Professor Rita Singh
from being mere workhorses to being our friends
I make machines understand human speech, and the world of sounds around them in general.
It remains unclear to scientists just how the speech signal carries information that is interpreted so effectively by the human brain. Current machine learning techniques for speech make up for this lack of understanding by analyzing large volumes of data and learning surface-level structure from them. Today's best automated speech recognition systems are trained on much more speech than most humans will hear in their lifetime. Yet their performance remains much worse than that of humans. For the most part, my research tries to reduce this gap by building on a better understanding of the speech signal itself.
On the other hand, it’'s not inconceivable that such understanding is irrelevant: a purely superficial data-driven approach may suffice. Taking the latter view to the extreme, it may be possible to teach machines to recognize speech well by simply exposing them to the ever increasing amounts of data from sources such as YouTube, news broadcasts etc. These data are largely unlabeled and untagged -- we only have the audio, but no written record of what words were spoken. The challenge here is to devise mechanisms that enable machines to learn effectively from very large amounts of unlabeled data. A part of my research also addresses this challenge.
Speech becomes harder to recognize when it is distorted, reverberated or corrupted by interfering sounds. As noise levels increase humans rely increasingly on visual and contextual cues to understand speech. At high noise levels it becomes difficult for humans to carry out a conversation without shouting, and at some point conversation becomes impossible. Machines respond in much the same way; however the interference levels at which it becomes impossible for machines to recognize speech is much lower than the levels that affect humans similarly. Machines fail where human beings might not even notice the interferences. I also work on mitigating the effects of noise, and other corrupting influences on machine speech recognition.
For humans, "audition" goes beyond only recognizing speech. There is a world of sounds around us that we constantly hear, process and respond to. My research extends to enabling generic audition capabilities in machines. I work on developing techniques that help machines interpret the soundscapes around them.
I believe in the power of the collective mind. I collaborate with many experts in the areas I work on and other related areas. Together we hope to solve these mysteries of human speech and audition someday, and impart our speech abilities to machines. Someday machines will transition from being mere workhorses to being our friends and companions. I like to believe that we will have contributed to this transition.
to create a computer system that learns over time to read the web
During the masters study in India, I got interested in the area of Information Retrieval, Data Mining and Machine Learning. Subsequently, while working at Google, I came across some of the latest advancements in above mentioned fields along with upcoming challenges of harvesting web-scale data.
When I joined PhD program at LTI, I got exposure to entire spectrum of Language technologies and a variety of exciting research problems that researchers here are working on. The question that I find most fascinating is: "Can computers learn to read?"
"Read the Web" is a research project that attempts to create a computer system that learns over time to read the web. This system keeps reading the Web and generates a knowledge base of all the facts that it learns. My research in particular focuses on extracting useful information from semi-structured sources on the Web e.g. HTML tables. We believe that information extracted from these sources will help us enrich and expand the existing knowledge base. It will also help in improving performance of keyword search.
My vision for the future of language technologies is that computers would be able to understand the semantics of natural languages that humans speak. This can help language technologies in various ways. E.g. If machines can understand semantics, search engines can answer user queries based on query intent. Machine translation can generate translations which are semantically correct along with being grammatically correct. Speech synthesis systems can generate speech signals which have stress and intonation patterns according to mood and intension of the conversation.
Thuy Linh Nguyen
different cultures can keep their origins and still able to access to world knowledges
Word Segmentation Morphology Tokenization Machine Translation
I am a PhD candidate at Language Technologies Institute working in word segmentation and morphology tokenization for machine translation.
Languages express meaning differently, for example the translation of a word in some languages might be a long sentence in English. To get a better translation, my research focuses on how computers segment such long foreign words into smaller parts equivalent to English words.
Machine translation is a fun and rewarding research area. It helps bring people together. I believe in the future, people from different cultures can keep their origins and still able to access to world knowledges and machine translation will play an important role in that road.
an interactive learning system that learns to identify concepts expressed in text
A lot of text is produced and consumed over internet everyday. It is impossible for us to read everything and find the information we are looking for in real-time. With the advancement in computing, can computers process and understand large volumes of text available on the internet and help us find the information we are looking for, quickly?
Understanding natural language is a hard problem. Especially with informal language (with slangs, metaphors and spelling mistakes) used on social media sites such as Twitter and Facebook. However, the need to understand text is becoming more important every day. We don't buy a product unless we know its star rating, determined from customer reviews. We expect to find answers to our questions on the internet. These are the challenges and applications that excite me to work in the area of Language Technologies. Language Technologies lies at the intersection of Natural Language Processing, Machine Learning and Statistics.
Language technology applications that excite me the most are Question Answering, Sentiment Analysis and Information Extraction. In my thesis research, I am working on an interactive learning system that learns to identify concepts expressed in text. The goal is to use a mix of different forms of supervision from the user in order to minimize the total supervision cost required to achieve the desired performance.
this field of research is the tip of an iceberg
I'm a third year PhD student in LTI. My research is about understanding the language of biology. I work to discover the relationship between protein building blocks and how they fold, so that we can design better drugs. I also work on gene expression data, to understand which genes cause specific diseases. Whether it is the language that we speak or the language that our cells speak, LTI has great scientists that can help you understand it better, so I am so happy to be part of LTI. I think this field of research is the tip of an iceberg. The combination of technology, machine learning, and biological data will enable applications that make many people's lives better. I am so passionate about my field and will definitely stay in the area of research, either in academia or industry.
Professor Carolyn Rose
a clear path towards impacting students around the world
CSCLComputer Supported Cooperative Learning Education Machine Learning
Carolyn Penstein Rose is an Assistant Professor with a joint appointment between the Language Technologies Institute and the Human-Computer Interaction Institute. In that role, she leads several research projects in the area of Computer Supported Collaborative Learning and teaches courses in machine learning, discourse analysis, automatic summarization, computer supported collaborative learning, and research communication.
Dr. Rose's research group works towards advancing language technologies for use in supporting online collaboration. One of the major cross-cutting themes of her work is identification of conversational constructs that predict important group learning outcomes such as learning, identification, motivation and relationship formation. Several projects focus on development of technology for automatic analysis of conversational data as well as development of tool benches that put this technology in the hands of other researchers. One of the most exciting developments has been the effective adaptation of tutorial dialogue technology for the purpose of improving online collaboration. Results from numerous studies in a variety of learning domains show statistically significant improvements in student learning, exchange of help, productivity in brainstorming, and improvement of social climate using this technology.
As a faculty member here, she is strongly interdisciplinary and actively involved in an international network of researchers in the fields of Language Technologies, Sociolinguistics, Education, Health, and Psychology. It is her strong desire to see her research in computer supported collaborative learning make an impact in the world. To this end she is involved in efforts that have a clear path towards impacting students around the world and transforming how they learn on-line through discussion, including partnerships with The Math Forum, a major university based math service reaching millions of students each year, and dissemination through Worth Publishing's Psychology Portal, which is packaged with our country's most popular undergraduate psychology textbook.
Beyond her research, Dr. Rose is pursuing her mission through spearheading The Internship Program for Technology Supported Education, which is a partnership between faculty at the International Institute for Information Technology in Hyderabad (IIIT-H) and the School of Computer Science at Carnegie Mellon University (CMU). The goal is to reach out particularly to 3rd year undergraduates at Indian institutions, to expose them to research, and to start mentoring relationships with the students, and ultimately research partnerships with faculty at their home institutions.
open research areas that LTI could immediately have an effect
My interests are in computational biology and applied machine learning. Computational biology is not one of the main research areas of Language Technologies Institute (LTI) unlike information retrieval or machine translation. LTI research themes cover a broad array of research topics. We are lucky to have open minded and inspirational faculty that allows graduate students to explore different research directions.
In my case, as a PhD student I began conducting research in an NSF funded project titled 'The Biological Language Modeling project'. The idea which initiated this project was based on the assumption that biological sequences such as protein sequences from different organisms can be viewed as text written in different languages. Therefore techniques and tools that are applied and developed for analyzing languages can be utilized in analyzing biological sequences. In general, once you abstract the computational problems in a domain, it very often maps to a more general computer science or mathematical problem, therefore algorithms are often transferable across different domains. I worked for a while in protein sequence language. Later on I started working on the virus-host protein-protein interaction networks, which turned out to be the topic of my PhD thesis. I should mention that LTI is in close relationship with other departments such as Machine Learning Department of SCS or Biological departments both within CMU and UPITT, which allowed me to get access to different disciplines and collaborate with researchers from very diverse backgrounds.
The advent of the sequencing of the human genome has marked a start of a revolution in biology. The advances in high-throughput techniques in molecular biology continue to produce diverse experimental data e.g. genomic sequences, gene expression, and proteomics data. This biological dataset opens a window into how the cell works and what goes wrong when the cell malfunctions and provides a fragmentary, but a complementary view of the cell at different levels of resolution. The availability of such data sets exiting challenges and opportunities for computer scientists. In order to realize the full potential of this diverse and large-scale data, algorithms are needed to integrate diverse datasets and to gain insight in the molecular underpinnings of the diseases such as cancer or AIDS. Furthermore, the translation of the knowledge to the clinics, such as personalized therapies, requires accurate and robust computational and statistical algorithms.
LTI is already a place where a state of the art research is conducted for many areas of language technologies including nature language processing, machine translation and speech recognition. This makes me proud. I am most eager to see LTI deeping its expertise in biology. There are several open research areas, that LTI could immediately have an effect. For instance building ontologies to represent different biological information such as protein interactions, diseases is one area that could make the data scientists life easier. Similarly extracting biological information from unstructured text in an automated fashion is an area which LTI is expert of and can help biologist to leverage the literature in a more efficient way.
I have already graduated. I am currently a post-doctoral researcher in Microsoft Research (MSR) New England Lab. I like to continue to be a researcher either in industry or in academia.
technology that allows us to capture gene expression in live tissue samples
I work in applying machine learning to computational biology problems. One of the exciting problems that I work on is automatically inferring how genes interact by analyzing microscopic images of their expression. This problem lies at the intersection of computer vision, systems biology and machine learning, and over the next one year, we are also hoping to incorporate genetics knowledge in the analysis!
The most exciting development in my field has been the development of imaging technology that allows us to capture gene expression in live tissue samples. This allows researchers to track how genes interact in different parts of the tissue at different times, and will allow us to better understand the causes of various diseases. While current studies have been limited to fruit-flies and worms, I hope that over the next few years, we will develop new algorithms to allow similar studies to be possible in humans.
the level of inter-disciplinary research
Computational Proteomics Machine Learning
My research is in the broad area of Computational Proteomics, with particular emphasis on protein-protein interaction prediction in infectious diseases. I am involved in applying and developing machine learning methods in the study of how diseases work at the level of proteins, which is a part of the bigger puzzle that is the human immune system, and its response to various pathogens.
In the future, I want to continue working on problems in computational biology, as I find this field very challenging and it concerns problems which have a broader impact on the society.
According to me, the best part of LTI is the level of inter-disciplinary research that happens here and I wish that it continues to create new inroads on this front.
to help human makes sense or reasons over the wealth of information
My research area is Information Extraction and Knowledge Discovery. I am involved in the NELL (Never Ending Language Learning) project that aims to build a never-ending machine learning system that acquires the ability to extract structured information from unstructured web pages.
With the onset of the internet, more information is made available online. However, most of the information is in unstructured form, making it almost impossible (especially manually) to do further processing, analysis or reasoning over the billions of information that is available online. The goal of NELL project is to automatically read the abundant information on the web, extract and store it in a structured form that will make it easier for human or machine alike to browse or reason over it. There has been many development in the field of information extraction in the past 10 years. Most of the early systems were using human-constructed rules and patterns to extract information from unstructured text. Recent approaches use machine learning to automatically learns these patterns to extract information. NELL project is unique in that not only it attempts to learn these patterns automatically, it is also built to continuously improve its "reading competence" so that tomorrow it can extract more information from the web more accurately. In the future, it will be exciting to see the development of a machine that can read (extract information) as well as human and perhaps more: to help human makes sense or reasons over the wealth of information there is in the web.
Speech and language technologies have matured to a stage where they are actually usable
I am a second year Masters student, currently working on automatically assessing children's oral reading prosody in the context of Project LISTEN's Reading Tutor. Along with assessment, my other topic of research is the visualization of prosody, to enable the Reading Tutor to provide feedback on prosody to children in real time.
One of the most exciting developments in my field, in my opinion is that speech and language technologies have matured to a stage where they are actually usable and are being integrated into devices we use daily. I am eager to see more open source software come out of the LTI, so that people all over the world can build systems for their own languages. It is my dream to build products using these technologies that can be useful to people, especially in developing countries, and I hope to go back to industry and work on this after I graduate.