Describe ICCS dataset
Browse files
README.md
CHANGED
@@ -1,3 +1,42 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# IPCC Confidence in Climate Statements
|
2 |
+
_What do LLMs know about climate? Let's find out!_
|
3 |
+
|
4 |
+
### ICCS Dataset
|
5 |
+
|
6 |
+
We introduce the **ICCS dataset (IPCC Confidence in Climate Statements)** is a novel, curated, expert-labeled, natural language dataset of 8094 statements extracted or paraphrased from the IPCC Assessment Report 6: [Working Group I report](https://www.ipcc.ch/report/ar6/wg1/), [Working Group II report](https://www.ipcc.ch/report/ar6/wg2/), and [Working Group III report](https://www.ipcc.ch/report/ar6/wg3/), respectively.
|
7 |
+
|
8 |
+
Each statement is labeled with the corresponding IPCC report source, the page number in the report PDF, and the corresponding confidence level (, along with their associated confidence levels (`low`, `medium`, `high`, or `very high`) as assessed by IPCC climate scientists based on available evidence and agreement among their peers.
|
9 |
+
|
10 |
+
### Dataset Construction
|
11 |
+
|
12 |
+
To construct the dataset, we retrieved the complete raw text from each of the three IPCC report PDFs that are available online using an open-source library [pypdf2](https://pypi.org/project/PyPDF2/). We then normalized the whitespace, tokenized the text into sentences using [NLTK](https://www.nltk.org/) , and used regex search to filter for complete sentences including a parenthetical confidence label at the end of the statement, of the form _sentence (low|medium|high|very high confidence)_. The final ICCS dataset contains 8094 labeled sentences.
|
13 |
+
|
14 |
+
From the full 8094 labeled sentences, we further selected **300 statements to form a smaller and more tractable test dataset**. We performed a random selection of sentences within each report and confidence category, with the following objectives:
|
15 |
+
- Making the test set distribution representative of the confidence class distribution in the overall train set and within each report;
|
16 |
+
- Making the breakdown between source reports representative of the number of statements from each report;
|
17 |
+
- Making sure the test set contains at least 5 sentences from each class and from each source, to ensure our results are statistically robust.
|
18 |
+
|
19 |
+
Then, we manually reviewed and cleaned each sentence in the test set to provide for a fairer assessment of model capacity.
|
20 |
+
- We removed 26 extraneous references to figures, call-outs, boxes, footnotes, or subscript typos (`CO 2');
|
21 |
+
- We split 19 compound statements with conflicting confidence sub-labels, and removed 6 extraneous mid-sentence labels of the same category as the end-of-sentence label;
|
22 |
+
- We added light context to 23 sentences, and replaced 5 sentences by others when they were meaningless outside of a longer paragraph;
|
23 |
+
- We removed qualifiers at the beginning of 29 sentences to avoid biasing classification (e.g. 'But...', 'In summary...', 'However...').
|
24 |
+
\end{itemize}
|
25 |
+
|
26 |
+
**The remaining 7794 sentences not allocated to the test split form our train split.**
|
27 |
+
|
28 |
+
Of note: while the IPCC report uses a 5 levels scale for confidence, almost no `very low confidence` statement makes it through the peer review process to the final reports, such that no statement of the form _sentence (very low confidence)_ was retrievable. Therefore, we chose to build our data set with only statements labeled as `low`, `medium`, `high` and `very high` confidence.
|
29 |
+
|
30 |
+
### Code Download
|
31 |
+
|
32 |
+
The code to reproduce dataset collection and our LLM benchmarking experiments is avalaible on [GitHub](https://github.com/rlacombe/Climate-LLMs).
|
33 |
+
|
34 |
+
### Paper ###
|
35 |
+
|
36 |
+
We use this dataset to evaluate how recent LLMs fare at classifying the scientific confidence associated with each statement in a statistically representative, carefully constructed test split of the dataset.
|
37 |
+
|
38 |
+
We show that `gpt3.5-turbo` and `gpt4` assess the correct confidence level with reasonable accuracy even in the zero-shot setting; but that, along with other language models we tested, they consistently overstate the certainty level associated with low and medium confidence labels. Models generally perform better on reports before their knowledge cutoff, and demonstrate intuitive classifications on a baseline of non-climate statements. However, we caution it is still not fully clear why these models perform well, and whether they may also pick up on linguistic cues within the climate statements and not just prior exposure to climate knowledge and/or IPCC reports.
|
39 |
+
|
40 |
+
Our results have implications for climate communications and the use of generative language models in knowledge retrieval systems. We hope the ICCS dataset provides the NLP and climate sciences communities with a valuable tool with which to evaluate and improve model performance in this critical domain of human knowledge.
|
41 |
+
|
42 |
+
Pre-print upcomping.
|