Ayush Kumar

Ayush Kumar, Vijit Malik, Jithendra Vepa

Intent Detection is one of the core tasks of dialog systems. Few-shot Intent Detection is challenging due to limited number of annotated utterances for novel classes. Generalized Few-shot intent detection is more realistic but challenging setup which aims to discriminate the joint label space of both novel intents which have few examples each and existing intents consisting of enough labeled data. Large label spaces and fewer number of shots increase the complexity of the task. In this work, we employ a simple and effective method based on Natural Language Inference that leverages the semantics in the class-label names to learn and predict the novel classes. Our method achieves state-of-the-art results on 1-shot and 5-shot intent detection task with gains ranging from 2-8\% points in F1 score on four benchmark datasets. Our method also outperforms existing approaches on a more practical setting of generalized few-shot intent detection with gains up to 20% F1 score. We show that the suggested approach performs well across single and multi domain datasets with the number of class labels from as few as 7 to as high as 150.

Ayush Kumar, Mukuntha Narayanan Sundararaman, Jithendra Vepa

Language Models (LMs) have been ubiquitously leveraged in various tasks including spoken language understanding (SLU). Spoken language requires careful understanding of speaker interactions, dialog states and speech induced multimodal behaviors to generate a meaningful representation of the conversation. In this work, we propose to dissect SLU into three representative properties:conversational (disfluency, pause, overtalk), channel (speaker-type, turn-tasks) and ASR (insertion, deletion,substitution). We probe BERT based language models (BERT, RoBERTa) trained on spoken transcripts to investigate its ability to understand multifarious properties in absence of any speech cues. Empirical results indicate that LM is surprisingly good at capturing conversational properties such as pause prediction and overtalk detection from lexical tokens. On the downsides, the LM scores low on turn-tasks and ASR errors predictions. Additionally, pre-training the LM on spoken transcripts restrain its linguistic understanding. Finally, we establish the efficacy and transferability of the mentioned properties on two benchmark datasets: Switchboard Dialog Act and Disfluency datasets.

Ayush Kumar, Jithendra Vepa

Multimodal sentiment analysis has recently gained popularity because of its relevance to social media posts, customer service calls and video blogs. In this paper, we address three aspects of multimodal sentiment analysis; 1. Cross modal interaction learning, i.e. how multiple modalities contribute to the sentiment, 2. Learning long-term dependencies in multimodal interactions and 3. Fusion of unimodal and cross modal cues. Out of these three, we find that learning cross modal interactions is beneficial for this problem. We perform experiments on two benchmark datasets, CMU Multimodal Opinion level Sentiment Intensity (CMU-MOSI) and CMU Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI) corpus. Our approach on both these tasks yields accuracies of 83.9% and 81.1% respectively, which is 1.6% and 1.34% absolute improvement over current state-of-the-art.

Ayush Kumar, Rudolf Netzel, Michael Burch et al.

We present an algorithmic and visual grouping of participants and eye-tracking metrics derived from recorded eye-tracking data. Our method utilizes two well-established visualization concepts. First, parallel coordinates are used to provide an overview of the used metrics, their interactions, and similarities, which helps select suitable metrics that describe characteristics of the eye-tracking data. Furthermore, parallel coordinates plots enable an analyst to test the effects of creating a combination of a subset of metrics resulting in a newly derived eye-tracking metric. Second, a similarity matrix visualization is used to visually represent the affine combination of metrics utilizing an algorithmic grouping of subjects that leads to distinct visual groups of similar behavior. To keep the diagrams of the matrix visualization simple and understandable, we visually encode our eye-tracking data into the cells of a similarity matrix of participants. The algorithmic grouping is performed with a clustering based on the affine combination of metrics, which is also the basis for the similarity value computation of the similarity matrix. To illustrate the usefulness of our visualization, we applied it to an eye-tracking data set involving the reading behavior of metro maps of up to 40 participants. Finally, we discuss limitations and scalability issues of the approach focusing on visual and perceptual issues.

Ayush Kumar, Teng Joon Lim

Many security issues have come to the fore with the increasingly widespread adoption of Internet-of-Things (IoT) devices. The Mirai attack on Dyn DNS service, in which vulnerable IoT devices such as IP cameras, DVRs and routers were infected and used to propagate large-scale DDoS attacks, is one of the more prominent recent examples. IoT botnets, consisting of hundreds-of-thousands of bots, are currently present ``in-the-wild'' at least and are only expected to grow in the future, with the potential to cause significant network downtimes and financial losses to network companies. We propose, therefore, to build testbeds for evaluating IoT botnets and design suitable mitigation techniques against them. A DETERlab-based IoT botnet testbed is presented in this work. The testbed is built in a secure contained environment and includes ancillary services such as DHCP, DNS as well as botnet infrastructure including CnC and scanListen/loading servers. Developing an IoT botnet testbed presented us with some unique challenges which are different from those encountered in non-IoT botnet testbeds and we highlight them in this paper. Further, we point out the important features of our testbed and illustrate some of its capabilities through experimental results.