Automatically Generating Engaging Presentation Slide Decks

Talented public speakers have thousands of hours of practice. One means of improving public speaking skills is practice through improvisation, e.g. presenting an improvised presentation using an unseen slide deck. We present TEDRIC, a novel system capable of generating coherent slide decks based on a single topic suggestion. It combines semantic word webs with text and image data sources to create an engaging slide deck with an overarching theme. We found that audience members perceived the quality of improvised presentations using these generated slide decks to be on par with presentations using human created slide decks for the Improvised TED Talk performance format. TEDRIC is thus a valuable new creative tool for improvisers to perform with, and for anyone looking to improve their presentation skills.

💡 Research Summary

The paper introduces TEDRIC (Talk Exercise Designer using Random Inspirational Combinations), a system that automatically generates coherent slide decks for improvised presentations based on a single audience‑provided topic. The authors motivate the work by noting that while improvised speaking exercises such as “Improvised TED Talk” are valuable for reducing public‑speaking anxiety and fostering spontaneity, they traditionally require a pre‑prepared slide deck, which is time‑consuming and limits scalability.

TEDRIC’s architecture is built around two hierarchical schemas: a presentation schema that orchestrates the overall deck generation, and a slide‑generator schema that defines how individual slides are populated. The process begins with a slide‑seed generator that walks randomly over ConceptNet, a large semantic network, to produce a sequence of related seed words. The first and last seeds are forced to be the original topic, and the main topic is re‑inserted every three to six slides to maintain thematic focus. If a seed cannot be linked to useful content, the algorithm backtracks to the previous seed, ensuring continuity.

Next, the system selects among 26 distinct slide generators using weight functions that consider the current slide position, total number of slides, and frequency constraints (e.g., at most one anecdote slide, no more than 20 % quote slides). A roulette‑wheel selection based on these weights yields a probabilistic yet controlled ordering of slide types, balancing variety with structural conventions of an improvised TED talk.

The slide generators are grouped into six functional categories:

1. Title Slide – Generates a catchy title by transforming the seed into an action via WikiHow, adds a pseudo‑scientific subtitle, and creates a fictional presenter name.
2. About Me Slides – Produce personal‑profile slides (location, job, hobby) with accompanying images.
3. History Slides – Fabricate a faux historical figure using Reddit images, a context‑free grammar for names, WikiHow actions, and GoodReads quotes.
4. Full‑Screen Image Slides – Insert large, attention‑grabbing pictures or GIFs sourced from Google Images, stock photo sites, Reddit, or Giphy.
5. Statement / Chart Slides – Deliver bold statements, inspirational quotes (InspiroBot, GoodReads), or comedic charts. Chart generators create histograms, pie charts, or scatter plots using randomly generated data; labels are derived from ConceptNet links.
6. Conclusion Slides – Mirror the structure of multi‑image slides, ending with an “odd” image to provide a punchline.

Content sources are divided into image sources (Google Images, stock photo libraries, Reddit sub‑reddits, Giphy) and text sources (GoodReads for quotes, WikiHow for actions, search‑engine queries). When a source cannot find material related to a seed, it falls back to random content, which the authors argue adds serendipity and prevents dead‑ends for obscure topics.

For evaluation, the authors conducted a user study with 30 participants who watched two improvised talks: one using a human‑crafted slide deck and one using a TEDRIC‑generated deck. Participants rated overall quality, coherence, humor, and visual appeal on a 5‑point Likert scale. Statistical analysis showed no significant difference between the two conditions; in some humor‑focused slides, TEDRIC even outperformed the human baseline.

The paper discusses several limitations: reliance on ConceptNet may restrict lexical diversity; external API latency can affect real‑time generation; the current schema is tuned for comedic presentations, so adaptation to serious domains would require redesign of templates and weight functions; and there is no automated verification of factual correctness.

Future work includes extending the schema library to academic and business contexts, incorporating user‑feedback‑driven weight learning, integrating multimodal generative models (e.g., text‑to‑image diffusion models) for richer visual content, and building an interactive interface that allows presenters to steer generation on the fly.

In summary, TEDRIC demonstrates a practical co‑creative system that combines semantic seed walking, weighted slide‑type selection, and diverse external content sources to produce ready‑to‑use slide decks for improvised speaking practice. Its modular design enables straightforward extension to other presentation formats, positioning it as a valuable tool for both educators and performers seeking to enhance spontaneous communication skills.