Conference 041818.Pdf

Conference 041818.Pdf

“© 2018 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.” Information Enhancement for Travelogues via a Hybrid Clustering Model Lu Zhang, Jingsong Xu, Jian Zhang, Yongshun Gong Global Big Data Technologies Centre, University of Technology Sydney, Sydney, Australia Email: fLu.Zhang-5@student., Jingsong.Xu@, Jian.Zhang@, [email protected] uts.edu.au Abstract—Travelogues consist of textual information shared by tourists through web forums or other social media which often lack illustrations (images). In image sharing websites like Flicker, Travelogue Data Package Image Data Package users can post images with rich textual information: ‘title’, Title: Best Zoo in Sydney ‘tag’ and ‘description’. The topics of travelogues usually revolve Poster: Good day to all, we are coming from NZ to visit Sydney, wanted around beautiful sceneries. Corresponding landscape images to know which zoo is the best, especially recommended to these travelogues can enhance the vividness of for hand feeding kangaroo and cuddle a reading. However, it is difficult to fuse such information because koala bear. Taronga Zoo - Koala Park Sanctuary - Wild life Sydney Zooor the text attached to each image has diverse meanings/views. others? thank you for your time and Title: Fraser Island In this paper, we propose an unsupervised Hybrid Multiple attention... Tag: Australia, Queensland, Fraser Kernel K-means (HMKKM) model to link images and trav- Comment: Taronga Park Zoo is top- Island, Fraser, Island, Coast, Beach... elogues through multiple views. Multi-view matrices are built rate and well worth visiting. The Koala Description: Perfect beach, perfect to reveal the correlations between several respects. For further Park in Pennant Hills is dismal and to be sky, perfect clouds. this is 75 Mile Beach improving the performance, we add a regularisation based on avoided . The Wild Life Sydney Zoo is a on Fraser Island, Queensland, Australia. tourist trap in Darling Harbour. The This stretch of golden sand on the east textual similarity. To evaluate the effectiveness of the proposed Featherdale Wildlife Park in western coast of Fraser Island is classified as an method, a dataset is constructed from TripAdvisor and Flicker Sydney is small but interesting... official... to find the related images for each travelogue. Experiment results demonstrate the superiority of the proposed model by comparison (a) (b) with other baselines. Fig. 1: Travelogue data examples. (a) Each travelogue is composed Index Terms—multiple kernel k-means, multi-view clustering, by three parts: ‘title’, ‘poster’ and ‘comment’. (b) Image data exam- multimedia analyses, information enhancement ple. Each image has three kinds of textual information: ‘title’, ‘tag’ and ‘description’. I. INTRODUCTION Due to the rapid development of the Internet, large volumes of multi-media data enrich travel experiences of people [1]. public datasets with images and texts, such as NUS-WIDE [2], A large number of trip sharing websites provide sufficient Pascal VOC 2007 [3], ImageNet [4], Wiki [5] and MIT travelogues for tourists who are planning their future trips. Place 205 [6], topics can be of very wide range. It could Most of such trip sharing websites (e.g. TripAdvisor and be easier to recognise topics between multiple fields, such as Wikitravel) only provide heaps of humdrum textual contents sports news and political news than to distinguish relatively without vividness and abundance. On the other hand, by some fine distinctions in one field. (3) Insufficiency of textual visual image sharing websites (e.g. Flickr and Pinterest) it is information. Texts in these public datasets are often quite much more convenient for people to get plenty of vivid images. short. A few tags or simple sentences related to the contents But people cannot gain precise and detailed information by of images can only be seen as the supplementary information mere images. Therefore, the cross-domain information can be in the same domain. connected by the same topic. These data have different em- We build a new multi-view dataset which contains both phases and formats, which can enhance the reading experience. textual travelogues and landscape images (crawled from Tri- With the illustrated travelogues, people could feel the majestic pAdvisor and Flickr), as shown in Fig. 1. We manually label sceneries directly rather than be confused about the tedious travelogues with matched images as our ground truth. Specif- words. ically, each image includes three views of explanatory textual In order to integrate vivid illustration into textual trav- information termed ‘title’, ‘tag’ and ‘description’. Every view elogues, we need to bridge gaps between different domain has its emphasis as complementary information of other views, information. In particular, there are three major difficulties. which expands textual description of images. (1) Heterogeneous features. Features are represented hetero- The objective is to link information (images and texts) be- geneously regarding their modalities, which brings difficulties tween two domains. One traditional solution is to learn trans- for mining inner correlations between images and texts. (2) formation matrices to maximise the similarity, like Canonical Restricted topics. Considering the specificity of travel sharing Correlation Analysis (CCA) [7]. However, there are already a data in some specific countries or places, topics are more lot of semantic gaps between the image content and textual restricted compared with public datasets. While in some other information attached to the images in Flicker. The problem is that it would be much more difficult to match textual Canonical Correlation Analysis (CCA) [7], [25] is a rep- travelogues and images by learning the map between them resentative statistic method for exploring the relationship be- directly. In turn, considering images from Flicker have multi- tween two sets of variables. It aims at maximizing the correla- view textual information (title, tag, and description), in this tion of two related multi-domain data. Similar to CCA, Partial paper, we propose to match travelogues and images through Least Squares (PLS) [26] is another classic method which aims textual information directly. Since title, tag and description to learn a linear projection that maps different domains into can represent the image from different views, Multiple Kernel a common latent subspace. Reference [26] utilised PLS for K-means (MKKM) [8] based model is adopted to discover multi-modal face recognition. In reality, the linear projection Top-N correlative images for travelogues. We apply different is not applicative in many cases which may lead to limited word embedding methods, such as Term Frequency-Inverse performance [27]. Document Frequency [9]–[11] and Word2Vec [12], [13] to Some deep learning based methods [28], [29] have shown build kernel matrices and propose a hybrid MKKM model. At their excellent performance. Reference [27] proposed a cross- first, a multi-view similarity framework is built to reveal the modal correlation learning approach with multi-grained fusion correlations from several perspectives/views. The HMKKM by a hierarchical network. It divided the task into two steps method is used to mine potential associations among these and used deep belief network for coarse-grained learning. views. To improve the performance of this unsupervised learn- Reference [16] proposed an approach for text illustration from ing process, a regularisation is also introduced to construct a tagged repository. Reference [30] proposed a generalized a hybrid model. We conduct experiments to evaluate the semi-supervised structured subspace learning model for cross- proposed method and compare with other baselines on a modal retrieval. However, these supervised or semi-supervised dataset constructed from TripAdvisor and Flickr. The results learning methods need paired data for training which are not show the superiority of the proposed method against other applicable to our small dataset. compared models. Different from above-mentioned methods, in this paper, we The rest of the paper is organised as follows. In Section 2, take the advantage of image properties in Flicker, where each we review some related work. Section 3 introduces our cross- image has abundant textual information: title, tag, and descrip- domain hybrid model. Section 4 focuses on the experiments tion. Each property can represent the image in different view. settings and results. In section 5, we make a conclusion. The problem of mining relations between textual travelogues and landscape images can be seen as mining relations between II. RELATED WORK textual information through multiple views. We build a hybrid multiple kernel k-means model using different textual features Several text representation methods have been proposed in for clustering travelogues and images without a large number previous work [10]–[14]. Term Frequency-Inverse Document of training samples. Our proposed model can reveal the direct Frequency (TF-IDF) [9]–[11], [14] is one of the most classic relations among texts by fusing multi-view features. We also algorithms. TF-IDF directly represents one article by TF-IDF show that the

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    9 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us