Search

Social networks play a key role in providing a new approach to collecting information regarding mobility and transportation services. To study this information, sentiment analysis can make decent observations to support intelligent transportation systems (ITSs) in examining traffic control and management systems. However, sentiment analysis faces technical challenges: extracting meaningful information from social network platforms, and the transformation of extracted data into valuable information. In addition, accurate topic modeling and document representation are other challenging tasks in sentiment analysis. We propose an ontology and latent Dirichlet allocation (OLDA)-based topic modeling and word embedding approach for sentiment classification. The proposed system retrieves transportation content from social networks, removes irrelevant content to extract meaningful information, and generates topics and features from extracted data using OLDA. It also represents documents using word embedding techniques, and then employs lexicon-based approaches to enhance the accuracy of the word embedding model. The proposed ontology and the intelligent model are developed using Web Ontology Language and Java, respectively. Machine learning classifiers are used to evaluate the proposed word embedding system. The method achieves accuracy of 93%, which shows that the proposed approach is effective for sentiment classification.

Document representation plays an important role in the fields of text mining, natural language processing, and information retrieval. Traditional approaches to document representation may suffer from the disregard of the correlations or order of words in a document, due to unrealistic assumption of word independence or exchangeability. Recently, long–short-term memory (LSTM) based recurrent neural networks have been shown effective in preserving local contextual sequential patterns of words in a document, but using the LSTM model alone may not be adequate to capture global topical semantics for learning document representation. In this work, we propose a new topic-enhanced LSTM model to deal with the document representation problem. We first employ an attention-based LSTM model to generate hidden representation of word sequence in a given document. Then, we introduce a latent topic modeling layer with similarity constraint on the local hidden representation, and build a tree-structured LSTM on top of the topic layer for generating semantic representation of the document. We evaluate our model in typical text mining applications, i.e., document classification, topic detection, information retrieval, and document clustering. Experimental results on real-world datasets show the benefit of our innovations over state-of-the-art baseline methods.

Nowadays, most recommendation approaches used to suggest appreciate items for individual users. However, due to the social nature of human beings, group activities have become an integral part of our daily life, thus motivating the study on group recommendation. Unfortunately, most existing approaches used in group recommender systems make recommendations through aggregating individual preferences or individual predictive results rather than comprehensively investigating the social features that govern user choices made within a group. As a result, such approaches often fail to detect many latent factors that could potentially improve the performance of the group recommender systems. Therefore, we propose a new approach, random walks based on a topic model (RTM), for group recommendations through combining an integrated probabilistic topic model − a User Topic Model (UTM) with the Random Walk with Restart (RWR) method. The goal of the work in this paper is better identify group preferences by comprehensively detecting the latent associations among group members, in order to alleviate the data sparsity problem and improve the performance of group recommender systems. The UTM provides a latent framework of users, groups, and items by exploiting both the users’ preference profiles and the items’ content information, which together can describe group interests and item features in a more complete manner. This latent framework is then combined with RWR to predict the preference degrees of groups to unrated items. In particular, we develop two different recommendation strategies based on the proposed approach, and design a special random walk path for each developed recommendation strategy to comprehensively detect various latent associations. Finally, we conduct experiments to evaluate our approach and compare it with other state-of-the-art approaches using the real-world CAMRa2011 dataset. The results demonstrate the advantage of our approach over comparative ones.

Learning topics from short texts has become a critical and fundamental task for understanding the widely-spread streaming social messages, e.g., tweets, snippets and questions/answers. Up to date, there are two distinctive topic learning schemes: generative probabilistic graphical models and geometrically linear algebra approaches, with LDA and NMF being the representative works, respectively. Since these two methods both could uncover the latent topics hidden in the unstructured short texts, some interesting doubts are coming to our minds that which one is better and why? Are there any other more effective extensions? In order to explore valuable insights between LDA and NMF based learning schemes, we comprehensively conduct a series of experiments into two parts. Specifically, the basic LDA and NMF are compared with different experimental settings on several public short text datasets in the first part which would exhibit that NMF tends to perform better than LDA; in the second part, we propose a novel model called “Knowledge-guided Non-negative Matrix Factorization for Better Short Text Topic Mining” (abbreviated as KGNMF), which leverages external knowledge as a semantic regulator with low-rank formalizations, yielding up a time-efficient algorithm. Extensive experiments are conducted on three representative corpora with currently typical short text topic models to demonstrate the effectiveness of our proposed KGNMF. Overall, learning with NMF-based schemes is another effective manner in short text topic mining in addition to the popular LDA-based paradigms.

To address the problem of query-focused multi-document summarisation, we present a novel unsupervised pattern-enhanced approach for representing coherent topics across documents, as well as the query relevance, in order to generate topically coherent summaries that meet the information needs of users. The proposed model employs not only a pattern-enhanced topic model to generate discriminative and semantic rich representations for topics and documents, but also a pattern-based relevance model for the query relevance of sentences. With these dual pattern-based representations for sentences, we are able to integrate various indicative metrics, such as rational coverage of document topics and sentence relevance, into a unified model. When evaluated on the datasets of the document understanding conferences of 2006 and 2007, the proposed approach shows a performance improvement as compared to a number of state-of-the-art methods and unsupervised baseline systems.

Social media web sites have become major media platforms to share personal information, news, photos, videos and more. Users can even share live streams whenever they want to reach out to many other. This prevalent usage of social media attracted companies, data scientists, and researchers who are trying to infer meaningful information from this vast amount of data. Information diffusion and maximizing the spread of words is one of the most important focus for researchers working on social media. This information can serve many purposes such as; user or content recommendation, viral marketing, and user modeling. In this research, finding topical influential/authority users on Twitter is addressed. Since Twitter is a good platform to spread knowledge as a word of mouth approach and it has many more public profiles than protected ones, it is a target media for marketers. In this paper, we introduce a novel methodology, called Personalized PageRank, that integrates both the information obtained from network topology and the information obtained from user actions and activities in Twitter. The proposed approach aims to determine the topical influencers who are experts on a specific topic. Experimental results on a large dataset consisting of Turkish tweets show that using user specific features like topical focus rate, activeness, authenticity and speed of getting reaction on specific topics positively affects identifying influencers and lead to higher information diffusion. Algorithms are implemented on a distributed computing environment which makes high-cost graph processing more efficient.

Nowadays plenty of user-generated posts, e.g., sina weibos, are published on the social media. The posts contain the public’s sentiments (i.e., positive or negative) towards various topics. Bursty sentiment-aware topics from these posts reveal sentiment-aware events which have attracted much attention. To detect sentiment-aware topics, we attempt to utilize Joint Sentiment/Topic models, these models are achieved with Latent Dirichlet Allocation (LDA) based models. However, most of the existing sentiment/topic models cannot be directly utilized to detect sentiment-aware topics on the posts, since applying the models to the posts directly suffers from the context sparsity problem. In this paper, we propose a Time-User Sentiment/Topic Latent Dirichlet Allocation (TUS-LDA) which simultaneously models sentiments and topics for posts. Thereinto, TUS-LDA aggregates posts in the same timeslices or from the same users as pseudo-documents to alleviate the context sparsity problem. Based on TUS-LDA, we further design an approach to detect bursty sentiment-aware topics and these sentiment-ware topics can reflect bursty real-world events. Experiments on the Chinese sina weibos show that TUS-LDA outperforms previous models in the tasks of sentiment classification and burst detection in sentiment-aware topics. Finally, we visualize the bursty sentiment-aware topics discovered by TUS-LDA.

With the rapid growth of population on social networks, people are confronted with information overload problem. This clearly makes filtering the targeted users a demanding and key research task. Uni-directional social networks are the scenarios where users provide limited follow or not binary features. Related works prefer to utilize these follower-followee relations for recommendation. However, a major problem of these methods is that they assume every follower-followee user pairs are equally likely, and this leads to the coarse user following preferences inferring. Intuitively, a user’s adoption of others as followees may be motivated by her interests as well as social connections, hence a good recommender should be able to separate the two situations and take both factors into account for better recommendation results. In this regard, we propose a new user recommendation framework namely UIS-MF in this work. UIS-MF can well capture user preferences by involving both interest and social factors in prediction, and targeted to recommend Top- N followees who have similar interest and close social connection relevant to a target user. Specifically, we first present a unified probabilistic topic model on follower-followee relations, namely UIS-LDA , and it employs Generalized Polya Urn (GPU) models on mutual-following relations for discovering interest topics and social topics of users. Next we propose a community-based method for user recommendation, it organizes social communities and interest communities based on the estimation of topics obtained from UIS-LDA , and then performs Matrix Factorization (MF) method on each community to generate N most likely followees for individual user. Systematic experiments on Twitter, Sina Weibo and Epinions datasets have not only revealed the significant effect of our UIS-LDA model for the extraction of interest and social topics of users in improving recommending accuracy, but also demonstrated the advantage of our proposed recommendation framework over competitive baselines by large margins.

Abstractive text summarization (ATS) often fails to capture salient information and preserve the original meaning of the content in the generated summaries due to a lack of background knowledge. We present a method to provide the topic information based on the background knowledge of documents to a deep learning-based summarization model. This method comprises a topic knowledge base (TKB) and convolutional sequence network-based text summarization model with knowledge-powered topic level attention (KTOPAS). TKB employs conceptualization to retrieve the semantic salient knowledge of documents and the knowledge-powered topic model (KPTopicM) to generate coherent and meaningful topic information by utilizing the knowledge that represents the documents well. KTOPAS obtains knowledge-powered topic information (also called topic knowledge) from TKB and incorporates the topic knowledge into the convolutional sequence network through a high-level topic level attention to resolve the existing issues in ATS. KTOPAS introduces a tri-attention channel to jointly learn the attention of the source elements over the summary elements, the source elements over topic knowledge, and topic knowledge over the summary elements to present the contextual alignment information from three aspects and combine them using the softmax function to generate the final probability distribution which enables the model to produce coherent, concise, and human-like summaries with word diversity. By conducting experiments on datasets, namely CNN/Daily Mail and Gigaword, the results show that our proposed method consistently outperforms the competing baselines. Moreover, TKB improves the effectiveness of the resulting summaries by providing topic knowledge to KTOPAS and demonstrates the quality of the proposed method.

Topic modeling is a growing field within the area of text analysis that extracts underlying topics from document collections. Several objectives can be simultaneously considered when designing an approach for topic modeling. A multi-objective optimization approach based on the swarm intelligence of a bee colony (MOABC, Multi-Objective Artificial Bee Colony) has been designed, implemented, and tested. This new approach has been evaluated by using documents from the Reuters-21578 and TagMyNews datasets. Three objective functions (coherence, coverage, and perplexity) and three multi-objective metrics (hypervolume, set coverage, and distance to the ideal point) have been considered in two topic scenarios. Results show that MOABC provides relevant improvements with respect to LDA (Latent Dirichlet Allocation, the most used approach for topic modeling) and MOEA/D (Multi-Objective Evolutionary Algorithm based on Decomposition, the only multi-objective approach published to date). This demonstrates that the multi-criteria nature of topic modeling should be exploited with multi-objective optimization approaches.

Because of its efficiency, word embedding has been widely used in many natural language processing and text modeling tasks. It aims to represent each word by a vector so such that the geometry between these vectors can capture the semantic correlations between words. An ambiguous word can often have diverse meanings in different contexts, a quality which is called polysemy. The bulk of studies aimed to generate only one single embedding for each word, whereas a few studies have made a small number of embeddings to present different meanings of each word. However, it is hard to determine the exact number of senses for each word, as meanings depend on contexts. To address this problem, this paper proposes a novel adaptive cross-contextual word embedding (ACWE) method for capturing the word polysemy in different contexts based on topic modeling, in which the word polysemy is defined over a latent interpretable semantic space. The proposed ACWE consists of two main parts, in the first of which an unsupervised cross-contextual probabilistic word embedding model is designed to obtain the global word embeddings, and each word is represented by an embedding in the unified latent semantic space. Based on the global word embeddings, an adaptive cross-contextual word embedding process is then devised in the second part to learn the local embeddings for each polysemous word in different contexts. In fact, a word embedding is adaptively adjusted and updated with respect to different contexts to generate different word embeddings tailored to the corresponding contexts. The proposed ACWE is validated on two datasets collected from Wikipedia and IMDb on different tasks including word similarity, polysemy induction, semantic interpretability, and text classification. Experimental results indicate that ACWE does not only outperform the established word embedding methods, which consider word polysemy on six popular benchmark datasets, but it also yields competitive performance compared with state-of-the-art deep learning-based approaches without considering polysemy. Moreover, the proposed ACWE significantly improves the performances of text classification both in precision and F1, and the visualizations of the semantics of words demonstrate the feasibility and advantage of the proposed ACWE model on polysemy.

Due to the popularity of social networks, such as microblogs and Twitter, a vast amount of short text data is created every day. Much recent research in short text becomes increasingly significant, such as topic inference for short text. Biterm topic model (BTM) benefits from the word co-occurrence patterns of the corpus, which makes it perform better than conventional topic models in uncovering latent semantic relevance for short text. However, BTM resorts to Gibbs sampling to infer topics, which is very time consuming, especially for large-scale datasets or when the number of topics is extremely large. It requires O ( K ) operations per sample for K topics, where K denotes the number of topics in the corpus. In this paper, we propose an acceleration algorithm of BTM, FastBTM, using an efficient sampling method for BTM, which converges much faster than BTM without degrading topic quality. FastBTM is based on Metropolis-Hastings and alias method, both of which have been widely adopted in Latent Dirichlet Allocation (LDA) model and achieved outstanding speedup. Our FastBTM can effectively reduce the sampling complexity of biterm topic model from O ( K ) to O (1) amortized time. We carry out a number of experiments on three datasets including two short text datasets, Tweets2011 Collection dataset and Yahoo! Answers dataset, and one long document dataset, Enron dataset. Our experimental results show that when the number of topics K increases, the gap in running time speed between FastBTM and BTM gets especially larger. In addition, our FastBTM is effective for both short text datasets and long document datasets.

The popularity of online forums provides a good opportunity to learn user interests which can be used in many business scenarios, such as product or news recommendation. There exist many approaches to infer forum topics and users interests. Among them, Author-Topic (AT) like models are most popular. But a thread in online forum is composed of a root post and some response posts which may be relevant or irrelevant to the root post. So the assumption of AT that response posts are generated from users interest topics is not comprehensive. In this paper, we distinguish users serious and unserious interest topics and argue that the topic of a relevant response post is jointly determined by its authors serious interest topics and the topics of its root post, while the topic of irrelevant response post is only determined by its authors unserious interest topics. Based on these assumptions, we propose Forum-LDA to model the generative process of root post, relevant and irrelevant response posts jointly. Therefore, our model can not only learn more coherent topics and serious interests, but also identify unserious users who publish many irrelevant posts. Extensive experiments on real forum dataset demonstrate the advantages of our model in tasks such as user interest and unserious user discovery.

Named entity disambiguation (NED) refers to the task of mapping entity mentions in running texts to the correct entries in a specific knowledge base (e.g., Wikipedia). Although there has been a lot of work on NED for long and formal texts like Wikipedia and news, the task is not well studied for questions in community question answering (CQA). The challenges of the task include little context for mentions in questions, lack of ground truth for learning, and language gaps between CQA and knowledge bases. To overcome these problems, we propose a topic modelling approach to NED for questions. Our model performs learning in an unsupervised manner, but can take advantage of weak supervision signals estimated from the metadata of CQA and knowledge bases. The signals can enrich the context of mentions in questions, and bridge the language gaps between CQA and knowledge bases. Besides these advantages, our model simulates peoples behavior in CQA and thus is intuitively interpretable. We conduct experiments on both Chinese and English CQA data. The experimental results show that our method can significantly outperform state-of-the-art methods when we apply them to questions in CQA.

Topic model is a hot research topic which is attracting attentions from many fields. Recently, several studies have applied topic model to ASR (audio scene recognition). Among these studies, most of them use the document-word co-occurrence matrix for topic analysis. In this work, we propose a new ASR algorithm based on audio events and topic model, which uses the document-event co-occurrence matrix for topic analysis. Our work is based on the hypothesis that: for an audio document, compared with its word distribution, its event distribution is more in line with humans way of thinking, and then the topic distribution obtained based on the document-event co-occurrence matrix can represent the audio document better. The contribution of this work lies in that: (1) we propose an ASR algorithm which uses document-event co-occurrence matrix for topic analysis. Compared with the current studies which use document-word co-occurrence matrix for topic analysis, the proposed algorithm can extract the topic distribution which can express the audio documents better, and then can get better recognition results; (2) we propose a much easier method to obtain the document-event co-occurrence matrix; (3) we propose a method to weight the event distribution of audio documents; this weighting method can emphasize the audio events that are important in reflecting the unique topics of the audio documents, and can suppress the audio events that are common to many topics. Experimental results on two public datasets verify the effectiveness of the proposed ASR algorithm, and also verify the necessity and effectiveness of the proposed weighting method. The innovative ideas in this work are not limited to ASR, but can be extended to many other fields, such as the video classification etc.

Rapid advancements in internet and social media technologies have made information overload a rampant and widespread problem. Complex subjects, histories, or issues break down into branches, side stories, and intertwining narratives; a topic evolution map can assist in joining together and clarifying these disparate parts of an unfamiliar territory. This paper reviews the extant research on topic evolution map based on text and cross-media corpora over the past decade. We first define a series of necessary terms, then go on to describe the traditional topic evolution map per 1) topic evolution over time, based on the probabilistic generative model, and 2) topic evolution from a non-probabilistic perspective. Next, we discuss the current state of research on topic evolution map based on the cross-media corpus, including some open questions and possible future research directions. The main contribution of this review is in its construction of an evolution map that can be used to visualize and integrate the extant studies on topic modeling specifically in regards to cross-media research.

Topic models, such as probabilistic latent semantic analysis (PLSA) and latent Dirichlet allocation (LDA), have shown impressive success in many fields. Recently, multi-view learning via probabilistic latent semantic analysis (MVPLSA), is also designed for multi-view topic modeling. These approaches are instances of generative model, whereas they all ignore the manifold structure of data distribution, which is generally useful for preserving the nonlinear information. In this paper, we propose a novel multiple graph regularized generative model to exploit the manifold structure in multiple views. Specifically, we construct a nearest neighbor graph for each view to encode its corresponding manifold information. A multiple graph ensemble regularization framework is proposed to learn the optimal intrinsic manifold. Then, the manifold regularization term is incorporated into a multi-view topic model, resulting in a unified objective function. The solutions are derived based on the Expectation Maximization optimization framework. Experimental results on real-world multi-view data sets demonstrate the effectiveness of our approach.

This paper presents HBIN-LBD, a novel literature-based discovery (LBD) method that exploits the lexico-citation structures within the heterogeneous bibliographic information network (HBIN) graphs. Unlike other existing LBD methods, HBIN-LBD harnesses the metapath features found in HBIN graphs for discovering the latent associations between scientific papers published in otherwise disconnected research areas. Further, this paper investigates the effects of incorporating semantic and topic modeling components into the proposed models. Using time-sliced historical bibliographic data, we demonstrate the performance of our method by reconstructing two LBD hypotheses: the Fish Oil and Raynaud’s Syndrome hypothesis and the Migraine and Magnesium hypothesis. The proposed method is capable of predicting the future co-citation links between research papers of these previously disconnected research areas with up to 88.86% accuracy and 0.89 F-measure.

Creating links among online encyclopedia articles in different languages is crucial in the construction and integration of large multilingual knowledge bases. Most research to date has focused on linking among different language versions of Wikipedia, yet other large online encyclopedias in a variety of languages exist. In this work, we present a cross-language article-linking method using a bilingual topic model and translation features based on an SVM model to link articles in English Wikipedia and Chinese Baidu Baike, the most widely used Wiki-like encyclopedia in China. To evaluate our approach, we compile data sets from Baidu Baike articles and their corresponding English Wikipedia articles. The evaluation results show that our approach achieves at most 0.8158 in MRR, outperforming the baseline system by 0.1328 (+19.44%) in MRR. Our method does not heavily depend on linguistic characteristics, and it can be easily extended to generate cross-language article links among different online encyclopedias in other languages.

Recommender system has been recognized as a superior way for solving personal information overload problem. Rating, as an evaluation criteria revealing how much a customer likes a product, has been a foundation of recommender systems for a long period based on the popular latent factor models. However, review texts as the valuable user generated content have been neglected all the time. Recently, models integrating ratings and review texts as training sources have attracted a lot of attention, which may model review texts by topic model or its variants and then link latent factor vectors to topic distribution of review texts. For that, the integrated models need complicated optimization algorithms to fuse the heterogeneous sources, that may cause greater errors.In this work, we aim to propose a novel model, called Aspect-based Latent Factor Model (ALFM) to integrate ratings and review texts via latent factor model, in which by integrating rating matrix, user-review matrix and item-attribute matrix, the user latent factors and item latent factors with word latent factors can be derived. Our proposed model aggregates all review texts of the same user on the respective items and builds a user-review matrix by word frequencies. Similarly, an item's review is considered as all review texts of the same item collected from respective users. According to different information abstracted from review texts, we introduce two different kinds of item-attribute matrix to integrate the item-word frequencies and polarity scores of corresponding words. Experimental results on real-world data sets from amazon.com illustrate that our model can not only perform better than traditional models and art-of-state models on rating prediction task, but also accomplish cross-domain task through transferring word embedding.

Topic modeling discovers the hidden topics in a document collection. Most of the existing topic models focus only on word usage and generate the topics based on the word frequency and co-occurrence without considering the meaning of the text. In this paper, we propose a novel approach to generate a semantic pattern-based topic representation based on the meaning of the text to represent the topics in a document collection. The proposed approach considers both the semantics and co-occurrence of words to generate a set of frequent semantic patterns to represent each topic. The semantics are captured by matching the words in each topic with concepts in the Probase ontology. A set of frequent semantic patterns in each topic is generated based on the co-occurrence of the matched words to represent the topic. Hence, our approach differs from traditional topic models because of the meaningful frequent semantic patterns generated based on the ontology. The proposed topic representation was evaluated in terms of topic quality and information filtering performance against a set of state-of-the-art systems. Perplexity, coherence, and topic word distribution were examined in the topic quality evaluation. The generated frequent semantic patterns were used as features for the information filtering evaluation. Our topic representation outperformed in all the evaluations.

Review-based recommendation algorithms can alleviate the data sparsity issue in collaborative filtering by combining user ratings and reviews in model learning. However, most existing methods simplify the feature extraction process from reviews by assuming that different granularities of information (e.g., word, review, and feature) are equally important, which cannot optimally leverage the most important information and thus achieves suboptimal recommendation accuracy. Besides, many existing works directly regard text features as users or items representations, which may not be enough to make precise representations due to the large amount of redundant information in reviews. To tackle the two problems mentioned above, we propose a deep learning-based method named Hierarchical Attention Network Oriented Towards Crowd Intelligence (HANCI). First, HANCI replaces the commonly-used topic models or CNN text processor with an RNN text processor in review feature extraction, which can fully exploit the advantages of the sequential dependencies of reviews by using the whole hidden layers of the bidirectional LSTM as outputs. Second, HANCI weighs the importance of features guided by crowd intelligence to more accurately represent each user on each item, and vice versa. Third, HANCI utilizes a hierarchical attention network based on multi-level review text analysis to extract more precise user preferences and item latent features, so that HANCI can explore the importance of words, the usefulness of reviews and the importance of features to achieve more accurate recommendation. Extensive experiments on three public datasets show that HANCI outperforms the state-of-the-art review-based recommendation algorithms in accuracy and meanwhile provides insightful explanations.

An extremely competitive business environment requires every company to monitor its competitors and anticipate future opportunities and risks, creating a dire need for competitive intelligence. In response to this need, foresight study became a prominent field, especially the concept of weak signal detection. This research area has been widely studied for its utility, but it is limited by the need of human expert judgments on these signals. Moreover, the increase in the volume of information on the Internet through blogs and web news has made the detection process difficult, which has created a need for automation. Recent studies have attempted topic modeling techniques, specifically latent Dirichlet allocation (LDA), for automating the weak signal detection process; however, these approaches do not cover all parts of the process. In this study, we propose a fully automatic LDA-based weak signal detection method, consisting of two filtering functions: the weakness function aimed at filtering topics, which potentially contains weak signals, and the potential warning function, which helps to extract only early warning signs from the previously filtered topics. We took this approach with a famous daily web news dataset, and we could detect the risk of the COVID19 pandemic at an early stage.

Nowadays, social network sites (SNSs) have been significant platforms for content sharing in our daily life. With the emergence of different kinds of social network sites and users’ diverse needs for content sharing, their content sharing practices are generally taken place in multiple SNSs. To construct models that can characterize users’ content sharing practices in a composite context constituted by multiple social network sites (cross-site user generated content modeling) has been an emerging research topic in web data mining and human behavior research. However, previous methods such as Dirichlet Multinomial Mixture model (DMM), Biterm Topic Model (BTM), Twitter-LDA and MultiLDA have limited representation ability or are based on unreliable assumption, which cannot characterize the user generated content (UGC) accurately from the perspective of multiple SNSs. In this paper, we first conduct an empirical study to investigate the characteristics of users’ content sharing practices in cross-site context, based on which we propose a more reliable cross-site UGC model named CrossSite-LDA (C-LDA). We then evaluate the performances of the C-LDA model with four state-of-the-art models based on the two data sets sampled from Weibo–Douban and Facebook–Twitter. Results show that the C-LDA has better performances in perplexity, word coherence, topic KL divergence, UCI and UMass metrics compared with existing models, which suggests its superior accuracy on modeling users’ content characteristics in cross-site context.

Predicting the popularity of web content is widely regarded as an important but challenging task. Online news articles are typical examples of this. In particular, owing to their time-sensitive nature, it is preferable to predict the popularity of news articles before their publication. To achieve this, this study proposes a named entity topic model (NETM) to extract the textual factors that can drive popularity growth. Here, each named entity is assumed to have a popularity-gain distribution over all semantic topics. The popularity of a news article is considered as the accumulation of popularity gains generated by its named entities (NEs) over all the topics. By learning the popularity-gain matrix for each named entity, the popularity of any news article can be predicted. Experiments on two collections of news articles demonstrate that the proposed NETM can outperform existing models in terms of accuracy. Additionally, the popularity-gain matrix learned by the NETM can be used to effectively explain the popularity of specific news articles.

Twitter is an online social networking microblogging service that allows registered users to broadcast 140-character messages called tweets . The service has gained worldwide popularity since it was created in March 2006, with more than 316 million monthly active users in June 2015 who posted 500 million tweets per day. As the number of available tweets grows, the problem of managing tweets becomes extremely difficult, which could lead to information overload. To avoid this problem, people use the hashtag symbol # before a relevant keyword or phrase in their tweets to categorize those tweets and help them show more easily in each Twitter search. Furthermore, hashtags can be used to collect public opinions on events and their ideas at the individual, community or even the world level. Incorporating hashtags to obtain better performance such as sentiment classification and breaking events detection also has attracted considerable research attention in recent years. However, there are very few tweets containing hashtags, which impedes the quality of search results and their further usage in various applications. Therefore, hashtag recommendation has become a particularly important research problem. In this paper, we first propose a novel model, namely online Twitter-User LDA to learn Twitter users’ dynamic interests. Then considering the shortness, sparsity, and high volume of tweets, we introduce an effective method to discover the latent topics of streaming tweet content, which uses recently proposed incremental biterm topic model (IBTM). We finally design an automatic hashtag recommendation method called User-IBTM by combining the online Twitter-User LDA and IBTM. As shown in the experimental results on real world data from Twitter, our design method based on dynamic user interests and streaming tweet content significantly outperforms several other baseline methods and can suggest more precise hashtags.

Most current book recommendation and marketing strategies in online social media are implemented by creating topics or posting advertisements for the brand. They do not precisely target the audiences who are interested in these books, so the recommendation or marketing quality is not guaranteed. In order to solve this problem, we propose an effective audience detection method based on Group-based Latent Dirichlet Allocation (Group-LDA) in order to precisely detect book audiences. Group-LDA is a new probabilistic topic model derived from Latent Dirichlet Allocation (LDA), which introduces a new latent concept of group to describe the topic relevance among documents by incorporating book module and book chapter information into the model. Group-LDA is evaluated on Weibo.com with fifty popular books randomly sampled from the reading channel on Douban.com. According to the evaluation results, Group-LDA can effectively detect different types of readers for most categories of books. It outperforms LSA, LDA, author-topic model (ATM) and some other collaborative filtering methods in terms of precision, recall, F1-score and MAP for book audience detection.

We exploit the knowledge from a topic-consistent corpus for topic modeling and use the topics to enrich the corpus and the short texts.We learn the vector representations of both words and topics interactively on the enriched corpus.We use the vectors of the words and topics to represent the features of short texts for training and classification.Our method performs better than many baselines. This paper presents a general framework for short text classification by learning vector representations of both words and hidden topics together. We refer to a large-scale external data collection named "corpus" which is topic consistent with short texts to be classified and then use the corpus to build topic model with Latent Dirichlet Allocation (LDA). For all the texts of the corpus and short texts, topics of words are viewed as new words and integrated into texts for data enriching. On the enriched corpus, we can learn vector representations of both words and topics. In this way, feature representations of short texts can be performed based on vectors of both words and topics for training and classification. On an open short text classification data set, learning vectors of both words and topics can significantly help reduce the classification error comparing with learning only word vectors. We also compared the proposed classification method with various baselines and experimental results justified the effectiveness of our word/topic vector representations.

Explicit Semantic Analysis (ESA) is a knowledge-based method which builds the semantic representation of the words depending on the textual description of the concepts in the certain knowledge source. Due to its simplicity and success, ESA has received wide attention from researchers in the computational linguistics and information retrieval. However, the representation vectors formed by ESA method are generally very excessive, high dimensional, and may contain many redundant concepts. In this paper, we introduce a reduced semantic representation method that constructs the semantic interpretation of the words as the vectors over the latent topics from the original ESA representation vectors. For modeling the latent topics, the Latent Dirichlet Allocation (LDA) is adapted to the ESA vectors for extracting the topics as the probability distributions over the concepts rather than the words in the traditional model. The proposed method is applied to the wide knowledge sources used in the computational semantic analysis: WordNet and Wikipedia. For evaluation, we use the proposed method in two natural language processing tasks: measuring the semantic relatedness between words/texts and text clustering. The experimental results indicate that the proposed method overcomes the limitations of the representation of the ESA method.

Electronic Medical Record (EMR) has established itself as a valuable resource for large scale analysis of health data. A hospital EMR dataset typically consists of medical records of hospitalized patients. A medical record contains diagnostic information (diagnosis codes), procedures performed (procedure codes) and admission details. Traditional topic models, such as latent Dirichlet allocation (LDA) and hierarchical Dirichlet process (HDP), can be employed to discover disease topics from EMR data by treating patients as documents and diagnosis codes as words. This topic modeling helps to understand the constitution of patient diseases and offers a tool for better planning of treatment. In this paper, we propose a novel and flexible hierarchical Bayesian nonparametric model, the word distance dependent Chinese restaurant franchise (wddCRF), which incorporates word-to-word distances to discover semantically-coherent disease topics. We are motivated by the fact that diagnosis codes are connected in the form of ICD-10 tree structure which presents semantic relationships between codes. We exploit a decay function to incorporate distances between words at the bottom level of wddCRF. Efficient inference is derived for the wddCRF by using MCMC technique. Furthermore, since procedure codes are often correlated with diagnosis codes, we develop the correspondence wddCRF (Corr-wddCRF) to explore conditional relationships of procedure codes for a given disease pattern. Efficient collapsed Gibbs sampling is derived for the Corr-wddCRF. We evaluate the proposed models on two real-world medical datasets - PolyVascular disease and Acute Myocardial Infarction disease. We demonstrate that the Corr-wddCRF model discovers more coherent topics than the Corr-HDP. We also use disease topic proportions as new features and show that using features from the Corr-wddCRF outperforms the baselines on 14-days readmission prediction. Beside these, the prediction for procedure codes based on the Corr-wddCRF also shows considerable accuracy.

Social media systems provide ever-growing huge volumes of information for dissemination and communication among communities of users, while recommender systems aim to mitigate information overload by filtering and providing users the most attractive and relevant items from information-sea. This paper aims at providing compound recommendation engine for social media systems, and focuses on exploiting multi-sourced information (e.g. social networks, item contents and user feedbacks) to predict the ratings of users to items and make recommendations. For this, we suppose the users' decisions on adopting item are affected both by their tastes and the favors of trusted friends, and extend Collaborative Topic Regression to jointly incorporates social trust ensemble, topic modeling and probabilistic matrix factorization. We propose corresponding approaches to learning the latent factors both of users and items, as well as additional parameters to be estimated. Empirical experiments on Lastfm and Delicious datasets show that our model is better and more robust than the state-of-the-art methods on making recommendations in term of accuracy. Experiments results also reveal some useful findings to enlighten the development of recommender systems in social media.

We extract co-occurrence term relations using IdeaGraph.We extract semantic term relations using topic model.We fuse multiple types of relations to form a coupled term graph.We extract topics from the graph using a graph analytical approach. Topic detection as a tool to detect topics from online media attracts much attention. Generally, a topic is characterized by a set of informative keywords/terms. Traditional approaches are usually based on various topic models, such as Latent Dirichlet Allocation (LDA). They cluster terms into a topic by mining semantic relations between terms. However, co-occurrence relations across the document are commonly neglected, which leads to the detection of incomplete information. Furthermore, the inability to discover latent co-occurrence relations via the context or other bridge terms prevents the important but rare topics from being detected.To tackle this issue, we propose a hybrid relations analysis approach to integrate semantic relations and co-occurrence relations for topic detection. Specifically, the approach fuses multiple relations into a term graph and detects topics from the graph using a graph analytical method. It can not only detect topics more effectively by combing mutually complementary relations, but also mine important rare topics by leveraging latent co-occurrence relations. Extensive experiments demonstrate the advantage of our approach over several benchmarks.

As news events on the same subject occur, our knowledge about the subject will accumulate and become more comprehensive. In this paper, we formally define the problem of incremental knowledge learning from similar news events on the same subject, where each event consists of a set of news articles reporting about it. The knowledge is represented by a topic hierarchy presenting topics at different levels of granularity. Though topic (hierarchy) mining from text has been researched a lot, incremental learning from similar events remains under developed. In this paper, we propose a scalable two-phase framework to incrementally learn a topic hierarchy for a subject from events on the subject as the events occur. First, we recursively construct a topic hierarchy for each event based on a novel topic model considering the named entities and entity types in news articles. Second, we incrementally merge the topic hierarchies through top-down hierarchical topic alignment. Extensive experimental results on real datasets demonstrate the effectiveness and efficiency of the proposed framework in terms of both qualitative and quantitative measures.

In recent years, topic modeling is gaining significant momentum in information retrieval (IR). Researchers have found that utilizing the topic information generated through topic modeling together with traditional TF-IDF information generates superior results in document retrieval. However, in order to apply this idea to real-life IR systems, some critical problems need to be solved: how to store the topic information and how to utilize it with the TF-IDF information for efficient document retrieval. In this paper, we propose the Topic Enhanced Inverted Index (TEII) to incorporate the topic information into the inverted index for efficient top- k document retrieval. Specifically, we explore two different types of TEIIs. We first propose the incremental TEII, which includes the topic information into the traditional inverted index by adding topic-based inverted lists. The incremental TEII is beneficial for legacy IR systems, since it does not change the existing TF-IDF-based inverted lists. As a more flexible alternative, we propose the hybrid TEII to incorporate the topic information into each posting of the inverted index. In the hybrid TEII, two relaxation methods are proposed to support dynamic estimation of the upper bound impact of each posting. The hybrid TEII is highly extensible for incorporating different ranking factors and we show an extension of the hybrid TEII by considering the static quality of the documents in the corpus. Based on the incremental and hybrid TEIIs, we develop several query processing algorithms to support efficient top- k document retrieval on TEIIs. Empirical evaluation on the TREC dataset verifies the effectiveness and efficiency of the proposed index structures and query processing algorithms.

Collaborative topic modeling is powerful to alleviate data sparsity in recommender systems owing to the incorporation of collaborative filtering and topic models. However sufficient textual data is not always available. On the other hand, tags serving as supplementary description of items can reflect users’ interests in item attributes. But previous works only mine the effect of tags on ratings in one domain and ignore that in related domains items can be related in attributes. Tags encode similar properties of items and can be transferred across domains to mutually benefit recommendations for both domains. In this study we propose a TagCDCTR (Tag-informed Cross Domain Collaborative Topic Regression) model, which exploits shared tags as bridges to link related domains through an extended collaborative topic modeling framework. The model exploits the inter-domain relations by encoding cross domain item–item similarity based on common tags and jointly learning a shared set of topics from all domains together. Collectively factorizing the rating matrices of multiple domains into common user latent factors and domain-specific item latent factors, so that the learned item latent factors are linked through the inter-domain relations, helping to capture the items more comprehensively. The rich information reused in multiple domains alleviates data sparsity and the semantic advantage of topics and tags provides a better interpretability of recommendations. The experiments conducted on three datasets demonstrate that TagCDCTR outperforms state-of-the-art collaborative-topic-based models and cross-domain-based models.