 Hello everyone. I am Neeraja Kirjane and today I am going to talk about our project Hidden Voices. This is a joint work by me, Anurag Shankar, Chelsea Jain, Ganesh Katrapati, Sentha Mezanvi, Rajeev Bhaskaran and Balraman Ravindran. Our partners are the Robert Bosch Centre for Data Science and AI, Super Bloom Studios, the IITM Alumni Association, Denver Data Works and Free Software Movement of India. To give you an overview, there is a lack of articles of notable women on Wikipedia as compared to notable men, especially in the STEM fields. Our goal is to generate automated Wiki-like biographies for women in the STEM fields to reduce the disparity on Wikipedia. Our pipeline is divided into three parts. First, we gather and compile pre-existing information about notable women. Then, we distill the knowledge into human verifiable intermediate knowledge representation in the form of factoids. Then, we generate text which has a Wiki-like style from the intermediate data that we gathered. Our pipeline has a human in the loop architecture to ensure the accuracy of the factoids that are generated and the text that is being generated from the factoids. The architecture is as follows. Now, we extract the pre-existing information as follows. We first scrape multiple sources to find relevant articles. We narrow down meaningful articles using our ranking systems. Attributes pertaining to STEM fields were used to enhance the search results. Google News section was also an important source of articles. Wikipedia has a Notability Criteria which makes sure that certain sources are reliable and certain sources are not. The Notability Criteria on Wikipedia is as follows. A person is notable enough if they have a sufficient online presence, have reliable secondary sources and have a significant online coverage. Online coverage of notable women is very less in India. This narrows down the number of individuals who pass the Notability Criteria on Wikipedia. Therefore, Wikipedia's Criteria of Notability is a flawed metric when determining if the article should be published or not. The intermediate representation is done using the following methods. We used triples, knowledge graphs and factoids. Pattern and rule-based methods in ML were explored. While the pattern-based methods were accurate, they detected only simple relations whereas the ML techniques would often predict inaccurate relations. Therefore, we used GPT-3 which is an instruction-tuned model to generate factoids from the script data that we have got. These are the ways in which we tried the rule-based methods. This is the knowledge graph generation. Now, coming to the tech generation, our idea is to leverage generative large language models to produce text from an intermediate knowledge representation that we just got. Experiments with table-to-text models, just wiki-tablet and GPT-j model, fine-tuned on the wiki-biode dataset, produced significant hallucination. To explain what hallucination means, it means that the text kept on repeating itself and the text was not accurate and it pushed out garbage information. The problem with pre-existing intermediate knowledge representation is that it is heavily dependent on wiki-data which is often inspires an incomplete version of the article it represents. The model is thus forced to fill in the gaps which leads to the hallucination. These are the hallucinations which are generated. As we can see in the output, it keeps on repeating the same sentence again and again. To ensure that the model doesn't hallucinate, we have a human in the loop architecture. It is a two-step process. From the script text, the intermediate knowledge generation, we have a human verification method in the middle to ensure that all the factoids that are generated are present in the script text. After that, we generate the wiki article and also let the human check if the article is correct or not. Our idea is to use the LLMs but in a controlled manner to generate the text. A human helps in the controlled generation. Our limitations till now are that large language models are hallucinating. Getting intermediate data from scraping is hard unless we used models like GPT-3. Actually getting the article published on wikipedia is difficult because of the reliability and the notability issues that I discussed before. We created a UI for our project. First is the scraping phase where we enter a person's name and we get the scraped data with the source. After that is the factoid generation phase where we generate factoids from the scraped data. And every factoid is linked to its source. The third is the generation phase where from the factoids we summarize the content as shown on the right. The conclusion is that we plan to replicate the experiments that we have carried on GPT-3 to publicly available models for our task. We plan to use GPT-J and GPT-JT and other similar models for both knowledge extraction and text generation while keeping a human in the loop. Thank you. If you have any more questions, feel free to contact us and if you want to be a part of this project, please visit our website. Thank you so much for listening.