This is what Prof. Ido Dagan says in a conversation with the science website, that the natural language processing laboratory where he researches recently received a grant of NIS 1.5 million for joint research with Intel
Prof. Yoav Goldberg - Head of the Language Processing and Deep Learning Laboratory, Bar-Ilan, Mariana Waxman - Director of Academic Relations, Intel Israel, Moshe Weserblat - Head of the Language Processing and Deep Learning Group, Intel AI Product Group. Photo: Intel Israel Spokesperson
Intel Israel announced support for the natural language processing and deep learning laboratory at Bar-Ilan University. Natural language processing is a field of research and technology that deals with the computer analysis of texts in human language, using methods that combine algorithms from the world of computer science, such as machine learning and deep learning with linguistic knowledge. The laboratory in Bar-Ilan, led by Prof. Ido Dagan and Prof. Yoav Goldberg, is the largest research group in Israel and the world leader in the field.
This is a project for a joint research program for three years, with funding of about one and a half million shekels. The research focuses on the analysis of multiple texts, natural semantic representation of sentences and information unification, and their applications in a variety of fields such as summarizing and interactive information search.
Open knowledge graphs
In a conversation with the Hidan website, Prof. Dagan explains: "The joint research with Intel is part of a growing relationship, there are a number of studies and especially one large study that interests them."
"In my group we are developing an approach to representing knowledge in texts, how it is possible to process many texts and arrive at their combined meaning. In every subject there are many texts, some overlapping, some complementary, some contradictory, how can one master many texts. The system will make it possible to research, to summarize."
According to him, this is the development of the concept of a knowledge graph. Companies like Google and Facebook use it to bring the appropriate results to the particular user. However, a knowledge graph is a scheme that allows to represent knowledge in a limited area, for example to understand from studies on drugs what are the symptoms of a certain disease, what are the drugs, what are the side effects and so on. But if you want to understand what is happening in many texts that, as mentioned, are partly repeated, some complement each other and some contradict each other, the computer needs to build a knowledge graph automatically that can represent all the knowledge present in the texts.
An example of this is the use of AI to analyze scientific studies in areas where the accumulated knowledge cannot be mastered, and to summarize them for the scientists and doctors.
"In the Intel laboratories, they want to understand in which directions the research in artificial intelligence is progressing so that they can meet the needs of the hardware." Prof. Dagan says. "Two engineers from Intel work with us in the lab, among other things, on joint articles, and we collaborate on research done in the Intel lab. Working on joint articles, also collaborating on their research in the lab at Intel. The research grant is for three years and reflects our strategic relationship with the Intel group in Israel."
Getting an explanation from the neural network how it reached the result
Goldberg researches in the field of NLP neural networks, deep learning, more infrastructural problems, atomic information in texts, and in some of the things he also collaborates with the people of the Intel research laboratory.
He was also interviewed for this article and explained: "The main field I am involved in is understanding natural language. In recent years, the field has developed with the help of machine learning and deep learning. Our goal is to understand which neural network architectures are suitable for language learning and why and also how they learn. Today artificial intelligence is a kind of black box and we want to know what they learned, and why the decision was made. Scientists and engineers want to know what is happening to improve the networks and thus understand what they can do."
Another area we deal with is the biases and how it is possible to build a fair model, a model that does not discriminate against people by mistake. In this field, everything related to the texts, but to refer only to the content, and not to refer to characteristics such as age or gender or origin. For example a CV, I want to be able to tell the artificial intelligence to ignore this data, so also in the case of a neural network that decides on granting loans, we would not want to discriminate based on place of residence for example, the question is how to make sure that this data does not reach the decision-making process indirectly.
This is an important feature and related to the EU's request for an explanation. Next we investigate what are the infrastructural tools needed to understand a text. I build a kind of building blocks that people like Ido use. The question is what building blocks are needed and how do you make building blocks that will be useful to many people, whether they are experts like Ido or people who are not experts and want a tool that will help them make decisions.
Today there are quite a lot of neural networks and deep learning running on NVIDIA infrastructures - graphics processors. They want to understand the loads in language processing that are different from those in image analysis, and adapt Intel hardware. We are creating the building blocks of natural language learning, they will need to build appropriate hardware, assuming a lot of people will use it, not just a few academics. Intel also provides services in the field of artificial intelligence and needs these capabilities as well.
In terms of understanding language, we focus on the structure of sentences, the subject matter, the subject matter, also to bring the meaning of the words and sentences even in things that are not directly said in the sentence and need to be deduced from the sentence. People understand the meaning automatically, it is more difficult for a computer.
Has a startup come out of your labs?
We do not work on the establishment of startups, we will not run an incubator, but many graduates of the laboratory have integrated into prominent companies and academia, there are graduates in the research laboratories of Google, Facebook and of course Intel. I am also at the same time a research center - the Allen Institute for Artificial Intelligence, a fairly new association. where we develop infrastructure for the benefit of the public.
Yaniv Garti from the head of Intel Israel said: "The support for Bar-Ilan University, which has first-class researchers, will yield important research in the field of natural language processing, which is an innovative field of research and technology. Academia plays an important part in the next developments of AI and I believe that combining forces with Intel in a variety of activities will give all parties substantial advantages."
Moshe Weserbalt, director of the research group in language processing and deep learning at Intel in the AI product group, said: "The AI laboratory at Intel is happy to expand the academic relationship in Israel through joint research with one of the world's leading research groups in the field of natural language processing. The close relationship between our developers and the researchers at Bar-Ilan is a fruitful relationship that will continue to yield innovative developments in the future as well."
Intel has extensive activity in the field of academic research, joint projects and advanced technology contributions to academic teaching laboratories in the field of AI, which include: the AI Research Center at the Technion, the Natural Language Processing and Deep Learning Laboratory at Bar-Ilan University, the Deep Learning Laboratory at Ben Gurion University, the SCIENCE DATA Center And the BIG DATA ANALYTICS center at the Hebrew University of Jerusalem - a collaboration that includes courses, lectures, projects and in addition a connection to Intel's huge DATA CENTER in the USA for remote research on Intel's largest servers.
