The MULTISENSOR project – Development of Multimedia Content Integration Technologies for Journalism, Media Monitoring and International Exporting Decision Support Dimitris Liparas1 , Stefanos Vrochidis1 , Ioannis Kompatsiaris1 , Gerard Casamayor2 , Leo Wanner2 , Ioannis Arapakis3 , David García Soriano3 , Reinhard Busch4 , Boris Vaisman4 , Boyan Simeonov5 , Vladimir Alexiev5 , Andrey Belous6 , Emmanuel Jamin6 , Nicolaus Heise7 , Tilman Wagner7 , Michael Jugov8 , Mirja Eckhoff8 , Teresa Forrellat9 and Martí Puigbó9 1 Centre for Research and Technology Hellas, 2 Pompeu Fabra University, 3 Eurecat, 4 Linguatec, 5 Ontotext, 6 everis, 7 Deutsche Welle, 8 pressrelations, 9 PIMEC Abstract. The rapid development of digital technologies has led to other areas in order to adjust the own. a great increase in the availability of multimedia content. The con- To break this isolation, there is a need for technologies capable sumption of such large amounts of content regardless of its reliability to capture, interpret and relate economic information and news from and cross-validation can have important consequences on the society various subjective views as disseminated via TV, radio, newspapers, and especially on journalism, media monitoring and international in- blogs and social media. On top of this, semantic integration of het- vestments. In this context, MULTISENSOR has researched and de- erogeneous media, including computer-mediated interaction, is re- veloped tools that provide unified access to multilingual and multi- quired to gain a usable understanding based on social intelligence, cultural economic, news story material across borders, that ensure while a correlation with relevant incidents with different spatiotem- its context-aware, spatiotemporal, sentiment-oriented and semantic poral characteristics would allow for extracting hidden meaning. interpretation, and that correlate and summarise the content into a In the MULTISENSOR (Mining and Understanding of multilin- coherent whole. The goal of the MULTISENSOR project is to pro- guaL contenT for Intelligent Sentiment Enriched coNtext and Social vide a platform, which allows for an integrated view of heteroge- Oriented inteRpretation) project1 , we have developed a unified plat- neous resources sensing the world (i.e. sensors), such as international form for enabling the multidimensional content integration from TV, newspapers, radio and social media. Three demonstrators have heterogeneous sensors, with a view to providing end-user services been developed, indicating the potential of the platform and provid- such as journalism, commercial media monitoring and decision sup- ing end-user services such as journalism, commercial media monitor- port for SME (Small and Medium Enterprises) internationalisation. ing and decision support for SME (Small and Medium Enterprises) More specifically, potential investors can benefit from integration internationalisation. and context-aware interpretation of complementary and contradict- ing multilingual and multimedia information and get decision sup- port for international investments. Media companies and archives 1 Introduction can also benefit from the spatiotemporal integration and sentiment- Nowadays, the extensive availability of multilingual and multimedia oriented interpretation of heterogeneous content both for media mon- content worldwide is a result of the advances in digital technologies itoring and for journalism purposes. Finally, the European public can during the past decade, as well as the low cost of recording media. benefit from this integration and context-aware interpretation in the In the best case, this content is repetitive or complementary across sense that it learns and comes to understand the views, fears and political, cultural, or linguistic borders. However, the reality shows worries of the citizens all over Europe and get support for forming that it is also often contradictive and in some cases unreliable, some- an objective opinion with respect to the state of affairs. thing that can greatly impact its consumption. An indicative example The approach of MULTISENSOR builds upon the multidimen- is the current crisis of the financial markets in Europe, which has sional content integration concept (Figure 1) by considering the fol- created an extremely unstable ground for economic transactions and lowing dimensions for mining, linking, understanding and summaris- caused insecurity in the population. The consequence is an extreme ing heterogeneous material: language, multimedia, semantics, con- uncertainty and nervousness of politics and economy on the one side, text, emotion, time and location. which makes national and international investments really risky, and on the other side, the inability of journalism and media monitoring to equally consider all the media resources leaves the population in each of these encapsulated areas in its own perspective–without the 1 FP7-ICT-2013-10: http://www.multisensorproject.eu/ realistic opportunity to understand the perspective developed in the well as impact on its condensed presentation along with the con- tent summary. 3 User perspective Within MULTISENSOR, three pilot use cases (UC) were defined and specific requirements were extracted for each one of them: UC1: Journalism: Journalists need to master large heterogeneous amounts of multimedia and multilingual data when writing a new ar- ticle. On the basis of a market analysis that was conducted and from a journalistic point of view, MULTISENSOR should be able to pro- vide an automatic summarisation of heterogeneous and multilingual digital information. The platform should also automatically suggest related content and information that allows journalists to enrich their Figure 1. The MULTISENSOR concept coverage of a specific topic. UC2: Commercial media monitoring: Professional clients of me- 2 Project objectives dia monitoring portals require direct access to comprehensive and targeted business and consumer information. This could include in- In the context of MULTISENSOR, the following scientific objec- formation on consumption habits, competitors and opinions. From a tives with respect to the individual research areas of the project are media monitoring point of view, it is important that the MULTISEN- addressed: SOR system follows the usual workflow for the creation of a media analysis. In a first step, the user needs to define the sources and time • Mining and content distillation of unstructured heterogeneous frame that is to be monitored, along with the search terms he wants and distributed multimedia and multilingual data: In this ob- to use. In a second step, the search results need to be curated and jective, MULTISENSOR attempts to facilitate the data mining validated. The MULTISENSOR system should present the results of from several international resources, including news articles, au- these queries in different output formats and visualisations. diovisual content (TV, radio), blogs and social media and provide UC3: SME (Small and Medium Enterprises) internationalisa- intelligent mechanisms for the distillation of information. This ob- tion: This UC deals with SME internationalisation, which refers to jective includes low- and high-level content analysis. small or medium-sized companies that want to start or are in the pro- • User- and context-centric analysis of heterogeneous multime- cess of expanding from a regional or a national market to a new and dia and multilingual content: Here, the focus is on analysing foreign market in order to increase turnover and profit. This process content from the user perspective to extract sentiment and context, is of particular importance, as it is often the only option to achieve analysing computer-mediated interaction in the web and specifi- growth. But it is also aligned with considerable challenges, such as cally in social media, as well as generating high-level information a lack in knowledge about market conditions or the spoken language based on the outcome of the previously mentioned objective. The in the targeted countries. From the aforementioned, in order for the aim is to develop and integrate into the MULTISENSOR platform MULTISENSOR platform to be fully helpful in SME internation- research techniques for context extraction, sentiment extraction alisation cases and improve the decision-making process, it should and social media mining (influential user detection and commu- provide information about several related indicators, regarding the nity detection). condition of the market, the political and financial situation of the • Semantic integration and context-aware interpretation over countries, potential competitors, consumption habits, etc. Further- the spatiotemporal and psychological dimension of heteroge- more, two very important requirements from this UC are summarisa- neous and spatiotemporally distributed multimedia and mul- tion (to reduce the amount of information that the internationalisation tilingual data: This includes multidimensional content correla- expert will need to read and study) and automatic language detection tion and alignment based on reasoning techniques, as well as and translation. on multimodal vector-based representation and topic-based mod- elling. The multimodal integration is performed on top of the low- and high-level content extracted in the two aforementioned objec- 4 MULTISENSOR framework tives. • Semantic reasoning and intelligent decision support services: The architecture of the MULTISENSOR framework is depicted in The purpose here is to make sense of very large amounts of het- Figure 2. In this architecture, a periodic process of content harvest- erogeneous data by providing diverse analytics, contextualised ing takes place, which retrieves source material by crawling a set of decision-making support for different situations to enable view of sources for news, multimedia and social network content. Next, the the information from multiple perspectives. In this context, MUL- different components of the framework, as well as the functionality TISENSOR has researched and developed advanced reasoning of the modules that they contain and provide are described. techniques that abide to requirements for scalability and usabil- ity. • Context-aware multimodal aggregation, multilingual sum- 4.1 Multimedia content extraction marisation and adequate presentation of the information to the user: This objective also includes context-aware interpreta- This component aims at extracting knowledge from multimedia input tion of news by examining their impact on the news consumers in data and presenting the extracted knowledge in a way that subsequent the light of cultural aspects, user experience and engagement, as components can operate on it. It includes the following technologies: Figure 2. Architecture of the MULTISENSOR framework 1. Language Identification: Before a text is stored in the repository, glish and German. it is analysed in which language it is written and the text is anno- 6. Multimedia concept and event detection: This module receives tated accordingly. The languages considered in MULTISENSOR as input a multimedia file (i.e. image or video) and computes de- are English, German, Spanish, Bulgarian and French. grees of confidence for a predefined set of visual concepts. The 2. Named entities extraction: This module aims at identifying module performs video decoding (applicable for video files only), names (named entities) in texts. Names are words which uniquely feature extraction and classification in order to assign a confidence identify objects, like ‘Berlin‘, ‘Siemens‘, etc. The module incor- value for a concept or event existence in an image or video shot porates two linguistic components that allow all analysis modules [3]. to operate on the same input: sentence segmentation and tokenisa- 7. Machine translation: Automatic machine translation (MT) has tion. two main goals: to provide the translation of the summarisation 3. Concept extraction from text: Concept extraction starts from the results in the end of the content analysis and summarisation chain results of the named entities extraction task. The goal of this mod- and to enable full-text translation on-demand during the develop- ule is to identify in the text mentions to concepts that belong to the ment of language-dependent analysis tools in the project, in case project domains. Candidate concepts are identified through analy- a subset of required languages is not supported by these tools. sis of multilingual corpora. When processing new documents, the module attempts disambiguation of mentions of concepts against relevant ontologies and datasets. 4.2 User- and context-centric analysis 4. Concept linking and relations: This module aims at identifying in texts relations between mentions of named entities and con- The objectives of this component are to model and represent con- cepts. Two relation types are considered: i) coreference relations textual, sentiment and online social interaction features, as well as i.e. several mentions make reference to the same entity, and ii) n- deploy linguistic processing at different levels of accuracy and com- ary relations describing situations and events involving multiple pleteness. entities and concepts. To this end, a deep dependency parser [1] that delivers deep-syntactic dependency structures from sentences 1. Extraction of contextual features: This module provides a set of in nature language has been developed. This parser uses the output contextual indicators characterising the content items and a frame- of an optimised dependency parser [2] as input. work for measuring their impact in the context of the use cases. 5. Audio recognition and analysis: Automatic speech recognition Moreover, it provides representation techniques to be used in ef- (ASR) is employed in order to provide a channel for analysis of fective context-based search. spoken language in audio and video files. The transcripts produced 2. Polarity and sentiment extraction: The polarity and sentiment follow the same analysis procedure as the input from other text extraction module aims at modelling a robust opinion mining sys- sources. The languages covered by the ASR component are En- tem that is based on linguistic analysis and is applicable to large datasets. Moreover, models that take into account the presence of named entities in different sentences have been designed within been developed: hybrid reasoning (consists of a combination of the module. forward and backward chaining), multi-threaded reasoning (par- 3. Social interaction analysis: The social interaction analysis task allel inference calculation), temporal reasoning (inference based involves a set of processes related to analysis of social network on temporal entities and sequence in time) and geo-spatial reason- data stored into the MULTISENSOR repositories, namely crawled ing (ability to reason based on latitude, longitude and altitude of a Twitter data. Two modules have been developed in the context of given location). Additionally, a reasoning-based recommendation this task, namely the influential user detection and community de- system with two main functionalities has been developed: firstly, it tection modules. First, a topic-dependent network of contributors determines relevant facts by navigating the graph and secondly, it based on the mentions in the set of monitored tweets is built and advises the user by interpreting these facts through the use of the next, retweet probabilities between users in this network are com- aforementioned hybrid reasoning techniques and the assignment puted. The goal of the influential user detection module is to pro- of relevance weights for each selected fact [7]. vide a ranked list of users by decreasing order of influence based on the aforementioned network of mentions and retweet probabil- 4.5 Content summarisation ities. The goal of the community detection module is to make use of crawled Twitter posts in order to detect online dynamic commu- The content summarisation component implements procedures for nities by means of an appropriate community detection algorithm, producing multilingual briefings. Two established strategies in the which is applied to each graph snapshot defined by the user net- field of text summarisation are considered in MULTISENSOR: work of mentions. 1. Extractive summarisation: Text-to-text summarisation, where the relevance of sentences in the original documents is assessed 4.3 Multidimensional content integration and based on shallow linguistic features in order to decide on its in- retrieval clusion of a summary. A module following this strategy is used in order to establish a basic infrastructure for summarisation services The objective of this component is to achieve integration and retrieval and implement a fall-back method. of content along different dimensions. 2. Abstractive summarisation: Documents are analysed and the in- 1. Multimodal indexing and retrieval: In this module, a multime- formation extracted from them is used to generate a summary that dia data representation framework that allows for the efficient stor- is not composed of fragments of the original documents, but is age and retrieval of socially connected multimedia objects is de- generated directly from data. A module implementing abstractive veloped. The representation model is called SIMMO (Socially In- methods operates on the semantic layer in order to select contents terconnected MultiMedia-enriched Objects) [4] and has the ability extracted from multimedia documents and also coming from other to fully capture all the content information of interconnected mul- datasets integrated into the MULTISENSOR system. Contents are timedia objects, while at the same time avoiding the complexity selected and organised into a text plan that guarantees the coherent of previously proposed models. presentation of information. A multilingual linguistic generation 2. Topic-based modelling: In this module, two subtasks are consid- system renders text plans into the final summaries. ered: a) category-based classification and b) topic-event detection. The module receives as input multimodal features that are created 5 Use cases applications in the multimedia content extraction component and provides as During the three years of the project’s lifetime, three applications output the degree of confidence of a number of categories for a have been developed based on MULTISENSOR technologies, with specific content item (for category-based classification) [5] or a each application addressing one of the three use cases considered grouping for a list of content items based on the existence or not in MULTISENSOR. The first one provides search and exploratory of a number of topics / events (for topic-event detection) [6]. functionalities for journalists2 , the second one aims at supporting a media monitoring company to monitor specific profiles for their 4.4 Semantic representation and reasoning clients3 , while the third one provides decision support for SME in- MULTISENSOR includes a semantic layer in order to represent in a ternationalisation4 . unified way heterogeneous content. The following technologies are involved: 5.1 Journalism use case application 1. Semantic representation: This representation includes a number The journalism use case demonstrator is an application that assists of ontologies that are integrated in a common framework, such as media professionals (e.g. journalists, media experts) in finding rele- DBpedia, GeoNames and FreeBase. vant information in different formats, coming from different sources, 2. Ontology alignment: The ontology alignment module discovers and according to the social activities that were produced around. candidate semantic correspondences between heterogeneous in- Figure3 shows the results section of the application, which dis- formation descriptions and terminologies and verifies the correct- plays the results of a search query that the user can make, based on a ness and consistency of the discovered mappings in an automatic selection of keywords and filtering criteria. On the left side, search- way. related entities are displayed. By clicking on an entity it will be added 3. Content alignment: This module deals with the semantic pro- to the search query. Then these entities can be used to extend the cessing of the multimodal content, in order to identify near dupli- search query. On the right side, the following information per article cate and contradictory information relying on semantic technolo- is displayed: gies. 2 http://grinder1.multisensorproject.eu/uc1/ 4. Hybrid reasoning and decision support: In the hybrid reason- 3 http://grinder1.multisensorproject.eu/uc2/ 4 http://grinder1.multisensorproject.eu/uc3/ ing and decision support module, four reasoning techniques have Figure 3. Journalism use case application – Results section • Context: Contextual features per article (title, source, etc.). charts are shown for all articles that have been marked as relevant • Summarisation: Display of the output of the summarisation mod- in the search section. Finally, in the influencer section, the informa- ule. tion that is extracted from the social interaction analysis modules of • Translation: The online machine translation service operates on MULTISENSOR (influential user detection and community detec- this functionality in order to translate a summary to one of the tion) is displayed. available languages of the MULTISENSOR project. • In-depth semantic analysis: Displays semantic page view. On 5.3 SME internationalisation use case application this page, more information extracted from the text is displayed (list of named entities, sentiment polarity, cloud of specific con- The SME internationalisation use case application supports SMEs cepts and related articles). in order to start a process of internationalisation with any kind • Article to portfolio: The link to add an article to the “portfolio” of product. Relevant information related to the countries, the eco- (a folder that contains the user’s favourite documents), for further nomic situation of the market, the legal information, and the expor- analysis. This analysis generates the aggregated analytical view of tation/importation conditions can be retrieved easily to support deci- the portfolio content. sion making. The application supports a number of sectors and products. When 5.2 Media monitoring use case application a user selects a specific sector, articles about that sector are shown. After the selection of a product, the search will contain specific infor- The media monitoring use case application replicates the workflow mation about it. Another important functionality of this application is of a media monitoring professional to execute an analysis for a client. browsing specific information to a certain country for international- This includes checking articles for relevance by various indicators isation support purposes, based on a number of indicators. The con- and saving the relevant articles for a client’s profile. The relevant sidered indicators have been selected and organised by categories to articles can then be analysed, so that conclusions can be drawn from depict the relevant information related to a target country: Politics, this analysis. Economy, Society and Culture. Figure4 depicts the search section of the application, where the Furthermore, the SME professionals are interested in targeting user is presented with a view to search for articles, based on key- specific countries to establish new commercial activities. For this, words and language/country filters. Alternatively, the user can select the application offers the comparison of several indicators between a profile, which has settings stored for recurring searches in order to two targeted countries through the decision support system of MUL- quickly populate the search mask. TISENSOR, which is depicted in Figure5. In order to evaluate whether an article is relevant for the client, the user can use additional functionalities, such as calling the summari- 6 Conclusions sation and/or translation service. In addition, he can take a look at the entities extracted from the text and read the article’s full text. There In this paper, an overview of the successful MULTISENSOR project is also the analysis section, where visual results in the form of bar is provided. The project has developed a platform that supports a) Figure 4. Media monitoring use case application – Search section Figure 5. SME internationalisation use case application – Decision support interface journalists in mastering heterogeneous content in order to prepare ar- ticles and identify topics, as well as have access to multilingual sum- maries; b) commercial media monitoring companies in summarising the opinions of people for specific products and c) SMEs that want to internationalise by providing market analysis, product reports and decision support services. This platform integrates and makes use of innovative modules, which could be separately exploited. MULTISENSOR technologies will have a big impact from several perspectives. First, they will actively support the journalists (profes- sional and amateurs), commercial media monitoring companies and the international investments by SMEs. Second, the SMEs in the ICT domain will benefit from the open source tools and technologies de- veloped in MULTISENSOR, in order to improve their existing prod- ucts and offer new services to their clients. Third, the development of such tools will boost the competitiveness specifically in the media monitoring domain and in Europe, since the mobility of SMEs will be facilitated. Finally, the social impact of MULTISENSOR refers to the production of cross-validated news articles and the presenta- tion of news stories from different cultural, political and linguistic perspectives. 7 Acknowledgements This work was supported by MULTISENSOR project5 , partially funded by the European Commission, under the contract number FP7-610411. REFERENCES [1] M. Ballesteros, B. Bohnet, S. Mille and L. Wanner, “Deep-Syntactic Parsing,” COLING 2014, Dublin, Ireland, 2014. [2] M. Ballesteros and B. Bohnet, “Automatic Feature Selection for Agenda-Based Dependency Parsing,” COLING 2014, Dublin, Ireland, 2014. [3] N. Gkalelis, F. Markatopoulou, A. Moumtzidou, D. Galanopoulos, K. Avgerinakis, N. Pittaras, S. Vrochidis, V. Mezaris, I. Kompatsiaris and I. Patras, “ITI-CERTH participation to TRECVID 2014,” Proc. TRECVID 2014 Workshop, Orlando, FL, USA, November 2014. [4] T. Tsikrika, K. Andreadou, A. Moumtzidou, E. Schinas, S. Papadopou- los, S. Vrochidis and I. Kompatsiaris, “A Unified Model for Socially Interconnected Multimedia-Enriched Objects,” MultiMedia Modelling, pp. 372-384, Springer International Publishing, January 2015. [5] D. Liparas, Y. Hacohen-Kerner, A. Moumtzidou, S. Vrochidis and I. Kompatsiaris, “News articles classification using Random Forests and weighted multimodal features”, 3rd Open Interdisciplinary MU- MIA Conference and 7th Information Retrieval Facility Conference (IRFC2014), Copenhagen, Denmark, November 10-12, 2014. [6] I. Gialampoukidis, S. Vrochidis and I. Kompatsiaris, “A hybrid frame- work for news clustering based on the DBSCAN-Martingale and LDA”, 12th International Conference on Machine Learning and Data Mining, New York, 16-21 July 2016. [7] B. Simeonov, V. Alexiev, D. Liparas, M. Puigbo, S. Vrochidis, E. Jamin and I. Kompatsiaris, “Semantic integration of web data for interna- tional investment decision support”, In International Conference on In- ternet Science, pp. 205-217, Springer International Publishing, Septem- ber 2016. 5 http://www.multisensorproject.eu/