Kavi Mahesh

Semantics in Action

The paper presents a concise description of a comprehensive approach to computational lexical semantics and focuses on the treatment of events. We reason about the semantic information that should be encoded in a lexicon entry to support the twin tasks of constructing Text Meaning Representations (TMRs) for input texts and generating texts off TMRs. As static knowledge sources cannot be expected to cover all textual inputs, we describe and illustrate how lexical entries can be changed dynamically to fit the textual context at processing time. On the very important issue of knowledge acquisition, our experience shows that determining the meaning of lexical items is not a trivial task for a team of human acquirers (who are, we believe, absolutely indispensable for the more complex decisions in lexical knowledge acquisition). We illustrate how one can overcome the subjectivity of acquirers partly through advanced methodology and partly by having the lexical-semantic model account for some of the combinatory and (semi-)productive principles of natural language.

Measuring semantic coverage

The development of natural language processing systems is currently driven to a large extent by measures of knowledge-base size and coverage of individual phenomena relative to a corpus. While these measures have led to significant advances for knowledge-lean applications, they do not adequately motivate progress in computational semantics leading to the development of large-scale, general purpose NLP systems. In this article, we argue that depth of semantic representation is essential for covering a broad range of phenomena in the computational treatment of language and propose depth as an important additional dimension for measuring the semantic coverage of NLP systems. We propose an operationalization of this measure and show how to characterize an NLP system along the dimensions of size, corpus coverage, and depth. The proposed framework is illustrated using several prominent NLP systems. We hope the preliminary proposals made in this article will lead to prolonged debates in the field and will continue to be refined.

Automatic Recognition of Sign Language Images

The objective of the research presented in this chapter is to enable communication between people with hearing impairment and those with visual impairment. Computer recognition of sign language snapshots is one of the most challenging research problems in this area. This chapter presents an efficient and fast algorithm for identification of the number of fingers opened in a gesture representing an alphabet of the American Sign Language. Finger detection is accomplished based on the concept of boundary tracing and finger tip detection. A significant feature of the solution is that it does not require the hand to be perfectly aligned to the camera or use any special markers or input gloves.

Visual Analytics of Terrorism Data

Terrorist attacks are on the rise in the last few decades. As we look for alternative methods to address this problem, Social Network Analysis (SNA) offers promising techniques among which visual analytics combines the best of statistical analysis with human n ability to visually recognize hidden patterns and trends in terrorist events and networks. In this paper, we present our work on visual analytics of terrorism in India. Since structured data is not readily available for such analysis, we show how data extracted from online news articles and tweets can be integrated with the Global Terrorism Database (GTD) to produce promising results.

Visualizing Textbook Concepts: Beyond Word Co-occurrences

In this paper, we present a simple and elegant algorithm to extract and visualize various concept relationships present in sections of a textbook. This can be easily extended to develop visualizations of entire chapters or textbooks, thereby opening up opportunities for developing a range of visual applications for e-learning and education in general. Our algorithm creates visualizations by mining relationships between concepts present in a text by applying the idea of transitive closure rather than merely counting co-occurrences of terms. It does not require any thesaurus or ontology of concepts. We applied the algorithm to two textbooks - Theory of Computation and Machine Learning - to extract and visualize concept relationships from their sections. Our findings show that the algorithm is capable of capturing deep-set relationships between concepts which could not have been found by using a term co-occurrence approach.

Terrorism analytics: Learning to predict the perpetrator

Data about terrorist attacks in India was analysed. Several machine learning algorithms were trained on the Indian subset of the Global Terrorism Database to learn to predict the perpetrator of a terrorist attack, given data about the types of attack, target and weapon in addition to the location, year and other attributes of the event. It was found that Support Vector Machine technique gave accuracy higher than 75% in predicting the perpetrators. This approach has the potential to aid investigating agencies and carries significant implications for national and international security.

Stringing Subtitles in Sign Language

This paper presents a system designed as a sign language teaching aid. This system is capable of generating the sign-language equivalent of an input phrase or sentence in textual or audio form, with automated processing for language constructs such as tenses and plurals. For words or phrases, especially named entities not present in the repository, the system strings together images or video clips of alphabets to transcribe the input. The system also has the provision for a user to record their version of phrases and words as video clips (or still images) to augment the systems repository. This is particularly useful for pedagogy in Indian sign languages since most schools have their own vocabulary. The system was designed with inputs and continuous feedback from faculty at the Mathru Center for the Deaf, Dumb and Differently-Abled in Bengaluru. We envisage continuing to test the system and augment features to eventually open this up as a useful avenue for the pedagogy of sign languages.

Farmer's Analytical Assistant

About half of the population of India depends on agriculture for its livelihood, but its contribution towards the GDP of India is only 14 per cent. One possible reason for this is the lack of adequate crop planning by farmers. There is no system in place to advise farmers what crops to grow. In this paper we present an attempt to predict crop yield and price that a farmer can obtain from his land, by analysing patterns in past data. We make use of a sliding window non-linear regression technique to predict based on different factors affecting agricultural production such as rainfall, temperature, market prices, area of land and past yield of a crop. The analysis is done for several districts of the state of Karnataka, India. Our system intends to suggest the best crop choices for a farmer in order to address the prevailing socio-economic crisis facing many farmers today.

Curating a Semantic Bibliographic Catalog

Large online libraries are now common across the Web. However these libraries are usually fragmented and are not semantically connected thereby making their search and access difficult. Also there are no large bibliographic Linked Open Datasets available for use in research and analytics. This paper shows how to create a large, comprehensive, RDF triple store of semantic data about books. The primary aim of this work is to establish a linked relation between all the available books in the world and connect them to the already available linked datasets. The Open Library dataset, which has over 45 million records is first serialized by converting it into a triple format and then linked together using predicates from different ontologies. A simple endpoint for a semantic book search engine called BookLOD is created to demonstrate the utility of the dataset.

ResearchAssist: Analyzing author interests

The typical predicament that haunts a researcher is whether to explore new research domains or specialize in their current domain. The prevalent methodology in scientific research is to do emerging trend detection by analyzing topics from scientific texts using a topic modeling approach. In this paper, we propose an alternative methodology. Initially, the researcher makes a fundamental choice of specializing in a specific area or experimenting in related areas. A rule based labelling algorithm is proposed to categorize authors based on their inclination to experimentation or specialization. Once the individual choice is made, research domains of the authors having similar dispositions can be analyzed as the researchers themselves are the best proxy for emerging trends. Finally, this paper explains the possible applications of this methodology for a new researcher, researchers looking to explore or specialize, formation of collaborations and interdisciplinary teams.