Aishwarya Naresh Reganti

Code mixing/switching (for the scope of this project, we use the two terms interchangeably) is the phenomenon of mixing two or more languages in speech/text communication. Over the past few years, code mixing has become the default code of communication in geographical regions like India where there is a high percentage of multilingual individuals. Although state-of-the-art conversational systems and language models produce good quality monolingual utterances, they struggle with code-mixed data… Show more

Code mixing/switching (for the scope of this project, we use the two terms interchangeably) is the phenomenon of mixing two or more languages in speech/text communication. Over the past few years, code mixing has become the default code of communication in geographical regions like India where there is a high percentage of multilingual individuals. Although state-of-the-art conversational systems and language models produce good quality monolingual utterances, they struggle with code-mixed data generation because of lack of data compounded by the increased confusability about the language structure and switching points. In this work, we aim to focus on building a goal-oriented Hinglish (Hindi and English code-mixed) dialogue system trained on data curated from code-mixed conversations about Hindi movies. We also aim to make the system more socially engaging and context-aware as compared to mainstream conversational agents. The contributions of our work include: (i) developing a new corpus for code-mixed movie conversations, (ii) building language models that can generate grammatically correct code-mixed utterances and finally, (iii) compiling an end-to-end goal-oriented dialogue system that can have engaging conversations about Hindi movies with prospective users.
Show less

Terrorist Organisations make extensive use of online social media platforms to promulgate their ideologies and evoke many individuals, particularly young people, to support the organization, and take part in their war. The sociological backgrounds of an individual play a very crucial role in determining his/her vulnerability of being lured into joining the organisation and indulge in terrorist activities. In this work, we analyse several sociological aspects (Values & Ethics, Personality… Show more

Terrorist Organisations make extensive use of online social media platforms to promulgate their ideologies and evoke many individuals, particularly young people, to support the organization, and take part in their war. The sociological backgrounds of an individual play a very crucial role in determining his/her vulnerability of being lured into joining the organisation and indulge in terrorist activities. In this work, we analyse several sociological aspects (Values & Ethics, Personality, Optimism /Pessimism, Age & Gender, Veridicality, Life Satisfaction, Mental Health, Implication Analysis) to understand the Sociological Vulnerability of individuals over the social media platform ’Twitter’ using their links with Terrorist Representatives and their Social media Posts. The sociological patterns thus obtained, can be used to determine the vulnerability of any given twitter user by using these patterns as metrics to gauge the the possibility of recruitment into the organisation Show less

Figurative language is language that uses words or expressions with a meaning that is different from the literal interpretation. Figurative language is used with a meaning that is different from the basic meaning and that expresses an idea in an interesting way by using language that usually describes something else. Therefore, one of the greatest challenges in computational linguistics is figurative language processing, since the words or expressions used possess a meaning that is different… Show more

Figurative language is language that uses words or expressions with a meaning that is different from the literal interpretation. Figurative language is used with a meaning that is different from the basic meaning and that expresses an idea in an interesting way by using language that usually describes something else. Therefore, one of the greatest challenges in computational linguistics is figurative language processing, since the words or expressions used possess a meaning that is different from the literal interpretation. In my summer internship at NTU Singapore, I worked on analysis of various elements of figurative language. My team also developed an automatic satire detection system, we published a research paper on the works carried out in ICDM-Sentire-2016
Show less

In network science, a community is considered to be a group of nodes densely connected internally and sparsely connected externally. However, the semantic interpretation of a community is hardly studied. In this project, my team attempts to understand whether individuals in a community possess similar Personalities, Values and Ethical background. Finally, we show that Personality and Values models could be used as features to discover more accurate community structure compared to the one… Show more

In network science, a community is considered to be a group of nodes densely connected internally and sparsely connected externally. However, the semantic interpretation of a community is hardly studied. In this project, my team attempts to understand whether individuals in a community possess similar Personalities, Values and Ethical background. Finally, we show that Personality and Values models could be used as features to discover more accurate community structure compared to the one obtained from only network information. Show less

To understand how someone’s personality and intrinsic values change with geolocation and city we intend to perform several experiments, the final outcome of which will be a map to represent geo-specific values.In order to create the World map, we intend to collect data from 40 most popular cities around the world. We will also be collecting the network structure of atleast 2000 users from each city and determine the Values and Personality and checking community Variations all over the world… Show more

To understand how someone’s personality and intrinsic values change with geolocation and city we intend to perform several experiments, the final outcome of which will be a map to represent geo-specific values.In order to create the World map, we intend to collect data from 40 most popular cities around the world. We will also be collecting the network structure of atleast 2000 users from each city and determine the Values and Personality and checking community Variations all over the world. This values/ethics map would provide an overview of the kind of values & Personalities possessed by people from different regions and community structure.
Show less

According to statistics Facebook is the 2nd most popular and Twitter is the 10th most popular website now! Probably the meaning space of social-status and Facebook/Twitter status is coming closer day by day. There could be a perpetual debate on whether digital representations of us on Facebook/Twitter can capture much about human social relations, but the increasing popularity of these sites and data made urgency to develop technology to manage this information more intelligently than ever… Show more

According to statistics Facebook is the 2nd most popular and Twitter is the 10th most popular website now! Probably the meaning space of social-status and Facebook/Twitter status is coming closer day by day. There could be a perpetual debate on whether digital representations of us on Facebook/Twitter can capture much about human social relations, but the increasing popularity of these sites and data made urgency to develop technology to manage this information more intelligently than ever. With that necessity in mind the goal of my present research is to assess personality(Openness (O),Conscientiousness (C),Extraversion (E), Agreeableness (A), Neuroticism (N)called Big Five Model)of any user from his/her Facebook/Twitter interactions. Show less

Sentiment Analysis seeks to identify the opinions and viewpoints communicated in a given piece of data which is generally in the form of text. In the recent years, there have been many attempts to classify texts from various sources based on their polarity. However, a major challenge in analyzing textual data is Code-Mixing. Especially, in a multilingual country like India where about 22 official languages exist, Code-Mixing is very prominent. For example, many native languages are Code-Mixed… Show more

Sentiment Analysis seeks to identify the opinions and viewpoints communicated in a given piece of data which is generally in the form of text. In the recent years, there have been many attempts to classify texts from various sources based on their polarity. However, a major challenge in analyzing textual data is Code-Mixing. Especially, in a multilingual country like India where about 22 official languages exist, Code-Mixing is very prominent. For example, many native languages are Code-Mixed in English script. In this project, my team attempts to provide a sentiment analysis of Telugu, Tamil and Hindi social media textual content obtained from various kind of social media sources like Twitter, Facebook e.t.c. The model will classify a given text into positive, negative and neutral. Show less

Developed a Hostel portal for the University to facilitate hostelers using web2py framework and SQLite database. User Interface was developed using Java-Script, HTML and bootstrap.

Developed a prototype Robotic Arm for assisting aged and Physically Challenged people. The Arm uses Wi-Fi and Bluetooth as communication networks and works on four control mechanisms i.e., Remote, Smart Phone Tilt, Voice and Hand gesture recognition. User Interface is provided by self developed Android Application

Neural networks have been widely applied to a variety of natural language tasks. However, their application to creative language generation has been limited so far. While current state-of-the-art neural models produce unprecedented quality of jokes, they are still far from human quality. We identify two major drawbacks in existing joke generation models namely - syntactic correctness and semantic coherence and take small steps to better them. We use a pointer-generator network and multi-task… Show more

Neural networks have been widely applied to a variety of natural language tasks. However, their application to creative language generation has been limited so far. While current state-of-the-art neural models produce unprecedented quality of jokes, they are still far from human quality. We identify two major drawbacks in existing joke generation models namely - syntactic correctness and semantic coherence and take small steps to better them. We use a pointer-generator network and multi-task learning with the primary objective of predicting a word at the current time-step and secondary objective of predicting the Part-Of-Speech (POS) tag for the next time-step. Our model outperforms the state of the art on the given dataset in the domain of joke generation. Show less

State of the art text summarization models work notably well for standard news datasets like CNN/DailyMail. However, they struggle to produce reasonable results with new domains like video transcript summarization. This can be attributed to the drastic shift in sentence and summary structures. Moreover, most summarization systems do not use multimodal attributes which have proved to help produce better output quality in other similar language tasks. In this project, we develop a multimodal… Show more

State of the art text summarization models work notably well for standard news datasets like CNN/DailyMail. However, they struggle to produce reasonable results with new domains like video transcript summarization. This can be attributed to the drastic shift in sentence and summary structures. Moreover, most summarization systems do not use multimodal attributes which have proved to help produce better output quality in other similar language tasks. In this project, we develop a multimodal summarization model using the How2 instructional video dataset using two successive summarization components, an extractor that extracts key summary sentences and an abstractor that paraphrases key sentences coherently. We use RL based policy update method to fine tune these components. Our model obtains a ROUGE-L of 47.3, which is an improvement over other tried pointer-generator based models. Human evaluation shows that our model is able to produce summaries containing key information from the video transcripts using video features Show less

In recent years, aggressive posts on social media have grown substantially which has serious implications for all social media users. While many previous research works have looked at aggression detection for monolingual data, there are very few works for code-mixed data. In this paper, we address the challenge of automatically identifying aggression in code-mixed (Hinglish) social media posts. We present a novel hierarchical attention model for aggression classification. The model outperforms… Show more

In recent years, aggressive posts on social media have grown substantially which has serious implications for all social media users. While many previous research works have looked at aggression detection for monolingual data, there are very few works for code-mixed data. In this paper, we address the challenge of automatically identifying aggression in code-mixed (Hinglish) social media posts. We present a novel hierarchical attention model for aggression classification. The model outperforms the state-of-the-art for this task and also helps improve the explainability of classification. Show less

Devised a Home Automation System which validates user on the basis of face recognition using MATLAB and Raspberry Pi