 Can you see my slide? Great, then I will get started. Hi, I'm Isaac Johnson. I'm a research scientist at the Wikimedia Foundation. I'll be presenting my paper, Language, Agnostic, Topic Classification for Wikipedia, which is work that I did alongside Martin Garlock and Diego Saez Trumper. We're also researchers in the Wikimedia Foundation research team. So let's start with the topic part for this project. What are topics for Wikipedia article? That is, how do we describe what any Wikipedia article is about using a very high level and relatively small vocabulary. So we can do things like understand trends and page views or edits or help people find relevant content to their interests. So let's start with an example. This is the English Wikipedia article for the Scarlet Baddest, an adorable little fish. How do we describe the topic of this article in a high level way? Well, luckily for us in English Wikipedia, there's an extensive ecosystem of what are known as wiki projects, groups of editors who work on content under a specific topic domain. And they go around tagging articles among other things that are relevant to their wiki project. So in this case, wiki project fishes has tagged the article for Scarlet Baddest, but fish is still a pretty specific topic and we'd like to go even higher level than that. Again, luckily for us, English Wikipedia has just this mapping that relates to these wiki projects, one or more of 64 different kind of high level topics. And this was an approach that was devised by a son and half occur in their 2018 CSCW paper, and we adopted here. So this is all well and good. But what about articles that haven't been tagged by English wiki projects or articles in other languages. And this is the problem we tackle in this paper. How to model Wikipedia article, such that you can predict its topics. And this isn't a new problem and there are many ways to approach it. So I'm going to walk you through a few. So to be more specific, how do you represent a given Wikipedia article such that you can input it into a machine learning model of some sort that can then tell you what topics it predicts pertaining to the article. And the most obvious approach, I think, perhaps is to use the text of the article. And this is what a son and half occur did in their 2018 paper. They use a simple approach, quite effective. You could throw Bert or any much more complicated models over top to, but regardless, the text should carry plenty of signal regarding what an article is about, and every article has text. You could take an outlink based approach so just use the links in the article, as I'll explain shortly this is the approach we take here to ignore the text just extract the other articles that it links to. You may be throughout a bunch of data, but the links are nicely fixed for that vocabulary. And the really important part is that while you could use a language specific links so for instance represent the scarlet baddest article as linking to the English Wikipedia article for freshwater fish. You can also treat the links as pointing to wiki data items, the Q IDs you see in that third step. And this makes the model language agnostic and that in English Wikipedia article that links to freshwater fish is represented the exact same way in the model as a Spanish article that links to a Spanish language version of the article for freshwater fish, and so on. So the model can make predictions for language additions it is never seen, as long as the links are to known wiki data items. So they just did this in their 2020 paper and they showed it to be quite effective. And there are other approaches wiki database so ignore the Wikipedia article entirely and said look at articles associated wiki data item and the statements it contains use those for the modeling. We can keep doing this exercise for at least other a few other approaches. So you can see there are a lot of ways that one might and many have approached this problem. So how do we choose the outlinks based approach. Well, we did it based upon the following guiding principles. Our first principle was coverage, we wanted a model that would work for almost any Wikipedia article in any language, even if it was a fairly new article. And we're focusing on three approaches here outlinks text and wiki data, because they all do pretty well here, all articles have text, almost all have links, and almost all are linked to a wiki data item. So how many model parameters, how long to train the model, does it store comfortably in memory, things like that. Here the outlinks in the wiki days, wiki data based models do well. They use a relatively small vocabulary that's based on wiki data covers every language of Wikipedia. They don't require language parsing. This means you can build a single model for all languages. The first base approach requires individual pre processing models, probably word embeddings for each language, and this is far harder to scale feedback. Is there a clear path to improving the predictions for outlinks and text start argue yes, if there's topics that are missing, you add content to the article that is relevant wiki data kind of breaks this feedback loop though because the topic predictions are based on the wiki data item, not the individual article and so there's this kind of disconnect. And finally performance performance is generally the foremost principle that paper is used in declaring success with the model, we purposely put it last to emphasize the other considerations take precedence here. That said, all three models actually perform quite well outlinks and text a bit better than wiki data, but in the same area. Based upon this we chose a topic classification model based upon article outlinks other certainly room for improvement I think this model is unique in its ability to satisfy all these principles at once. Some quick specifics on how the outlinks model actually does. So coverage, we're close to 100% of articles on Wikipedia having at least one outlink so that's great. That's in simplicity. It takes about 10 minutes to train the model and we use the fast text Python library here. And it's less than a gigabyte for the entire model on disk. So that's including the 50 dimensional embeddings of which there's about 4 million for this model. Feedback, because the model so simple, we've already identified some areas of low recall ways that we might improve the data here. So that's good. And then performance. So in the standard kind of held out test set. We found that the text based approach or the link based approach was within within 1% of the text based approach for precision close to 90% recall close to 80% f one and average precision curve these are all micro statistics for macro statistics it was a little bit lower but still in the same area. And then we also gathered feedback from Wikipedia's on 10 examples of each topic and five different languages. So these are examples that we might not have had ground truth data for and here we beat the text based approach by 5 to 10%, depending on the language. So this shows that this link based model does transfer well across languages. Finally, a few thoughts on what's next. There's an API and dumps of all the model prediction so I'm going to paste these links at the end but I encourage everyone to go kind of explore test it out. And then there's frame more types of language agnostic models using this framework or similar frameworks. There's always model improvements you can do things like imputing links that haven't been added yet which is something that party and West doing their work, as well as different model architectures things like graph based methods. And finally, and perhaps most important is question of addressing biases. So the precision is quite high, but we do see areas of low recall and we're looking into those and trying to understand what's going on there. So this is just a bigger question of whether this is the right taxonomy of topics. So I mentioned it's based upon English Wikipedia. But it's not clear that that actually applies to all the other languages, even if you can build a model that will make good predictions based upon that taxonomy. And with that, I will close my presentation if we have time we'll take questions otherwise we'll be happy to answer them in the chat and Martin Diego my co authors will be able to field questions as well later in the poster session. Thank you, Isaac. I think we have time for a question to turn is our Q&A manager. Do we have questions on the top. Yeah, there is a question from Akhil maybe Akhil if you want to ask directly the question for Hi, Isaac. Thanks for the nice talk. Just curious. Did you guys try using the pre trained embeddings available via fast text or even sentence from summers. There are multiple multilingual embeddings available. And how why do you still think they are resource inefficient if you have pre trained models and can be used for a downstream task. Yeah, it's a good question we did consider and I've done some work around this. Two reasons why we don't use kind of this text based approach. One is that the embeddings rarely are actually available for all 300 languages. And when they are languages that don't have that much content the embeddings tend to be relatively low quality. So there's that kind of piece of it. And then even if you do have these pre trained embeddings you're still training kind of a specific model for each language. And while you know 10 models not so bad as you get up to 300 models that ends up being a lot of models that you have to be kind of monitoring and training and and a lot of disk space so thanks thanks for the answer. Okay then there is another question. Maybe you want to ask directly or. Oh yeah for the feedback and for the quality check. You're researching text five languages from the English and Vietnamese and French is, but are there any criteria for that I mean, usually, is it based on the quality of the results or based on the number of the editors was etc. Yeah, for those languages that was based upon the availability of folks who could go you know wikipedia in those language editions who would go through the predictions and accurately say yes this topic applies or no this topic doesn't. We also do when we do the held out test set. We're using all English level ground truth that we do propagated across the languages so for instance for a subset articles and pretty much every language. We also do the held out test set that we can test against to and show that we get good performance still. Thanks for the answer. Thank you. Then there is a one last question from Tina about the other requirements, particularly memory wise. And with a small memory fingerprint. Yes, I, one of the reasons I use the fast text library so much as I found it to be really forgiving about a lot of things. It's designed to run fully on CPUs. It's very simple so trains very quickly. The whole model, like I said is under a gigabyte so if you have essentially gigabyte gigabyte and a half or so of Ram, you can store it all in memory and make predictions quickly. So we use it on the servers but I've used fast text on my laptop to great effect to. So I highly recommend it. It should run on full force. Yes. I think we can go to the next presentation. There is another question. I don't know if there is time. There is no much time. Okay, we can. Isaac, would you mind answering it in the chat. Yep. All right, we have a second oral presentation which is from. Yes. Are you here. Are you here great. Please share your screen you should have the permission to share the screen go ahead. Yes. Can you see my screen. Okay. I'm a PhD student and I'm studying University of South Florida. Today I am going to talk about opinion dynamics and group decision making process in Wikipedia content discussion. And this work was done by me and my supervisor, you can look at Shampaya. In each Wikipedia, every day, 1000 of articles are created, but not all of the articles can meet the guideline of the Wikipedia properly to monitoring the quality of the article there is a process called article for deletion in which editor and reviewer can have discussion and can recommend delete or keep for the article. And based on the consensus decision is made. So, today I will walk through our research questions such as are there in biasing factor that correlate with the voting patterns of the editors, and how, how do these biasing factors evolve over time and form among the editors. And also, can the estimated votes can be regarded as the predictive factor of the final outcomes. So, in prior work may field and black have proposed a predictive model of individual vote and outcome of the discussions and their model was based on the natural language processing, and they have used the AFD discussion log from 2005 to 2018. In our analysis we also have used the same data set. So our first question is, are there any biasing factors that can explain the voting patterns such as the degree of agreement or disagreement with her peers. At first we have measured the bias or the preferences of each individual editors by calculating the probability of keep and these queries between zero to one zero is the indicator of being full deletionist who often recommend delete for the article. One is the indicator of being full inclusionist. And then we feed Gaussian mixture model so that we can get the major group of the editor in prior work terribly and champagne have found two major groups who we chair deletionists and inclusionist. In our recent analysis we have found more than two groups major groups, and the four major groups are strong deletionist who are who have the score goes up to zero then moderate deletionist, then moderate inclusionist and the strong inclusionist. Then for capturing the agreement or disagreement level. At first we have created a bipartite network among the set of the editors and the set of the AFD discussions or the articles. So in this bipartite network, each age signifies the recommend they have casted for example keep or delete. And then from this bipartite network, we have built a signed network among the editors. Here, each age between the editors is the indicator of participating in the same discussion and the sign the positive sign is the indicator of a recall of recommending the same action and negative sign is a indicator that most of the cases that the record their recommendation is different. So, then we feed community detection model moving and from this model we have found four major groups who are based on the connections of the voters. Then we have analyzed K core, K core is the maximum sub network of a network such that each every node has degree at least K. So the higher degree core we have we have found the most central user. So, we can find the most active user or they have participated in the most popular article discussion. So, at this point we have found two types of faction, which are GMM faction based on the preference of the bias and the Louvain faction based on joining to the same discussion, and then we calculated the spearman correlation between the faction and the age information such as the weight and the weight. And we have found that GMM faction has the higher correlation. That means that individual level bias explains more the pattern of agreement among the editors. And we also have found that at the higher core, the higher correlation that means the core of the network is dominated by like minded editors. Then our second question is how do these preferences on content, content inclusion or deletion form among the editors and how do they evolve over time. For that, at first we have selected the most central editors from the sign network, and then we split the discussion they have participated in 10 equal sized beans, and then we compute the probability of keep from each bean. And then we obtain the preference trajectory designs. And again, we fit the Gaussian mixture module. In this case, we also have found qualitatively similar kind of a similar number of groups, such as strong inclusionist moderately inclusionist strong inclusionist and moderate deletionist. And we also found that the preference are relatively, relatively stable over time, but there is a substantial narrowing of the opening at the early period of the reviewers. This can be this is an evidence of social learning due to the imitation of the year reviewer. And also we have found that the deletionist are highly resistant to change their opinion. And the last question is, can the estimation of the votes of the editors be regarded as the predictive factor of the outcomes. So, at first, we, we used latent factor model, so that we can estimate the hidden, hidden factor of the editors and the articles Then we use the predicted or estimated ratings as features and use the logistic regression for building the predictive model for APT outcome. So, and as the features are the continuous value so we also introduce a threshold above which we assign positive one, which is the indicator of keep for otherwise we assign negative one which is the indicator of delete for. So, this is the result of our outcome predictive model architecture. So, at threshold minus point for we have found the highest accuracy which is 82%. And we also have found weighted precision and weighted a points code. And at the threshold minus point for we have found everything highest every performance metric highest at the right hand side, you can also see the RLC score. And we also got the highest RLC score at the threshold minus point for which is 82%. And in summary we can, we can say that we have found the voting pattern of the editors can be explained by the degree of agreement and disagreement of the peers. And also inclusionist are more open minded than the delusionist. And it is possible to predict or infer the outcome of the APT discussion simply from the knowledge of the composition of the APT group, such as the bias of the editors or bias of the article. And we also have found that active and experienced editors are from the early phase of the APT project. And though they are the cohorts from 2005 to 2008, and we have discussed in details in our paper, and you can check out our data repository to reconstruct our findings. And that's all. Thank you very much and I'm happy to answer your question. Thank you so much. So we don't have a question in the chat. Okay, so what we can do, while people think about questions, maybe you know this, this was an amazing presentation so maybe there is a little bit of thinking to do before I formulate the questions, we can maybe start the live in talks and you will answer the question in the chat Okay, yeah, sorry we are a little bit late with the workshop so I think if that works for you we're going to go to the to the lighting talk session. Okay, thank you so much. It's not thank you so much for monitoring the chat. All right. I believe you are the first speaker for our second night in talk session. So, please get ready. I'm going to share the screen. Yes. I think we will get there. Okay. So, this is you right this is your person. Yes, yes. Thank you. Thank you. Perfect. Perfect. All right. So remember you have three minutes each please. Let's keep it on time. Okay, so first, please. Okay first place the link in the chat. Okay, so my name is a lot of R.D. I'm a PhD student at the university under the guidance of Professor Lev Muchnik. In my research about public trends and content consumption I wanted to use the traffic data for Wikipedia. Wikipedia is sharing the traffic dumps but not in a way that is efficient to run big queries as it gives you the traffic per hour but not per subject. Online tools were lacking critical features. They were not well maintained and other major disadvantages. For example, they didn't allow to search the Wikipedia's entire traffic history. So I downloaded the entire dumps but using it required a lot of effort. So that many other researchers use the traffic dump again and again from scratch. To solve this we created Wikishark. You can go to the next slide. Thank you. Wikishark is an online tool that it's easy to use, is hourly updated, allows querying all of the traffic data back to 2008. It also enables searching and comparing unlimited titles, view charts and export the results as data files or image files. There is also a browser extension that let you see the traffic directly on Wikipedia. And we also developed a trend engine surfacing new trends and hot topics. In the next slide I will briefly explain how is the data stored. So after downloading all of the data dumps we discovered that over 90% of the titles had never more than 255 pages within an hour. Over 90% we saved that in an efficient binary way, which means only one byte per hour. And for the rest we used only three bytes per hour. Moreover the data was saved in a sequence and it allowed much less disk access request. The query to the whole data since 2008, which means about 13 years, takes only a few milliseconds. And this way it allows us to run big data queries on the page with data. Finally, if you are interested in collaborating with us on either data or analysis, we are happy to explore such opportunities. If you want to learn more, come to meet us. We are at room number four. Thank you very much. Thank you very much. Thank you. Yeah. Hi. So hello everyone. So the paper vision going to discuss talks about the reorganization of information and Wikipedia articles. In the next slide we can see that by reorganization of information, I mean how factoids are rearranged in an article and the level of actoid in this paper we have taken a sentences. So in each with each revision sentences gets edited new sentences are getting order inserted. They are reordered. So do they do it progressively I mean with each revision. Does it happen. Does it impact the semantic semantic meaning of that article positively or just some random thing happens in Wikipedia articles. So we wanted to evaluate this. And for this we collected a sample of Wikipedia articles using certified sampling, so that you can represent the actual class distribution of articles. So, yes, so we calculated. The semantic similarity between consecutive sentences, and then average semantic similarity of a revision and then we observed how it varies with each revision. So, in the observation section we can see that in the first observation we found out that the overall semantic similarity increases as we go with the subsequent revisions. In the Pearson correlation density, it was a positive value and does it implies that the overall semantic similarity within sentences or factoids here, the semantic similarity increases. And then the observation to be tried to see. So, we know that Wikipedia articles are divided in certain classes, starts to FHA good articles. So, so they are supposed to be of varying quality. And we wanted to check how the semantic similarity varies within these articles within these classes. So we observed that, yeah, according to our hypothesis, the average semantic similarity was increasing as we go with the higher quality or higher quality class of Wikipedia articles. So, yeah, so users are not randomly shuffling the sentences editing the sentences they are positively contributing in the development of the chart. So this is our observation for this. Thank you so much. Thank you. Now, you can hear me. Yes. Okay. Perfect. I will speak about the role of local content in Wikipedia. And especially how it relates to editor and reader engagement. So basically, what is local content. So local content is like the content which corresponds to the cultural context of a Wikipedia language edition like geographical places of events, particular figures, etc, which correspond to this language community. And this is about a quarter of all Wikipedia language edition, all of the content is a local content. Okay. And we, we now look at does this local content reflect the higher level of editor and reader engagement. This is the question we're looking at in our next slide. We see a nice figure for several language edition so on on the top the top bars, the dark green bar corresponds to the percentage of Wikipedia pages which are which are local content. And then we see that if you look at the page views which is the light green bar we already have a higher percentage of page views which are directed which these pages attract. And if you further look on the edits we have the orange bar which is edit by registered editors and the red bar is the edits by anonymous editors and we see that the further we go down there the more the highest proportion especially it's especially high for anonymous editors. Okay, so then in slide in the next slide. This leads us to the hypothesis or the conclusions that the local content is more engaging for readers, and especially for editors, and anonymous edit we see even a higher proportion of it so that somehow leads us to the hypothesis that these editors the editors interest in the local Wikipedia project is more likely to be ignited by this local content so this is actually might be a way to get new editors to Wikipedia, promoting this kind of content. And basically, instead of considering this type of a bias because every language edition favors of course this. It's a local content related to its language edition it should be seen as relevant and important and for both extending the diversity and the content of Wikipedia and attracting new editors. And if you are more want to see more we are in session for room for. Thank you. Hi. This is Shubham. I'm actually the second author for it. I will be taking a bit. Yeah. Yeah, absolutely. Actually, we all know right that Wikipedia is an open source and anyone from the public can create the account and read the article so obviously we all agree that there needs to be some rating mechanism right. The rating mechanism via which we can create the articles is some we the idea is to SDSM no article grades to each and every article right. So that's what the core idea is and I think it's already been taken care but the one of the prominent ways via which we were no grading the articles is the quality grades right is the manual interventions we are using some other other rating systems or everything people have tried to automate it they have tried some national language processing methods or some social networking methods. So our idea is very much based on the social network analysis we are using the social network properties to train the models and predict the quality grades of articles. Next slide can you please. Yeah. So what we have done is actually we have collected the for each and every note of the Wikipedia articles we have calculated these network properties over them. That is in degree out degree between a centrality cat centrality page rank clustering coefficient hub scores and shell number h index and the recent properties right which have been mentioned. So we have tried to plot the graphs on the seeing that how no and try to observe that know how these stars and articles are behaving depending upon these properties. Then we collected all the Wikipedia articles names and their corresponding quality grades must both these informations. In the block diagram you can see that we have much both this information to create the data set for ourselves right. So the data set what we are doing we are trying to train the if we try to train different machine learning models, the classification machine learning models to predict the qualities over these properties. And yeah, the best results we are getting with with the random forest, yeah. The random forest is giving us the best results as we can see in the confusion matrix and the difference course for the random forest when we trained it over this, this data set. And then we try to compare our results with it. There were the two types of the people's two types of researchers who have tried to perform the similar kind of classifications one was the multi classifications like they are trying to cover the whole spectrum. And the other was the people who were trying to do the binary kind of classification. We have compared our accuracy along with them. One thing like at the end that I would like to mention is that a few papers have some eliminated the a class or a class, they haven't expressed the reasons that why they are eliminating all this class. But we have tried to focus on all the whole spectrum of the seven classes. That's it. Thank you. Thank you so much. All right, hello everyone. My name is Sebastian and together with my two colleagues Florian lemmerich and Marco Strummeier we tackled inferring socio demographic attributes of Wikipedia editors by looking at several state of the art approaches and possible implications for editor privacy. Next slide please. This is the overview of our approach in total illustrated for a specific user and here we can see all the different parts of the standard profile page. For example, the profile text is located in the middle of the page below that we have the categories a user chose to associate himself with. And on the right we have the different user boxes. And as you can see editors tend to display lots of different information about themselves like in this case that he was part of the military the time that has passed since he joined Wikipedia his gender is age and so on. And as we can see the profile text is used as input feature after applying some kind of embedding. And also the extracted socio demographic attributes are also uses labels and then both are given to different classifiers with the goal of predicting an editor's gender age, education and religion. Next slide please. And the prediction results we can see here are for our different models based on precision recall and F one score. So that bird outperforms TF IDF and doctor back for all socio demographic attributes and predictions for the attribute gender are more accurate than the predictions for the remainder of the attributes. And very important aspect deals with the implications for editor privacy because we should only utilize the acquired labels on an aggregated level to find general disparities in the Wikipedia editing community, which was our initial goal, is using this kind of information on an individual level, for example for personalized recommendation would you wrote editor privacy and would therefore most likely result in unethical applications. But the problem is that not disclosing personal attributes is not enough since as long as large enough sets of editors still decide to disclose some socio demographic attributes. So this would enable predictions for editors who might have might have explicitly decided against it. So, if someone is really concerned with the privacy set person should probably refrain from using the user page, or at least shouldn't use their user boxes to make automated processing and predictions more difficult. And now I'm handing over to Oscar I believe. Thank you. I don't see the chat so if there is. Oh, sorry, I was muted. Good. Please go ahead. Okay, sorry. Hello, all again. I'm Oscar and I'm presenting to you the research we carry out together with Lorenzo gatti and Kiriaki Calimeri. In this work, our aim is to model delivery moral foundations from text and access like please medium. We have a multi vehicle with which we express our beliefs and values. In this study, we operate as an allies morality via the moral foundation theory. Originally, the theory presents five foundations that have to pull our opposites by some virtue. The final model, this model, the creators of the theory propose a six dimension, the liberty dimension individuals that value strongly this moral value are consistently less concerned about individual level concerts, such as harm benevolence and altruism. There are also much less concerts with with group level moral issues for for instance a conformity loyalty tradition that are typically associated with conservative morality. The moral foundation theory is commonly used in computational science and why for the man for the main five dimensions that are lexical resources for the liberty dimension is still there is none. A million can you pass to the next slide. Here, our goal is to assess how people express the moral values of liberty through text. Since, since these faves the way to understand conflict conflicting social issues, such as vaccine hesitancy and subset subset 30 to conspire to conspiracy theories. Since according to the moral foundation theory device and virtue of the liberty scale relate to the conservative versus liberal political spectrum. We consider in the week we consider the Wikipedia pages and their conservative counterparts as a natural experiment. We collected all the conservative pages and we try to align them with the respective Wikipedia ones. 30,000 37,000 aligned pages, the figure shows the frequency distribution among the Wikipedia and conservative documents. At the top left, the blue dots, we found the terms most most commonly, most commonly using conservapedia, for example, Antifa or homosexual agenda in contracts at the bottom right we find the terms not using Wikipedia for example, affordable care. This is, this is in the red dots. Then we use a lexicon generation method, including the use of word embedding strain from our data set. Using this derived resource, we finally obtained an outdated vocabulary that aims to model the liberty dimension. This is a preliminary and novel work. Currently, we are working on refining the obtained dictionary to accurately model the liberty dimension. We are evaluating the quality of the end resource on value on various real life scenarios such as news and Twitter debates. Well, thank you very much for your attention. And if you have any questions, I will, I will be happy to take them. Also, I will be in session five. I think it carries a work on I think sorry if I can't work sorry. It's okay. Thanks. This is Chang-Wook Jung from Christ and IBS, South Korea. I'm presenting the study about information flow on COVID-19 of Wikipedia, which is part of the topic knowledge creation and diffusion. This research has been done by Christ, IBS, Post Ed, Sung-Shi University Wikimedia Foundation, and Max Planck Institute. Next slide. During the COVID-19 pandemic, misinformation killed people. Another incident was a rumor that methanol might cure the coronavirus. This misinformation spreads worldwide and homes more than 700 people just in Iran. How can you stop it? Misinformation can be blocked by the facts at the right time, right place from credible information sources such as Wikipedia. In this study, we analyzed the pattern of information generation and consumption by languages to provide the information needed in each language services. With the analysis, you can see the regional and cultural similarities and differences of the pattern in Wikipedia language services. We chose 11 target languages and collect COVID-19 related items in Wikipedia using Wikidata relationship. Then view counts and edit counts are collected and we categorize the items to see the topic shifts. Next slide, please. With the collected items, we ranked all 11 lists. Also, we made correlation metrics and clusters hierarchically based on the list to check the likeness between the languages. The clusters in first figure indicates similarity in reader's interests with view counts. The results reveal that European languages come together while East Asian languages make a cluster. Wikipedia items were categorized into four topics with biomedical information and reason, people and others. Such categorizations makes it easy to analyze how the pattern differs from other languages. The second figure shows the number of coronavirus pandemic related documents and the amount of access were examined by category. Although the number of items in the biomed category is just three. Coronavirus pandemic, coronavirus 19, and SARS COVID-19, the view counts is relatively high and indicating that there are high demand. In a nutshell, we found that information generation and consumption in Wikipedia are related to cultural and geographical connections. To establish the basis of the information provided strategies for the next outbreak or next pandemic, the information generation and consumption, speed and item network developments should be examined. We published our data set on ficture to promote more analysis from other researchers and also our data collecting method code is available on GitHub. I will post the links on the chat. Thanks. I will wait for the comments in room four. And next will be Karthik Madanagopa will present the marvelous research about knowledge gaps. Hello everyone. I'm Karthik Madanagopa from Texas A&M University. I'm going to talk about my current research with Dr. Kavali towards ongoing detection of linguistic bias on Wikipedia. First of all, what is neutral point of view? Next slide please. It is a Wikipedia guideline that expects all articles to be returned fairly proportionately and as far as possible without editorial bias. And yet knowingly or unknowingly, some form of subjective bias is injected into the subject to treatment of facts in Wikipedia articles. Let's take a look at these two example statements that are extracted from Wikipedia. The use of the highlighted words in these two statements introduces a subtle form of subjective bias that can influence the reader's perspective on those topics. Not all articles are reviewed for NPOV. Only articles that are tagged by readers or editors are disputed and NPOV board. Especially in case of current events like 2021 storming of United States Capitol. As the event unfolds, these articles were viewed by millions of viewers before it being revised by the NPOV board. It may be possible that some of the readers might get influenced by the subject to content present in those articles. There is a need to build a robust bias detection algorithm in case of these crowd source encyclopedias like Wikipedia. Various researchers in the past have built classification models that can detect bias statements as the writing styles of Wikipedia editors change over time. And some editors try to actually evade these kinds of bias detection models by using different language styles. These models struggle to detect more subtle and emerging forms of bias. This is mainly attributed to the fact that these models were solely built using Wikipedia data alone. The primary focus of my research is to build bias detection models that are more resilient that can self adapt over time. Such models should be robust to changes in editor behaviors and new subjective writing styles that has never seen before by the Wikipedia community. Next slide please. In our initial investigation, we built a series of bias detection models solely using Wikipedia based statements alone. And it kind of validated the fact that just by using Wikipedia alone, we won't be able to build a model that can have consistent performance over time. Even though the initial model had a better position of 77% its performance kind of degraded over time. To enrich our model performance we started exploring through cross-domain transfer learning approach by leveraging bias statements from various subject to rich domains such as political speeches and product reviews. We have observed some encouraging results like 89% accuracy with Roberta-based classifier and also across different topic areas like politics, language and literature-based articles. In order to extend our models performance in other topics, we have started working on generating more statements using generative adversarial networks. The generative data can be used to augment our current cross-domain bias detection approach to further improve its classification accuracy. Thanks for the opportunity. The next paper is about accessing the quality of health related Wikipedia articles by Luis Coato and Carla Lopez. Well, thank you, Kartik. So, I am Luis Coato, me and Carla Teixeira-Lopes. We did a research work about assessing the quality of health related articles in Wikipedia using generic and specific metrics. Next slide, please. There are many works about assessing the Wikipedia quality using metrics based on generic features that we used. We wanted to research if there are specific features of Wikipedia that would be used to assess the quality of health articles and which of them are the most important. We also wanted to propose specific metrics based on these features and finally we wanted to know if these metrics are better to those existing generic features for the health domain. To achieve these objectives, we used the methodology represented in this diagram. So, we began by exploring the health related content features in articles from health and medicine areas. Then we collected the top 1,000 most-viewed health articles, at least provided by the Wikipedia Project Medicine. After that, we analyzed the quality of generic features proposed by Svilje and the specific features we proposed on assessing the collected articles. From that, we proposed health-specific metrics that I will talk about soon. In the end, we analyzed the quality of the proposed metrics, comparing to the generic metrics proposed by Svilje. The results are present in this slide. So, as health-specific features, we proposed those present in the first table of the slide, which are the number of external links with a reported source, the number of sections in the article that belong to the list of recommended sections in the Wikiproject Medicine guidelines. Another proposed feature is the presence or not of the article in the healthcare translation task force list. The next feature is the share of additions made by the Wikiproject Medicine admins. Next, we proposed the number of health-related templates present in the articles, such as that one in the second figure of the slide. We also proposed the number of medical code classifications present in the templates, as represented also in the second figure. The next feature is the number of health-related info boxes as represented in the first figure. And finally, we proposed the number of images included in the info boxes. We adapted the generic metrics of Svilje, adding or replacing the proposed features, creating the specific metrics shown in the second table of the slide, which are health authority, health completeness, health informativeness, and health consistency. The respective formulas are shown on the slide. And as you can see in the table, we achieved good results with the proposed features and metrics. The most relevant features were the number of recruited links, the number of recommended sections, and the articles translated by the task force. We improved all the generic metrics, although we get the best results in the authority and marginal improvements in consistency metrics. With that, I finish my presentation. I'll be available later in shot room 3. And next, we will have the presentation of Bhuvana. Thank you. Thank you. This is Bhuvana Meenakshi and I'll be presenting the research study on Virginia Gender Gap in English Wikimedia Communities. And this particular research study was done with association with the Center for Internet and Society and the Access to Knowledge program team. This study was basically to document and analyze the gender bias in the Indian language Wikimedia Communities. So next slide, please. Thanks. In this, we analyzed the previous research that's done globally and also the research is focused on Indian language communities itself that have been done in the past. And in a 2018 survey from the Wikimedia page on gender bias states that there have been 90% of the contributors across various versions of the Wikipedia. And among which the 90% each constitutes of male participants and 8.8% each of female and 1% as non-binary gender who have been editing and actively participating in Wikipedia project specifically. And other studies since 2011 mostly focused on English Wikipedia having a percentage of female editors as up to 20% each. And this gave us a thought about understanding how much of the percentage of the female contributors exist in the Indian language communities. And especially across various Wikimedia projects and also to basically understand how the gender bias has been perceived in the local communities. And therefore, we brought this study into three thematic areas, which is mainly about online participation that includes content created by women, content about women and their online engagement with communities. And also understanding offline participation by women across various Indian language communities and the strategies to remove barriers to sustain participation of women contributors and mapping the diversity of Wikimedia projects. And how women are involved with sustaining the participation across the projects. So in this study we had 15 interviewees from 13.