May 29th, 2018
Understanding the Logical and Semantic Structure of Large Documents
Muhammad Mahbubur Rahman
11:00am Wednesday, 30 May 2018, ITE 325b
Understanding and extracting of information from large documents, such as business opportunities, academic articles, medical documents and technical reports poses challenges not present in short documents. The reasons behind this challenge are that large documents may be multi-themed, complex, noisy and cover diverse topics. This dissertation describes a framework that can analyze large documents, and help people and computer systems locate desired information in them. It aims to automatically identify and classify different sections of documents and understand their purpose within the document. A key contribution of this research is modeling and extracting the logical and semantic structure of electronic documents using deep learning techniques. The effectiveness and robustness of ?the framework is evaluated through extensive experiments on arXiv and requests for proposals datasets.
Committee Members: Drs. Tim Finin (Chair), Anupam Joshi, Tim Oates, Cynthia Matuszek, James Mayfield (JHU)
April 22nd, 2018
Preventing Poisoning Attacks on Threat Intelligence Systems
Nitika Khurana, Graduate Student, UMBC
11:00-12:00 Monday, 23 April 2018, ITE346, UMBC
As AI systems become more ubiquitous, securing them becomes an emerging challenge. Over the years, with the surge in online social media use and the data available for analysis, AI systems have been built to extract, represent and use this information. The credibility of this information extracted from open sources, however, can often be questionable. Malicious or incorrect information can cause a loss of money, reputation, and resources; and in certain situations, pose a threat to human life. In this paper, we determine the credibility of Reddit posts by estimating their reputation score to ensure the validity of information ingested by AI systems. We also maintain the provenance of the output generated to ensure information and source reliability and identify the background data that caused an attack. We demonstrate our approach in the cybersecurity domain, where security analysts utilize these systems to determine possible threats by analyzing the data scattered on social media websites, forums, blogs, etc.
April 21st, 2018
UMBC at SemEval-2018 Task 8: Understanding Text about Malware
Ankur Padia, Arpita Roy, Taneeya Satyapanich, Francis Ferraro, Shimei Pan, Anupam Joshi and Tim Finin, UMBC at SemEval-2018 Task 8: Understanding Text about Malware
, Int. Workshop on Semantic Evaluation (collocated with NAACL-HLT), New Orleans, LA, June 2018.
We describe the systems developed by the UMBC team for 2018 SemEval Task 8, SecureNLP (Semantic Extraction from CybersecUrity REports using Natural Language Processing). We participated in three of the sub-tasks: (1) classifying sentences as being relevant or irrelevant to malware, (2) predicting token labels for sentences, and (4) predicting attribute labels from the Malware Attribute Enumeration and Characterization vocabulary for defining malware characteristics. We achieved F1 scores of 50.34/18.0 (dev/test), 22.23 (test-data), and 31.98 (test-data) for Task1, Task2 and Task2 respectively. We also make our cybersecurity embeddings publicly available at https://bit.ly/cybr2vec.
April 15th, 2018
Cognitively Rich Framework to Automate Extraction and Representation of Legal Knowledge
Srishty Saha, UMBC
11-12 Monday, 16 April 2018, ITE 346
With the explosive growth in cloud-based services, businesses are increasingly maintaining large datasets containing information about their consumers to provide a seamless user experience. To ensure privacy and security of these datasets, regulatory bodies have specified rules and compliance policies that must be adhered to by organizations. These regulatory policies are currently available as text documents that are not machine processable and so require extensive manual effort to monitor them continuously to ensure data compliance. We have developed a cognitive framework to automatically parse and extract knowledge from legal documents and represent it using an Ontology. The legal ontology captures key-entities and their relations, the provenance of legal-policy and cross-referenced semantically similar legal facts and rules. We have applied this framework to the United States government’s Code of Federal Regulations (CFR) which includes facts and rules for individuals and organizations seeking to do business with the US Federal government.
April 11th, 2018
2018 Mid-Atlantic Student Colloquium on Speech, Language and Learning
The 2018 Mid-Atlantic Student Colloquium on Speech, Language and Learning (MASC-SLL) is a student-run, one-day event on speech, language & machine learning research to be held at the University of Maryland, Baltimore County (UMBC) from 10:00am to 6:00pm on Saturday May 12. There is no registration charge and lunch and refreshments will be provided. Students, postdocs, faculty and researchers from universities & industry are invited to participate and network with other researchers working in related fields.
Students and postdocs are encouraged to submit abstracts describing ongoing, planned, or completed research projects, including previously published results and negative results. Research in any field applying computational methods to any aspect of human language, including speech and learning, from all areas of computer science, linguistics, engineering, neuroscience, information science, and related fields is welcome. Submissions and presentations must be made by students or postdocs. Accepted submissions will be presented as either posters or talks.
Important Dates are:
- Submission deadline (abstracts): April 16
- Decisions announced: April 21
- Registration opens: April 10
- Registration closes: May 6
- Colloquium: May 12
January 27th, 2018
Cleaning Noisy Knowledge Graphs
My dissertation research is developing an approach to identify and explain errors in a knowledge graph constructed by extracting entities and relations from text. Information extraction systems can automatically construct knowledge graphs from a large collection of documents, which might be drawn from news articles, Web pages, social media posts or discussion forums. The language understanding task is challenging and current extraction systems introduce many kinds of errors. Previous work on improving the quality of knowledge graphs uses additional evidence from background knowledge bases or Web searches. Such approaches are diffuclt to apply when emerging entities are present and/or only one knowledge graph is available. In order to address the problem I am using multiple complementary techniques including entitylinking, common sense reasoning, and linguistic analysis.
December 12th, 2017
Jennifer Sleeman receives AI for Earth grant from Microsoft
Visiting Assistant Professor Jennifer Sleeman (Ph.D. ’17) has been awarded a grant from Microsoft as part of its ‘AI for Earth’ program. Dr. Sleeman will use the grant to continue her research on developing algorithms to model how scientific disciplines such as climate change evolve and predict future trends by analyzing the text of articles and reports and the papers they cite.
AI for Earth is a Microsoft program aimed at empowering people and organizations to solve global environmental challenges by increasing access to AI tools and educational opportunities, while accelerating innovation. Via the Azure for Research AI for Earth award program, Microsoft provides selected researchers and organizations access to its cloud and AI computing resources to accelerate, improve and expand work on climate change, agriculture, biodiversity and/or water challenges.
UMBC is among the first grant recipients of AI for Earth, first launched in July 2017. The grant process was a competitive and selective process and was awarded in recognition of the potential of the work and power of AI to accelerate progress.
As part of her dissertation research, Dr. Sleeman developed algorithms using dynamic topic modeling to understand influence and predict future trends in a scientific discipline. She applied this to the field of climate change and used assessment reports of the Intergovernmental Panel on Climate Change (IPCC) and the papers they cite. Since 1990, an IPCC report has been published every five years that includes four separate volumes, each of which has many chapters. Each report cites tens of thousands of research papers, which comprise a correlated dataset of temporally grounded documents. Her custom dynamic topic modeling algorithm identified topics for both datasets and apply cross-domain analytics to identify the correlations between the IPCC chapters and their cited documents. The approach reveals both the influence of the cited research on the reports and how previous research citations have evolved over time.
Dr. Sleeman’s award is part of an inaugural set of 35 grants in more than ten countries for access to Microsoft Azure and AI technology platforms, services and training. In an post on Monday, AI for Earth can be a game-changer for our planet, Microsoft announced its intent to put $50 million over five years into the program, enabling grant-making and educational trainings possible at a much larger scale.
More information about AI for Earth can be found on the Microsoft AI for Earth website.
November 28th, 2017
Automated Knowledge Extraction from the Federal Acquisition Regulations System (FARS)
With increasing regulation of Big Data, it is becoming essential for organizations to ensure compliance with various data protection standards. The Federal Acquisition Regulations System (FARS) within the Code of Federal Regulations (CFR) includes facts and rules for individuals and organizations seeking to do business with the US Federal government. Parsing and gathering knowledge from such lengthy regulation documents is currently done manually and is time and human intensive.Hence, developing a cognitive assistant for automated analysis of such legal documents has become a necessity. We have developed semantically rich approach to automate the analysis of legal documents and have implemented a system to capture various facts and rules contributing towards building an ef?cient legal knowledge base that contains details of the relationships between various legal elements, semantically similar terminologies, deontic expressions and cross-referenced legal facts and rules. In this paper, we describe our framework along with the results of automating knowledge extraction from the FARS document (Title48, CFR). Our approach can be used by Big Data Users to automate knowledge extraction from Large Legal documents.
November 17th, 2017
Discovering Scientific Influence using Cross-Domain Dynamic Topic Modeling
We describe an approach using dynamic topic modeling to model influence and predict future trends in a scientific discipline. Our study focuses on climate change and uses assessment reports of the Intergovernmental Panel on Climate Change (IPCC) and the papers they cite. Since 1990, an IPCC report has been published every five years that includes four separate volumes, each of which has many chapters. Each report cites tens of thousands of research papers, which comprise a correlated dataset of temporally grounded documents. We use a custom dynamic topic modeling algorithm to generate topics for both datasets and apply crossdomain analytics to identify the correlations between the IPCC chapters and their cited documents. The approach reveals both the influence of the cited research on the reports and how previous research citations have evolved over time. For the IPCC use case, the report topic model used 410 documents and a vocabulary of 5911 terms while the citations topic model was based on 200K research papers and a vocabulary more than 25K terms. We show that our approach can predict the importance of its extracted topics on future IPCC assessments through the use of cross domain correlations, Jensen-Shannon divergences and cluster analytics.
September 5th, 2017
Cognitive Assistance for Automating the Analysis of the Federal Acquisition Regulations System
Government regulations are critical to understanding how to do business with a government entity and receive other bene?ts. However, government regulations are also notoriously long and organized in ways that can be confusing for novice users. Developing cognitive assistance tools that remove some of the burden from human users is of potential bene?t to a variety of users. The volume of data found in United States federal government regulation suggests a multiple-step approach to process the data into machine readable text, create an automated legal knowledge base capturing various facts and rules, and eventually building a legal question and answer system to acquire understanding from various regulations and provisions. Our work discussed in this paper represents our initial efforts to build a framework for Federal Acquisition Regulations System (Title 48, Code of Federal Regulations) in order to create an efficient legal knowledge base representing relationships between various legal elements, semantically similar terminologies, deontic expressions and cross-referenced legal facts and rules.
July 16th, 2017
Deep Representation of Lyrical Style and Semantics for Music Recommendation
Abhay L. Kashyap
11:00-1:00 Thursday, 20 July 2017, ITE 346
In the age of music streaming, the need for effective recommendations is important for music discovery and a personalized user experience. Collaborative filtering based recommenders suffer from popularity bias and cold-start which is commonly mitigated by content features. For music, research in content based methods have mainly been focused in the acoustic domain while lyrical content has received little attention. Lyrics contain information about a song’s topic and sentiment that cannot be easily extracted from the audio. This is especially important for lyrics-centric genres like Rap, which was the most streamed genre in 2016. The goal of this dissertation is to explore and evaluate different lyrical content features that could be useful for content, context and emotion based models for music recommendation systems.
With Rap as the primary use case, this dissertation focuses on featurizing two main aspects of lyrics; its artistic style of composition and its semantic content. For lyrical style, a suite of high level rhyme density features are extracted in addition to literary features like the use of figurative language, profanity and vocabulary strength. In contrast to these engineered features, Convolutional Neural Networks (CNN) are used to automatically learn rhyme patterns and other relevant features. For semantics, lyrics are represented using both traditional IR techniques and the more recent neural embedding methods.
These lyrical features are evaluated for artist identification and compared with artist and song similarity measures from a real-world collaborative filtering based recommendation system from Last.fm. It is shown that both rhyme and literary features serve as strong indicators to characterize artists with feature learning methods like CNNs achieving comparable results. For artist and song similarity, a strong relationship was observed between these features and the way users consume music while neural embedding methods significantly outperformed LSA. Finally, this work is accompanied by a web-application, Rapalytics.com, that is dedicated to visualizing all these lyrical features and has been featured on a number of media outlets, most notably, Vox, attn: and Metro.
Committee: Drs. Tim Finin (chair), Anupam Joshi, Tim Oates, Cynthia Matuszek and Pranam Kolari (Walmart Labs)
June 27th, 2017
Ph.D. Dissertation Defense
Dynamic Data Assimilation for Topic Modeling
9:00am Thursday, 29 June 2017, ITE 325b, UMBC
Understanding how a particular discipline such as climate science evolves over time has received renewed interest. By understanding this evolution, predicting the future direction of that discipline becomes more achievable. Dynamic Topic Modeling (DTM) has been applied to a number of disciplines to model topic evolution as a means to learn how a particular scientific discipline and its underlying concepts are changing. Understanding how a discipline evolves, and its internal and external influences, can be complicated by how the information retrieved over time is integrated. There are different techniques used to integrate sources of information, however, less research has been dedicated to understanding how to integrate these sources over time. The method of data assimilation is commonly used in a number of scientific disciplines to both understand and make predictions of various phenomena, using numerical models and assimilated observational data over time.
In this dissertation, I introduce a novel algorithm for scientific data assimilation, called Dynamic Data Assimilation for Topic Modeling (DDATM), which uses a new cross-domain divergence method (CDDM) and DTM. By using DDATM, observational data in the form of full-text research papers can be assimilated over time starting from an initial model. DDATM can be used as a way to integrate data from multiple sources and, due to its robustness, can exploit the assimilating observational information to better tolerate missing model information. When compared with a DTM model, the assimilated model is shown to have better performance using standard topic modeling measures, including perplexity and topic coherence. The DDATM method is suitable for prediction and results in higher likelihood for subsequent documents. DDATM is able to overcome missing information during the assimilation process when compared with a DTM model. CDDM generalizes as a method that can also bring together multiple disciplines into one cohesive model enabling the identification of related concepts and documents across disciplines and time periods. Finally, grounding the topic modeling process with an ontology improves the quality of the topics and enables a more granular understanding of concept relatedness and cross-domain influence.
The results of this dissertation are demonstrated and evaluated by applying DDATM to 30 years of reports from the Intergovernmental Panel on Climate Change (IPCC) along with more than 150,000 documents that they cite to show the evolution of the physical basis of climate change.
Committee Members: Drs. Tim Finin (co-advisor), Milton Halem (co-advisor), Anupam Joshi, Tim Oates, Cynthia Matuszek, Mark Cane, Rafael Alonso