1 Introduction and Motivation

Real-time access to the latest news any time and from any location has become possible with widespread adaptation of mobile networks and devices. Increasingly, users of such devices expect custom-made, native applications to access their information. In this article we describe 3XL News: a system stemming from the XLike EU project, offering X(cross)Lingual analysis of eXtra Large News. 3XL News is an iOS application targeting news professionals and the general public. It shows how semantic technologies can be used in a real-world scenario to provide real-time global news monitoring and analysis across several languages. The main novelty is the linking of stories across six languages using semantic data derived from multi-lingual entity detection and cross-language news linking.

Related Work: A multitude of news monitoring mobile applications is available on the market, broadly divided into two groups: publishers’ own applications (like BBC, Al-Jazeera and RTV Slovenija) and news aggregators (like News Republic, Yahoo News, EMM mobile app [7] and iDiversiNews [8]). Those services reduce the overwhelming amount of information by summarizing news events and filtering them according to predefined user preferences or reading habits, however, they do not link news stories across languages. There are systems that perform cross-language linking such as NewsReader [10] and SPIGA [4] but support only four and two languages respectively and lack the capability to explore news along other dimensions (sentiment, location etc.).

2 Technological Background

XLike Project. XLike stands for Cross-LIngual Knowledge Extraction. The main goal of the projectFootnote 1 was to develop technology to monitor and aggregate knowledge spread across mainstream and social media, as well as across different languages. This is achieved by applying computational linguistics and semantic technologies to extract formal knowledge from multilingual texts.

Developed methods were integrated into an extensive linguistic and semantic text analysis pipeline [2], which provides input data for 3XL News. News is gathered from the Internet [9], annotated with semantic data [2], similar articles are linked across languages [5], news articles are clustered into stories and finally, news stories are linked across languages using content (i.e. text) and semantic data. We have performed manual evaluation of the obtained clusters [1] but we omit it here due to lack of space.

Cross-Lingual Similarity Function. To compute similarities between documents written in different languages, we model them in a latent, language-independent vector space [6]; projections into the latent space are inferred using techniques from linear algebra and statistics. The method [5] is related to Canonical Correlation Analysis (CCA) [3], which we apply on a multilingual corpus of documents obtained from WikipediaFootnote 2.

Semantic Data Extraction. Semantic annotation consists of named entity recognition [2] and Wikipedia Miner Wikifier [11]. The former detects named entities (persons, organisations, locations), while the later relates them to Wikipedia entries. For each entity a list of identifiers across all supported languages is provided. Thus, semantic annotation is used not only to better present data to users, but also to identify the same entity across different languages.

Linking News Stories Across Languages. Automatic linking of news stories (clusters of closely related news articles describing about the same event) across languages is an important addition to existing aggregation approaches. The main goal is to interconnect many influential and often related news that are reported constantly by numerous news outlets in different languages. We use CCA and semantic data (entities) to link stories as described in [1].

3 3XL News Application

The application consists of the following views:

Languages Overview (Fig. 1, left). The graph represents stories grouped by language (colour-coded) with images of the top-mentioned entities for each language. Node size corresponds to the number of articles in each language, while the width of each edge is computed from the overall similarity of stories in the two corresponding languages. News sources are shown on the world map, in which the display of any language can be switched on or off.

Fig. 1.
figure 1

Overview screens: by language and by entities (Color figure online)

Full size image
Fig. 2.
figure 2

NASA story exploration (left) and Middle East story exploration (right)

Full size image

Entity Overview (Fig. 1, right). Top-mentioned entities are aggregated across all languages, with possible filtering for major categories. Furthermore, users can choose a single entity to read related stories. Selected entities are shown on the world map below where news sources can be compared.

News Stories List. Contains a list of news stories filtered by language or entity and ordered by relevance with basic information for each story: title, summary, photo, publication date, number of news articles and related stories per language. The user can select a story and explore it.

News Story Exploration (Fig. 2). Components representing different aspects of a story are shown. The graph represents related stories across languages, with node sizes corresponding to the number of articles in a story and edge widths corresponding to the similarity of given two stories. The geographic selector sorts articles according to sources, while the last section gives a quick overview of the story by giving its top-mentioned entities, summary and keywords.

Articles List. Contains a list of articles for a selected story, ordered by user selection, is shown on a separate screen. Each article can be selected and the original web page, from which the text was extracted from, is displayed.

4 Demonstration

We will demonstrate a typical user session by selecting and analysing few news stories as shown in the figures.

Katja checks the overview screen (Fig. 1, left) and top-mentioned entities (Fig. 1, right). She then compares several entities by reporting locations. Being interested in space-related news, she selects “NASA” to view a list of related news stories (screen-shot not provided).

Katja now explores the main story she selected, together with its semantically related stories. A thick connection shows strongly related stories, while thin connections represent weak relations. For example, stories related to NASA astronauts space-walk are strongly related, while stories about Russian plans to get to the moon are weakly connected (Fig. 2, left).

Switching to earthly events (Fig. 2, right), Katja finds an abundance of interconnected stories from all languages. Compared to space-related news, current events are more widely covered by the world media. Finally, she can list all the articles making up the current story and read them in the integrated browser (screen-shot not provided).

5 Conclusion

In this article we presented 3XL News, an iPad news aggregation app that delivers a global view on the news media landscape by applying advanced computational linguistics and semantic approaches. The main advantage compared to similar systems is semantic linking of stories across six languages.

3XL News currently supports six languages, but the system is being extended to twelve languages with more planned. Through xLiMeFootnote 3, a research project dedicated to fusing the knowledge from different media content in different modalities, we plan to include annotated video and audio materials.

The demonstration video is available from the 3XL News homepageFootnote 4. The application will also be submitted to the Apple App Store and available free of charge.