Preeti Bhargava, PhD

Recently much progress has been made in entity disambiguation and linking systems (EDL). Given a piece of text, EDL links words and phrases to entities in a knowledge base, where each entity defines a specific concept. Although extracted entities are informative, they are often too specific to be used directly by many applications. These applications usually require text content to be represented with a smaller set of predefined concepts or topics, belonging to a topical taxonomy, that matches… Show more

Recently much progress has been made in entity disambiguation and linking systems (EDL). Given a piece of text, EDL links words and phrases to entities in a knowledge base, where each entity defines a specific concept. Although extracted entities are informative, they are often too specific to be used directly by many applications. These applications usually require text content to be represented with a smaller set of predefined concepts or topics, belonging to a topical taxonomy, that matches their exact needs. In this study, we aim to build a system that maps Wikidata entities to such predefined topics. We explore a wide range of methods that map entities to topics, including GloVe similarity, Wikidata predicates, Wikipedia entity definitions, and entity-topic co-occurrences. These methods often predict entity-topic mappings that are reliable, i.e., have high precision, but tend to miss most of the mappings, i.e., have low recall. Therefore, we propose an ensemble system that effectively combines individual methods and yields much better performance, comparable with human annotators. Show less

Social media serves as a unified platform for users to express their thoughts on subjects ranging from their daily lives to their opinion on consumer brands and products. These users wield an enormous influence in shaping the opinions of other consumers and influence brand perception, brand loyalty and brand advocacy. In this paper, we analyze the opinion of 19M Twitter users towards 62 popular industries, encompassing 12,898 enterprise and consumer brands, as well as associated subject matter… Show more

Social media serves as a unified platform for users to express their thoughts on subjects ranging from their daily lives to their opinion on consumer brands and products. These users wield an enormous influence in shaping the opinions of other consumers and influence brand perception, brand loyalty and brand advocacy. In this paper, we analyze the opinion of 19M Twitter users towards 62 popular industries, encompassing 12,898 enterprise and consumer brands, as well as associated subject matter topics, via sentiment analysis of 330M tweets over a period spanning a month. We find that users tend to be most positive towards manufacturing and most negative towards service industries. In addition, they tend to be more positive or negative when interacting with brands than generally on Twitter. We also find that sentiment towards brands within an industry varies greatly and we demonstrate this using two industries as use cases. In addition, we discover that there is no strong correlation between topic sentiments of different industries, demonstrating that topic sentiments are highly dependent on the context of the industry that they are mentioned in. We demonstrate the value of such an analysis in order to assess the impact of brands on social media. We hope that this initial study will prove valuable for both researchers and companies in understanding users’ perception of industries, brands and associated topics and encourage more research in this field. Show less

In this paper, we describe the Lithium Natural Language Processing (NLP) system - a resource-constrained, high-throughput and language-agnostic system for information extraction from noisy user generated text on social media. Lithium NLP extracts a rich set of information including entities, topics, hashtags and sentiment from text. We discuss several real world applications of the system currently incorporated in Lithium products. We also compare our system with existing commercial and… Show more

In this paper, we describe the Lithium Natural Language Processing (NLP) system - a resource-constrained, high-throughput and language-agnostic system for information extraction from noisy user generated text on social media. Lithium NLP extracts a rich set of information including entities, topics, hashtags and sentiment from text. We discuss several real world applications of the system currently incorporated in Lithium products. We also compare our system with existing commercial and academic NLP systems in terms of performance, information extracted and languages supported. We show that Lithium NLP is at par with and in some cases, outperforms state-of-the-art commercial NLP systems. Show less

In this work, we open up the DAWT dataset - Densely Annotated Wikipedia Texts across multiple languages. The annotations include labeled text mentions mapping to entities (represented by their Freebase machine ids) as well as the type of the entity. The data set contains total of 13.6M articles, 5.0B tokens, 13.8M mention entity co-occurrences. DAWT contains 4.8 times more anchor text to entity links than originally present in the Wikipedia markup. Moreover, it spans several languages including… Show more

In this work, we open up the DAWT dataset - Densely Annotated Wikipedia Texts across multiple languages. The annotations include labeled text mentions mapping to entities (represented by their Freebase machine ids) as well as the type of the entity. The data set contains total of 13.6M articles, 5.0B tokens, 13.8M mention entity co-occurrences. DAWT contains 4.8 times more anchor text to entity links than originally present in the Wikipedia markup. Moreover, it spans several languages including English, Spanish, Italian, German, French and Arabic. We also present the methodology used to generate the dataset which enriches Wikipedia markup in order to increase number of links. In addition to the main dataset, we open up several derived datasets including mention entity co-occurrence counts and entity embeddings, as well as mappings between Freebase ids and Wikidata item ids. We also discuss two applications of these datasets and hope that opening them up would prove useful for the Natural Language Processing and Information Retrieval communities, as well as facilitate multi-lingual research. Show less

The Entity Disambiguation and Linking (EDL) task matches entity mentions in text to a unique Knowledge Base (KB) identifier such as a Wikipedia or Freebase id. It plays a critical role in the construction of a high quality information network, and can be further leveraged for a variety of information retrieval and NLP tasks such as text categorization and document tagging. EDL is a complex and challenging problem due to ambiguity of the mentions and real world text being multi-lingual… Show more

The Entity Disambiguation and Linking (EDL) task matches entity mentions in text to a unique Knowledge Base (KB) identifier such as a Wikipedia or Freebase id. It plays a critical role in the construction of a high quality information network, and can be further leveraged for a variety of information retrieval and NLP tasks such as text categorization and document tagging. EDL is a complex and challenging problem due to ambiguity of the mentions and real world text being multi-lingual. Moreover, EDL systems need to have high throughput and should be lightweight in order to scale to large datasets and run on off-the-shelf machines. More importantly, these systems need to be able to extract and disambiguate dense annotations from the data in order to enable an Information Retrieval or Extraction task running on the data to be more efficient and accurate. In order to address all these challenges, we present the Lithium EDL system and algorithm - a high-throughput, lightweight, language-agnostic EDL system that extracts and correctly disambiguates 75% more entities than state-of-the-art EDL systems and is significantly faster than them. Show less

A large volume of research in ubiquitous systems has been devoted to using data, that has been sensed from users’ smartphones, to infer their current high level context and activities. However, mining users’diverse longitudinal behavioral patterns, which can enable exciting new context-aware applications, has not received much attention. In this paper, we focus on learning and identifying such behavioral patterns from large-scale data collected from users’ smartphones. To this end, we develop a… Show more

A large volume of research in ubiquitous systems has been devoted to using data, that has been sensed from users’ smartphones, to infer their current high level context and activities. However, mining users’diverse longitudinal behavioral patterns, which can enable exciting new context-aware applications, has not received much attention. In this paper, we focus on learning and identifying such behavioral patterns from large-scale data collected from users’ smartphones. To this end, we develop a unified infrastructure and implement several novel approaches for building diverse behavioral models of users. Examples of generated models include classifying users’ semantic places and predicting their availability for accepting calls etc. We evaluate our work on real-world data of 200 users, from the DeviceAnalyzer dataset, consisting of 365 million data points and show that our algorithms and approaches can model user behavior with high accuracy and outperform existing approaches. Show less

Due to increasing proliferation of smart devices, many users store a significant proportion of personal data on them. Thus, personal sensing applications that sense a user’s context via his smart device have significant privacy implications. In this paper, we conduct an exploratory study of privacy, trust, risks and other concerns of users with smart phone based context-aware personal sensing systems and applications. Our study results show that users are concerned that their sensed data can be… Show more

Due to increasing proliferation of smart devices, many users store a significant proportion of personal data on them. Thus, personal sensing applications that sense a user’s context via his smart device have significant privacy implications. In this paper, we conduct an exploratory study of privacy, trust, risks and other concerns of users with smart phone based context-aware personal sensing systems and applications. Our study results show that users are concerned that their sensed data can be misused, used for personal identification and tracking or for commercial purposes. However, they are willing to trade privacy for additional benefits if their sensed information is used for effective and beneficial causes. Furthermore, they are willing to trust reputed technology companies, with their data, if the benefits are significant. Based on these results, we propose a few design guidelines for designers of personal sensing apps and outline some interesting directions for future research. Show less

PhD dissertation, Department of Computer Science, University of Maryland

A context-aware system uses context to provide relevant information and services to the user, where relevancy depends on the user’s situation. This relevant information could include a wide range of heterogeneous content. Many existing context-aware systems determine this information based on pre-defined ontologies or rules. In addition, they rely on users’ context history to filter it. Moreover, they often provide domain-specific information. Such systems are not applicable to a large and… Show more

A context-aware system uses context to provide relevant information and services to the user, where relevancy depends on the user’s situation. This relevant information could include a wide range of heterogeneous content. Many existing context-aware systems determine this information based on pre-defined ontologies or rules. In addition, they rely on users’ context history to filter it. Moreover, they often provide domain-specific information. Such systems are not applicable to a large and varied set of user situations and information needs, and may suffer from cold start for new users. In this paper, we address these limitations and propose a novel, general and flexible approach for bootstrapped discovery and ranking of heterogeneous relevant services and information in context-aware systems. We design and implement four variations of a base algorithm that ranks candidate relevant services, and the information to be retrieved from them, based on the semantic relatedness between the information provided by the services and the user’s situation description. We conduct a live deployment with 14 subjects to evaluate the efficacy of our algorithms. We demonstrate that they have strong positive correlation with human supplied relevance rankings and can be used as an effective means to discover and rank relevant
services and information. We also show that our approach is applicable to a wide set of users’ situations and to new users without requiring any user interaction history Show less

Today’s context-aware systems tend to be reactive or ‘pull’ based - the user requests or queries for some information and the system responds with the requested information. However, none of the systems anticipate the user’s intent and behavior, or take into account his current events and activities to pro-actively ‘push’ relevant information to the user. On the other hand, Proactive context-aware systems can predict and anticipate user intent and behavior, and act proactively on the users’… Show more

Today’s context-aware systems tend to be reactive or ‘pull’ based - the user requests or queries for some information and the system responds with the requested information. However, none of the systems anticipate the user’s intent and behavior, or take into account his current events and activities to pro-actively ‘push’ relevant information to the user. On the other hand, Proactive context-aware systems can predict and anticipate user intent and behavior, and act proactively on the users’ behalf without explicit requests from them. Two fundamental capabilities of such systems are: prediction and autonomy. In this paper, we address the second capability required by a context-aware system to act proactively i.e. acting autonomously without an explicit user request. To address it, we present a new paradigm for enabling proactivity in context-aware
middleware systems by means of a Planning Framework based on HTN planning. We present the design of a Planning Framework within the infrastructure of our intelligent context-aware middleware called Rover II. We also implement this framework and evaluate its utility with several use cases. We also highlight the benefits of using such a framework in dynamic ubiquitous systems.
Show less

Given the abundance of online information available to mobile users, particularly tourists and weekend travelers, recommender systems that effectively filter this information and suggest interesting participatory opportunities will become increasingly important. Previous work has explored recommending interesting locations; however, users would also benefit from recommendations for activities in which to participate at those locations along with suitable
times and days. Thus, systems that… Show more

Given the abundance of online information available to mobile users, particularly tourists and weekend travelers, recommender systems that effectively filter this information and suggest interesting participatory opportunities will become increasingly important. Previous work has explored recommending interesting locations; however, users would also benefit from recommendations for activities in which to participate at those locations along with suitable
times and days. Thus, systems that provide collaborative recommendations involving multiple dimensions such as location, activities and time would enhance the overall experience of users.The relationship among these dimensions can be modeled by higher order matrices called tensors which are then solved by tensor factorization. However, these tensors can be extremely sparse. In this paper, we present a system and an approach for performing multidimensional
collaborative recommendations for Who (User), What (Activity), When (Time) and Where (Location), using tensor factorization on sparse user-generated data. We formulate an objective
function which simultaneously factorizes coupled tensors and matrices constructed from heterogeneous data sources. We evaluate our system and approach on large-scale real world data sets consisting of 588,000 Flickr photos collected from three major metro regions in USA. We compare our approach with several state-ofthe-art baselines and demonstrate that it outperforms all of them. Show less

With the advent of the internet, online communication media and social networks have become increasingly popular among users for interaction and communication. Integrating these online communications with other sources of a user’s context can help improve his interaction with context-aware systems as it enables the systems to provide highly personalized content to both
individual and groups of users. To this end, a user’s communication context (such as the people he communicates with often… Show more

With the advent of the internet, online communication media and social networks have become increasingly popular among users for interaction and communication. Integrating these online communications with other sources of a user’s context can help improve his interaction with context-aware systems as it enables the systems to provide highly personalized content to both
individual and groups of users. To this end, a user’s communication context (such as the people he communicates with often, and the topics he discusses frequently) becomes an important aspect of his context model and new frameworks and methodologies are required for extracting
and representing it. In this paper, we present a hybrid framework derived from traditional graph based and object oriented models that employs various Natural Language Processing techniques for extracting and representing users’ communication context from their aggregated online communications. We also evaluate the framework using the email communication log of a user. Show less

User interest profiles have become essential for personalizing information streams and services, and user interfaces and experiences. In today’s world, social networks such as Facebook or Twitter provide users with a powerful platform for interest expression and can, thus, act as a
rich content source for automated user interest modeling. This, however, poses significant challenges because the user generated content on them consists of free unstructured text. In addition, users may not… Show more

User interest profiles have become essential for personalizing information streams and services, and user interfaces and experiences. In today’s world, social networks such as Facebook or Twitter provide users with a powerful platform for interest expression and can, thus, act as a
rich content source for automated user interest modeling. This, however, poses significant challenges because the user generated content on them consists of free unstructured text. In addition, users may not explicitly post or tweet about everything that interests them. Moreover,
their interests evolve over time. In this paper, we propose a novel unsupervised algorithm and system that addresses these challenges. It models a broad range of an individual user’s explicit and implicit interests from her social network profile and activities without any user input. We perform extensive evaluation of our system, and algorithm, with a dataset consisting of 488 active
Facebook users’ profiles and demonstrate that it can accurately estimate a user’s interests in practice. Show less

A fundamental goal of indoor localisation technology is to achieve the milestone of combining minimal cost with accuracy sufficient enough for general consumer applications. To achieve this, current indoor positioning systems need either extensive calibration or expensive hardware. Moreover, very few systems built so far have addressed floor determination in multi-story buildings. In this paper, we explain a Wi-fi-based indoor localisation, tracking and navigation system for multi-story… Show more

A fundamental goal of indoor localisation technology is to achieve the milestone of combining minimal cost with accuracy sufficient enough for general consumer applications. To achieve this, current indoor positioning systems need either extensive calibration or expensive hardware. Moreover, very few systems built so far have addressed floor determination in multi-story buildings. In this paper, we explain a Wi-fi-based indoor localisation, tracking and navigation system for multi-story buildings called Locus. Locus determines a device’s floor as well as location on that floor using existing knowledge of infrastructure, and without requiring any calibration or proprietary hardware. It is an inexpensive solution with minimum set-up and maintenance expenses, is scalable, readily deployable and robust to environmental changes. Experimental results in three different buildings spanning multiple floors show that it can determine the floor with 95.33% accuracy and the location on the floor with an error of 6.49m on an average in real-life practical environments. We also demonstrate its utility via two location-based applications for indoor navigation and tracking in emergency scenarios. Show less

In order to make context-aware systems more effective and provide timely, personalized and relevant information to a user, the context or situation of the user must be clearly defined along several dimensions. To this end, the system needs to simultaneously recognize multiple dimensions of the user’s situation such as location, physical activity etc. in an automated and unobtrusive manner. In this paper, we present SenseMe - a system that leverages a user’s
smartphone and its multiple… Show more

In order to make context-aware systems more effective and provide timely, personalized and relevant information to a user, the context or situation of the user must be clearly defined along several dimensions. To this end, the system needs to simultaneously recognize multiple dimensions of the user’s situation such as location, physical activity etc. in an automated and unobtrusive manner. In this paper, we present SenseMe - a system that leverages a user’s
smartphone and its multiple sensors in order to perform continuous, on-device, and multi-dimensional context and activity recognition. It recognizes five dimensions of a user’s situation in a robust, automated, scalable, power efficient and non-invasive manner to paint a context-rich picture of the user. We evaluate SenseMe against several metrics with the aid of 2 two-week long live deployments involving 15 participants. We demonstrate improved or comparable accuracy with respect to existing systems without requiring any user calibration or input. Show less

The holy grail in indoor location technology is to achieve the milestone of combining minimal cost with accuracy, for general consumer applications. A low-cost system should be inexpensive both to install and maintain, requiring only available consumer hardware to operate and its accuracy should be room-level or better. To achieve this, current systems require either extensive calibration or expensive hardware. Moreover, very few systems built so far have addressed localization in multi-story… Show more

The holy grail in indoor location technology is to achieve the milestone of combining minimal cost with accuracy, for general consumer applications. A low-cost system should be inexpensive both to install and maintain, requiring only available consumer hardware to operate and its accuracy should be room-level or better. To achieve this, current systems require either extensive calibration or expensive hardware. Moreover, very few systems built so far have addressed localization in multi-story buildings. We explain a heuristics based indoor localization, tracking and navigation system for multi-story buildings called Locus that determines floor and location by using the locations of infrastructure points, and without the need for radio maps or calibration. It is an inexpensive solution with minimum setup and maintenance expenses. Initial experimental results in an indoor space spanning 175,000 square feet, show that it can determine the floor with 99.97% accuracy and the location with an average location error of 7m. Show less

A major challenge of context models is to balance simplicity, generality, usability and extensibility. It is also important that the model be practical and implementable. In pursuit of this goal, this paper proposes a context model, Rover Context Model (RoCoM), structured around four primitives that can be used to represent and model any situation and activity: entities, events, relationships, and activities. It introduces the notion of templates of context for each primitive and describes… Show more

A major challenge of context models is to balance simplicity, generality, usability and extensibility. It is also important that the model be practical and implementable. In pursuit of this goal, this paper proposes a context model, Rover Context Model (RoCoM), structured around four primitives that can be used to represent and model any situation and activity: entities, events, relationships, and activities. It introduces the notion of templates of context for each primitive and describes, albeit briefly, the RoCoM Ontology (RoCoMO). It also describes the design and architecture of an abstract, generic and intelligent context-aware middleware called Rover II. We propose this framework as a solution to address the context problem as a whole, and be usable in many domains. We also illustrate its application with the aid of a context-aware public safety application that is deployed in the UMD campus. Show less

We describe an abstract, generic and extensible ontology, the Rover Context Model Ontology (RoCoMO), which is currently being designed and developed to model and represent context in an intelligent context-aware middleware system, called Rover II. The Rover Context Model (RoCoM) is the underlying context model for Rover II and is centered on four primitives that can be used to represent a situation: entity, event, activity and relationship. The ontology is expressed using the Web Ontology… Show more

We describe an abstract, generic and extensible ontology, the Rover Context Model Ontology (RoCoMO), which is currently being designed and developed to model and represent context in an intelligent context-aware middleware system, called Rover II. The Rover Context Model (RoCoM) is the underlying context model for Rover II and is centered on four primitives that can be used to represent a situation: entity, event, activity and relationship. The ontology is expressed using the Web Ontology Language (OWL) and includes two components – RoCoM Core Ontology and the RoCoM Application Ontology. We also illustrate its usage with the aid of a public safety application called M-Urgency that is currently deployed at the UMD campus. Show less

The importance of information for the appropriate handling of any situation is well recognized. We, as human beings, use context not only to interpret available information, but also to seek additional relevant, missing information. Most information systems, which support human decision-making, process context inadequately. Moreover, when some context is considered, the structure is often so rigid that, but for very restricted applications, such systems are not usable. In order to integrate… Show more

The importance of information for the appropriate handling of any situation is well recognized. We, as human beings, use context not only to interpret available information, but also to seek additional relevant, missing information. Most information systems, which support human decision-making, process context inadequately. Moreover, when some context is considered, the structure is often so rigid that, but for very restricted applications, such systems are not usable. In order to integrate context-handling capabilities with an information integration system, we need a context model for representing context that can efficiently acquire, maintain and integrate contextual information and make it available to applications on demand. In this paper, we present such a model and describe the design of Rover II, a situation-handling platform which supports integration of context-aware applications. The system manages context by constructing a situation graph which dynamically reflects the relevant situation information. We illustrate the use of Rover II with M-Urgency, a public safety application providing audio and video support for emergency help. This system is being deployed for the University of Maryland campus community of 44 000 students, faculty and staff. Show less