-webpage-keyword structure to further refine the keyword-layer network communities. In the third method, a website topic model implicitly captures the mapping relations among the websites, webpages, and keywords. The performance of three proposed methods in website interest mining is compared using a bespoke evaluation metric. The experimental results show that the iterative procedure designed in the second method is able to improve website interest mining performance, and the website topic model in the third method achieves the best performance among the three methods.

Recently user clustering has become an increasingly important subject because of the high popularity of social medias like Twitter, Weibo and Facebook. The state-of-the-art algorithms of user clustering are all focused on the long or short text streams without considering factors of social network information, making them either unable to capture the social connectivity from short text streams or unable to conform to the sparsity, high-dimensionality and dynamically changing topics of short text streams. To address these issues, we propose a user clustering method named dynamic social network topic model (DSM) in this paper to cluster users by modeling their topics with dynamic features and social connectivity in short text streams. Experimental results show our topic model outperforms the state-of-the-art methods in the context of short text streams with social network information.

A framework to generate pseudo-documents suitable for topic modeling is proposed.An instantiation of the framework based on word vector representations is presented.Results of NPMI and F1 obtained are better than those of state-of-the art methods. Short texts are everywhere in the Web, including messages posted in social media, status messages and blog comments, and uncovering the topics of this type of messages is crucial to a wide range of applications, e.g., context analysis and user characterization. Extracting topics from short text is challenging because of the dependence of conventional methods, such as Latent Dirichlet Allocation, in words co-occurrence, which in short text is rare and make these methods suffer from severe data sparsity. This paper proposes a general framework for topic modeling of short text by creating larger pseudo-document representations from the original documents. In the framework, document components (e.g., words or bigrams) are defined over a metric space, which provides information about the similarity between them. We present two simple, effective and efficient methods that specialize our general framework to create larger pseudo-documents. While the first method considers word co-occurrence to define the metric space, the second relies on distributed word vector representations. The pseudo-documents generated can be given as input to any topic modeling algorithm. Experiments run in seven datasets and compared against state-of-the-art methods for extracting topics by generating pseudo-documents or modifying current topic modeling methods for short text show the methods significantly improve results in terms of normalized pointwise mutual information. A classification task was also used to evaluate the quality of the topics in terms of document representation, where improvements in F1 varied from 1.5 to 15%.

Reviews are collaboratively generated by users on items and generally contain rich information than ratings in a recommender system scenario. Ratings are modeled successfully with latent space models by capturing interaction between users and items. However, only a few models collaboratively deal with documents such as reviews. In this study, by modeling reviews as a three-order tensor, we propose a refined tensor topic model (TTM) for text tensors inspired by Tucker decomposition. User and item dimensions are co-reduced with vocabulary space, and interactions between users and items are captured using a core tensor in dimension-reduced form. TTM is proposed to obtain low-rank representations of words as well as of users and items. Furthermore, general rules are developed to transform a decomposition model into a probabilistic model. TTM is augmented further to predict ratings with the assistance of a low-dimensional representation of users and items obtained by TTM. This augmented model is called matrix factorization by learning a bilinear map. A core regularized version is further developed to incorporate additional information from the TTM. Encouraging experimental results not only show that the TTM outperforms existing topic models in modeling texts with a user-item-word structure, but also show that our proposed rating prediction models outperform state-of-the-art approaches.

© 2016 Elsevier Inc. Automatic linking of online content improves navigation possibilities for end users. We focus on linking content generated by users to other relevant sites. In particular, we study the problem of linking information between different usages of the same language, e.g., colloquial and formal idioms or the language of consumers versus the language of sellers. The challenge is that the same items are described using very distinct vocabularies. As a case study, we investigate a new task of linking textual Pinterest.com pins (colloquial) to online webshops (formal). Given this task, our key insight is that we can learn associations between formal and informal language by utilizing aligned data and probabilistic modeling. Specifically, we thoroughly evaluate three different modeling paradigms based on probabilistic topic modeling: monolingual latent Dirichlet allocation (LDA), bilingual LDA (BiLDA) and a novel multi-idiomatic LDA model (MiLDA). We compare these to the unigram model with Dirichlet prior. Our results for all three topic models reveal the usefulness of modeling the hidden thematic structure of the data through topics, as opposed to the linking model based solely on the standard unigram. Moreover, our proposed MiLDA model is able to deal with intrinsic multi-idiomatic data by considering the shared vocabulary between the aligned document pairs. The proposed MiLDA obtains the largest stability (less variation with changes in parameters) and highest mean average precision scores in the linking task. publisher: Elsevier articletitle: Latent Dirichlet allocation for linking user-generated content and e-commerce data journaltitle: Information Sciences articlelink: http://dx.doi.org/10.1016/j.ins.2016.05.047 content_type: article copyright: © 2016 Elsevier Inc. All rights reserved. ispartof: Information Sciences vol:367 pages:573-599 status: published

Word representation is crucial to lexical features used in Twitter sentiment analysis models. Recent work has demonstrated that dense, low-dimensional and real-valued word embedding gives competitive performance for Twitter sentiment classification. We follow this line of work, and propose a topic-enhanced word embedding for the task, which is generally neglected in previous work. Firstly, we exploit a recursive autoencoder framework to learn topic-enhanced word embedding, where topic information is generated through topic modeling based on an effective implementation of Latent Dirichlet Allocation (LDA). Then we use a uniform framework by adopting Support Vector Machine (SVM) classifier, to compare existing word representation methods with our method. Experimental results on the dataset show that topic-enhanced word embedding is very effective for Twitter sentiment classification.

Novel interactive activity recognition method using a topic modeling technique.Pose-based spatio-temporal relation features for intra- and inter-person.Two types of codeword corresponding to joint distance and angle.Four-level Pachinko Topic Model (PAM) for flexibly modeling interactions.Outperformance of recognition accuracy to state-of-the-art methods. In this paper, we go beyond the problem of recognizing video-based human interactive activities. We propose a novel approach that permits to deeply understand complex person-person activities based on the knowledge coming from human pose analysis. The joint coordinates of interactive objects are first located by an efficient human pose estimation algorithm. The relation features consisting of the intra and inter-person features of joint distance and angle, are suggested to use for describing the relationships between body components of the individual persons and the interacting two participants in the spatio-temporal dimension. These features are then provided to the codebook construction process, in which two types of codeword are generated corresponding to distance and angle features. In order to explain the relationships between poses, a flexible hierarchical topic model constructed by four layers is proposed using the Pachinko Allocation Model. The model is able to represent the full correlation between the relation features of body components as codewords, the interactive poselets as subtopics, and the interactive activities as super topics. Discrimination of complex activities presenting similar postures is further obtained by the proposed model. We validate our interaction recognition method on two practical data sets, the BIT-Interaction data set and the UT-Interaction data set. The experimental results demonstrate that the proposed approach outperforms recent interaction recognition approaches in terms of recognition accuracy.

In this paper, we present a new visual search system for finding similar images in a large database. However, there are a number of challenges regarding the robustness of the image representations and the efficiency of the retrieval framework. To tackle these challenges, we first propose an encoding technique based on soft-assignment of local features to convert an entire image into a single vector, which is a compact and discriminative representation. This encoded vector is suitable for most types of efficient indexing methods to produce an initial result. To compensate for the lack of incorporating geometric and object-related information during the encoding scheme, we then propose a probabilistic topic model to formalize the spatial structure among the local features. Moreover, the topic model allows us to effectively extract the object and background regions from the image. This is performed by a Markov Chain Monte Carlo algorithm for approximate inference. Finally, benefiting from the extracted objects in each image, we present a re-ranking scheme to automatically refine the initial search results. Our proposed retrieval framework has two major advantages: i) an aggregation strategy through soft-assignment improves the discriminative power of the representation, which has a determinative effect on the retrieval precision; and ii) the probabilistic latent topic model enables us to not only gain insight into the spatial structure of the image, but also handle a large variation in the object appearance. The experimental results from four benchmark datasets show that our approach provides competitive accuracy, and runs about ten times faster. Our studies also verify that proposed approach works effectively on large-scale databases of millions of images.

This paper proposes a novel topic-based model for identifying experimental mentions of Protein-Protein Interaction Method (PPIM) in the biomedical literature. The model combines topic-based classification models and some basic question-answering extraction techniques aiming at effectively detecting and identifying PPIM mentions on Protein-Protein Interactions. Unlike other state-of-the-art approaches, the approach captures underlying relationships within both input and output concept spaces by assuming the extraction task to be strongly driven by context provided by experts, usually in the form of a question to guide the search. Results indicate our topic-based question-driven approach obtained better results than other unsupervised learning probabilistic latent space models for detecting correct answers (PPIM mentions).

Recent research has focused on extracting personal interest data from social media. Although many methods have been developed, accurately estimating users' interests is often difficult because messages on social media are short and are not classified into any predefined categories. We propose a new method to overcome this problem by incorporating heterogeneous media, such as news. In our method, we first extract explicit features and implicit topics of categories using news media, where implicit topics are determined using a refined topic model. Next, we describe social media messages using these features and topics to estimate users' interests. Compared with several other approaches, our approach provides more accurate estimations of users' interests. We also demonstrate that the accuracy of friend recommendations is increased using the users' interests estimated by our method. Thus, we expect that the proposed approach could be helpful for enhancing the personalization of social media services.

The availability of electronic word-of-mouth, online consumer reviews, is increasing rapidly. Users frequently look for important aspects of a product or service in the reviews. They are typically interested in sentiment-oriented ratable aspects (i.e., semantic aspects). However, extracting semantic aspects across domains is challenging. We propose a domain-independent topic sentiment model called Joint Multi-grain Topic Sentiment (JMTS) to extract semantic aspects. JMTS effectively extracts quality semantic aspects automatically, thereby eliminating the requirement for manual probing. We conduct both qualitative and quantitative comparisons to evaluate JMTS. The experimental results confirm that JMTS generates semantic aspects with correlated top words and outperforms state-of-the-art models in several performance metrics.

This paper addresses deficiencies in current information retrieval models by integrating the concept of relevance into the generation model using various topical aspects of the query. The models are adapted from the latent Dirichlet allocation model, but differ in the way that the notation of query-document relevance is introduced in the modeling framework. In the first method, query terms are added to relevant documents in the training of the latent Dirichlet allocation model. In the second method, the latent Dirichlet allocation model is expanded to deal with relevant query terms. The topic of each term within a given document may be sampled using either the normal document-specific mixture weights in LDA using query-specific mixture weights. We also developed an efficient method based on the Gibbs sampling technique for parameter estimation. Experiment results based on the Text REtrieval Conference Corpus (TREC) demonstrate the superiority of the proposed models.

The rapid development of social media services has facilitated the communication of opinions through online news, blogs, microblogs/tweets, instant-messages, and so forth. This article concentrates on the mining of readers' emotions evoked by social media materials. Compared to the classical sentiment analysis from writers' perspective, sentiment analysis of readers is sometimes more meaningful in social media. We propose two sentiment topic models to associate latent topics with evoked emotions of readers. The first model which is an extension of the existing Supervised Topic Model, generates a set of topics from words firstly, followed by sampling emotions from each topic. The second model generates topics from social emotions directly. Both models can be applied to social emotion classification and generate social emotion lexicons. Evaluation on social emotion classification verifies the effectiveness of the proposed models. The generated social emotion lexicon samples further show that our models can discover meaningful latent topics exhibiting emotion focus.