Travel survey data are vital for urban mobility assessments—but they’re often costly and difficult to collect. Could LLMs help us synthesize such data? Our latest work, in collaboration with @anas_ant and @dpfoser, and the Best Paper Winner🥇at #SIGSPATIAL24, shows they can! 🧵👇

Work done in collaboration with @Harrod_Karlyn and @anas_ant . Read more in our paper: https://t.co/ketmTaiULJ 📄

Our geotagging results show that LLMs with access to global region databases enable precise geotagging to finer administrative levels.

Our human evaluations reveal that while 55% of the extracted data contain errors, the LLMs still produce over three times more accurate data than manual extraction methods.

We tested our approach on Rift Valley Fever data across 22+ countries, evaluating accuracy in extracting outbreak details and assigning correct geolocations. Our automatic evals show that LLMs extract more data than humans.

We propose a two-step LLM-based approach: 1️⃣ Extract data from reports. 2️⃣ Geotag the data using global admin region databases.

Epidemiological data is crucial for public health, but extracting and geotagging it from documents is challenging. Our work to be presented at #EMNLP2024’s 3rd NLP4PI Workshop, “From Text to Maps: LLM-Driven Extraction and Geotagging of Epidemiological Data”, tackles this. 🧵

More Details and analysis in the paper: https://t.co/ovvxGqrz65 Code and Data: https://t.co/YXA0SoBXFx Llama-2-trained model:

One key insight? Open source LLMs such as Llama-2, when trained even with a limited amount of actual travel data (Llama-2-trained in the plot), can generate quality synthetic travel surveys to facilitate urban mobility assessment.

We developed a robust evaluation framework to compare LLM-generated surveys with actual data across three granularities: 1) Pattern-level (overall survey metrics), 2) Trip-level (transition probability norms), and 3) Activity Chain-level (details in the table below).

Our approach uses LLMs to synthesize urban mobility data, creating scalable, cost-effective insights for smarter urban planning. By generating synthetic travel survey responses, we support comprehensive, data-driven mobility assessments!

One of THE largest personas datasets dropped in @huggingface 🤯 21 Million rows and 142GB 🤯 FinePersonas contains detailed personas for creating customized, realistic synthetic data. So now you can integrate unique persona traits into text generation apps. ------- What's a

Bayer Leverkusen fans, on fire 🔴⚫️🔥

Today, with @Tim_Dettmers, @huggingface, & @mobius_labs, we're releasing FSDP/QLoRA, a new project that lets you efficiently train very large (70b) models on a home computer with consumer gaming GPUs. 1/🧵 https://t.co/UAsWOLtn7a

The racism behind chatGPT that we aren't talking about... This year, I learned that students use chatGPT because they believe it helps them sound more respectable. And I learned that it absolutely does not work. A thread. 🧵

https://t.co/byOK0xE3Hf

To assess LLMs for geospatial reasoning, we devise an MDS-based experiment to predict a city’s location relative to other cities. Our findings indicate that the model’s performance is within an order of magnitude of what MDS could predict if we had access to actual distance.

We also prompt for geospatial “awareness”, the perception of space, using geospatial prepositions 'near,' 'close to,' and 'far from,' along with the control word 'and'. LLMs demonstrate an understanding of what “near” or “far” means. See maps visualizing the responses.

To assess geospatial knowledge encoded, we prompt to predict geo-coordinates of cities. LLMs are better at this task than older models, although predicted values are still quite far off. This discrepancy might be due to lack of enough geospatial data in pre-training datasets.

Are large language models geospatially knowledgeable? Our SIGSPATIAL paper examines the extent of geospatial knowledge encoded in LLMs, as well as their geospatial awareness and application in reasoning tasks related to geospatial data. Paper: