A Web Service to Generate Intelligent Previews of Web Links Amit Sarkar Samsung R&D Institute Bangalore, India amit.srkr@samsung.com Joy Bose Samsung R&D Institute Bangalore, India joy.bose@samsung.com Abstract— Previews of web links are typically generated based on the metadata captured from the URL content. Sometimes, the preview sentences are extracted by means of content summarization. Such web link previews can be seen in different apps like the web browser, chat app, messaging or email apps etc. These previews are static in nature and do not change with respect to changing context. Therefore, they may not be particularly relevant to the receiver of the link. In this paper, we present a web service for generating intelligent previews in a chat application, which captures the local intent of the user from the chat content and uses it to display only relevant content extracted from the previewed URL. Since the user intent can change dynamically, our system generated previews are also dynamic, which change on the fly if it detects a change of topic being discussed in the current chat. We describe details of a prototype web service implementation, with three methods for preview generation based on TF-IDF and Word2Vec word embedding. We also present results of an evaluation using shared URLs from a private real-world chat group as well as a sample chat app with a few users to determine the accuracy of the preview generation system. Keywords- User intent modelling; web previews; chat application; web service I. INTRODUCTION Most mobile applications, such as chat, messaging services like WhatsApp, web browser, web cards, social networking apps etc. have the ability to generate previews of web links. Such previews make it easy for the user to quickly visualize the content of the link. The web link preview consists of an image extracted from the URL content along with some text. The text is typically extracted from the URL’s metadata. In absence of enough metadata, the text can represent the most important sentences from the article. Web link previews are static, since they are extracted from the web content without considering any external context. The extracted information shown in the web preview may not be relevant to the user, if the user is interested in a specific part of the URL content. For example, if the user is reading a Wikipedia article on Mexico, the preview may only give the web page name and few lines related to main theme of the content, while the user may really be interested in Mexican food which is also mentioned in the same page. In such a case, it would be useful if the system could infer the topic of the user’s interest or intention, and display the extracted web content relevant to the topic. Fig. 1 shows static as well as dynamic web preview generation for a chat application on a mobile device. In this paper, we develop a web service for generating dynamic web previews that are relevant to the user. Our system customizes the web preview by extracting only information that the user is likely to be interested in, based on the chat topics. We expect such a system will improve the quality of the user experience and user engagement and also save the user’s time. (a) (b) Figure 1. Illustration of a chat application showing (a) a normal preview and (b) an intelligent preview generated by capturing the user’s intent or topic of interest during the chat session. We implement our intent detection based web preview generation service on a chat application running on a mobile device. However, our system can in theory be used in any app to generate relevant web link previews. The rest of this paper is structured as follows: in the next section, we survey related work in the area of generation of dynamic previews. Section 3 gives an overview of our model for intent capturing and preview generation. Section 4 gives implementation details for a proof of concept. Section 5 describes a test to find which sentences users find most relevant within a given URL content, and correlate the user generated results with those generated by our algorithm. Section 6 concludes the paper. II. RELATED WORK In this section we survey related work in the area of web link previews and their automatic generation.