 Hi everyone, I am Manoj Shalvi. In this paper, we will present you a Wikipedia real-time updates recommendation system with other co-authors, Jhishankaj and Vikram Podi. As we have known that Wikipedia is growing rapidly, so with the rapid growth of the Wikipedia, regular maintenance of its content require a corresponding growth in the laborious work of the Wikipedia, especially for the categories that need real-time updates like actors and politicians etc. Generally, Wikipedians follow the certain guidelines given by the Wikipedia to update the pages like searching relevant information online, checking the reliability of the sources and finally defining the worthiness of the information. So, this is not a very hard task, but time-consuming for the pages which need real-time updates. So, they need an automatic real-time updates recommendation system to help update the Wikipedia pages with the refresh and reliable content. To design a such system, we use the fact that reliability and worthiness are subjective and depend on the individual perspective. In our case, we should study it with respect to Wikipedia. So finally, to solve this problem, we have designed a system that presents a solution for each component for reliability and worthiness differently and help to reduce their workload. So, we divided our system into two components. The first component determines what to get, which determines the reliable sources that provide credible or reliable information. As we know that reliability of web domain exists on a spectrum and indicates the degree to which the information presented on the website is accurate or free from the error or bias. So, we presented a reliability scheme to solve this problem, what to get, where to get, and the next component is what to get. So, this component will extract reliable and worthy sentences which can be added into Wikipedia by cross-verifying across sources. We defined the worthiness of the sentence as how good it is to be added in a Wikipedia. Sometimes reliable sentences are not worthy enough to be for inclusion. So, to solve this component, we have trained a binary classification and to check the worthiness or not. Now, let's look at the architecture of the model or system. So, it's seen that there are some future extracts and algorithms are there. We extracted some features and we generate the dataset for classification. We use the Wikipedia pages and the reference pages. We will look them into more detail. So, we defined the reliability, how we defined the reliability. So, we mentioned that to define the reliability scheme, we extracted features from the read log history and the reference pages of the Wikipedia pages. These features are extracted for each Wikipedia page of the given category. The features include editor counts which is which is extracted based on the human trust, the coverage relevance score based on the what reliable content we have, URL domain count based on what how many times the domain has been cited and diversity score based on the how many times on how many unique pages it have been cited. So, to calculate the final reliability score for the domain, we have to assign some weight to the future. So, there's been lack in the ground through the different unsupervised feature impotence algorithm has been applied including the laplacian feature impotence algorithm is practical feature impotence algorithm and PC. So, after defining the reliability scheme, we have to extract the reliable sentence and check their worthiness to so getting it is crucial for the Wikipedia updates. So, we use the Google search API to extract the latest information about the entity. The total reliable score for the sentence is calculated across multiple sources by checking the presence of a sentence on that page. So, we use a cosine simulator to check the presence of the sentence on the multiple pages to cross verify the worthiness of the reliable sentence we trade. We already trained the binary classification where Wikipedia pages sentence are taken as a positive sample while the negative sample are taken from the reference pages whose cosine simulator with the Wikipedia is less than a certain threshold. And this is this sampling is done using the fact that if the particular sentence from the reference page is not included in a Wikipedia page, then it is it might not be worthy enough. So, as we already mentioned, we have to assign the future weights to all the features. So, after analyzing all the algorithm we mentioned, we get to know that the editor count should get the highest weight or we can look it into a with a different perspective like human test can't be measured. So, we should give a return count as higher weight and followed by the URL domain count coverage relevancy score and diversity score for generating the sample for the classification data set we experimented with the differences threshold. We got our best result for point four point four and the sensitivity or recall hopper model is around 98 percent. We gave priority to recall rather than giving priority to accuracy because we assume that upper positive samples which are or we can say the sentence which are worthy enough should not be classified as unworthy. Also, we discarded all the data set is data set which are unbalanced like and these are generated when we put the threshold points six or above it. That's all. Thank you very much. If you have any query reach out to us at following ML IDs. Thank you.