 Like in many other fields, artificial intelligence and machine learning have a very important role to play in the field of media and communication as well. So in today's presentation, we'll be looking at how various newsrooms across the world have adopted some of these AI applications, what are the pitfalls of these applications, what are the changing roles of journalists and what actually is artificial intelligence and machine learning. So in the next few slides, we'll first talk about what is machine learning. So unlike science fiction, machine learning is not about machines having a brain made out of metal. It is not an artificial brain that machines are being provided. It's about training those machines to performing a single specific task according to a specified measure that the user has defined. So basically, we are training the machine into certain algorithm to predict and perform certain tasks and we'll explain what these tasks are and how that training goes off. So before we begin discussing about the machine learning applications, this is a very recent survey about what journalists feel are the most enabling technologies for journalism. And no surprises, almost 7 out of 10 journalists feel that artificial intelligence is going to play a very important role in future journalism. And of course, other technologies as 5G and there are new devices also. But artificial intelligence for a number of years have been the most important technology that journalists have been discussing. So let's talk about how AI has been used across the world in many of these newsrooms. So the Peruvian news outlet Oho Publico, for example, has a tool which can spot potential patterns of corruption in government procurement contracts. So whenever they have these contract documents, they pass it through their algorithm and they try and look for any pattern of corruption there. The BBC of course has been testing that artificial intelligence powered chatbot to answer questions about coronavirus using their own reporting resources and summaries from official sources. So this again is a very useful resource where all the answers are provided based on some very authentic evidence. The South China Morning Post is also using artificial intelligence to identify lookalike audiences so that when they want to target new subscribers, those new subscribers will be very similar to the subscribers they already have. So that's where artificial intelligence also has a very important role to perform. And of course, the Reuters news agency has used this speech-to-text technology to add time-coded transcripts to their archives of videos. So that makes the videos much more searchable and much more easier to find the key moments or anything that the researcher or the communicator wants to find out from the huge video resource beginning from 1896. So as we said in the beginning, artificial intelligence makes it possible for machines to learn from experience. So we provide data sets to machines and the more data points we provide to the machines, the better they become at predicting and performing certain tasks. So it makes machines to learn from experience and we'll find out how that experience is reinforced and all. But generally, two of the things they rely on is deep learning and natural language processing. And we'll see how natural language processing and natural language generation has a very important role to play in everyday journalism. So using some of these techniques of deep learning and natural language processing, computers can be trained to accomplish specific tasks. So we provide them with a large amount of data and they can look for patterns in that data in a matter of few seconds. So we train these machines for certain tasks and that's a very important thing. So machine learning is a branch of artificial intelligence that learns from experience as we said and this learning can take many forms. It can range learning from example, learning by analogy and at times it can be autonomous learning. But the important thing is that we are training the machine with millions of data points if possible. And when the machine is fed with newer data, it can look for the patterns it has been trained to recognize. So we first of all make the machine aware of what these patterns are and then the machine will be able to look for those patterns when we give them a fresh data set. So at its very basis, this is what machine learning is. So it's about learning some properties of a data set and applying them to new data and for that the common practice in machine learning is to split the machine learning into two sets. So for example if we provide them with 100,000 data points and we split that into two different sets, 50,000 each. The first is the training set on which we make the machine learn those data properties. So for example, we could be feeding them with thousands and thousands of images and telling them that what those images are. For example, we could be feeding them with images of different vehicles and train the machine to recognize bicycles for example. And after we have trained the machine onto those training sets, we then provide a testing set on which we ask the machine to recognize those bicycles for example. And based on the machine performance, we can streamline or we can add further data or fine tune it to recognize bicycles for example when we feed them with thousands and thousands of photographs at a later point of time. So machine learning in this sense is telling the machine to recognize these patterns or these pictures and then whenever we provide them with a new kind of a picture, the machine will be able to find that out and predict that okay this is what this image has or it can just in a matter of seconds it can tell us that okay there is a bicycle in this photograph. So there are three different processes of machine learning the first is the supervised learning where along with the training input we also tell them what these classifications are. So we tell the machine that okay this is positive and this is negative and this is neutral. So this is a supervised learning where we are telling the machine that these are the categories or these are the categories that we are training you to. So if I give them a particular word and tell them that this is positive later on it will be expected to recognize similar words as positive or similar words as neutral if we train them as neutral words. So this is known as supervised learning where we provide the machines with the categories and later on the machine can find out those categories from a larger data set. The other kind of machine learning process is the unsupervised learning. So unlike unsupervised learning we provide no labels to the learning algorithm. We let the machine find out the patterns and then you know we might give those patterns or those categories certain names but the machine itself finds out those patterns and we do not give categories or we do not train the machine into the categories that we have identified before. So we let the machine identify those hidden patterns or it can look for those categories by itself. The other point is the reinforcement where the computer program interacts with the dynamic environment. So it must perform a certain goal. So it could be like driving a vehicle or it could be playing a game against an opponent or it could be playing a chess game for example. And as it learns or as it interacts with that dynamic environment the environment which keeps changing the program is provided feedback in terms of rewards and punishment. So if it does a right thing it is rewarded if it does a wrong thing it is given a negative feedback. And when based on these feedbacks the machine reinforces or it knows that what needs to be done so this is what is reinforcement learning. So machine learning is generally of these three types the supervised learning the unsupervised learning and the reinforcement. So for the impact of artificial intelligence on newsrooms I have drawn on this book by Francisco Marconi this is a very useful book which talks about artificial intelligence and future of journalism. So generally the traditional journalism model is a linear model where the news gathering the production and the distribution process it takes in a linear fashion. But when we use algorithms then you know it could be a circular pattern or it's a lot more dynamic. So there are different ways in which artificial intelligence works in these three different processes. So we will see that in the news gathering processes we can use algorithms to mine data from the social media for example and also mine data using AI sensors. In the production part there are different ways in which we can write stories or construct stories from data using these machine learning applications. And even the distribution of content can be helped in a big way by using these AI applications. So in the next few slides we are going to talk about these processes of news production news gathering and distribution and how AI can help us in that or does help in that. So one of the first instances of automated news was from Los Angeles Times in March 17, 2014 where there was a 4.7 magnitude earthquake in Los Angeles. And within three minutes of the earthquake subsiding the first news accounts appeared on the website of Los Angeles Times. So based on data that was fed into these AI programs the program could construct a news story out of that. So that is one area where certain kind of news stories can be constructed with the help of AI applications. So the machine learning process which helps us or the AI process which helps us with writing news stories is a subset of the natural language processing. So as we know natural language processing is an application used to help computers understand the human's natural language. So one of the objectives of NLP is for computers to read, decipher, understand and make sense of the human language. And we'll see that it's not only just understanding and deciphering human language but also constructing words out of data with their knowledge of how humans interact or how humans communicate. So the traditional uses of NLP are things that we all are aware of or we might have seen in our everyday usage. So we have these language translation applications as in Google Translate. We have these processes like Microsoft Word or even Grammarly which checks the grammatical accuracy of text and that's where AI applications are used. Even in IVR applications interactive voice responses are the ones which we use to call call centers. So it's based on AI applications and based on these speech-to-text recognition. It recognizes our tone and our voice and our diction and tries to help us with that. And a lot of these personal assistant applications like Siri, Alexa, Cortana, they use natural language processing in a big way. So these are the applications that we already see in our everyday life. In the newsroom, the NLP can help us to scan huge documents. So when we have lots and lots of papers or pictures or documents and we want to find out certain kinds of words or we want to find out whether there is a pattern or whether there is some wrongdoing somewhere. So AI-driven artificial intelligence can create summaries by ranking the relevance of phrases and whether they connect with each other or whether they appear with each other and so many other analyses that can be drawn from there. So a lot of critical information can be gathered from these big documents just by using artificial intelligence applications. And one of the subset of NLP is the natural language generation. And it's the process of producing meaningful phrases and sentences in the form of natural languages. So it basically the input could be just structured data maybe in form of tables or in the earthquake example as we just saw. So we just provide them with data and they can bring out with tailor-made content in a matter of seconds. Thousands of pages in a matter of seconds and that would be somewhat like a human-like communication. And we'll see that readers are not even aware of the fact that they might be reading some content produced by an AI program. So it's so much precise. So this natural language generation has reached that level of precision where the language is very similar to what humans would write with a number of other advantages. So energy as we saw the natural language generation is just a subset of the larger natural language processing and there's also this concept of natural language understanding. So these processes are used in the newsrooms to generate stories out of data and we'll see some of these applications in the next few slides. So one of these commercial energy applications is this area and this is from their website. It says that it's a form of artificial intelligence that transforms structured data into natural language. So through this knowledge automation language generation, it can provide tailor-made content and the tailor-made means that it can be from the perspective of both teams. For example, if there is a sporting event and there are two teams there and there are supporters of these two teams who obviously don't agree on many things. So we might have content tailor-made for these two different kinds of supporters. And also it can change in terms of any other template that we might want to provide them and it can be done on a massive scale and in a matter of seconds. So it's just the matter of creating a template and providing the structured data and letting the algorithm work out. It's the magic of creating stories or creating text that makes sense and text that is tailor-made for the audience and for the time of day and something that can be updated also. So one of the things is narrative that is also an AI application. So it uses natural language generation and it can write more than 10 matches on a single day in just a matter of seconds. So the journalist's task of repetitive writing is taken care of. So the journalist doesn't have to focus on the main stories because they are generated just from the data that are provided to the AI application and it frees up the journalist to focus on analysis. So it's not that journalists are becoming redundant but it's the fact that the journalistic work is freed up from these repetitive tasks into more analytical tasks or into more soft stories or human interest kind of tasks. This is again from a narrative web page and this is where they talk about this was created in 2015 and they've been generating automatic content from any field. It could be from the field of economy, from the field of health, from the field of science, from the field of sport. So we just need to know the template that has to be created and we just have to provide the right kind of data to the AI application and it can create up those kind of stories in a matter of seconds as I said. Also AI can help us with very many different kinds of storytelling so it's not just one kind of storytelling but the same data can be used to provide these 12 different kinds of stories at the same time. So you could be getting photo galleries or you could be having timelines or you could be having short videos or you could be having VR or AR stories out of that or newsletters or even chatbots as we saw some time back. It could be a long form story, it could be listicles or it could be visualized data or data visualizations, it could be an audio experience and there could be explorable stories and even alerts and notifications. So the same data can be used in multiple ways without much of an effort. So the simple information can be provided to the consumers at a very swift rate into very many different formats and even tailor made according to their own requirements according to what they would want to read or what they would want to see. So this is a very important thing about AI applications that this can tailor made content or create content according to the readers, varying personalities and according to the location, the time of the day and so many other conditions there. BBC is already working on turning text into visual stories as suitable for mobile phones with pullcores and animations. So again the same thing providing data and harnessing technology for providing a format of a story which the end news consumer finds very relevant. And you can also provide summary of those news items and maybe even audio versions of their stories using those synthetic voices. So there are many other users of artificial intelligence as well so using new sensors for example rather than relying on government figures about how many people turned up in a meeting for example it can provide you with a more precise head count at a particular meeting. It can scan through photographs and provide you with the kind of people you are looking for or it can even provide us with causal links and correlations within data. But as I keep emphasizing the humans will always be needed. They will not be redundant because the task of human journalists will be much more different than it is in a non-AI environment. And as I said this is a very important finding by Christopher Clairwell in 2014. This was published in a journalism practice and it was found that the news audience could not distinguish between automated journalism and human journalism. And that is what drives some of these AI applications because they know that the consumers would not be able to distinguish whether it has been written by a human being or whether it has been created out of an artificial intelligence program. And as I said before it's just another tool in the journalist's armory and unlike in many places where people have been thinking even in the past when computers came a lot of people thought that journalists should be out of work or if we travel back a few centuries it was thought that even printing machines would make people lazier. But we've seen that it's just a change of culture it's a change of how we use that technology into our everyday work. So jobs that involve creativity and the jobs that involve ideation or even empathy they will not be automated. For them we will always require a human creative mind there. So artificial intelligence just accelerates the process of collecting and contextualizing data. So it provides the right context to the data and it can very swiftly collect that data as well. So now we are going to talk about how AI can help in the process of news gathering. Even in the news gathering process AI applications are extremely useful. So there is this news tracer at Reuters and it finds events that are breaking on Twitter in a matter of seconds. So it is scanning all the Twitter timelines and whenever there is an event breaking on Twitter it can trace that event and it even assigns them a newsworthiness score and also gives them a confidence score about how likely it is that those events are true. So instead of journalists scouting for information manually it is this application which helps journalists save those precious seconds in reporting the event very quickly. So it is about scanning the social media environment to find out newsworthy events giving them a newsworthiness score and also a confidence score but it has been found out that this news tracer is very good at finding out only certain kinds of events quicker than the mainstream news organizations. So there are a lot of these applications. This is from Graphics and it can take as it says it helps us develop, test and validate the hypothesis in matter of minutes. So we will be talking about many such applications. The other is the Data Miner which helps newsrooms mine data from various sources. So whenever there is a breaking news and Data Miner as they say on their website is used by 650 newsrooms around the world 24-7 to provide earliest tips on breaking news and breaking events. Then there is Factiva which does a very similar kind of thing. And there is this ICIJ which has its custom built search engine to pour over millions of documents and to classify them and find out patterns in them and to look for stories in them. So as we see that this machine learning algorithms are revolutionizing the news gathering process as well both in terms of speed and in terms of the huge volume they can handle to predict and provide insights into stories. So it is not just the news gathering process but also the distribution process and this is one study done by native and they found out that a lot of these distribution processes that that bus feed uses. So it is not only just the social media account and not just their own website but there are lots and lots of distribution channels that they can find out through these artificial intelligence applications and these applications help a news organization find out where the possible news consumers could be. Media cloud again is a very important open source platform for studying the media ecosystem. So it can allow researchers to track how the stories and ideas spread through the media. It helps us understand how when a story is out in the news environment how it finally spreads through the various media and how people are interacting with those things as well. So as we can understand there is a lot more technological help for journalists not only in finding out where the stories are or writing the stories but also the impact of the stories also. There is another application which is known as True Anthem based in San Francisco, California where they use artificial intelligence to identify the right content for the social audiences and post it to them at the right time. So identifying who are our potential consumers. As I said in the distribution part also artificial intelligence has a very important role to play in identifying who your potential consumers could be. There is this application News Whip which again tracks the impact of millions of stories and it empowers the news and communications professionals. So it tells us about the impact of the news story within a matter of seconds and predict whether it will have an impact for a certain kind of audience also. So as we talk about the positives of artificial intelligence very important to understand the question of ethics there and the question of responsibility because if we treat artificial intelligence as just a black box where certain kinds of information goes in and then we get out certain information out of that then there will be a problem. So this question of transparency and the explainability of algorithms explaining people how a certain algorithm was built up and what kind of training data was used there and what are the methodological issues involved there what are the questions of algorithmic errors or bias. So all these things have to be disclosed to the end user or to the larger academic and the communication fraternity. So there are many things or many kinds of jobs that will come up with the new applications and one of them is that of automation editors. So as we've discussed about these natural language generation it's important to create certain kinds of templates. So these automation editors would write these templates and even think about the branches or the possible variation in terms of audience and in terms of location in terms of time and all that. So that template has to be humanly decided by these automation editors. They would create these templates and then use the AI applications to create content but this template has to be a dynamic process and it has to be done by a human and after those stories are generated if there is any additional information required or if there is any additional explanation required that also has to be done by the automation editor. So although the AI applications speed up the machine learning speed up the news creation process there are a lot of things that journalists will still have to do. Also there will be a new class of data journalists whose job would be to run sophisticated analysis on available data using all the traditional means of data science and to collaborate with reporters who have domain expertise in a specific area of coverage to make sense of that kind of data. So these kind of data journalists and local reporter teams will become even more common in days to come. As we said these ethical questions are extremely important and Kathy O'Neill brilliantly summarized it by suggesting that algorithms are opinions embedded in code. So as we saw at the beginning that when we are talking of algorithms we are talking of training machine learning programs using human coders to identify certain patterns or even tell them that what those patterns mean. So there is always an element of opinion involved in these algorithmic processes. So it's important for us to bring these things to the forefront as journalists and as communicators and even as people who are using these AI applications about these algorithmic decisions, about what does the algorithm do. So not taking algorithms just as a black box but explaining all these questions about category. What does it do? What was its goal? What was it trying to maximize its time on? What data was the algorithm based on? And is this made clear to people who have been using it and people whose data have been used? Is there an oversight by humans to make decisions and tweak the algorithm? If the algorithm is not doing well, so whether there is a human element there as well and as we explained earlier whether it's explainable and whether it is interpretable and whether how errors are taken care of under all such things. One of the problems with artificial intelligence is this problem of deep fake where these images or audio files are altered with the help of AI applications to dupe an audience into thinking that they're real. So we could be having someone else speaking something else and it is not even observable and people can't even make out what reality is. So these problems of deep fake are real and we have these open source systems like Deep Face Lab where you can use face swapping for example to make someone else say something which they didn't say in the first place. So there are lots of these examples. Channel 4 for example created a queen lookalike recently to deliver an alternative Christmas message wearing a coronavirus-shaped bruise and making jokes about the members of our own family. And Synthesis even released a deep fake Santa that delivered personalized greetings. So this AI in communication is not only just for applications but also in the theoretical field there will be a lot more changes because most of our communication theories are about humans as subjects humans as subjects of communication. But when we have AI devices where machine subjects who are doing the communication or people are communicating with machines rather than through machines so that's where a lot of these communication theories will also have to be re-looked. Thank you very much for your patience and thank you for joining in.