<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Archiving and Interchange DTD v1.0 20120330//EN" "JATS-archivearticle1.dtd">
<article xmlns:xlink="http://www.w3.org/1999/xlink">
  <front>
    <journal-meta />
    <article-meta>
      <title-group>
        <article-title>Direct and Indirect Annotation with Generative AI: A Case Study into Finding Animals and Plants in Historical Text</article-title>
      </title-group>
      <contrib-group>
        <contrib contrib-type="author">
          <string-name>Arjan van Dalfsen</string-name>
          <email>j.a.vandalfsen@uu.nl</email>
          <xref ref-type="aff" rid="aff0">0</xref>
        </contrib>
        <contrib contrib-type="author">
          <string-name>FolgertKarsdorp</string-name>
          <xref ref-type="aff" rid="aff2">2</xref>
        </contrib>
        <contrib contrib-type="author">
          <string-name>Ayoub Bagheri</string-name>
          <email>a.bagheri@uu.nl</email>
          <xref ref-type="aff" rid="aff1">1</xref>
        </contrib>
        <contrib contrib-type="author">
          <string-name>DieuwertjeMentink</string-name>
          <xref ref-type="aff" rid="aff0">0</xref>
        </contrib>
        <contrib contrib-type="author">
          <string-name>Thirza vanEngelen</string-name>
          <xref ref-type="aff" rid="aff0">0</xref>
        </contrib>
        <contrib contrib-type="author">
          <string-name>Els Stronks</string-name>
          <email>e.stronks@uu.nl</email>
          <xref ref-type="aff" rid="aff0">0</xref>
        </contrib>
        <aff id="aff0">
          <label>0</label>
          <institution>Department of Language, Literature and Communication, Utrecht University</institution>
          ,
          <addr-line>Trans 10, Utrecht, 3512 JK</addr-line>
          ,
          <country country="NL">The Netherlands</country>
        </aff>
        <aff id="aff1">
          <label>1</label>
          <institution>Department of Methods and Statistics, Utrecht University</institution>
          ,
          <addr-line>Padualaan 14, 3584 CH, Utrecht</addr-line>
          ,
          <country country="NL">The Netherlands</country>
        </aff>
        <aff id="aff2">
          <label>2</label>
          <institution>KNAW Meertens Instituut</institution>
          ,
          <addr-line>Oudezijds Achterburgwal 185, 1012 DK Amsterdam</addr-line>
          ,
          <country country="NL">The Netherlands</country>
        </aff>
      </contrib-group>
      <fpage>1053</fpage>
      <lpage>1074</lpage>
      <abstract>
        <p>This study explores the use of generative AI (GenAI) for annotation in the humanities, comparing direct and indirect annotation approaches with human annotations. Direct annotation involves using GenAI to annotate the entire corpus, while indirect annotation uses GenAI to create training data for a specialized model. The research investigates zero-shot and few-shot methods for direct annotation, alongside an indirect approach incorporating active learning, few-shotting, and k-NN example retrieval. The task focuses on identifying words (also referred to as entities) related to plants and animals in Early Modern Dutch texts. Results show that indirect annotation outperforms zero-shot direct annotation in mimicking human annotations. However, with just a few examples, direct annotation catches up, achieving similar performance to indirect annotation. Analysis of confusion matrices reveals that GenAI annotators make similar types of mistakes, such as confusing parts and products or failing to identify entities, which are broader than those made by humans. Manual error analysis indicates that each annotation method (human, direct, and indirect) has some unique errors. Given the limited scale of this study, it is worthwhile to further explore the relative afordances of direct and indirect GenAI annotation methods.</p>
      </abstract>
      <kwd-group>
        <kwd>eol&gt;large language models</kwd>
        <kwd>natural language processing</kwd>
        <kwd>historical text</kwd>
        <kwd>token classification</kwd>
        <kwd>environmental humanities</kwd>
      </kwd-group>
    </article-meta>
  </front>
  <body>
    <sec id="sec-1">
      <title>1. Introduction</title>
      <p>
        The introduction of advanced generative AI (GenAI) models has sparked interest among
humanities scholars in leveraging these tools to extract structured information from t3e,x2ts3,[
        <xref ref-type="bibr" rid="ref16 ref2 ref22 ref4 ref5">2, 12, 25, 18, 19, 4, 5</xref>
        ]. So far, the use of GenAI in the humanities has primarily involved “direct
annotation”, where GenAI analyzes a corpus without further interference. This approach has
shown promise, potentially “supercharging the humanities1”1[].
      </p>
      <p>
        Researchers in Natural Language Processing (NLP) have proposed an alternative “indirect
annotation” framework. This two-step process involves GenAI generating training data, which
is then used to train a specialized model. This approach ofers potential cost and performance
advantages over direct annotation3[
        <xref ref-type="bibr" rid="ref1 ref28">1, 32</xref>
        ]. However, indirect annotation’s efectiveness has
primarily been demonstrated on languages well-represented in GenAI training data, raising
questions about its applicability to texts from smaller languages or historical variants often
encountered in humanities research.
      </p>
      <p>As a first exploration into its usability in the humanities, our study tests GenAI as an indirect
annotator for nature-entities in historical Dutch texts. We employ the LLMaAA (Large
Language Models as Active Annotators) framework32[], which combines few-shotting, k-Nearest
Neighbors (k-NN) example retrieval, and active learning. Our research compares the
performance of indirect annotation, i.e. an LLMaAA-derived model, against both human annotations
and direct annotation with GenAI. We find that our proposed method of indirect GenAI
annotation performs better than fully-unsupervised direct GenAI annotation. However, we also find
that providing direct annotation with demonstrations (i.e., examples of annotations) results in
similar performance. Moreover, our study reveals that humans, direct GenAI annotators, and
indirect GenAI annotators each have unique weaknesses and strengths.</p>
      <p>This study is structured as follows: We first examine the broader context of using GenAI for
annotation in humanities research. We then provide an overview of current research on direct
and indirect annotators in NLP. Subsequently, we introduce our specific use-case: identifying
animals and plants in historical texts. Finally, we detail our methodology for comparing the
performances of human annotation, direct GenAI annotation, and indirect GenAI annotation.</p>
    </sec>
    <sec id="sec-2">
      <title>2. Related Work</title>
      <sec id="sec-2-1">
        <title>2.1. GenAI annotations in humanities</title>
        <p>
          In the humanities, research on GenAI annotation has primarily focused on direct annotation
experiments. Studies have compared GenAI methods with traditional approaches and human
annotators across various tasks, including sentiment analysi2s, [
          <xref ref-type="bibr" rid="ref22 ref5">25, 5</xref>
          ], topic detection [
          <xref ref-type="bibr" rid="ref16">18</xref>
          ],
and text classification [ 19]. Findings generally suggest that while GenAI often outperforms
dictionary-based methods, it typically falls short of specialized models. However, Ka1r2j]us [
reports human-level annotations by GenAI across diverse tasks and languages, proposing a
machine-assisted mixed methods approach. These studies underscore the potential of GenAI
in humanities research, while also highlighting the need to explore both direct and indirect
annotation approaches to fully leverage its capabilities.
        </p>
      </sec>
      <sec id="sec-2-2">
        <title>2.2. GenAI as direct annotators in NLP</title>
        <p>
          Direct GenAI annotation involves prompting GenAI to annotate a dataset for immediate use.
Studies assessing this approach have found that while GenAI generally lags behind
state-of-theart models 9[
          <xref ref-type="bibr" rid="ref14 ref21 ref29">, 16, 24, 33</xref>
          ], it often equals or outperforms crowd-worker3s0[
          <xref ref-type="bibr" rid="ref29 ref9">, 33, 10</xref>
          ]. Challenges
in direct GenAI annotation include difÏculties with long-tail target types, irrelevant context,
and specific tasks like sequence tagging [
          <xref ref-type="bibr" rid="ref14 ref21">16, 24</xref>
          ]. These limitations have led to the exploration
of indirect annotation methods, which aim to address these shortcomings by integrating GenAI
in a more targeted manner.
        </p>
      </sec>
      <sec id="sec-2-3">
        <title>2.3. GenAI as indirect annotator in NLP</title>
        <p>In what we coin the indirect GenAI annotation framework, GenAI is not employed to perform
the entire annotation task on a given dataset. Instead, it is used to annotate a specific subset of
the dataset. This annotated subset is then subjected to further fine tuning by another model,
such as a BERT model.</p>
        <p>
          Wang et al. [
          <xref ref-type="bibr" rid="ref27">31</xref>
          ] found models trained on GenAI-annotated data equal to human-annotated
models and outperforming direct use of GenAI. Ding et a7l]. [largely echo this but also
highlight a practical problem when it comes to textual analysis: GenAI is good at finding entities,
but oftentimes struggles with defining the boundaries of these entities. Li et al. [
          <xref ref-type="bibr" rid="ref17">20</xref>
          ]
propose a CoAnnotating framework, in which GenAI output-uncertainty is measured and
annotations with the highest uncertainty (i.e., a lack of result robustness when confronted with small
prompt perturbations) are sent to a human annotator. While they report promising results,
there is the disadvantage of higher costs. With Large Language Models as Active Annotators
(LLMaAA) by Zhang et al. [
          <xref ref-type="bibr" rid="ref28">32</xref>
          ], the idea is to use active annotation to make the downstream
specific model better. It includes:
• Few-shotting: putting exemplary annotations in the prompt for the GenAI (this helps
        </p>
        <p>
          GenAI to annotate [
          <xref ref-type="bibr" rid="ref18">21</xref>
          ]);
• k-NN example retrieval: sequence embeddings of the text to annotate and the examples
are used to select the examples closest to the new text for few-shotting;
• Training cycles: doing step-by-step training, where first a specific model is trained, new
data is annotated by the GenAI, and the specific model is trained again;
• Active learning: selecting examples for indirect annotation for which the current model
struggle;
• Automatic reweighting: assigning learnable weights to the annotated training
samples [27] (this makes it possible to reduce the impact of noisy labeling by GenAI).
The authors test their method for NER and Information Extraction (modern Chinese and
modern English) and find that the resulting model strongly outperforms zero-shot direct annotation
with GenAI. In comparison to few-shot GenAI annotation (with k-NN optimized examples),
LLMaAA shows a marginal performance advantage, besides the obvious advantages when it
comes to robustness, costs, and speed, making it promising for humanities research.
        </p>
      </sec>
      <sec id="sec-2-4">
        <title>2.4. Plants and animals</title>
        <p>
          In this study, we research the detection of plants and animals in historical texts, which can be
seen as the traditional NLP task of token (in sequence) classification or NER. Roughly starting
with publications asMan and the Natural World [
          <xref ref-type="bibr" rid="ref25">29</xref>
          ] and The Animal Estate [28], humanities’
scholarly interest in nature has skyrocketed. This is only natural, considering a widely shared
sense of humanity being in environmental, ecological, and climate crises. For cultural
historians, the main question has been how humans observed, interpreted, and represented and thus
perceived nature 2[
          <xref ref-type="bibr" rid="ref2">2</xref>
          ]. Research on this topic has virtually exclusively been done qualitatively.
Although this approach provides illuminative insights, a qualitative approach will always work
with a relatively narrow scope because of the sheer size of the historical record, thus resulting
in a less precise large-scale overview of the studied phenomenon. It has, therefore, great
potential to complement qualitative studies with quantitative research.
        </p>
      </sec>
    </sec>
    <sec id="sec-3">
      <title>3. Methods</title>
      <p>In this section, we describe the methodology employed to compare direct and indirect GenAI
annotation strategies for identifying plants and animals in historical Dutch texts. First we
explain the data parsing used to prepare our dataset. Following this, we describe the annotation
procedure undertaken by human annotators. Then, we address the dataset creation.
Subsequently, we describe the token classification. After this, we describe our prompts. Then, we
document the training process of the indirect annotation models. Finally, we outline the ways
in which we compare the annotation approaches.</p>
      <sec id="sec-3-1">
        <title>3.1. Data parsing</title>
        <p>
          Our study used theDigitale Bibliotheek voor de Nederlandse Letteren (DBNL) [6], comprising
about 1,500 diverse Dutch texts. After preprocessing, the corpus yielded approximately 7
million unique sentences 1[
          <xref ref-type="bibr" rid="ref6 ref7">7, 8</xref>
          ].
        </p>
      </sec>
      <sec id="sec-3-2">
        <title>3.2. Annotation procedure</title>
        <p>
          Two texts from the 1750s were selected for manual annotation by two Early Modern Dutch
literature experts. 200 sentences, parsed to have a minimal length of 10 and a maximal length
of 100 words, were annotated using theINCEpTION tool [
          <xref ref-type="bibr" rid="ref13">15</xref>
          ] (cf. Fig. 1), following iteratively
developed guidelines (AppendixA). The annotation schema tagged entities on three levels:
Category (Plants/Animals), Type (Organism/Part/Product/Collective), and Usage
(Literal/Symbolical/Petrified). For example, in “The bear grabbed an apple with its claw”, “bear” would be
tagged as Animals-Organisms-Literal.
        </p>
      </sec>
      <sec id="sec-3-3">
        <title>3.3. Task description</title>
        <p>The annotated sentences were split into demonstration, validation, and test sets. We
conceptualized the detection of animals and plants as a token classification task. Prompts for both
direct and indirect annotators included the annotation schema, with technical details omitted
to improve performance (full prompts in AppendBi,xmodel settings in AppendixC).</p>
      </sec>
      <sec id="sec-3-4">
        <title>3.4. Training indirect annotation models</title>
        <p>For indirect annotation, we adapted the LLMaAA framework by Zhang et al3.2[],
integrating it with Huggingface and OpenAI ecosystems. We used GysBERT 1[] for historical Dutch,
applied k-NN few-shot selection (with the paraphrase-multilingual-mpnet-base-v2 sentence
transformer 2[6]), and confidence-based active learning. Automatic reweighting was not
included. GPT4o served as the LLM backbone. To address the scarcity of plant and animal
entities, we employed a pre-filtering strategy using GPT-3.5. The specialized model underwent 10
training rounds with 10 epochs each, adding 50 example sentences per round (25 pre-filtered
sentences + 25 sentences with lowest confidence). This process was repeated for two datasets,
using five distinct random seeds, resulting in 10 indirect models.</p>
        <p>It is important to point out that the indirect annotators have not seen any of the human
annotations in their training regime. However, the 500 sentences they are trained on are
annotated with the help of the human-annotated demonstration set and the performance of the
model is determined by its score on a human-annotated validation set.</p>
      </sec>
      <sec id="sec-3-5">
        <title>3.5. Comparing annotation strategies</title>
        <p>All of the analyses below were done on the annotations of the various strategies on the held-out
test set. Note that the sentences in this set came from the same documents as the demonstration
and validation.</p>
        <p>To assess the inter-annotator agreement among human annotators, direct GenAI, and
indirect GenAI approaches, we conducted inter-annotator agreement analysis, confusion matrix
analyses, and performed a manual error analysis. We compare each human annotator’s results
against each automatic system’s output:
1. Human Annotations
• Human1: Annotations from the first human annotator
• Human2: Annotations from the second human annotator
2. Direct GenAI Annotations
• Direct zero-shot: Zero-shot direct annotation (without examples)
• Direct few-shot 1: Few-shot direct annotation using examples from Human1
• Direct few-shot 2: few-shot direct annotation using examples from Human2</p>
        <sec id="sec-3-5-1">
          <title>3. Indirect GenAI Annotations</title>
          <p>• Indirect1: Indirect annotation using examples from Human1
• Indirect2: Indirect annotation using examples from Human2</p>
          <p>For the inter-annotator agreement, positives-only weighted F1 is used as a metric. The
positives-only weighted F1 is the weighted average of all harmonious means of precision and
recall of labeled entities. Thus, words that were not labeled as referring to a plants or animals
related word, which is by far the most common category in this situation, are disregarded. For
all GenAI annotations (direct and indirect), predictions were done five times. The average and
standard deviation of the inter-annotator agreements were calculated from these iterations.
The observed low variance across models and approaches suggests that these results are likely
stable, despite the relatively small number of iterations.</p>
          <p>The confusion matrices were made by comparing the human annotations to single instances
of the other strategies. It should be noted thaHt uman1 had labeled one example as
”Nonelabel” (a token was tagged but no label was chosen), which was removed later for the process
of making the confusion matrices. In addition to the confusion matrix analysis, we performed
a manual error analysis on the annotations for the held-out test set.</p>
        </sec>
      </sec>
    </sec>
    <sec id="sec-4">
      <title>4. Results</title>
      <sec id="sec-4-1">
        <title>4.1. Inter-annotator agreement</title>
        <p>1. Annotators of the same type resemble each other best (e.g., Human1 is closest to
Human2).
2. Zero-shot direct annotation demonstrates lower internal coherence (F1 = 0.74) compared
to few-shot direct (F1 = 0.9 and 0.88) and indirect annotation (F1 = 0.81 and 0.86).
3. Direct zero-shot annotation consistently underperforms, while annotations from the
other human annotator achieve the highest agreement.
4. Few-shot direct and indirect annotations perform similarly, falling between zero-shot
and human performance.
5. GenAI models don’t simply mimic the specific human annotator they were trained on,
but generalize from the examples.</p>
        <p>These results suggest that indirect annotation and few-shot direct annotation are more
reliable methods for replicating human-like annotations compared to zero-shot approaches. The
choice between these methods may depend on factors beyond performance, such as ease of
implementation or specific task requirements.
Method
human1
human2
direct zero-shot
direct few-shot1
direct few-shot2
indirect1
indirect2</p>
      </sec>
      <sec id="sec-4-2">
        <title>4.2. Confusion matrices</title>
        <p>While F1 scores provide an overall measure of performance, they do not ofer insights into
how well the annotation methods perform f oinrdividual labels. To gain a more nuanced
understanding of the agreement (or lack thereof) between the annotations, we turn to confusion
matrices (Fig.2), which provide insights into individual label performance across annotation
methods:
1. Human2’s annotations most closely align with Human1’s.
2. Zero-shot direct annotation shows low recall, with many entities remaining unlabeled.
3. All GenAI annotators struggle with both precision and recall:
• Precision errors: confusion between “Animals Parts Literal” and “Animals Products</p>
        <p>Literal”.</p>
        <p>• Recall errors: suggesting “No Label” for entities labeled by Human1.
4. GenAI methods, including indirect approach, produce labels not found by human
annotators, demonstrating higher label diversity.</p>
        <p>These patterns highlight the strengths and weaknesses of each annotation method,
emphasizing the need for careful selection and potential combination of approaches in annotation tasks.</p>
      </sec>
      <sec id="sec-4-3">
        <title>4.3. Manual Error Analysis</title>
        <p>Our manual examination of the annotations (AppendiDx) reveals distinct error patterns across
diferent annotation strategies. First, human annotators occasionally overlook words in a
sentence. For instance, in sentence 10, Human1 tagged the wordvleesch (meat) only twice out
of its three occurrences. Such errors likely stem from simple oversight rather than
misunderstanding (although misunderstandings occur too). Second, few-shot direct annotation
strategies sometimes struggle with entity aggregation. A notable example is sentence 41, where
(a) Human1 vs. Human2
(c) Human1 vs. Direct Few-Shot1
(d) Human1 vs. Direct Few-Shot2
(e) Human1 vs. Indirect1
(f) Human1 vs. Indirect2
“nek van het Varken” (neck of the pig) is incorrectly labeled as a single entity, instead of
recognizing “neck” and “pig” as separate entities (with distinct labels “Animals Products Literal”
and “Animals Organisms Literal”, respectively). Importantly, these errors are incidental, not
systematic, suggesting they are unlikely to be consistently repeated. Third and finally, while
indirect annotation models are (once trained) deterministic, and therefore not susceptibilne-to
cidental mistakes, they can produce counter-intuitive systematic errors. A revealing example
is sentence 31, where “salt” is misclassified as “Plant Product Literal”. This error likely stems
from the proximity of salt to spices like pepper and nutmeg in the transformer model’s vector
space.</p>
      </sec>
    </sec>
    <sec id="sec-5">
      <title>5. Discussion</title>
      <p>This study compared direct and indirect annotation with GenAI in a humanities context,
focusing on identifying plant and animal-related words in Early Modern Dutch texts. While we
studied a specific case, we believe that this method can be used for a wide range of
applications. Our findings reveal both the potential and limitations of various annotation strategies
for historical humanities studies.</p>
      <p>Indirect annotation demonstrates clear advantages over fully-unsupervised zero-shot direct
annotation, particularly in terms of recall. However, few-shot direct annotation achieves
comparable performance to indirect annotation, suggesting that both approaches have merit in
diferent contexts. Based on these results, we advise against using zero-shot direct
annotations for historical humanities research. Its significantly lower recall compared to the
alternatives means that many relevant entities are likely to be missed, potentially skewing research
outcomes. The choice between few-shot direct annotation and indirect annotation is less
clearcut, as both display similar F1 scores. Here, time, cost, and technical considerations should be
considered.</p>
      <p>The unique error patterns suggest two important points. First, it is crucial to investigate
shortcomings of chosen methods on a micro-level to be aware of specific pitfalls. Second,
there’s potential for stacking annotation methods: human, direct, and indirect annotation can
be applied to the same texts, after which points of contention can be analyzed. In this way,
they may bundle their strengths and cover each other’s weaknesses.</p>
      <p>Regarding the generalizability of this explorative study, several points should be noted.
Token labeling is a specific task, and the behavior of direct and indirect GenAI annotators may
difer for tasks of another nature. The prompts used for direct annotation have not been
systematically tested, and it’s possible that especially zero-shot direct annotation would have better
results with more guidance regarding the output format. The held-out test set was small and
from the same document (i.e., not the same data) as the training data, which might have
influenced the results. The training of the indirect annotation model has been done with just 500
examples, a typically low number for fine-tuning its underlying transformer model.
Additionally, during the training of the indirect model, automatic reweighting was not applied (as we
deemed its efects in the LLMaAA paper to be marginal), but integrating it might approve the
model.</p>
    </sec>
    <sec id="sec-6">
      <title>6. Conclusion</title>
      <p>Despite these limitations, this study shows potential for applying GenAI as indirect
annotators in humanities research. However, there are are notable diferences to other annotation
strategies. Future research should address questions about indirect GenAI annotation’s
performance on other tasks (e.g., text classification), the impact of prompt optimizing frameworks
(e.g., DSPy [13, 14]), and the potential of combining human and GenAI annotations to check
on each other.</p>
    </sec>
    <sec id="sec-7">
      <title>Author Contributions</title>
      <p>Conceptualization: Arjan van Dalfsen, Folgert Karsdorp, Ayoub Bagheri, Els Stronks; Data
Curation: Thirza van Engelen, Dieuwertje Mentink; Investigation: Arjan van Dalfsen;
Methodology: Arjan van Dalfsen; Writing - Original Draft: Arjan van Dalfsen; Writing - Review &amp;
Editing: Folgert Karsdorp, Ayoub Bagheri, Els Stronks; Visualization: Arjan van Dalfsen;
Supervision: Folgert Karsdorp, Ayoub Bagheri, Els Stronks.</p>
    </sec>
    <sec id="sec-8">
      <title>Acknowledgments</title>
      <p>This research would not have been possible without the financial support from Utrecht
University AI Labs, the Meertens Instituut, and the Utrecht University focus area Advanced Data
Science. Their generous contributions provided the necessary resources to conduct this study.
Additionally, we would like to extend our gratitude to SURF for providing cloud computing
services, which were instrumental in the analysis and processing of our data.</p>
    </sec>
    <sec id="sec-9">
      <title>7. Appendices</title>
    </sec>
    <sec id="sec-10">
      <title>A. Annotation Guidelines</title>
      <sec id="sec-10-1">
        <title>A.1. Annotation Schema</title>
        <sec id="sec-10-1-1">
          <title>Category</title>
          <p>• Animals: A living thing that can move around to search for food. It usually has ways to
see, hear, smell, taste, and feel the world around it.
• Plants: A living thing that usually stays in one place. It creates its own food using
sunlight, water, and air.</p>
          <p>Type
• Organisms: A whole, living animal or plant. Think of it like one complete cat, or one
whole oak tree.
• Parts: A piece of an animal or plant. Things like a bird’s wing, a flower petal, or a bear’s
claw.
• Products: Something we get from a plant or animal that we use. Only first-order products
count (i.e. it’s the first “product” that comes from the plant/animal, not a product of an
earlier product). Examples are milk from a cow, honey from bees, or apples from a tree.
• Collective: Something is collective if the word refers to a heterogenous multitude of
plants/animals. Nature explicitly and inherently is a prominent part of, but it is not
100% clear what kinds of nature. If the collective might belong to both categories (you
choose the best or least-wrong category). Examples are: weide, grastapijt, bos, woud,
vee, kudde.</p>
        </sec>
        <sec id="sec-10-1-2">
          <title>Usage</title>
          <p>• Literal: When the word means exactly the animal, plant, part, or product itself. If you
envision the text, you should see it. (”The bear ate a fish.”)
• Symbolical: When the word is used as a symbol or metaphor, representing something
else. If you envision the text, you should not see it. (”His heart was as cold as a snake.”)
Pictures are symbolic. Nicknames are probably symbolical.</p>
          <p>• Petrified: if the plants/animals word is the name of something or someone.</p>
        </sec>
      </sec>
      <sec id="sec-10-2">
        <title>A.2. Technicalities</title>
        <p>General rule Textual context is always dominant in annotating.</p>
        <p>Discontinuous annotations Sometimes an annotation is discontinuous, meaning that there
are words between the parts to be annotated. An example is: “esschen- en pijnhout”.
Here, “esschen-” and “hout” should be annotated. This can be done by annotating both
(here, meaning that you should also make a separate annotation for hout!) and then
defining a relationship. Step-by-step guide: 1. Annotate “esschen-” and “pijnhout” and
“hout”. 2. Select the first part (“esschen-“). 3. Right click the second part (“hout”). 4.</p>
        <p>Click “Link to” and select “discontinuous entity”.</p>
        <p>Part – Whole constructions Sometimes part-whole constructions occur, e.g. “de wortel van
de brem”. Here, it is important to look at the parts that are separately referential (wortel
and brem, here). If the text has “bremwortel” there is just one separate referential entity.
Syntactic head Concerning compound words, we annotate based on the syntactic head. You
can find the syntactic head by doing a reference test: to what pard of the compound can
you refer? (“hazenpad”: not annotated; “padhaas”: annotated). In Dutch, the syntactic
head is normally on the right side of the word.</p>
        <p>Co-references Co-references to entities are not tagged. (in “de wolf is blij, hij eet graag
haas”, “hij” should not be annotated). Likewise, words that in a specific context refer
to plants/animals should not be annotated, unless the plant/animal aspect is inherent.
“Veulen” and “kalf” are names for young animals and should be annotated, however,
“jong”, “wijfje”, “mannetje”, “wederhelft”, “lichaam”, are not.</p>
        <p>Adjectives As a general rule, adjectives are not annotated. There are a few exceptions: 1. If
the adjective is part of the name of a plant/animal, it should be annotated (e.g. “blauwe”
in “blauwe vinvis” and “kruipende” in “kruipende boterbloem”). 2. Sometimes a word
looks like an adjective, but it is used as a substantive. In that case, annotate it.
Foreign languages When plants/animals/nature-locations are in a non-Dutch language, they
should still be annotated. There are two exceptions: 1. If the whole text is in a diferent
language, it should not be annotated; 2. If the entities name is in a non-Latin script (e.g.,
Arabic, Greek, Hebrew), it shouldn’t be annotated.</p>
      </sec>
    </sec>
    <sec id="sec-11">
      <title>B. Used Prompts</title>
      <sec id="sec-11-1">
        <title>A ChatPromptTemplate is used. Therefore, ”chat messages” are provided between</title>
        <p>parentheses; inside the parentheses the ”sender” of the message and the message itself
are divided with a comma. Parts in italics dependent on the texts that is annotated.</p>
      </sec>
      <sec id="sec-11-2">
        <title>Here, it is only indicated that these parts exist.</title>
        <sec id="sec-11-2-1">
          <title>B.1. Pre-filtering prompt</title>
          <p>(
System,
You are a helpful assistant. You'll get a historical Dutch text.
It's your task to tell whether (non-human)
animals or plants are directly present in this text. You do this
by reasoning step by step, and then end by completing: 'I deem
the statement that literal animals are present in this text to
be:' with True or False. I know you can do it!
),
(
User,
\textit{Text to pre-filter}
)</p>
        </sec>
        <sec id="sec-11-2-2">
          <title>B.2. Few-shot direct annotation prompt</title>
          <p>(
System,
You are a highly intelligent and accurate nature domain information extraction
system. I'll provide a small text, written in historical Dutch. Your task is
to recognize and extract all entities related to plants or animals. If you
have found anything that falls into that category, you should annotate it on
three levels: 1. Category; 2. Type; 3. Usage.</p>
          <p>For Category, there are two possibilities: Plants and Animals.
* Animals: A living thing that can move around to search for food. It usually
has ways to see, hear, smell, taste, and feel the world around it.
* Plants: A living thing that usually stays in one place. It creates its own
food using sunlight, water, and air.</p>
          <p>For Type, there are four possibilities: Organisms, Parts, Products, Collective.
* Organisms: A whole, living animal or plant. Think of it like one complete cat,
or one whole oak tree.
* Parts: A piece of an animal or plant. Things like a bird's wing, a flower
petal, or a bear's claw.
* Products: Something we get from a plant or animal that we use. Only
first-order products count (i.e. it's the first 'product' that comes from the
plant/animal, not a product of an earlier product). Examples are milk from a
cow, honey from bees, or apples from a tree.
* Collective: Something is collective if the word refers to a heterogeneous
multitude of plants/animals. Nature explicitly and inherently is a prominent
part of, but it is not 100\% clear what kinds of nature. If the collective
might belong to both categories (you choose the best or least-wrong category).</p>
          <p>Examples are: weide, grastapijt, bos, woud, vee, kudde.</p>
          <p>For Usage there are three possibilities: Literal, Symbolical, Petrified.
* Literal: When the word means exactly the animal, plant, part, or product
itself. If you envision the text, you should see it. ('The bear ate a fish.')
* Symbolical: When the word is used as a symbol or metaphor, representing
something else. If you envision the text, you should not see it. ('His heart
was as cold as a snake.') Pictures are symbolic. Nicknames are probably
symbolical.
* Petrified: if the plants/animals word is the name of something or someone.
To summarize, you should detect all plant and animal related words and tag them
according to this schema. So, for each found entity you annotate its category
(Plant/Animal), its Type (Organisms/Parts/Products/Collective), and its Usage
(Literal/Symbolical/Petrified).</p>
          <p>It is extremely important that you work precise. Therefore, you should explain
step by step why you make a choice. Also extremely important: the annotation you
do should be in the form of a list with dictionaries. You should also do an
explanation, but your ultimate annotation should be in that format. So you
should always have output like this:
Very important: if you don't find any entities, your annotation should be an
empty dictionary in a list:
otherwise the postprocess script will get in trouble.</p>
          <p>Good luck, I count on you!
)
(
System,</p>
          <p>The span must be exactly the same as in the original text, including white
spaces.
Please now annotate the following input:
Input: \textit{Text to annotate.}
)</p>
        </sec>
        <sec id="sec-11-2-3">
          <title>B.3. Zero-shot direct annotation prompt</title>
          <p>Same prompt as above, but without the examples.</p>
        </sec>
      </sec>
    </sec>
    <sec id="sec-12">
      <title>C. Model Settings</title>
      <p>OpenAI API parameters temperature = 1; top_p = 1; frequency_penalty = 0;
presence_penalty = 0; gpt-4-o version: gpt-4o-2024-05-13. gpt-3.5-turbo version:
gpt-3.5-turbo-0125. ‘2023-03-15-preview’
GysBERT parameters archictecture: BertForTokenClassification; optimizer: Adam;
learning_rate: 2e-5.</p>
      <p>D. Held-out Test Set
1. Doch in kommerlyke tyden word dit kruid, een weinig geroost, door de menschen ten
spyze gebruikt.
2. Al de gedroogde visch, die zich toen op het eiland bevond, werd daar van geheel zwart
en onbruikbaar, ja in de twe naastvolgende jaren werden door die assche, of veeleer door
de ’er mede vermengde scherpachtige rotsbrokjes of zand, gelyk boven by den brand op
Jan Mayen eiland aangemerkt is, zo verre het
3. Als men het Varken, in ’t midden aan weeder zyden van de rugge-graad,
doorgesneden heeft, zo laat men ieder helft, even onder de schouwder nog eens doorsnyden in de
breedte.
4. Het gerookt vleesch moet ook acht dagen in het zout leggen, en dan in zakken genaait
in de rook gehangen worden, en moet drie of wel vier maanden rooken.
5. Zouten van Spek, Hammen en Ossen-Vleesch, hoe daar mede te handelen.
6. Dan legt men alles aldus in de kuip om in order te gebruiken: 1. de 6 klapstukken
van de buik onder in, want ze konnen het langste duuren: 2. de staartstukken: 3. de
schouwderbladeren: 4. de twee borststukken: 5. de twee beste ribben: 6. de vier
andere ribben: de twee ongeschikte ribben die by de schouwders zitten: 7. de huspot zo
men wil boven op; maar men moet zorg dragen dat de stukken wel vast in malkanderen
sluiten, en de openingen moeten met zout gevuld worden, en wat zout ’er boven op, en
eerst onder op den bodem gespreid; ook moet de kuip eerst schoon uitgebroeid en met
kruidnagels gedroogt worden.
7. Dit alles te zaamen in een groote pan of styfsel-kom of hakkebord gedaan, en 6 tinne
kommetjes met Osse-vleesch-nat of ander vleesch-nat, of warm water daar op gegooten
en digt toegedekt en altemets eens omgeroert. en zo een nagt over, op de warme plaat
laaten staan weken; en dan stopt men ze gelyk Leverbeulingen; dog om dat de gort sterk
zwelt maar half vol, en dan zynze half vol als men ze plat duuwt: Als ze gestopt zyn laat
men ze zeer zagt kooken dat het water maar even beweegt omtrent een half uurtje, en
men prikt ze ondertussen met een doorntje om niet te barsten en uittekooken, en dan
zyn ze heel goed. 4.
8. En schoon veele staande houden dat het eeten van dit vleesch geen quaad aan de
menschen doedt, zo zyn fatsoendelyke lieden nogtans beschroomd om het te gebruiken: om
dit met zeekerheid te weeten zo kan men daar deeze proeve van neemen.
9. §. XXXI. De Koemelk word tot artzeny gebruikt. De Melk is de voornaamste artzeny
der Yslanders, en word daarom ook, zodra zy van de koe koomt, door gene anderen, dan
alleen kranken, genoten.
10. Die het beter willen maken, en ’er de middelen toe hebben, kopen een weinig zout,
snyden, als het ge-slagt dier noch onafgehakt hangt, op drie of vier plaatsen een diepe snede
in het vleesch, en doen in iedere opening een kleine hand vol zout, zich verbeeldende, dat
het dus zelf, zo veel nodig is, door het gantsche beest trekt, en het vleesch, wanneer ’er
vervolgens wind en rook by koomt, zeer wel bewaard word Op de beide gezegde wyzen
handelen de ingezetenen ook met het schapenvleesch, als zy het voor hun huisgezin
slagten.
11. Zeeusche Pens en Hoofdvleesch, hoe men die maaken zal.
12. Ossen en Koeyen vallen niet groter dan het kleinst geestvee in Duitsland; hebben, gelyk
bereids gezegt is, gene Hoornen, en genieten alleen het voorrecht, door de huis lieden
in den winter mede onder ’t dak genomen en met het zo kommerlyk gewonnen hooy,
of, by mangel van het zelve, met het gedroogd zeegewas Zeenestel spaarzaam gevoed te
worden.
13. Men stopt de beulingen maar half vol om dat die anders te ligt uitkooken of barsten; en
men bind ze met een touwtje onder en boven toe, en dan wordenze op een schootel plat
nedergelegt, tot dat ze gekookt worden: Voor al moet men niet vergeeten genoeg vet
daar in te doen, want anders zyn de Leverbeulingen te droog.
14. weshalven de boeren ’er aldaar meer acht op geven. Dezen jagen alleen de Hamels in ’t
gebergte; doch houden de Oyen zo veel by huis, als doenlyk is.
15. Als men zo veel moeiten niet doen wil om Rolpens en Hoofd-vleesch te maaken, zo snyd
men de pens in stukken, en men kookt het met de kop tot dat alles gaar is, en dan legt
men het vleesch met de pens door een, met wat zout en heele peper, in den azyn, in een
keulse aarde pot, is heel goed om met appelen des winters gebakken te eeten. 17.
16. De Boter kaarnen de meesten voor en na zo hairig, als zy uit ongereinigde melk in een
zamengenaaide schapenvacht gemolken is, en leggen dezelve dus op; weshalven een
vreemdeling die Boter niet ligtelyk door de keel zoude konnen krygen.
17. Dan neemt men een groote vleesch keetel en men hangt ze vol regen water over het
vuur, en als het water kookt doet men de beulingen daar in, dat die regt uit en niet op
malkanderen leggen, daarom mag men niet meer als anderhalf douzyn beulingen te gelyk
kooken; en ze moeten heel zagtjes kooken, omtrent een half uur lang.
18. Afhakken van ’t vleesch in de Slacht-tyd, en hoe men de stukken best en ten meesten
voordeelen zal gebruiken, en hoe men verder met alles in de Slacht-tyd, moet handelen.
1.
19. Neemt by de 20 ponden, gehakt redelyk vet, varkens vleesch, anderhalf loot of twee
loot nootemuscaten; twee loot nagelen; twee loot zwarte peeper, dit alles ter deegen fyn
gestooten zynde, zo roert men het onder anderhalf vierendeel zout, en men kneed het
door het gekapte Varkensvleesch heen; en men laat het zo een nacht met een schoone
doek bedekt staan doortrekken.
20. Hunne vellen vallen in den winter, als zy het meeste en vastste hair hebben, het best;
weshalven de Yslanders dezelve dan naarstig vangen, en wel, uit aangebore afschuuw
van schietgeweer, met uitgezette netten of vangyzers, die gelyk een kleermakersschaar
gevormt, en met een dood lam ten lokaas voorzien zyn.
21. geweld der uitbrekende en uitgezette lucht een groot gedeelte van den berg, ’t geen te
zwaar was, om opgeligt te worden, op zyde en niet slegts een gantsche myl wegs langs het
eiland tot aan het strand, maar zelfs noch een myl verr’ in zee voortgeschoven, en aldaar
neder gezet wierd, alwaar het, onaangezien de diepte, in den beginne wel 60 vademen
boven het water uitstak, en aldaar merendeels noch staat e.
22. Neemt voor het vleesch, het geen men daar in legt, het vleesch van de schouwder van
een Os dat het malste is; of anders een van de platte billen.
23. Ja zy zyn het zelven, die gemeenlyk het begin der aardbranden veroorzaken.
24. §. XXXIV. Hebben geen Zwynen, maar wel Honden en Katten.
25. Doch wat de eigentlyke en natuurlyke oorzaak dezer zeldzaamheid zyn mag, is niet zeer
ligt te besefen w.
26. Reusel, hoe men die wel zal smelten.
27. Van harde of Coraalachtige Zeegewassen wist myn berichter te zeggen, dat enigen van
dezelven op de gronden gevonden wierden; doch konde hen niet noemen of beschryven,
nadien hy, volgens zyne eigen belydenis, ’er nooit naar gezien had.
28. Dezen zyn de Snoriper op de lappische Alpen, die zich a steeds op het land houden, meer
lopen dan vliegen, en mitsdien niet bezwaarlyk te vangen zyn.
29. Men moet zich verwonderen, wat zy konnen uitstaan; doch zy worden wel degelyk door
de ongemakken verhard, nadien zy jaar uit jaar in in het open veld onder den bloten
Hemel blyven, en ’s winters onder de sneeuw zowel, als ’s zomers, hun voeder zelven
moeten zoeken, waar toe zy alleen de weldaad van de natuur genieten, dat zy met
byzondere styve, lange en dikke hairen, allermeest tegen den wintertyd, bedekt zyn.
30. Vervol-gens bragt men het zieke volk aan land, ’t geen, ofschoon het, behalven enig
Lepelblad, niet als Zuring in warme Melk en een weinig Schapenvleesch nuttigde, nochtans
velen binnen acht en de anderen binnen veertien dagen zo fris en gezond werden, dat zy
huppelden en sprongen, en in minder dan vier weken na hun komst weder scheep gaan,
zelven hun anker lichten, en die lange en bezwaarlyke reize voorts vrolyk voleinden
konden.
31. Het vleesch snyd men eerst aan stukken als Ossekarbenaden; en dan snyd men het aan
lange reepen omtrent een vinger dik en vierkant; men snyd het vet ook aan zulke
langwerpige stukken; en dan bestrooid men de pens met wat geprepareerd zout en kruit, gelyk
ik boven gezegt heb.
32. Men neemt 3 loot bruine peper, en een halfvierendeel nagelen; dit te zaamen eerst fyn
gestoten en in een aarde schootel gedaan, en een hand vol gedroogde Saly, die men op
den haart wat te droogen legt en die klein gewreven is, en een hand vol of vier zout daar
onder geroert, tot men denkt dat men genoeg zal hebben; want den een doet het wel wat
hartiger dan den ander.
33. Men behoefd ’er geen Sukade nog Amandelen in te doen als men niet wil, en is evenwel
goed maar zo lekker niet.
34. De Harsten laat men een dag of vyf in het zout leggen en men moet ze niet te groot laaten
hakken, om dat ze anders te ongeschikt zyn, en ieder een doet dit naa de groote van zyn
huisgezin, ook zyn die Harsten dus zeer goed om in den Oven gezet en gebraaden te
worden.
35. Het vleesch om in de Kuip in te zouten, daar neemt men toe de zes klapstukken van
de buyk, de twee staartstukken, de schouwderbladeren, de twee borststukken, de vier
andere ribben, als men de twee beste ribben wil in de rook hangen, anders kan men ook
de Paterstukken inzouten, en dan nog de twee ongeschikte ribben die by de schouders
zitten; en men laat die stukken groot of klein hakken naa dat men het wil hebben en het
huisgezin groot is.
36. Mitsdien ziet men zelden op Ysland andere, dan uitgebrande bergen, aan en om welke
men bequaam de werkingen en overgebleven tekenen van een vorigen brand bespeuren
kan.
37. Buiten dien tyd leggen de inwoonders, nadien de Vossen de schapen zeer schadelyk
zyn, kraanogen (nuces vomicae) in honig geweekt, die zy, anders niets zoets te eten
bekomende, zeer begerig inzwelgen.
38. Neemt 4 kop Gort schoon afgewasschen: 4 pond korenten die wel verlezen en schoon
gewassen zyn: 8 loot gestoote kaneel: 1 loot gestoote nagelen: 3 loot gestoote
notemuscaaten; 1/2 pond poeijer-zuiker: 1 pond gepelde amandelen in stukjes gesneden: 6 sukade
schellen aan stukjes gesneden: Een hand vol zout: 10 pond of daar omtrent Osse-niervet
aan dobbelsteentjes gesneden.
39. Het zoude gezwellen verwekken, en, als men ’er veel van eet, sterk openende zyn.
40. de Ravens verjagen; doch het Lam, vermits het, zyn voeder niet konnende zoeken,
elendig omkomen moet, slagten, en het het zachte vel afstropen, ’t geen de peltery geeft,
die in Denmarken en Holstein onder den naam van Schmaaskin of Schmaasken x verkogt
en zeer veel door lieden van een middelbaar vermogen gedragen word.
41. Neemt een van de grootste Kalfskoppen, en reinigt die, en wascht ze vier of vyfmalen
ter degen schoon af, en laatze een nacht in schoon regen water staan te trekken, dat ’er
de slym en het bloedige wel schoon af is, en hangt de kop met schoon regen-water over
het vuur; en doet ook in de keetel, de nek van het Varken, en de twee ooren met wat
veel zwoort dat ’er genoeg is om het vleesch in het hakkebord van booven en onderen te
bedekken; en als men te veel zwoort en niet genoeg vleesch heeft, zo doet men ’er wel
een of twee van de vleesigste stukken van het varken by, en men laat het te zaamen een
uur of drie kooken, na dat men het alvorens wel schoon geschuimt heeft, en het moet
zeer gaar zyn tot dat het vleesch van de beenen af valt, en dan schept men het uit op een
aarde vergiettest of doorslag.
42. Hunne manier, om het Rundvee te slagten, heeft ook iets byzonders, Zy kollen het niet
voor den kop, menende, dat daar door het bloed in ’t vleesch stremt, en mitsdien niet
lopen kan; maar steken het een dun penmes diep in den nek, waar door het ter aarde
valt; als dan trekken zy de poten gezwind met strikken zamen, en openen de keel, op
dat al het bloed zoude uitvlieten Het ingewand word door de Yslanders allereerst, zonder
veel te reinigen, genuttigt, en het dier zelf afgehakt.
43. Neemt de Lever van het Varken en wascht die schoon, en laat die op een aarde schotel
leggen; doet daar zo raauw de vellen en spieren met een mes ter degen schoon uit, en
doet het in een schoon tobbetje.
44. Voor een geheele pens heeft men omtrent 20 pond vleesch noodig, behalven het vet dat
men daar by gebruikt, dat nog omtrent 10 ponden is.
45. Laat dan een ketel of twee met regenwater kooken en laat het Koud worden; en als het
koud is neemt dan schaars drie kommetjes van dat water tegen ruim een kommetje wyn
azyn, en mengt dat te zaamen onder malkanderen zo veel tot dat de pens, als he daar over
gegooten is, kan onderleggen, en zet ze dan zo open weg daar ze niet te vogtig staan, is
heel goed om met appelen gebakken, of gestooft met wyn te eeten. 10.
46. Saucysen of Worst van Varkenvleesch, hoe men die maaken zal.
47. De stukken worden niet met zout gewreven, maar slegts twemaal door zeewater gehaalt,
en dan in de lucht, op dat zy winddroog zouden worden, en vervolgens in hunne hutten
over hunne haardsteden gehangen, om dezelve te roken, en te meer te doen drogen Dus
behandelen zy hun geslagt half verrot en half stinkend vleesch, tot zy het voorts opeten.
48. Het vleesch om in de rook te hangen daar toe neemt men de Paterstukken, de twee andere
platte billen; en de twee beste ribben; en de spieren, die achter tusschen de beenen van
de ribben inzitten, moeten daar schoon uitgedaan worden, om dat daar door ligt verderf
ontstaan kan.
49. Dan doet men twee geraspte nootemuscaten, met wat gestoote foelie en met wat zout
daar in, en men hakt het te zaamen onder een tot het redelyk klein, maar niet al te klein
is.
50. Alsdan begeeft een harder zich met de afgerichte honden op een heuvel, en geeft met zyn
hoorn een teken, waarop de honden zich verdelen, en de Schapen van alle kanten uit de
klippen en wildernissen in een zekere omtuining of staketzel dryven, ’t geen vooraan
wyd uitgezet is; doch, op dat zy niet zouden konnen ontvluchten, naar achter allengs
enger word.
In humanities, costs are often an important consideration. For all strategies, there’s the cost of
establishing annotation guidelines and making something of a test set. After that point:
• Human annotation costs €0.7 per sentence;
• Direct annotation costs €0.007 per sentence (for GPT-4o, directly via OpenAI), with
zeroshotting being slightly cheaper due to omitting examples in the prompt;
• Indirect annotation costs €4.50 to train the model, and nothing per sentence.</p>
      <p>It is important to emphasize that costs per strategy are likely to change over time since
GenAI models are getting cheaper and that human annotations might difer significantly per
country or institution. Also, multiple factors should be considered, such as available hardware
and environmental efects.</p>
    </sec>
    <sec id="sec-13">
      <title>F. Online Resources</title>
      <p>Code and data used in this study can be found here:
• Data: https://www.dbnl.org/letterkunde/pd/index.ph,p
• Code and annotationsGitHub Repositor y.</p>
    </sec>
  </body>
  <back>
    <ref-list>
      <ref id="ref1">
        <mixed-citation>
          [1]
          <string-name>
            <given-names>E. M.</given-names>
            <surname>Arevalo</surname>
          </string-name>
          and
          <string-name>
            <given-names>L.</given-names>
            <surname>Fonteyn</surname>
          </string-name>
          . “
          <article-title>Non-Parametric Word Sense Disambiguation for Historical Languages”</article-title>
          .
          <source>In:Proceedings of the 2nd International Workshop on Natural Language Processing for Digital Humanities. Taipei, Taiwan</source>
          ,
          <year>2022</year>
          . url: https://aclanthology.
          <source>org/20 22.nlp4dh-1</source>
          .
          <fpage>16</fpage>
          .
        </mixed-citation>
      </ref>
      <ref id="ref2">
        <mixed-citation>
          [2]
          <string-name>
            <given-names>J.</given-names>
            <surname>Borst</surname>
          </string-name>
          ,
          <string-name>
            <given-names>J.</given-names>
            <surname>Klähn</surname>
          </string-name>
          , and
          <string-name>
            <given-names>M.</given-names>
            <surname>Burghardt</surname>
          </string-name>
          . “
          <article-title>Death of the Dictionary?- The Rise of ZeroShot Sentiment Classification”</article-title>
          .
          <source>In: Computational Humanities Research Conference (CHR)</source>
          . Paris, France,
          <year>2023</year>
          , pp.
          <fpage>303</fpage>
          -
          <lpage>319</lpage>
          . url:https://ceur-ws.
          <source>org/</source>
          Vol-
          <volume>3558</volume>
          /paper3130.pd.f
        </mixed-citation>
      </ref>
      <ref id="ref3">
        <mixed-citation>
          [3]
          <string-name>
            <surname>T. B. Brown</surname>
            ,
            <given-names>B.</given-names>
          </string-name>
          <string-name>
            <surname>Mann</surname>
            ,
            <given-names>N.</given-names>
          </string-name>
          <string-name>
            <surname>Ryder</surname>
            ,
            <given-names>M.</given-names>
          </string-name>
          <string-name>
            <surname>Subbiah</surname>
            ,
            <given-names>J.</given-names>
          </string-name>
          <string-name>
            <surname>Kaplan</surname>
            ,
            <given-names>P.</given-names>
          </string-name>
          <string-name>
            <surname>Dhariwal</surname>
            ,
            <given-names>A.</given-names>
          </string-name>
          <string-name>
            <surname>Neelakantan</surname>
            ,
            <given-names>P.</given-names>
          </string-name>
          <string-name>
            <surname>Shyam</surname>
            ,
            <given-names>G.</given-names>
          </string-name>
          <string-name>
            <surname>Sastry</surname>
            ,
            <given-names>A.</given-names>
          </string-name>
          <string-name>
            <surname>Askell</surname>
            ,
            <given-names>S.</given-names>
          </string-name>
          <string-name>
            <surname>Agarwal</surname>
            ,
            <given-names>A.</given-names>
          </string-name>
          <string-name>
            <surname>Herbert-Voss</surname>
            , G. Krueger,
            <given-names>T.</given-names>
          </string-name>
          <string-name>
            <surname>Henighan</surname>
            ,
            <given-names>R.</given-names>
          </string-name>
          <string-name>
            <surname>Child</surname>
            ,
            <given-names>A.</given-names>
          </string-name>
          <string-name>
            <surname>Ramesh</surname>
            ,
            <given-names>D. M.</given-names>
          </string-name>
          <string-name>
            <surname>Ziegler</surname>
            ,
            <given-names>J.</given-names>
          </string-name>
          <string-name>
            <surname>Wu</surname>
            ,
            <given-names>C.</given-names>
          </string-name>
          <string-name>
            <surname>Winter</surname>
            ,
            <given-names>C.</given-names>
          </string-name>
          <string-name>
            <surname>Hesse</surname>
            ,
            <given-names>M.</given-names>
          </string-name>
          <string-name>
            <surname>Chen</surname>
            , E. Sigler,
            <given-names>M.</given-names>
          </string-name>
          <string-name>
            <surname>Litwin</surname>
            ,
            <given-names>S.</given-names>
          </string-name>
          <string-name>
            <surname>Gray</surname>
            ,
            <given-names>B.</given-names>
          </string-name>
          <string-name>
            <surname>Chess</surname>
            ,
            <given-names>J.</given-names>
          </string-name>
          <string-name>
            <surname>Clark</surname>
            ,
            <given-names>C.</given-names>
          </string-name>
          <string-name>
            <surname>Berner</surname>
            ,
            <given-names>S.</given-names>
          </string-name>
          <string-name>
            <surname>McCandlish</surname>
            ,
            <given-names>A.</given-names>
          </string-name>
          <string-name>
            <surname>Radford</surname>
            ,
            <given-names>I. Sutskever</given-names>
          </string-name>
          , and
          <string-name>
            <given-names>D.</given-names>
            <surname>Amodei</surname>
          </string-name>
          .
          <article-title>Language Models are Few-Shot Learners</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2 005.14165.
          <year>2020</year>
          . doi:
          <volume>10</volume>
          .48550/arXiv.
          <year>2005</year>
          .
          <volume>14165</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref4">
        <mixed-citation>
          [4]
          <string-name>
            <given-names>Y.</given-names>
            <surname>Chen</surname>
          </string-name>
          ,
          <string-name>
            <given-names>S.</given-names>
            <surname>Li</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Li</surname>
          </string-name>
          , and
          <string-name>
            <given-names>M.</given-names>
            <surname>Atari</surname>
          </string-name>
          .
          <article-title>Surveying the Dead Minds: Historical-Psychological Text Analysis with Contextualized Construct Representation (CCR) for Classical Chinese</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2403.0050 9.
          <year>2024</year>
          . doi: https://doi.org/10.48550/arXiv.240 3.00509.
        </mixed-citation>
      </ref>
      <ref id="ref5">
        <mixed-citation>
          [5]
          <string-name>
            <given-names>T.</given-names>
            <surname>Dejaeghere</surname>
          </string-name>
          ,
          <string-name>
            <given-names>P.</given-names>
            <surname>Singh</surname>
          </string-name>
          ,
          <string-name>
            <given-names>E.</given-names>
            <surname>Lefever</surname>
          </string-name>
          , and
          <string-name>
            <given-names>J.</given-names>
            <surname>Birkholz</surname>
          </string-name>
          . “
          <article-title>Exploring Aspect-Based Sentiment Analysis Methodologies for Literary-Historical Research PurposesP”.rIonc:eedings of the Third Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA) LREC-COLING-2024</article-title>
          . Torino, Italia: ELRA and
          <string-name>
            <surname>ICCL</surname>
          </string-name>
          ,
          <year>2024</year>
          . urlh:ttps://acla nthology.
          <source>org/2024.lt4hala-1. 1</source>
          .
          <issue>6</issue>
          [6]
          <string-name>
            <given-names>Digitale</given-names>
            <surname>Bibliotheek voor de Nederlandse</surname>
          </string-name>
          <article-title>Letteren (DBNCLo).llectie publiek domein</article-title>
          . htt ps://www.dbnl.org/letterkunde/pd/index.
          <source>ph.2p023.</source>
        </mixed-citation>
      </ref>
      <ref id="ref6">
        <mixed-citation>
          [7]
          <string-name>
            <given-names>B.</given-names>
            <surname>Ding</surname>
          </string-name>
          ,
          <string-name>
            <given-names>C.</given-names>
            <surname>Qin</surname>
          </string-name>
          , L. Liu,
          <string-name>
            <given-names>Y. K.</given-names>
            <surname>Chia</surname>
          </string-name>
          ,
          <string-name>
            <given-names>B.</given-names>
            <surname>Li</surname>
          </string-name>
          ,
          <string-name>
            <given-names>S.</given-names>
            <surname>Joty</surname>
          </string-name>
          , and
          <string-name>
            <given-names>L.</given-names>
            <surname>Bing</surname>
          </string-name>
          . “
          <article-title>Is GPT-3 a Good Data Annotator?” In:Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</article-title>
          . Toronto, Canada,
          <year>2023</year>
          . doi:
          <volume>10</volume>
          .18653/v1/
          <year>2023</year>
          .
          <article-title>acl-lo ng</article-title>
          .
          <volume>626</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref7">
        <mixed-citation>
          [8]
          <string-name>
            <surname>M. van Gompel.</surname>
          </string-name>
          <article-title>python-ucto [computer software]</article-title>
          . https://languagemachines.github.io/u cto/.
          <year>2023</year>
          .
        </mixed-citation>
      </ref>
      <ref id="ref8">
        <mixed-citation>
          [9]
          <string-name>
            <given-names>R.</given-names>
            <surname>Han</surname>
          </string-name>
          ,
          <string-name>
            <surname>T</surname>
          </string-name>
          . Peng,
          <string-name>
            <given-names>C.</given-names>
            <surname>Yang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>B.</given-names>
            <surname>Wang</surname>
          </string-name>
          , L. Liu, and
          <string-name>
            <given-names>X.</given-names>
            <surname>Wan</surname>
          </string-name>
          .
          <article-title>Is Information Extraction Solved by ChatGPT? An Analysis of Performance, Evaluation Criteria, Robustness and Errors</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2305.1445 0.
          <year>2023</year>
          . doi: https://doi.org/10.48550/arXiv.230 5.01445.
        </mixed-citation>
      </ref>
      <ref id="ref9">
        <mixed-citation>
          [10]
          <string-name>
            <given-names>X.</given-names>
            <surname>He</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Z.</given-names>
            <surname>Lin</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Gong</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.-L.</given-names>
            <surname>Jin</surname>
          </string-name>
          ,
          <string-name>
            <given-names>H.</given-names>
            <surname>Zhang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>C.</given-names>
            <surname>Lin</surname>
          </string-name>
          ,
          <string-name>
            <given-names>J.</given-names>
            <surname>Jiao</surname>
          </string-name>
          ,
          <string-name>
            <given-names>S. M.</given-names>
            <surname>Yiu</surname>
          </string-name>
          ,
          <string-name>
            <given-names>N.</given-names>
            <surname>Duan</surname>
          </string-name>
          , and
          <string-name>
            <given-names>W.</given-names>
            <surname>Chen. AnnoLLM: Making Large Language</surname>
          </string-name>
          <article-title>Models to Be Better Crowdsourced Annotators</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2303.1685 4.
          <year>2024</year>
          . doi: https://doi.org/10.48550/arXiv.230 3.16854.
        </mixed-citation>
      </ref>
      <ref id="ref10">
        <mixed-citation>
          [11] [12] [13] [14]
          <string-name>
            <given-names>A.</given-names>
            <surname>Karjus</surname>
          </string-name>
          .
          <article-title>Large language models to supercharge humanities and cultural analytics research</article-title>
          . Poster presentation at CHR2023https://2023.
          <article-title>computational-humanities-researc h</article-title>
          .org/programme./
          <year>2023</year>
          .
        </mixed-citation>
      </ref>
      <ref id="ref11">
        <mixed-citation>
          <string-name>
            <given-names>A.</given-names>
            <surname>Karjus</surname>
          </string-name>
          .
          <article-title>Machine-assisted mixed methods: augmenting humanities and social sciences with artificial intelligence</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2309.1437 9.
          <year>2023</year>
          . doi: https://doi.org/10.48550/arXiv.2309.1437.9
          <string-name>
            <given-names>O.</given-names>
            <surname>Khattab</surname>
          </string-name>
          ,
          <string-name>
            <given-names>K.</given-names>
            <surname>Santhanam</surname>
          </string-name>
          ,
          <string-name>
            <given-names>X. L.</given-names>
            <surname>Li</surname>
          </string-name>
          ,
          <string-name>
            <given-names>D.</given-names>
            <surname>Hall</surname>
          </string-name>
          , percy Liang,
          <string-name>
            <given-names>C.</given-names>
            <surname>Potts</surname>
          </string-name>
          , and
          <string-name>
            <given-names>M.</given-names>
            <surname>Zaharia</surname>
          </string-name>
          .
          <article-title>Demonstrate-Search-Predict: Composing Retrieval and Language Models for KnowledgeIntensive NLP</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2212.1402 4.
          <year>2022</year>
          . doi: https://do i.
          <source>org/10</source>
          .48550/arXiv.2212.14024.
        </mixed-citation>
      </ref>
      <ref id="ref12">
        <mixed-citation>
          <string-name>
            <given-names>O.</given-names>
            <surname>Khattab</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Singhvi</surname>
          </string-name>
          ,
          <string-name>
            <given-names>P.</given-names>
            <surname>Maheshwari</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Z.</given-names>
            <surname>Zhang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>K.</given-names>
            <surname>Santhanam</surname>
          </string-name>
          ,
          <string-name>
            <given-names>S.</given-names>
            <surname>Vardhamanan</surname>
          </string-name>
          ,
          <string-name>
            <given-names>S.</given-names>
            <surname>Haq</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Sharma</surname>
          </string-name>
          , T. T. Joshi,
          <string-name>
            <given-names>H.</given-names>
            <surname>Moazam</surname>
          </string-name>
          ,
          <string-name>
            <given-names>H.</given-names>
            <surname>Miller</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Zaharia</surname>
          </string-name>
          , and
          <string-name>
            <given-names>C.</given-names>
            <surname>PottDs.SPy: Compiling Declarative Language Model</surname>
          </string-name>
          <article-title>Calls into Self-Improving Pipelines</article-title>
          . arXiv preprinthttps://ar xiv.
          <source>org/abs/2310.03714</source>
          .
          <year>2023</year>
          . doi: https://doi.org/10.48550/arXiv.2310.0371.4
        </mixed-citation>
      </ref>
      <ref id="ref13">
        <mixed-citation>
          [15]
          <string-name>
            <surname>J.-C. Klie</surname>
            ,
            <given-names>M.</given-names>
          </string-name>
          <string-name>
            <surname>Bugert</surname>
            ,
            <given-names>B.</given-names>
          </string-name>
          <string-name>
            <surname>Boullosa</surname>
            , R. E. de Castilho,
            <given-names>and I. Gurevych.</given-names>
          </string-name>
          “
          <article-title>The INCEpTION Platform: Machine-Assisted and Knowledge-Oriented Interactive Annotation”</article-title>
          .
          <source>PInro:- ceedings of the 27th International Conference on Computational Linguistics: System Demonstrations. Santa Fe</source>
          , New Mexico,
          <year>2018</year>
          . url: https://aclanthology.org/C18-200.
          <fpage>2</fpage>
        </mixed-citation>
      </ref>
      <ref id="ref14">
        <mixed-citation>
          [16]
          <string-name>
            <given-names>J.</given-names>
            <surname>Kocoń</surname>
          </string-name>
          ,
          <string-name>
            <given-names>I.</given-names>
            <surname>Cichecki</surname>
          </string-name>
          ,
          <string-name>
            <given-names>O.</given-names>
            <surname>Kaszyca</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Kochanek</surname>
          </string-name>
          ,
          <string-name>
            <given-names>D.</given-names>
            <surname>Szydło</surname>
          </string-name>
          ,
          <string-name>
            <given-names>J.</given-names>
            <surname>Baran</surname>
          </string-name>
          ,
          <string-name>
            <given-names>J.</given-names>
            <surname>Bielaniewicz</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Gruza</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Janz</surname>
          </string-name>
          ,
          <string-name>
            <given-names>K.</given-names>
            <surname>Kanclerz</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Kocoń</surname>
          </string-name>
          ,
          <string-name>
            <given-names>B.</given-names>
            <surname>Koptyra</surname>
          </string-name>
          ,
          <string-name>
            <given-names>W.</given-names>
            <surname>Mieleszczenko-Kowszewicz</surname>
          </string-name>
          ,
          <string-name>
            <given-names>P.</given-names>
            <surname>Miłkowski</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Oleksy</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Piasecki</surname>
          </string-name>
          , Ł. Radliński,
          <string-name>
            <given-names>K.</given-names>
            <surname>Wojtasik</surname>
          </string-name>
          ,
          <string-name>
            <given-names>S.</given-names>
            <surname>Woźniak</surname>
          </string-name>
          , and
          <string-name>
            <given-names>P.</given-names>
            <surname>Kazienko</surname>
          </string-name>
          . “
          <article-title>ChatGPT: Jack of all trades, master of none”</article-title>
          .
          <source>InI n:formation Fusion</source>
          <volume>99</volume>
          (
          <year>2023</year>
          ), p.
          <fpage>101861</fpage>
          . doi: https://doi.org/10.1016/j.inffus.
          <year>2023</year>
          .
          <volume>10186</volume>
          .1
        </mixed-citation>
      </ref>
      <ref id="ref15">
        <mixed-citation>
          [17]
          <article-title>Koninklijke BibliotheekO.ver ons - Diensten DBNL</article-title>
          . https://www.kb.nl/over-ons/dienst en/dbnl.
          <year>2024</year>
          .
        </mixed-citation>
      </ref>
      <ref id="ref16">
        <mixed-citation>
          [18]
          <string-name>
            <given-names>A.</given-names>
            <surname>Kosar</surname>
          </string-name>
          ,
          <string-name>
            <given-names>G. D.</given-names>
            <surname>Pauw</surname>
          </string-name>
          , and
          <string-name>
            <given-names>W.</given-names>
            <surname>Daelemans</surname>
          </string-name>
          . “
          <article-title>Comparative Evaluation of Topic Detection: Humans vs</article-title>
          .
          <source>LLMs”</source>
          .
          <source>In: Computational Linguistics in the Netherlands Journal</source>
          <volume>13</volume>
          (
          <year>2024</year>
          ), pp.
          <fpage>91</fpage>
          -
          <lpage>120</lpage>
          . url: https://www.clinjournal.org/clinj/article/view/.
          <volume>173</volume>
          [19]
          <string-name>
            <given-names>L. D.</given-names>
            <surname>Langhe</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Maladry</surname>
          </string-name>
          ,
          <string-name>
            <given-names>B.</given-names>
            <surname>Vanroy</surname>
          </string-name>
          ,
          <string-name>
            <given-names>L. D.</given-names>
            <surname>Bruyne</surname>
          </string-name>
          ,
          <string-name>
            <given-names>P.</given-names>
            <surname>SIngh</surname>
          </string-name>
          ,
          <string-name>
            <given-names>E.</given-names>
            <surname>Lefever</surname>
          </string-name>
          , and
          <string-name>
            <given-names>O. D.</given-names>
            <surname>Clercq</surname>
          </string-name>
          . “
          <article-title>Benchmarking Zero-Shot Text Classification for Dutch”</article-title>
          .
          <source>In:Computational Linguistics in the Netherlands Journal</source>
          <volume>13</volume>
          (
          <year>2024</year>
          ), pp.
          <fpage>63</fpage>
          -
          <lpage>90</lpage>
          . url: https://clinjournal.org/clinj/article /view/172.
        </mixed-citation>
      </ref>
      <ref id="ref17">
        <mixed-citation>
          [20]
          <string-name>
            <given-names>M.</given-names>
            <surname>Li</surname>
          </string-name>
          ,
          <string-name>
            <given-names>T.</given-names>
            <surname>Shi</surname>
          </string-name>
          ,
          <string-name>
            <given-names>C.</given-names>
            <surname>Ziems</surname>
          </string-name>
          , M.-
          <string-name>
            <given-names>Y.</given-names>
            <surname>Kan</surname>
          </string-name>
          ,
          <string-name>
            <given-names>N.</given-names>
            <surname>Chen</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Z.</given-names>
            <surname>Liu</surname>
          </string-name>
          , and
          <string-name>
            <given-names>D.</given-names>
            <surname>Yang</surname>
          </string-name>
          . “
          <article-title>CoAnnotating: Uncertainty-Guided Work Allocation between Human and Large Language Models for Data Annotation”</article-title>
          .
          <source>In:Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Singapore</source>
          ,
          <year>2023</year>
          . doi:
          <volume>10</volume>
          .18653/v1/
          <year>2023</year>
          .emnlp-main.
          <volume>92</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref18">
        <mixed-citation>
          [21]
          <string-name>
            <given-names>J.</given-names>
            <surname>Liu</surname>
          </string-name>
          ,
          <string-name>
            <given-names>D.</given-names>
            <surname>Shen</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Zhang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>B.</given-names>
            <surname>Dolan</surname>
          </string-name>
          ,
          <string-name>
            <given-names>L.</given-names>
            <surname>Carin</surname>
          </string-name>
          , and
          <string-name>
            <given-names>W.</given-names>
            <surname>Chen</surname>
          </string-name>
          . “
          <article-title>What Makes Good In-Context Examples for GPT-3?” In:Proceedings of Deep Learning Inside Out (DeeLIO</article-title>
          <year>2022</year>
          ):
          <article-title>The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures</article-title>
          . Dublin, Ireland and Online,
          <year>2022</year>
          . do1i:
          <fpage>0</fpage>
          .18653/v1/
          <year>2022</year>
          .deelio-
          <volume>1</volume>
          .
          <fpage>10</fpage>
          .
        </mixed-citation>
      </ref>
      <ref id="ref19">
        <mixed-citation>
          [22]
          <string-name>
            <given-names>L.</given-names>
            <surname>Molle</surname>
          </string-name>
          . “
          <article-title>Inleiding - Een geschiedenis van mensen en (andere) dieren”</article-title>
          .
          <source>ITni:jdschrift voor Geschiedenis</source>
          <volume>125</volume>
          (
          <year>2012</year>
          ), pp.
          <fpage>464</fpage>
          -
          <lpage>475</lpage>
          . doi:
          <volume>10</volume>
          .5117/tvgesch2012.4.moll.
        </mixed-citation>
      </ref>
      <ref id="ref20">
        <mixed-citation>
          [23]
          <string-name>
            <surname>OpenAI</surname>
          </string-name>
          . Introducing ChatGPT. https://openai.com/blog/chatgp.t2022.
        </mixed-citation>
      </ref>
      <ref id="ref21">
        <mixed-citation>
          [24]
          <string-name>
            <given-names>C.</given-names>
            <surname>Qin</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Zhang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Z.</given-names>
            <surname>Zhang</surname>
          </string-name>
          , J. Chen,
          <string-name>
            <given-names>M.</given-names>
            <surname>Yasunaga</surname>
          </string-name>
          , and
          <string-name>
            <given-names>D.</given-names>
            <surname>Yang. Is ChatGPT a GeneralPurpose Natural Language Processing Task Solver</surname>
          </string-name>
          ? arXiv preprinthttps://arxiv.org/abs/2 302.06476.
          <year>2023</year>
          . doi: https://doi.org/10.48550/arXiv.2302.0647.6
        </mixed-citation>
      </ref>
      <ref id="ref22">
        <mixed-citation>
          [25]
          <string-name>
            <given-names>S.</given-names>
            <surname>Rebora</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Lehmann</surname>
          </string-name>
          ,
          <string-name>
            <given-names>A.</given-names>
            <surname>Heumann</surname>
          </string-name>
          ,
          <string-name>
            <given-names>W.</given-names>
            <surname>Ding</surname>
          </string-name>
          , and
          <string-name>
            <surname>G. Lauer.</surname>
          </string-name>
          “
          <article-title>Comparing ChatGPT to Human Raters and Sentiment Analysis Tools for German Children's Literature”C</article-title>
          . oInm: - putational
          <source>Humanities Research Conference (CHR)</source>
          . Paris, France,
          <year>2023</year>
          . url:https://ceurws.org/Vol-
          <volume>3558</volume>
          /paper3340.pd. f [26]
          <string-name>
            <given-names>N.</given-names>
            <surname>Reimers</surname>
          </string-name>
          and
          <string-name>
            <surname>I. Gurevych.</surname>
          </string-name>
          “
          <article-title>Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks”</article-title>
          .
          <source>In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Hong Kong</source>
          ,
          <year>2019</year>
          , pp.
          <fpage>3982</fpage>
          -
          <lpage>3992</lpage>
          . doi: https://doi.org/10.48550 /arXiv.
          <year>1908</year>
          .
          <volume>10084</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref23">
        <mixed-citation>
          <string-name>
            <given-names>M.</given-names>
            <surname>Ren</surname>
          </string-name>
          ,
          <string-name>
            <given-names>W.</given-names>
            <surname>Zeng</surname>
          </string-name>
          ,
          <string-name>
            <given-names>B.</given-names>
            <surname>Yang</surname>
          </string-name>
          , and
          <string-name>
            <given-names>R.</given-names>
            <surname>Urtasun</surname>
          </string-name>
          .
          <article-title>Learning to Reweight Examples for Robust Deep Learning</article-title>
          . arXiv preprinthttps://arxiv.org/abs/
          <year>1803</year>
          .0905 0.
          <year>2019</year>
          . doi: https://doi.org/10 .48550/arXiv.
          <year>1803</year>
          .
          <volume>09050</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref24">
        <mixed-citation>
          <string-name>
            <given-names>H.</given-names>
            <surname>Ritvo</surname>
          </string-name>
          .
          <article-title>The Animal Estate: The English</article-title>
          and
          <article-title>Other Creatures in the Victorian Age</article-title>
          . New ed. Cambridge, MA: Harvard University Press,
          <year>1989</year>
          .
        </mixed-citation>
      </ref>
      <ref id="ref25">
        <mixed-citation>
          [29]
          <string-name>
            <given-names>K.</given-names>
            <surname>Thomas</surname>
          </string-name>
          .
          <source>Man and the Natural World: Changing Attitudes in England 1500-1800</source>
          .
          <article-title>New edition</article-title>
          . London, UK: Penguin Books Ltd,
          <year>1991</year>
          .
        </mixed-citation>
      </ref>
      <ref id="ref26">
        <mixed-citation>
          [30]
          <string-name>
            <given-names>P.</given-names>
            <surname>Törnberg.</surname>
          </string-name>
          ChatGPT-4
          <string-name>
            <given-names>Outperforms</given-names>
            <surname>Experts</surname>
          </string-name>
          and
          <article-title>Crowd Workers in Annotating Political Twitter Messages with Zero-Shot Learning</article-title>
          . arXiv preprinthttps://arxiv.org/abs/2304.
          <fpage>065</fpage>
          <lpage>88</lpage>
          .
          <year>2023</year>
          . doi: https://doi.org/10.48550/arXiv.2304.0658.8
        </mixed-citation>
      </ref>
      <ref id="ref27">
        <mixed-citation>
          [31]
          <string-name>
            <given-names>S.</given-names>
            <surname>Wang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Liu</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Xu</surname>
          </string-name>
          ,
          <string-name>
            <given-names>C.</given-names>
            <surname>Zhu</surname>
          </string-name>
          , and
          <string-name>
            <given-names>M.</given-names>
            <surname>Zeng</surname>
          </string-name>
          . “Want To Reduce Labeling Cost?
          <article-title>GPT-3 Can Help”</article-title>
          . In:
          <article-title>Findings of the Association for Computational Linguistics: EMNLP 2021</article-title>
          . Punta Cana, Dominican Republic,
          <year>2021</year>
          . doi:
          <volume>10</volume>
          .18653/v1/
          <year>2021</year>
          .findings-emnlp.
          <volume>354</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref28">
        <mixed-citation>
          [32]
          <string-name>
            <given-names>R.</given-names>
            <surname>Zhang</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Li</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Y.</given-names>
            <surname>Ma</surname>
          </string-name>
          ,
          <string-name>
            <given-names>M.</given-names>
            <surname>Zhou</surname>
          </string-name>
          , and
          <string-name>
            <given-names>L.</given-names>
            <surname>Zou</surname>
          </string-name>
          . “
          <article-title>LLMaAA: Making Large Language Models as Active Annotators”</article-title>
          . In:
          <article-title>Findings of the Association for Computational Linguistics: EMNLP 2023</article-title>
          . Singapore,
          <year>2023</year>
          . doi:
          <volume>10</volume>
          .18653/v1/
          <year>2023</year>
          .findings-emnlp.
          <volume>872</volume>
          .
        </mixed-citation>
      </ref>
      <ref id="ref29">
        <mixed-citation>
          [33]
          <string-name>
            <given-names>C.</given-names>
            <surname>Ziems</surname>
          </string-name>
          ,
          <string-name>
            <given-names>W.</given-names>
            <surname>Held</surname>
          </string-name>
          ,
          <string-name>
            <given-names>O.</given-names>
            <surname>Shaikh</surname>
          </string-name>
          ,
          <string-name>
            <given-names>J.</given-names>
            <surname>Chen</surname>
          </string-name>
          ,
          <string-name>
            <given-names>Z.</given-names>
            <surname>Zhang</surname>
          </string-name>
          , and
          <string-name>
            <given-names>D.</given-names>
            <surname>Yang.Can Large Language Models Transform Computational Social Science</surname>
          </string-name>
          ? arXiv preprinthttps://arxiv.org/abs/2305.0351 4.
          <year>2024</year>
          . doi: https://doi.org/10.48550/arXiv.2305.0351.4
        </mixed-citation>
      </ref>
    </ref-list>
  </back>
</article>