ICANN & PAVE LabsUniversity of South Carolina
Technical lab · since 2020

ICANN Lab

Intelligent Conversational Agents and Neural Networks

The ICANN Lab builds and stress-tests the machinery of conversational information access. We model how people talk to search systems, train and probe neural architectures for dialogue and retrieval, and construct the benchmarks that make claims about these systems checkable. The lab has run since 2020 and has mentored more than thirty students at the graduate, undergraduate, and high school levels.

Conversational searchNatural language processingInformation retrievalAI evaluationComputational social science

Focus areas

  1. Conversational search and spoken dialogue

    Clarification behavior, mixed-initiative dialogue, and result presentation when there is no screen to fall back on. We study what a system should ask before it answers, and what a conversational turn costs the person having it. Methods include Wizard-of-Oz studies and controlled dialogue experiments.

  2. Neural models for language, speech, and intent

    Speech act classification, spoken intent recognition, attention architectures for task-oriented dialogue, and natural language processing for code-mixed and low-resource text.

  3. Information retrieval and search behavior modeling

    Session-based search behavior in naturalistic settings, searching as learning, the effect of task complexity on strategy, and signals of question quality in community platforms.

  4. Evaluation infrastructure for AI systems

    Benchmarks, reusable harnesses, and methods for checking whether an automated evaluation measures what it claims to measure. A result you cannot reproduce is not a result.

  5. Computational social science at scale

    Sentiment and stance analysis, misinformation detection, and sample-efficient methods for reading public opinion from large social media corpora.

Ongoing projects

  • Simulated users in system evaluation

    Whether language models can stand in for human participants in interactive user studies, and where that substitution quietly breaks down. The output is a harness other groups can reuse rather than a single result.

    In progress

  • Linking speech to retrieval

    Connecting what a person says in a spoken request to what a system should go and find, extending our earlier work on speech acts to conversational digital assistants.

    Under review

  • Affect-aware conversational agents

    Agents that register the emotional register of a user's turn and can account for how that reading shaped the response they gave.

    Funded

  • Bias mitigation in generated dialogue

    Detecting and reducing bias in the language conversational systems produce, with explanations a person can inspect. Developed jointly with PAVE.

    In development

People

Dr. Souvick “Vic” Ghosh

Director and Principal Investigator · profile

Open positions

Graduate Research Assistant

Open, Fall 2026

Doctoral Researcher

Open

Student collaborators are credited as co-authors in the publications below. See the join page for how to apply.

Publications

  • Spoken conversational search: evaluating the effect of system clarifications on user experience through a Wizard-of-Oz study

    Ghosh, S., & Shah, C. (2025). Journal of the Association for Information Science and Technology, 76(5), 819–839. DOI

  • Unmasking public sentiment: a sample efficient approach to analyzing public opinion on US aid to Ukraine

    Ghosh, S., Ghosh, S., Dewitt, N., & McCoy, D. (2025). HICSS 58, 2459–2469.

  • Exploring the economics of conversational search sessions

    Ghosh, S., Gogoi, J., & Chua, K. (2024). Aslib Journal of Information Management, 76(4), 613–628.

  • Toward connecting speech acts and search actions in conversational search tasks

    Ghosh, S., Ghosh, S., & Shah, C. (2023). ACM/IEEE JCDL, 119–131. Vannevar Bush Best Paper nomination.

  • Sentiment-aware design of human–computer interactions

    Ghosh, S. (2023). In Computational Intelligence Applications for Text and Sentiment Data Analysis, 209–224. Academic Press.

  • “Don't Downvote”: an exploration of Reddit's advice communities

    Cannon, E., Crouse, B., Ghosh, S., Rihn, A., & Chua, K. (2022). HICSS 55.

  • The sound of music: from increased personalization to therapeutic values

    Caceres, I., & Ghosh, S. (2022). ISIC: The Information Behaviour Conference. Best Paper nomination.

  • Classifying speech acts using a multi-channel deep attention network for task-oriented conversational search agents

    Ghosh, S., & Ghosh, S. (2021). ACM CHIIR.

  • Predicting question deletion and assessing question quality in social Q&A sites using weakly supervised deep neural networks

    Ghosh, S. (2021). HICSS 54.

  • “Do users need human-like conversational agents?” Exploring conversational system design using a framework of human needs

    Ghosh, S., & Ghosh, S. (2021). DESIRES, Padova, Italy.

  • “Can you search for me?” Understanding and improving user-system dialogues for complex search tasks

    Ghosh, S. (2021). Proceedings of ASIS&T, 58(1), 173–184.

  • Designing human-computer conversational systems using needs hierarchy

    Ghosh, S. (2021). School of Information Student Research Journal, 11(1), 3.

  • Identifying citation sentiment and its influence while indexing scientific papers

    Ghosh, S., & Shah, C. (2020). HICSS 53.

  • Exploring intelligent functionalities of spoken conversational search systems

    Ghosh, S. (2020). Doctoral dissertation, Rutgers University.

  • Session-based search behavior in naturalistic settings for learning-related tasks

    Ghosh, S., & Shah, C. (2019). ACM CIKM.

  • Exploring the ideal depth of neural network when predicting question deletion on community question answering

    Ghosh, S., & Ghosh, S. (2019). FIRE. Springer.

  • Searching as learning: exploring search behavior and learning outcomes in learning-related tasks

    Ghosh, S., Rath, M., & Shah, C. (2018). ACM CHIIR, 22–31.

  • Towards automatic fake news classification

    Ghosh, S., & Shah, C. (2018). Proceedings of ASIS&T, 55(1), 805–807.

  • Exploring online and offline search behavior based on varying task complexity

    Rath, M., Ghosh, S., & Shah, C. (2018). ACM CHIIR, 285–288.

  • Toward multimodal cyberbullying detection

    Singh, V. K., Ghosh, S., & Jose, C. (2017). ACM CHI Extended Abstracts, 2090–2099.

  • Sentiment identification in code-mixed social media text

    Ghosh, S., Ghosh, S., & Das, D. (2017). CICLING, Budapest.

  • Complexity metric for code-mixed social media text

    Ghosh, S., Ghosh, S., & Das, D. (2017). Computación y Sistemas, 21(4), 693–701.

  • Part-of-speech tagging of code-mixed social media text

    Ghosh, S., Ghosh, S., & Das, D. (2016). Second Workshop on Computational Approaches to Code Switching, 90–97.

  • Determining sentiment in citation text and analyzing its impact on the proposed ranking index

    Ghosh, S., Das, D., & Chakraborty, T. (2016). CICLING, 292–306. Springer.

Both labs, all years →

Join the lab

What I am looking for

I am recruiting students who want to build things and then find out whether they actually work. Prior publication is not expected, and I would rather teach a careful person a method than inherit a fast person's habits.

  • Some Python, and comfort with PyTorch or an equivalent framework
  • Patience to run an experiment three times before believing the result
  • Interest in dialogue and retrieval modeling, evaluation tooling for AI systems, or large-scale text analysis

Email with the subject line ICANN Lab inquiry, a short note on what you want to work on, your CV, and a link to any code you have written. Full details are on the join page.

souvick.ghosh@sc.edu