aaaifss2022_12.txt

CLIMATE BERT: A Pretrained Language Model for Climate-Related Text
Nicolas Webersinke,1Mathias Kraus,1Julia Anna Bingler,2Markus Leippold3
1FAU Erlangen-Nuremberg, Germany
2ETH Zurich, Switzerland
3University of Zurich, Switzerland
[email protected], [email protected], [email protected], [email protected]
Abstract
Over the recent years, large pretrained language models (LM)
have revolutionized the field of natural language processing
(NLP). However, while pretraining on general language has
been shown to work very well for common language, it has
been observed that niche language poses problems. In par-
ticular, climate-related texts include specific language that
common LMs can not represent accurately. We argue that
this shortcoming of today’s LMs limits the applicability of
modern NLP to the broad field of text processing of climate-
related texts. As a remedy, we propose C LIMATE BERT, a
transformer-based language model that is further pretrained
on over 2 million paragraphs of climate-related texts, crawled
from various sources such as common news, research arti-
cles, and climate reporting of companies. We find that C LI-
MATE BERT leads to a 48% improvement on a masked lan-
guage model objective which, in turn, leads to lowering error
rates by 3.57% to 35.71% for various climate-related down-
stream tasks like text classification, sentiment analysis, and
fact-checking.
1 Introduction
Researchers working on climate change-related topics in-
creasingly use natural language processing (NLP) to auto-
matically extract relevant information from textual data. Ex-
amples include the sentiment or specificity of language used
by companies when discussing climate risks and measuring
corporate climate change exposure, which increases trans-
parency to help the public know where we stand on climate
change (e.g., Callaghan et al. 2021; Bingler et al. 2022b).
Many studies in this domain apply traditional NLP meth-
ods, such as dictionaries, bag-of-words approaches or sim-
ple extensions thereof (e.g., Gr ¨uning 2011; Sautner et al.
2022). However, such analyses face considerable limita-
tions, since climate-related wording could vary substan-
tially by source (Kim and Kang 2018). Deep learning tech-
niques that promise higher accuracy are gradually replacing
these approaches (e.g., K ¨olbel et al. 2020; Luccioni, Baylor,
and Duchene 2020; Bingler et al. 2022a; Callaghan et al.
2021; Wang, Chillrud, and McKeown 2021; Friederich et al.
2021). Indeed, it has been shown in related domains that
Copyright c
2022, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.deep learning in NLP allows for impressive results, outper-
forming traditional methods by large margins (Varini et al.
2020).
These deep learning-based approaches make use of lan-
guage models (LMs), which are trained on large amounts
of textual and unlabelled data. This training on unlabelled
data is called pretraining and leads to the model learning
representations of words and patterns of common language.
One of the most prominent language models is called BERT
(Bidirectional Encoder Representations from Transformers)
(Devlin et al. 2018) with its successors R OBERT A(Liu et al.
2019), Transformer-XL (Dai et al. 2019) and ELECTRA
(Clark et al. 2020). These models have been trained on huge
amounts of text which was crawled from an unprecedented
amount of online resources.
After the pretraining phase, most LMs are trained on addi-
tional tasks, the downstream task . For the downstream tasks,
the LM builds on and benefits from the word representations
and language patterns learned in the pretraining phase. The
pre-training benefit is especially large on downstream tasks
for which the collection of samples is difficult and, thus, the
resulting training datasets are small (hundreds or few thou-
sands of samples). Furthermore, it has been shown that a
model that was pretrained on the downstream task-specific
text exhibits better performance, compared to a model that
has been pretrained solely on general text (Araci 2019; Lee
et al. 2020).
Hence, a straightforward extension to the standard com-
bination of pretraining is the so-called domain-adaptive pre-
training (Gururangan et al. 2020). This approach has re-
cently been studied for various tasks and basically comes in
the form of pretraining multiple times — in particular pre-
training in the language domain of the downstream task, i.e.,
pretraining (general domain)
+domain-adaptive
pretraining (downstream domain)
+training (downstream task) :
To date, regardless of the increase in using NLP for cli-
mate change related research, a model with climate domain-
adaptive pretraining has not been publicly available, yet.
Research so far rather relied on models pretrained on gen-
eral language, and fine-tuned on the downstream task. To

fill this gap, our contribution is threefold. First, we in-
troduce C LIMATE BERT, a state-of-the-art language model
that is specifically pretrained on climate-related text cor-
pora of various sources, namely news, corporate disclosures,
and scientific articles. This language model is designed to
support researchers of various disciplines in obtaining bet-
ter performing NLP models for a manifold of downstream
tasks in the climate change domain. Second, to illustrate
the strength of C LIMATE BERT, we highlight the perfor-
mance improvements using C LIMATE BERT on three stan-
dard climate-related NLP downstream tasks. Third, to fur-
ther promote research at the intersection of climate change
and NLP, we make the training code and weights of all lan-
guage models publicly available at GitHub and Hugging
Face.12
2 Background
As illustrated in Figure 1, our LM training approach for C LI-
MATE BERTcomprises all three phases — using an LM pre-
trained on a general domain, the domain-adaptive pretrain-
ing on the climate domain, and the training phase on climate-
related downstream tasks.
Pretraining on General Domain
As of 2018, pretraining became the quasi-standard for learn-
ing NLP models. First, a neural language model, often with
millions of parameters, is trained on large unlabeled corpora
in a semi-supervised fashion. By learning on multiple levels
which words/word-sequences/sentences appear in the same
context, an LM can represent a semantically similar text by
similar vectors. Typical objectives for training LMs are the
prediction of masked words or the prediction of a label indi-
cating whether two sentences occurred consecutively in the
corpora (Devlin et al. 2018).
In the earlier NLP pretraining days, LMs tradition-
ally used regular or convolutional neural networks (Col-
lobert and Weston 2008), or later Long-Short-Term-Memory
(LSTM) networks to process text (Howard and Ruder 2018).
Todays LMs mostly build on transformer models (e.g., De-
vlin et al. 2018; Dai et al. 2019; Liu et al. 2019). One of
the latter is named R OBERT A(Liu et al. 2019) which was
trained on 160GB of various English-language corpora -
data from BOOKCORPUS (Zhu et al. 2015), WIKIPEDIA,
a portion of the CCNEWS dataset (Nagel 2016), OPEN-
WEBTEXT corpus of web content extracted from URLs
shared on Reddit (Gokaslan and Cohen 2019), and a sub-
set of CommonCrawl that is said to resemble the story-like
style of WINOGRAD schemas (Trinh and Le 2019). While
these sources are valuable to build a model working on gen-
eral language, it has been shown that domain-specific, niche
language (such as climate-related text) poses a problem to
current state-of-the-art language models (Araci 2019).
Domain-Specific Pretraining
As a remedy to inferior performance of general language
models when applied to niche topics, multiple language
1www.github.com/climatebert/language-model
2www.huggingface.co/climatebertmodels have been proposed which build on the pretrained
models but continue pretraining on their respective domains.
FinBERT, LegalBert, MedBert are just a few language mod-
els that have been further pretrained on the finance, legal, or
medical domain (Araci 2019; Chalkidis et al. 2020; Rasmy
et al. 2021). In general, this domain-adaptive pretraining
yields more accurate models on downstream tasks (Guru-
rangan et al. 2020).
Domain-specific pretraining requires a decision about
which samples to include in the training process. It is still an
open debate which sample strategy improves performance
best. Various strategies can be applied to extract the text
samples on which the LM is further pretrained. For exam-
ple, while traditional pretraining uses all samples from the
pretraining corpus, similar sample selection (S IM-SELECT )
uses only a subset of the corpus, in which the samples are
similar to the samples in the downstream task (Ruder and
Plank 2017). In contrast, diverse sample selection (D IV-
SELECT ) uses a subset of the corpus, which includes dissim-
ilar samples compared to the downstream dataset (Ruder and
Plank 2017). Previous research has investigated the benefit
of these approaches, yet no final conclusion about the effi-
ciency has been obtained. Consequently, we compare these
approaches in our experiments.
NLP on Climate-Related Text
In the past, climate-related textual analysis often used pre-
defined dictionaries of presumably relevant words and then
simply searched for these words within the documents.
For example, Cody et al. (2015) use such an approach
for climate-related tweets. Similarly, Sautner et al. (2022)
use a keyword-based approach to capture firm-level climate
change exposure. However, these methods do not account
for context. The lack of context is a significant drawback,
given the ambiguity of many climate-related words such
as ”environment,” ”sustainable,” or ”climate” itself (Varini
et al. 2020).
Only recently, BERT has been used for NLP in climate-
related text. The transformers-based BERT models are ca-
pable of accounting for the context of words and have out-
performed traditional approaches by large margins across
various climate-related datasets (K ¨olbel et al. 2020; Luc-
cioni, Baylor, and Duchene 2020; Varini et al. 2020; Bin-
gler et al. 2022a; Callaghan et al. 2021; Wang, Chillrud, and
McKeown 2021; Friederich et al. 2021; Stammbach et al.
2022). However, this research has also shown that extracting
climate-related information from textual sources is a chal-
lenge, as climate change is a complex, fast-moving, and of-
ten ambiguous topic with scarce resources for popular text-
based AI tasks.
While context-based algorithms like BERT can detect
a variety of complex and implicit topic patterns in addi-
tion to many trivial cases, there remains great potential
for improvement in several directions. To our knowledge,
none of the above cited work has examined the effects of
domain-adaptive pretraining on their specific downstream
tasks. Therefore, we investigate whether domain-adaptive
pretraining will improve performance for climate change-
related downstream tasks such as text classification, senti-

News Abstracts ReportsCommon
crawlPretraining (general domain)Domain-adaptive pretraining (climate
domain)Training (downstream tasks)
+ +- Text classification
- Sentiment analysis
- Fact-checkingFigure 1: Sequence of training phases. Our main contribution is the continued pretraining of language models on the climate
domain. In addition, we evaluate the obtained climate domain-specific language models on various downstream tasks.
ment analysis, and fact-checking.
3 C LIMATE BERT
In the following, we describe our approach to train C LI-
MATE BERT. We first list the underlying data sources before
describing our sample selection techniques and, finally, the
vocabulary augmentation we used for training the language
model.
Text Corpus
Our goal was to collect a large corpus of text, C ORP, that
included general and domain-specific climate-related lan-
guage. We decided to include the following three sources:
news articles, research abstracts, and corporate climate re-
ports. We decided not to include full research articles be-
cause this language is likely too specific and does not rep-
resent general climate language. We also did not include
Twitter data, as we assume that these texts are too noisy. In
total, we collected 2,046,523 paragraphs of climate-related
text (see Table 1).
The N EWS dataset is mainly retrieved from Refinitiv
Workspace and includes 135,391 articles tagged with cli-
mate change topics such as climate politics, climate actions,
and floods and droughts. In addition, we crawled climate-
related news articles from the web.
The A BSTRACTS dataset includes abstracts of climate-
related research articles crawled from the Web of Science,
primarily published between 2000 and 2019.
The R EPORTS dataset comprises corporate climate and
sustainability reports of more than 600 companies from the
years 2015-2020 retrieved from Refinitiv Workspace and the
respective company websites.
Given the nature of the datasets, we find a large het-
erogeneity between the paragraphs in terms of number
of words. Unsurprisingly, on average, the paragraphs with
the least words come from the N EWS and the R EPORTS
datasets. In contrast, A BSTRACTS includes paragraphs with
the most words. Table 1 lists these descriptives.
To estimate the benefit from domain-adaptive pretrain-
ing, we compare the similarity of our text corpus with the
one used for pretraining R OBERT A. Following Gururangan
et al. (2020), we consider the vocabulary overlap between
both corpora. The resulting overlap of 57.05% highlights the
dissimilarity between the two domains and the need to add
specific vocabularies. Therefore, we expect to see consid-
erable performance improvements of domain-adaptive pre-
training.Dataset Num. of Avg. num. of words
paragraphs Q1 Mean Q3
News 1,025,412 34 56 65
Abstracts 530,819 165 218 260
Reports 490,292 34 65 79
Total 2,046,523 36 107 168
Table 1: Corpus C ORP used for pretraining C LIMATE BERT.
Q1 and Q3 stand for the 0.25 and 0.75 quantiles, respec-
tively.
Sample Selection
Prior work has shown that specific selections of the samples
used for pretraining can foster the performance of the LM. In
particular, incorporating information from the downstream
task by selecting similar or diverse samples has been shown
to yield favorable results compared to using all samples from
the dataset. We follow both approaches and select samples
that are similar or diverse to climate-text using our text clas-
sification task (see 5). We experiment with three different
strategies from Ruder and Plank (2017) for the selection of
samples from our corpus:
In the most traditional sample selection strategy, F ULL-
SELECT , we use all paragraphs from C ORP to train
CLIMATE BERTF.
In S IM-SELECT , we select the 70% of samples from
CORP, which are most similar to the samples of our text
classification task. We use a Euclidean similarity met-
ric for this sample selection strategy. We call this LM
CLIMATE BERTS.
In D IV-SELECT , we select the 70% of samples from
CORP, which are most diverse compared to the samples
from our text classification task. We use the sum be-
tween the type-token-ratio and the Shannon-entropy for
measuring diversity (Ruder and Plank 2017). This LM is
named C LIMATE BERTD.
In D IV-SELECT + SIM-SELECT , we use the same diver-
sity and similarity metrics as before. We then compute
a composite score by summing over their scaled values.
We keep the 70% of the samples with the highest com-
posite score to train C LIMATE BERTD+S.

Downstream domain- Downstream tasks
adaptive pretraining training
Hyperparameter Value
Batch size 2016 32
Learning rate 5e-4 5e-5
Number of epochs 12 1000
Patience — 4
Class weight — Balanced
Feedforward nonlinearity — tanh
Feedforward layer — 1
Output neurons — Task dependent
Optimizer Adam
Adam epsilon 1e-6
Adam beta weights (0.9, 0.999)
Learning rate scheduler Warmup linear
Weight decay 0.01
Table 2: Hyperparameters used for the downstream domain-
adaptive pretraining and the downstream tasks training of
CLIMATE BERT.
Vocabulary Augmentation
We extend the existing vocabulary of the original model
to include domain-specific terminology. This allows C LI-
MATE BERT to explicitly learn representations of terminol-
ogy that frequently occur in a climate-related text but not in
the general domain. In particular, we add the 235 most com-
mon tokens as new tokens to the tokenizer, thereby extend-
ing the size of the vocabulary for our basis language model
(DistilR OBERT A) from 50,265 to 50,500. See Appendix C
for a list of all added tokens. We also experimented with
language models that do not use vocabulary augmentation
or add more tokens. However, overall we find improvements
using this technique and, thus, apply it to all language mod-
els which we pretrain on the climate domain.
Model Selection
For all our experiments, we use DistilR OBERT A, a distilled
version of R OBERT Afrom Huggingface,3as our starting
point for training (Sanh et al. 2019). All our language mod-
els are trained with a masked language modeling objective
(i.e., cross-entropy loss on predicting randomly masked to-
kens). We report all hyperparameters in Table 2. The large
batch size of 2016 for training the LM is achieved using gra-
dient accumulation.
Training on Downstream Task
After pretraining DistilR OBERT Aon C ORP, we follow
standard practice (Devlin et al. 2018) and pass the final layer
[CLS] token representation to a task-specific feedforward
layer for prediction. We report all hyperparameters of this
feedforward layer in Table 2.
4 Performance Analysis of Language Model
Table 3 lists the results after pretraining DistilR OBERT Aon
CORP with various sample selection strategies. For evalu-
ation, we split C ORP randomly into 80% training data and
20% validation data. The reported loss is the cross-entropy
3www.huggingface.co/distilroberta-baseloss on predicting randomly masked tokens from the valida-
tion data. We find that C LIMATE BERTFleads to the lowest
validation loss. This performance is followed by the other
CLIMATE BERT LMs, which all show similar results. Over-
all, we find that our domain-adaptive pretraining decreases
the cross-entropy loss by 46–48% compared to the basis Dis-
tilR OBERT A, cutting the loss almost in half.
Model Val. loss
DistilR OBERT A 2.238
CLIMATE BERTF 1.157
CLIMATE BERTS 1.205
CLIMATE BERTD 1.204
CLIMATE BERTD+S 1.203
Table 3: Loss of our language models on a validation set
from our text corpus C ORP.
5 Performance Analysis for Climate-Related
Downstream Tasks
For our experiments, we used the following downstream
tasks: text classification, sentiment analysis, and fact-
checking. Table 4 lists basic statistics about the downstream
tasks. We repeated the training and evaluation phase 60
times for each experiment, each time with a random 90%
set of samples for training and the remaining 10% set for
validation.
Downstream Num. of Labels Label
task samples distribution
Text classification 1220 climate-related: yes/no 1000/220
Sentiment analysis 1000 opportunity/neutral/risk 250/408/342
Fact-checking 2745 claim: support/refute 1943/802
Table 4: Overview of our downstream tasks used for evalu-
ating C LIMATE BERT.
Text Classification
For our text classification experiment, we use a dataset con-
sisting of hand-selected paragraphs from companies’ annual
reports or sustainability reports. All paragraphs were anno-
tated as yes(climate-related) or no(not climate-related) by at
least four experts from the field using the software prodigy.4
See Appendix B for our annotation guidelines. In case of a
close verdict or a tie between the annotators, the authors of
this paper discussed the paragraph in depth before reaching
an agreement.
In the following, Table 5 reports the results of the lan-
guage models when trained on our text classification task,
i.e., whether the text is climate-related or not. Overall,
we find that all C LIMATE BERT LMs outperform a non-
pre-trained DistilR OBERT Aacross both metrics for the
text classification task. Most notably, domain-adaptive pre-
training with similar samples to our downstream tasks
4www.prodi.gy

(CLIMATE BERTS) leads to improvements of 32.64% in
terms of cross-entropy loss and a reduction in the error rate
of the F1 score by 35.71%.
Text classification
Model Loss F1
DistilR OBERT A 0:242 0:171 0:986 0:010
CLIMATE BERTF 0:191 0:136 0:989 0:010
CLIMATE BERTS 0:163 0:132 0:991 0:008
CLIMATE BERTD 0:197 0:153 0:988 0:009
CLIMATE BERTD+S0:217 0:153 0:988 0:009
Table 5: Results on our text classification task. Reported are
the average cross-entropy loss and the average weighted F1
score on the validation sets across 60 evaluation runs. Value
subscripts report the standard deviations.
Sentiment Analysis
Our next task studies the sentiment behind the climate-
related paragraphs, using the same dataset as in the previ-
ous section. In our context, we use the term ‘sentiment’ to
distinguish whether an entity reports on climate-related de-
velopments as negative risk, as positive opportunity , or as
neutral .
Therefore, we created a second labeled dataset on climate-
related sentiment, for which we asked the annotators to label
the paragraphs by one of the three categories — risk,neutral ,
oropportunity . See Appendix B for our annotation guide-
lines. Similarly, as before, in case of a close verdict or a tie
between the annotators, the authors of this paper discussed
the paragraph in depth before reaching an agreement.
Table 6 shows the performance of our models in senti-
ment prediction. Again, all C LIMATE BERTLMs outperform
the DistilR OBERT Abaseline model in terms of F1 score and
average cross-entropy loss. The largest improvements can be
observed with C LIMATE BERTF, which amount to a 7.33%
lower cross-entropy loss and a 7.42% lower error rate in
terms of average F1 score compared to the DistilR OBERT A
baseline LM.
Sentiment analysis
Model Loss F1
DistilR OBERT A 0:150 0:069 0:825 0:046
CLIMATE BERTF 0:139 0:042 0:838 0:036
CLIMATE BERTS 0:140 0:057 0:836 0:033
CLIMATE BERTD 0:138 0:043 0:835 0:040
CLIMATE BERTD+S0:139 0:043 0:834 0:036
Table 6: Results on our sentiment analysis task in terms
of average validation loss and average weighted F1 score
across 60 evaluation runs. Subscripts report the standard de-
viations.Fact-Checking
We now turn to the fact-checking downstream task. We ap-
ply our model to a dataset that was proposed by Diggelmann
et al. (2020) and comprises 1.5k sentences that make a claim
about climate-related topics. This CLIMATE -FEVER dataset
is to the best of our knowledge to date the only dataset
that focuses on climate change fact-checking. CLIMATE -
FEVER adapts the methodology of FEVER , the largest dataset
of artificially designed claims, to real-life claims on cli-
mate change collected online. The authors of CLIMATE -
FEVER find that the surprising, subtle complexity of mod-
eling real-world climate-related claims provides a valuable
challenge for general natural language understanding. Work-
ing with this dataset, Wang, Chillrud, and McKeown (2021)
recently introduced a novel semi-supervised training method
to achieve a state-of-the-art (SotA) F1 score of 0.7182 on the
fact-checking dataset CLIMATE -FEVER .
Claim : 97% consensus on human-caused
global warming has been disproven.
Evidence
REFUTE: In a 2019 CBS poll, 64% of the US
population said that climate change
is a ””crisis”” or a ””serious prob-
lem””, with 44% saying human ac-
tivity was a significant contributor.
Claim : The melting Greenland ice sheet is
already a major contributor to ris-
ing sea level and if it was eventu-
ally lost entirely, the oceans would
rise by six metres around the world,
flooding many of the world’s largest
cities.
Evidence
SUPPORT: The Greenland ice sheet occupies
about 82% of the surface of Green-
land, and if melted would cause sea
levels to rise by 7.2 metres.
Table 7: Examples taken from CLIMATE -FEVER .
Each claim in CLIMATE -FEVER is supported or refuted by
evidence sentences (see Table 7), and an evidence sentence
can also be classified as giving not enough information. The
objective of the model is to classify an evidence sentence to
support orrefute a claim. To feed this combination of claim
and evidence into the model, we concatenate the claims with
the related evidence sentences, with a [SEP] token sepa-
rating them. As in Wang, Chillrud, and McKeown (2021),
and for comparison with their results, we filter out all evi-
dence sentences with the label NOT ENOUGH INFO in the
CLIMATE -FEVER dataset.
Table 8 lists the results of our experiments on the
CLIMATE -FEVER dataset. In line with our previous exper-
iments, we find similar or better results for all C LIMATE -
BERTLMs across all metrics. Our C LIMATE BERTD+SLM
achieves similar cross-entropy loss compared to the basis
DistilR OBERT Amodel, yet pushes the average F1 score
from 0.748 to 0.757, which outperforms Wang, Chillrud, and
McKeown (2021)’s previous SotA F1 score of 0.7182, and

is hence, to the best of our knowledge, the new SotA on this
dataset.
Fact-checking
Model Loss F1
DistilR OBERT A 0:135 0:017 0:748 0:036
CLIMATE BERTF 0:134 0:020 0:755 0:037
CLIMATE BERTS 0:133 0:017 0:753 0:042
CLIMATE BERTD 0:135 0:016 0:752 0:042
CLIMATE BERTD+S0:135 0:018 0:757 0:044
Table 8: Results on our fact-checking task on CLIMATE -
FEVER in terms of average validation loss and average
weighted F1 score across 60 evaluation runs. Subscripts re-
port the standard deviations.
6 Carbon Footprint
Training deep neural networks in general and large lan-
guage models in particular, has a significant carbon footprint
already today. If the LM research trends continue, this detri-
mental climate impact will increase considerably. The topic
of efficient NLP was also discussed by a working group
appointed by the ACL Executive Committee to promote
ways that the ACL community can reduce the computational
costs of model training (https://public.ukp.informatik.tu-
darmstadt.de/enlp/Efficient-NLP-policy-document.pdf).
We acknowledge that our work is part of this trend. In
total, training C LIMATE BERTcaused 115.15 kg CO2 emis-
sions. We use two energy efficient NVIDIA RTX A5000
GPUs: 0.7 kW (power consumption of GPU server) x 350
hours (combined training time of all experiments) x 470
gCO2e/kWh (emission factor in Germany in 2018 according
to www.umweltbundesamt.de/publikationen/entwicklung-
der-spezifischen-kohlendioxid-7) = 115,149 gCO2e. We
list all details about our climate impact in Table 9 in
Appendix A. Nevertheless, we decided to carry out this
project, as we see the high potential of NLP to support
action against climate change. Given our awareness of the
carbon footprint of our research, we address this sensitive
topic as follows:
1. We specifically decided to focus on DistilR OBERT A,
which is a considerably smaller model in terms of num-
ber of parameters compared to the non-distilled version
and, thus, requires less energy to train. Moreover, we do
not crawl huge amounts of data without considering the
quality. This way, we try to take into account the issues
mentioned by Bender et al. (2021).
2. Hyperparameter tuning yields considerably higher CO2
emissions in the training stage due to tens or hundreds
of different training runs. Note that our multiple train-
ing runs on the downstream task are not causing long
training times as the downstream datasets are very small
compared to the dataset used for training the language
model. We therefore refrain from exhaustive hyperpa-
rameter tuning. Rather, we build on previous findings.We systematically experimented with a few hyperparam-
eter combinations and found that the hyperparameters
proposed by Gururangan et al. (2020) lead to the best
results.
3. We would have liked to train and run our model on
servers powered by renewable energy. This first best op-
tion was unfortunately not available. In order to speed
up the energy system transformation required to achieve
the global climate targets, we contribute our part by do-
nating Euro 100 to atmosfair. atmosfair was founded in
2005 and is supported by the German Federal Environ-
ment Agency. atmosfair offsets carbon dioxide in more
than 20 locations: from efficient cookstoves in Nigeria,
Ethiopia and India to biogas plants in Nepal and Thai-
land to solar energy in Senegal and Brazil and renewable
energies in Tansania and Indonesia. See www.atmosfair.
de/en/offset/fix/. We explicitly refrain from calling this
donation a CO2 compensation, and we refrain from a so-
lution that is based on afforestation.
7 Conclusion
We propose C LIMATE BERT, the first language model that
was pretrained on a large scale dataset of over 2 mil-
lion climate-related paragraphs. We study various selec-
tion strategies to find samples from our corpus which are
most helpful for later tasks. Our experiments reveal that
our domain-adaptive pretraining leads to considerably lower
masked language modeling loss on our climate corpus. We
further find that this improvement is also reflected in predic-
tive performance across three essential downstream climate-
related NLP tasks: text classification, the analysis of risk and
opportunity statements by corporations, and fact-checking
climate-related claims.
Acknowledgments
We are very thankful to Jan Minx and Max Callaghan from
the Mercator Research Institute on Global Commons and
Climate Change (MCC) Berlin for providing us with the
data, which is a subset of the data they used in Berrang-Ford
et al. (2021) and Callaghan et al. (2021).
References
Araci, D. 2019. Finbert: Financial sentiment analy-
sis with pre-trained language models. arXiv preprint
arXiv:1908.10063 .
Bender, E. M.; Gebru, T.; McMillan-Major, A.; and
Shmitchell, S. 2021. On the Dangers of Stochastic Par-
rots: Can Language Models Be Too Big? In Proceedings
of the 2021 ACM Conference on Fairness, Accountability,
and Transparency , 610–623.
Berrang-Ford, L.; Sietsma, A. J.; Callaghan, M.; Minx, J. C.;
Scheelbeek, P. F.; Haddaway, N. R.; Haines, A.; and Dan-
gour, A. D. 2021. Systematic mapping of global research on
climate and health: a machine learning review. The Lancet
Planetary Health , 5(8): e514–e525.

Bingler, J. A.; Kraus, M.; Leippold, M.; and Webersinke,
N. 2022a. Cheap talk and cherry-picking: What C LIMATE -
BERT has to say on corporate climate risk disclosures. Fi-
nance Research Letters , 102776.
Bingler, J. A.; Kraus, M.; Leippold, M.; and Webersinke, N.
2022b. Cheap talk in corporate climate commitments: The
role of active institutional ownership, signaling, materiality,
and sentiment. Swiss Finance Institute Research Paper .
Callaghan, M.; Schleussner, C.-F.; Nath, S.; Lejeune, Q.;
Knutson, T. R.; Reichstein, M.; Hansen, G.; Theokritoff, E.;
Andrijevic, M.; Brecha, R. J.; et al. 2021. Machine-learning-
based evidence and attribution mapping of 100,000 climate
impact studies. Nature climate change , 11(11): 966–972.
Chalkidis, I.; Fergadiotis, M.; Malakasiotis, P.; Aletras, N.;
and Androutsopoulos, I. 2020. LEGAL -BERT : The muppets
straight out of law school. arXiv preprint arXiv:2010.02559 .
Clark, K.; Luong, M.-T.; Le, Q. V .; and Manning, C. D.
2020. ELECTRA: Pre-training Text Encoders as Discrimi-
nators Rather Than Generators. In International Conference
on Learning Representations .
Cody, E. M.; Reagan, A. J.; Mitchell, L.; Dodds, P. S.; and
Danforth, C. M. 2015. Climate Change Sentiment on Twit-
ter: An Unsolicited Public Opinion Poll. PLOS ONE , 10(8):
e0136092.
Collobert, R.; and Weston, J. 2008. A Unified Architecture
for Natural Language Processing: Deep Neural Networks
with Multitask Learning. In Proceedings of the 25th Inter-
national Conference on Machine Learning , 160–167.
Dai, Z.; Yang, Z.; Yang, Y .; Carbonell, J. G.; Le, Q.; and
Salakhutdinov, R. 2019. Transformer-XL: Attentive Lan-
guage Models beyond a Fixed-Length Context. In Proceed-
ings of the 57th Annual Meeting of the Association for Com-
putational Linguistics , 2978–2988.
Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2018.
Bert: Pre-training of deep bidirectional transformers for lan-
guage understanding. arXiv preprint arXiv:1810.04805 .
Diggelmann, T.; Boyd-Graber, J.; Bulian, J.; Ciaramita, M.;
and Leippold, M. 2020. CLIMATE-FEVER: A Dataset for
Verification of Real-World Climate Claims. arXiv preprint
arXiv:2012.00614 .
Friederich, D.; Kaack, L. H.; Luccioni, A.; and Steffen,
B. 2021. Automated Identification of Climate Risk Dis-
closures in Annual Corporate Reports. arXiv preprint
arXiv:2108.01415 .
Gokaslan, A.; and Cohen, V . 2019. OpenWebText Corpus.
Gr¨uning, M. 2011. Artificial intelligence measurement of
disclosure (AIMD). European Accounting Review , 20(3):
485–519.
Gururangan, S.; Marasovi ´c, A.; Swayamdipta, S.; Lo, K.;
Beltagy, I.; Downey, D.; and Smith, N. A. 2020. Don’t Stop
Pretraining: Adapt Language Models to Domains and Tasks.
InProceedings of the 58th Annual Meeting of the Associa-
tion for Computational Linguistics , 8342–8360.
Hershcovich, D.; Webersinke, N.; Kraus, M.; Bingler, J. A.;
and Leippold, M. 2022. Towards Climate Awareness in NLP
Research. arXiv preprint arXiv:2205.05071 .Howard, J.; and Ruder, S. 2018. Universal Language Model
Fine-tuning for Text Classification. In Proceedings of the
58th Annual Meeting of the Association for Computational
Linguistics , 328–339.
Kim, D.-Y .; and Kang, S.-W. 2018. Analysis of Recognition
of Climate Changes using Word2Vec. International Journal
of Pure and Applied Mathematics , 120(6): 5793–5807.
K¨olbel, J. F.; Leippold, M.; Rillaerts, J.; and Wang, Q. 2020.
Ask BERT: How regulatory disclosure of transition and
physical climate risks affects the CDS term structure. Avail-
able at SSRN 3616324 .
Lee, J.; Yoon, W.; Kim, S.; Kim, D.; Kim, S.; So, C. H.; and
Kang, J. 2020. BioBERT: a pre-trained biomedical language
representation model for biomedical text mining. Bioinfor-
matics , 36(4): 1234–1240.
Liu, Y .; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy,
O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V . 2019.
ROBERT A: A robustly optimized bert pretraining approach.
arXiv preprint arXiv:1907.11692 .
Luccioni, A.; Baylor, E.; and Duchene, N. 2020. Analyzing
Sustainability Reports Using Natural Language Processing.
arXiv preprint arXiv:2011.08073 .
Nagel, S. 2016. CC-NEWS.
Rasmy, L.; Xiang, Y .; Xie, Z.; Tao, C.; and Zhi, D. 2021.
Med-BERT: pretrained contextualized embeddings on large-
scale structured electronic health records for disease predic-
tion. NPJ digital medicine , 4(1): 1–13.
Ruder, S.; and Plank, B. 2017. Learning to select data for
transfer learning with Bayesian Optimization. In Proceed-
ings of the 2017 Conference on Empirical Methods in Natu-
ral Language Processing , 372–382.
Sanh, V .; Debut, L.; Chaumond, J.; and Wolf, T. 2019. Dis-
tilBERT, a distilled version of BERT: smaller, faster, cheaper
and lighter. arXiv preprint arXiv:1910.01108 .
Sautner, Z.; van Lent, L.; Vilkov, G.; and Zhang, R. 2022.
Firm-level Climate Change Exposure. Journal of Finance ,
forthcoming.
Stammbach, D.; Webersinke, N.; Bingler, J. A.; Kraus, M.;
and Leippold, M. 2022. A Dataset for Detecting Real-World
Environmental Claims. arXiv preprint arXiv:2209.00507 .
Trinh, T. H.; and Le, Q. V . 2019. A Simple Method for Com-
monsense Reasoning. arXiv preprint arXiv:1806.02847 .
Varini, F. S.; Boyd-Graber, J.; Ciaramita, M.; and Leippold,
M. 2020. ClimaText: A dataset for climate change topic de-
tection. In Tackling Climate Change with Machine Learning
(Climate Change AI) workshop at NeurIPS .
Wang, G.; Chillrud, L.; and McKeown, K. 2021. Evidence
based Automatic Fact-Checking for Climate Change Misin-
formation. International Workshop on Social Sensing on The
International AAAI Conference on Web and Social Media .
Zhu, Y .; Kiros, R.; Zemel, R.; Salakhutdinov, R.; Urtasun,
R.; Torralba, A.; and Fidler, S. 2015. Aligning books and
movies: Towards story-like visual explanations by watching
movies and reading books. In Proceedings of the IEEE in-
ternational conference on computer vision , 19–27.

Appendix
A Climate Performance Model Card
Table 9 shows our climate performance model card, follow-
ing Hershcovich et al. (2022).
ClimateBert
1. Model publicly available? Yes
2. Time to train final model 48 hours
3. Time for all experiments 350 hours
4. Power of GPU and CPU 0.7 kW
5. Location for computations Germany
6. Energy mix at location 470 gCO2eq/kWh
7. CO2eq for final model 15.79 kg
8. CO2eq for all experiments 115.15 kg
9. Average CO2eq for inference per sample 0.62 mg
Table 9: Climate performance model card for ClimateBert.
B Annotation Guidelines
For our annotation procedure, we implemented the fol-
lowing general rules. The annotators had to label climate-
relevant paragraphs. If the paragraph was climate-relevant,
then they had to attach a sentiment to a paragraph. Annota-
tors were asked to apply common sense, e.g., when a given
paragraph might not provide all the context, but the context
might seem obvious. Moreover, annotators were informed
that each annotation should be a 0-1 decision. Hence, if
an annotator was 70% certain, then this was rounded up to
100%. We asked, on average, five researchers to annotate the
same tasks to obtain some measure of dispersion. In case of
a close verdict or a tie between the annotators, the authors of
this paper discussed the paragraph in depth before reaching
an agreement.
Text classification
The first task was to label climate-relevant paragraphs. The
labels are YesorNo. As a general rule, we determined that
just discussing nature/environment can be sufficient, and
mentioning clean energy, emissions, fossil fuels, etc., can
also be sufficient. It is a Yes, if the paragraph includes some
wording on a climate change or environment related topic
(including transition and litigation risks, i.e., emission mit-
igation measures, energy consumption and energy sources
etc.; and physical risks, i.e., increase in risk of floods, coastal
area exposure, storms etc.). It is a No, if the paragraph is not
related to climate policy, climate change or an environmen-
tal topic at all. For some examples, see Table 10.
Sentiment Analysis
For the sentiment analysis, annotators had to provide la-
bels as to whether a (climate change-related) paragraph talks
about a Risk or threat that negatively impacts an entity of in-
terest, i.e. a company (negative sentiment), or whether an en-
tity is referring to some Opportunity arising due to climate
change (positive sentiment). The paragraph can also make
just a Neutral statement.Label Examples
Yes Sustainability: The Group is subject
to stringent and evolving laws, reg-
ulations, standards and best prac-
tices in the area of sustainabil-
ity (comprising corporate gover-
nance, environmental management
and climate change (specifically
capping of emissions), health and
safety management and social per-
formance) which may give rise
to increased ongoing remediation
and/or other compliance costs and
may adversely affect the Group’s
business, results of operations, fi-
nancial condition and/or prospects.
Yes Scope 3: Optional scope that in-
cludes indirect emissions associ-
ated with the goods and services
supply chain produced outside the
organization. Included are emis-
sions from the transport of products
from our logistics centres to stores
(downstream) performed by exter-
nal logistics operators (air, land
and sea transport) as well as the
emissions associated with electric-
ity consumption in franchise stores.
No Risk and risk management Opera-
tional risk and compliance risk Op-
erational risk is the risk of loss re-
sulting from inadequate or failed
internal processes, people and sys-
tems, or from external events in-
cluding legal risk but excluding
strategic and reputation risk. It also
includes, among other things, tech-
nology risk, model risk and out-
sourcing risk.
Table 10: Examples for the annotation task climate
(Yes/No).
To be more precise, we consider a paragraph relating to
risk, if the paragraph mainly talks about 1) business down-
side risks, potential losses and adverse developments detri-
mental to the entity 2) and/or about negative impact of an
entity’s activities on the society/environment 3) and/or asso-
ciates specific negative adjectives to the anticipated, past or
present developments and topics covered.
We consider a paragraph relating to opportunities, if the
paragraph mainly talks about 1) business opportunities aris-
ing from mitigating climate change, from adapting to cli-
mate change etc. which might be beneficial for a specific
entity 2) and/or about positive impact of an entity’s activi-
ties on the society/environment 3) and/or associates specific

positive adjectives to the anticipated, past or present devel-
opments and topics covered.
Lastly, we consider a paragraph as neutral if it mainly
states facts and developments 1) without putting them into
positive or negative perspective for a specific entity and/or
the society and/or the environment, 2) and/or does not as-
sociate specific positive or negative adjectives to the antic-
ipated, past or present facts stated and topics covered. For
some examples, see Table 11.
C Added Tokens
’CO2’, ’emissions’, ”’, ’temperature’, ’environmental’,
’soil’, ’increase’, ’conditions’, ’potential’, ’increased’, ’ar-
eas’, ’degrees’, ’across’, ’systems’, ’emission’, ’precipi-
tation’, ’impacts’, ’compared’, ’countries’, ’sustainable’,
’provide’, ’reduction’, ’annual’, ’reduce’, ’greenhouse’,
’approach’, ’processes’, ’factors’, ’observed’, ’renewable’,
’temperatures’, ’distribution’, ’studies’, ’variability’, ’sig-
nificantly’, ’–’, ’further’, ’regions’, ’addition’, ’showed’,
’“’, ’industry’, ’consumption’, ’regional’, ’risks’, ’atmo-
spheric’, ’supply’, ’companies’, ’plants’, ’biomass’, ’elec-
tricity’, ’respectively’, ’activities’, ’communities’, ’cli-
matic’, ’solar’, ’investment’, ’spatial’, ’rainfall’, ’ ’, ’sus-
tainability’, ’costs’, ’reduced’, ’2021’, ’influence’, ’vegeta-
tion’, ’sources’, ’possible’, ’ecosystem’, ’scenarios’, ’sum-
mer’, ’drought’, ’structure’, ’economy’, ’considered’, ’var-
ious’, ’atmosphere’, ’several’, ’technologies’, ’transition’,
’assessment’, ’dioxide’, ’ocean’, ’fossil’, ’patterns’, ’waste’,
’solutions’, ’transport’, ’strategy’, ’CH4’, ’policies’, ’un-
derstanding’, ’concentration’, ’customers’, ’methane’, ’ap-
plied’, ’increases’, ’estimated’, ’flood’, ’measured’, ’ther-
mal’, ’concentrations’, ’decrease’, ’greater’, ’following’,
’proposed’, ’trends’, ’basis’, ’provides’, ’operations’, ’dif-
ferences’, ’hydrogen’, ’adaptation’, ’methods’, ’capture’,
’variation’, ’reducing’, ’N2O’, ’parameters’, ’ecosystems’,
’investigated’, ’yield’, ’strategies’, ’indicate’, ’caused’, ’dy-
namics’, ’obtained’, ’efforts’, ’coastal’, ’become’, ’agri-
cultural’, ’decreased’, ’GHG’, ’materials’, ’mainly’, ’rela-
tionship’, ’ecological’, ’benefits’, ’+/-’, ’challenges’, ’nitro-
gen’, ’forests’, ’trend’, ’estimates’, ’towards’, ’Committee’,
’seasonal’, ’developing’, ’particular’, ’importance’, ’tropi-
cal’, ’ratio’, ’2030’, ’composition’, ’employees’, ’charac-
teristics’, ’scenario’, ’measurements’, ’plans’, ’fuels’, ’in-
frastructure’, ’overall’, ’responses’, ’presented’, ’least’, ’as-
sess’, ’diversity’, ’periods’, ’delta’, ’included’, ’already’,
’targets’, ’achieve’, ’affect’, ’conducted’, ’operating’, ’pop-
ulations’, ’variations’, ’studied’, ’additional’, ’construction’,
’northern’, ’variables’, ’soils’, ’ensure’, ’recovery’, ’com-
bined’, ’decision’, ’practices’, ’however’, ’determined’, ’re-
sulting’, ’mitigation’, ’conservation’, ’estimate’, ’identify’,
’observations’, ’losses’, ’productivity’, ’agreement’, ’mon-
itoring’, ’investments’, ’pollution’, ’contribution’, ’oppor-
tunities’, ’simulations’, ’gases’, ’statements’, ’planning’,
’shares’, ’sediment’, ’flux’, ’requirements’, ’trees’, ’tempo-
ral’, ’determine’, ’southern’, ’previous’, ’integrated’, ’rel-
atively’, ’analyses’, ’means’, ’2050’, ’”’, ’uncertainty’,
’pandemic’, ’fluxes’, ’findings’, ’moisture’, ’consistent’,
’decades’, ’snow’, ’performed’, ’contribute’, ’crisis’Label Examples
Opportunity Grid & Infrastructure and Retail – today represent
the energy world of tomorrow. We rank among Eu-
rope‘s market leaders in the grid and retail busi-
ness and have leading positions in renewables. We
intend to spend a total of between Euro 6.5 bil-
lion and Euro 7.0 billion in capital throughout the
Group from 2017 to 2019.
Opportunity We want to contribute to the transition to a circu-
lar economy. The linear economy is not sustain-
able. We discard a great deal (waste and there-
fore raw materials, experience, social capital and
knowledge) and are squandering value as a result.
This is not tenable from an economic and ecolog-
ical perspective. As investor we can ‘direct’ com-
panies and with our network, our scale and our in-
fluence we can help the movement towards a cir-
cular future (creating a sustainable society) further
along.
Neutral A similar approach could be used for allocating
emissions in the fossil fuel electricity supply chain
between coal miners, transporters and generators.
We don’t invest in fossil fuel companies, but those
investors who do should account properly for their
role in the production of dangerous emissions from
burning fossil fuels.
Neutral Omissions: Emissions associated with joint ven-
tures and investments are not included in the emis-
sions disclosure as they fall outside the scope of our
operational boundary. We do not have any emis-
sions associated with heat, steam or cooling. We
are not aware of any other material sources of omis-
sions from our emissions reporting.
Risk We estimated that between 36.5 and 52.9 per cent
of loans granted to our clients are exposed to tran-
sition risks. If the regulator decides to pass am-
bitious laws to accelerate the transition towards a
low-carbon economy, carbon-intensive companies
would incur in higher costs, which may prevent
them from repaying their debt. In turn, this would
weaken our bank’s balance sheets. .
Risk American National Insurance Company recognizes
that increased claims activity resulting from catas-
trophic events, whether natural or man-made, may
result in significant losses, and that climate change
may also affect the affordability and availability of
property and casualty insurance and the pricing for
such products.
Table 11: Examples for the annotation task sentiment (Op-
portunity/Neutral/Risk).