=Paper=
{{Paper
|id=Vol-3683/CEUR-Template-2col5
|storemode=property
|title=Exploring Spatial Representations in the Historical Lake District Texts with LLM-based Relation Extraction
|pdfUrl=https://ceur-ws.org/Vol-3683/paper9.pdf
|volume=Vol-3683
|authors=Erum Haris,Anthony G. Cohn,John G. Stell
|dblpUrl=https://dblp.org/rec/conf/ecir/Haris0S24
}}
==Exploring Spatial Representations in the Historical Lake District Texts with LLM-based Relation Extraction==
https://ceur-ws.org/Vol-3683/paper9.pdf
Exploring Spatial Representations in the Historical
Lake District Texts with LLM-based Relation
Extraction
Erum Haris1,∗ , Anthony G. Cohn1,2 and John G. Stell1
1
School of Computing, University of Leeds, UK
2
The Alan Turing Institute, London, UK
Abstract
Navigating historical narratives poses a challenge in unveiling the spatial intricacies of past landscapes.
The proposed work addresses this challenge within the context of the English Lake District, employing
the Corpus of the Lake District Writing. The method utilizes a generative pre-trained transformer model
to extract spatial relations from the textual descriptions in the corpus. The study applies this large
language model to understand the spatial dimensions inherent in historical narratives comprehensively.
The outcomes are presented as semantic triples, capturing the nuanced connections between entities and
locations, and visualized as a network, offering a graphical representation of the spatial narrative. The
study contributes to a deeper comprehension of the English Lake District’s spatial tapestry and provides
an approach to uncovering spatial relations within diverse historical contexts.
Keywords
spatial narratives, spatial relation extraction, large language models, semantic triples network
1. Introduction
Recent years have witnessed a growing interest in understanding geographies within textual
sources across various disciplines. In the realm of historical archives, the Corpus of the Lake
District Writings (CLDW) [1] has attracted many researchers to explore the English Lake District,
a region known for its cultural heritage, characterized by lakes and mountains in the northwest
of England. The collection contains spatial narratives with descriptions of landscapes, places
and routes. The writers’ experiences of geography, as depicted in these narratives, are subjective
and a detailed analysis is necessary to handle the qualitative aspects of spatial information.
A particular case is the representation of relative locations in narratives using qualitative
spatial relations. References such as “near”, “to the south”, or “a few hours drive from” are
often used by writers as they experience relative locations. However, owing to the ambiguity of
natural language, it is a challenge to precisely extract spatial relations between any two entities
from text data. Spatial relations are classified into different types [2], including topological,
direction and distance relations ; each category requires a contextual understanding of the
GeoExT 2024: Second International Workshop on Geographic Information Extraction from Texts at ECIR 2024, March 24,
2024, Glasgow, Scotland
∗
Corresponding author.
Envelope-Open e.haris@leeds.ac.uk (E. Haris); a.g.cohn@leeds.ac.uk (A. G. Cohn); j.g.stell@leeds.ac.uk (J. G. Stell)
Orcid 0000-0002-8012-8850 (E. Haris); 0000-0002-7652-8907 (A. G. Cohn); 0000-0001-9644-1908 (J. G. Stell)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
text for extraction. Moreover, relation arity, toponym matching, coreference resolution, long-
dependency relations and related issues make relation extraction, in general, a demanding
research problem.
Spatial relation extraction has been primarily addressed either using manually crafted rules
and patterns to be matched over sentence tokens or using supervised learning, which requires
an enormous amount of annotated training data. Web-scale open information extraction [3],
though, does not require patterns in advance, but it results in many redundant and incoherent
facts in output. The significant breakthroughs driven by Large Language Models (LLMs) [4]
in processing natural language text have led to potential use cases and research findings in
GIS and spatial analysis. LLMs are pre-trained on large-scale textual data, which enables
them to generalize to unseen data without annotated examples or requiring minimal examples.
Numerous studies have concluded their success in various downstream NLP tasks [5, 6]; hence,
they are also called foundation models [7].
In this paper, we integrate two problem domains and perform a holistic analysis. We aim
to unravel the spatial representations in the CLDW corpus using a Generative Pre-trained
Transformer (GPT) model [4], thereby exploring the potential of pre-trained models in extracting
spatial relations from historical accounts. Specifically, the experiment focuses on the extraction
of the spatial relation “near” between entities, and results are visualized as a network of semantic
triples.
2. Related Work
The proposed study overlaps with multiple areas, including GIS and NLP in spatial humanities,
particularly studies on the CLDW, spatial relation extraction and applications of LLMs in
geographical information science (GISc). Hence, a brief review is presented for each domain.
Donaldson et al. [1] propose an interdisciplinary methodology for analyzing historical text
corpora. The study integrates corpus analysis, automated geoparsing and GIS to examine the
spatial dimensions linked to significant aesthetic terms historically used to depict the English
Lake District. The approach, termed “geographical text analysis,” demonstrates its capacity to
advance the analysis of interconnections among literature, aesthetics, and physical geography.
Going beyond toponymic geography, Ezeani et al. [8] introduce an extensible framework to
study the CLDW using NLP, GISc, qualitative spatio-temporal representation and reasoning
(QSTR), and visual analytics. They demonstrate preliminary work using the framework on
extracting, analyzing, and visualizing the spatial elements that define the “location”, “locale”
and “sense of place” as referenced in the text.
Proximity holds a pivotal role in any comprehensive spatial ontology, [9, 10]. The spatial
concept of nearness or proximity serves to identify either the spatial relationship between two
entities, such as “near (street, village),” or an object’s relative location in space “street (near,
village),” encapsulating people’s psychological perceptions of distances [11]. However, inter-
preting the meaning of qualitative proximity statements remains challenging within GIS. Talmy
[12] describes that natural language expressions representing spatial relations often exhibit
asymmetry. Along this line, Worboys [9] explores nearness relations in environmental spaces,
considering nearness as a “conceptual” distance relation. Novel et al. [11] propose a context-
dependent model for defining “near,” leveraging Turney’s two dimensional contextualization
framework [13].
In spatial relation extraction, methods usually focus on a single aspect of spatial information,
or sometimes a few aspects, owing to the domain-dependent nature of the corpus language.
Drymonas and Pfoser [14] reconstruct route maps by extracting relative and absolute geospatial
information from travel guides using a rule-based extraction module and a location ontology.
Skoumas et al. [15] extract short route indicators including distance and topological relations
from crowdsourced geospatial data using custom-defined regular expression patterns and
syntactic rules. Cadorel et al. [16] build a structured Geospatial knowledge base with a spatial
information extraction pipeline including named entities and relations developed using recurrent
neural networks.
The application of LLMs in the geospatial domain has resulted in various research studies.
Hu et al. [7] integrate geo-knowledge from location descriptions with a GPT model resulting
in a geo-knowledge-guided GPT model, designed explicitly for extracting precise location
information from social media messages related to disasters. Mooney et al. [17] scrutinize
the capabilities of ChatGPT in specialized subject area like GIS, acknowledging the significant
divergence between human learning of spatial concepts and LLM training methodologies. The
evaluation entails subjecting ChatGPT to a real GIS exam assessing its performance and spatial
literacy. Addressing one of the facets of spatial relations, Ramrakhiyani et al. [18] focus on
extracting the orientation information of borders between countries from Wikipedia text using
large language models based on Natural Language Inference (NLI) technique, combined with
lexical patterns.
3. Data and Method
3.1. Corpus of the Lake District Writing
This work is part of a spatial exploration of the CLDW [19], a collection encompassing travel
writing and tourist literature that vividly depicts the English Lake District from 1622 to 1900.
Comprising 80 geoparsed texts with over 1.5 million words and encompassing various genres,
including travel writing and fiction, the corpus provides valuable accounts from the Lake Poets
like Wordsworth and Coleridge and other significant figures, alongside works from lesser-known
authors. The descriptions capture the essence of travelling, aiming to describe landscapes and
emotional responses.
We considered the instances of the term “near” in the corpus for this study. Table 1 presents
a statistical summary of the spatial content in the CLDW, where named place refers to any
occurrence of a toponym and geographic noun represents a geographical feature appearing in
the text. A list of 139 nouns (e.g. river, road, waterfall, etc) as well as their inflections (rivers,
roads, waterfalls) has been manually curated[8]. Spatial references are the different occurrences
of spatial relations, including spatial prepositions, locative adverbs and distances [8]. Table 2
lists keyword-in-context (KWIC) [20] examples for the relation “near” showing its left and right
context within text passages.
Table 1
Statistical information on spatial content of the CLDW
Parameter Count
T ext Files 80
W ords 1,524,718
U nique Word Forms 43,765
N amed Places 39,916
G eographic Nouns 69,103
S patial References 41,901
" Near” Relation Occurrences 1611
Figure 1: Proposed framework for relation extraction and visualization. Template style (https://learn.
microsoft.com/en-us/azure/architecture/ai-ml/guide/query-based-summarization).
3.2. Framework
Figure 1 illustrates the overall framework for relation extraction, which consists of two stages.
The first is the entity extraction stage, which provides place names and geographic nouns in the
CLDW using the named entity recognition (NER) module developed by Ezeani et al. [8]. These
lists and the plain corpus files are used to design prompts for the pre-trained model, details of
which are described in the experiment section. The method employs GPT-4 using the Azure
OpenAI services [21], which provide access to OpenAI’s powerful language models, some of
which have reached general availability and can be deployed to NLP and image understanding.
The prompts are developed and passed to the model through Python SDK. The resultant semantic
triples are compared with the gold-standard set of labelled sentences for precision calculation.
Finally, the correct triples are passed to the network visualization module for a graphical
illustration of the target place’s spatial connections.
Table 2
Keyword-in-context examples for spatial relation “near”
Left Term Right
s tation is at Coniston village near the head of the lake
s een in the church-yard near the road. We have now
e ntered among the luxuriant woods near Gowbarrow Park and dark Ulleswater’s
i n the coves of Wythburn near to Dunmail Raise, and, after
3.3. LLM Prompt Engineering for Relation Extraction
LLMs are powerful but can sometimes generate outputs not aligned with the specific intent.
Hence, prompt engineering takes effect [22], which refers to the deliberate design and for-
mulation of prompts or queries provided to LLMs, particularly in fine-tuning or guiding the
behaviour of these models to improve their usability and reliability. In practical terms, the
prompt serves as a mechanism that directs the model’s attention and influences its output. The
model adjusts its internal weights and parameters during the training or inference process
[21]. This adjustment aligns the model’s behavior to better suit the intended task, improving
its performance and accuracy in generating responses or predictions. A model’s behaviour
is highly sensitive to the prompt; therefore, formulating prompts is challenging and requires
considerable effort.
In our approach to spatial relation extraction, the prompt engineering stage essentially follows
these steps. The process starts with specifying the task instructions and giving content and
context to assist the model’s understanding, defining the desired output format and experi-
menting with different prompt structures without providing training examples. This method
is called zero-shot learning [6], in which the model works to correlate the provided task with
existing categories it has previously learned and formulates responses accordingly. Finally,
prompt engineering is an iterative process that involves testing and refining prompts based
on the model’s output errors until the desired results are achieved. We performed several
iterations following a top-down approach, starting with open-ended spatial relation extraction
to confining to only specific category of spatial relations, extracting relations as predicates and
then spatial prepositions, mentioning spatial entities to be looked for, providing single and
multiple paragraphs and as a minimum, changing instruction articulation in the prompt. After
performing various experiments, we finally resorted to starting with the most basic setting for
extracting spatial relations to see which places are nearby located to others. Hence, the term
“near” is chosen at first for extraction and analysis. It is apparent that terms like “close to”,
“surrounds”, “next to” and others also convey a sense of proximity. However, in this work, we
have explicitly focused on the keyword “near” to formulate the prompts at this stage better. We
describe the technique followed in the experiment section.
4. Experiment and Results
Efficient crafting of input prompts must consider the token count constraint, which influences
the model’s response latency and throughput [21]. Therefore, the corpus was first subjected to
Table 3
Zero-shot Prompt with response
Prompt Response
"content":"From the given passages numbered in
brackets (), extract spatial relation "near" ”content”: ”(1)
only if the entity 'keswick' is involved in a (2)
be in the form .(1)Carlisle railway conducts from Kendal near, Keswick>(3)
to Birthwaite, a mile to the north of Bowness, and (4)
coaches run to Ambleside and Keswick (2) is not (5)
view is taken from Castlehead, near Keswick. The (6)
immediately to the page (3) on the small scale (7)
lives at Greta Hall, near Keswick, with Skiddaw at ...
Coleridge, who married... "
some text-mining techniques before passing the text to the model. First, a KWIC analysis was
performed for the spatial relation “near” with a context window of 15 words before and after,
which is the maximum allowed window size in the text analytics software [23]. The search
resulted in 1609 contexts. As the research is in the initial phases of assessing the GPT model’s
performance, we attempted to refine the search further. Our next choice was to look for the
most recurring places in the list obtained from the NER stage. Hence, a frequency count of the
place names was retrieved, according to which the town of “Keswick” is the most frequently
occurring place name. Applying the filtering with the keyword “Keswick” on the relation “near”
contexts’ list resulted in 84 outcomes. Table 4 provides information on this analysis for the
most frequent place names in the CLDW.
4.1. Zero-shot Spatial Relation Extraction
Table 3 gives a glimpse of the descriptive prompt provided to the GPT-4 model and the generated
response. We explicitly specified three things inside the prompt in this zero-shot setting:
• () : the numbered appearance of the input passages
• near : the spatial relation to look for
• : the format of the output where “subject” and
“object” are the spatial entities, one of which must be the specified place name which is
“Keswick” in this context
We are interested in using the generalization power of the GPT model and assessing its
overall effectiveness in identifying the spatial relationships between entities in the text without
providing specific examples. This is desirable in cases where input data is very large and
pre-annotation incurs a high cost, as is the case of the CLDW. Hence, in the initial stages, it is
preferred that the model should achieve good precision on its task, which means that extracted
facts should be correct in comparison with total predictions, though not complete, as we have
not guided the model with example use-cases which can cover different patterns of target spatial
relation occurrences. On the other hand, the model may not provide high-precision results
since in zero-shot learning, the model leverages statistical patterns learned during its training to
determine what should be generated next and hence, it may not have a semantic understanding
of the problem [21].
Table 4 provides precision scores for extracting “near” spatial relation in different cases. The
resultant triples were compared with a gold standard set of sentences labeled with relation
occurrences and corresponding place names. Hence, the precision scores denote the fraction of
correct triples out of all the triples produced. The prompt was repeated for two iterations to
ensure the model generates consistent results, and average precision was computed. The results
show that the model can perform in a zero-shot setting with better precision. However, the
underlying textual details are highly influential. In many cases, the sentence structure was not
straight enough for the model to precisely extract relation predicates. The model was also unable
to interpret metaphorical references in some instances correctly. Moreover, the geographical
descriptions vary from one type of entity to another, which was the case of “Windermere”, a
lake, and its features and nearby landmarks have been described in comparatively complex
phrases. Similarly, “Ambleside” is primarily expressed in the context of roads around it. Despite
the textual nuances, the model tried to infer nearness in some cases, representing its ability
beyond explicit extraction.
It is imperative to mention that we specified the output format to the model in a zero-shot
setting. Without specification, prompting for extracting “near” relation between entities resulted
in large chunks of words extracted as subject or object, merging relation term with object and
other unexpected output forms. Next, there is a possibility that the terms “near” and “Keswick”,
in this scenario, do not hold any relation despite being present in the filtered context. In such
a case, the model may try to fabricate output that can be matched and pointed out but would
reduce the overall precision if left untreated. Hence, we also tried adding an instruction in the
prompt to respond “not found” if the model is not confident about the predicted relation. This
attempt resulted in a downgraded response because the model discarded most of the correct
triples it had previously extracted. Finally, the environment-specific parameters play crucial
role in generating deterministic responses. In Azure OpenAI Service, the prompt is passed to a
completion endpoint with other probability parameters, such as “temperature” which has been
kept to zero to achieve replicable results [21].
4.2. Network of Semantic Triples
For visualization, incorrect outputs were filtered from the model’s response. The remaining
were used to construct a network [24], shown in Figure 2, that visualizes the resultant semantic
triples for the place name “Keswick”. The nodes represent all places in resultant triples without
toponym matching to merge nodes referring to the same place. The edges are weighted according
to their output frequency in the triples set. Some relation instances occur more than once than
Figure 2: Network visualization of “near”semantic triples for place“Keswick”.
others reflected via thickened edges.
As described earlier, “near” is a vague term used subjectively. In most cases, the nearness
implies genuine proximity. Still, some destinations requiring more than two or three hours of
walking may be related as near to the source depending on the mode of travel, which should be
identified from textual descriptions. To strike a balance in our analysis, we explicitly checked
the proximity of place names when there was an ambiguity while labeling their relation. This
was needed owing to the expressions of nearness in the text challenging the gold standard
preparation.
5. Conclusion
This paper is part of an overarching research initiative that seeks to unveil the spatial and
temporal dynamics inherent in narratives [25]. The study aims to introduce a considerable
shift in our methodical approaches to exploring the geographies outlined in extensive historical
archives by harnessing the power of LLMs. We proposed a framework for extracting spatial
Table 4
Precision results for four selected places
Place name Frequency Context Count with “Near” Average Precision
K eswick 1452 84 (0.630, 0.702) = 0.666
A mbleside 900 43 (0.656, 0.647) = 0.652
W indermere 873 45 (0.543, 0.675) = 0.609
P enrith 714 53 (0.714, 0.606) = 0.660
relations from the CLDW and presented results for relation “near”. The results are visualized as a
network that depicts the target place, showing its nearby spatial entities. The proposed approach
complements the existing geographical analyses by introducing a distinctive computational
representation of place, thereby enhancing the capacity of social scientists and humanists to
interpret narrative depictions of location. For the future, we are working towards improving
the extraction performance by refining the zero-shot prompts, experimenting with few-shot
learning and extracting other qualitative spatial relations.
Acknowledgments
We acknowledge the support of the Economic and Social Research Council (ESRC) under grant
ES/W003473/1. We also thank Paul Rayson and Ignatius Ezeani for their comments on this
work, and the entire team of the Spatial Narratives project for their discussions on the CLDW.
References
[1] C. Donaldson, I. N. Gregory, J. E. Taylor, Locating the beautiful, picturesque, sublime and
majestic: spatially analysing the application of aesthetic terminology in descriptions of
the english lake district, Journal of Historical Geography 56 (2017) 43–60. doi:10.1016/j.
jhg.2017.01.006 .
[2] E. Haris, K. H. Gan, T.-P. Tan, Spatial information extraction from travel narratives:
Analysing the notion of co-occurrence indicating closeness of tourist places, Journal of
Information Science 46 (2020) 581–599. doi:10.1177/0165551519837188 .
[3] K. Detroja, C. Bhensdadia, B. S. Bhatt, A survey on relation extraction, Intelligent Systems
with Applications 19 (2023). doi:10.1016/j.iswa.2023.200244 .
[4] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan,
P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, in: Pro-
ceedings of the 34th International Conference on Neural Information Processing Systems,
NIPS’20, Curran Associates, Inc., Red Hook, NY, 2020.
[5] X. Wei, X. Cui, N. Cheng, X. Wang, X. Zhang, S. Huang, P. Xie, J. Xu, Y. Chen,
M. Zhang, et al., Zero-shot information extraction via chatting with chatgpt, 2023.
arXiv:2302.10205 .
[6] C. Yuan, Q. Xie, S. Ananiadou, Zero-shot temporal relation extraction with chatgpt, in:
The 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared
Tasks, ACL, 2023, pp. 92–102. doi:10.18653/v1/2023.bionlp- 1.7 .
[7] Y. Hu, G. Mai, C. Cundy, K. Choi, N. Lao, W. Liu, G. Lakhanpal, R. Z. Zhou, K. Joseph,
Geo-knowledge-guided gpt models improve the extraction of location descriptions from
disaster-related social media messages, International Journal of Geographical Information
Science 37 (2023) 2289–2318. doi:10.1080/13658816.2023.2266495 .
[8] I. Ezeani, P. Rayson, I. N. Gregory, E. Haris, A. G. Cohn, J. G. Stell, T. Cole, J. E. Taylor,
D. Bodenhamer, N. Devadasan, et al., Towards an extensible framework for understanding
spatial narratives, in: Proceedings of the 7th ACM SIGSPATIAL International Workshop
on Geospatial Humanities, GeoHumanities ’23, ACM Press, New York, NY, 2023, pp. 1–10.
doi:10.1145/3615887.3627761 .
[9] M. F. Worboys, Nearness relations in environmental space., Intelligent Systems with
Applications 15 (2001). doi:10.1080/13658810110061162 .
[10] J. Brennan, A framework for modelling spatial proximity, Ph.D. thesis, UNSW Sydney,
2009. doi:10.26190/unsworks/19586 .
[11] M. Novel, R. Grütter, H. Boley, A. Bernstein, Nearness as context-dependent expression: an
integrative review of modeling, measurement and contextual properties, Spatial Cognition
& Computation 20 (2020) 161–233. doi:10.1080/13875868.2020.1754832 .
[12] L. Talmy, How language structures space, in: Spatial orientation: Theory, research, and
application, Springer, 1983, pp. 225–282. doi:10.1007/978- 1- 4615- 9325- 6_11 .
[13] P. D. Turney, The management of context-sensitive features: A review of strategies, arXiv
preprint cs/0212037 (2002). arXiv:cs/0212037 .
[14] E. Drymonas, D. Pfoser, Geospatial route extraction from texts, in: Proceedings of the 1st
ACM SIGSPATIAL International Workshop on Data Mining for Geoinformatics, DMG ’10,
ACM Press, New York, NY, 2010, pp. 29––37. doi:10.1145/1869890.1869894 .
[15] G. Skoumas, K. A. Schmid, G. Jossé, M. Schubert, M. A. Nascimento, A. Züfle, M. Renz,
D. Pfoser, Knowledge-enriched route computation, in: Proceedings of the 14th In-
ternational Symposium on Advances in Spatial and Temporal Databases, SSTD ’15,
Springer International Publishing, Cham, 2015, pp. 157–176. doi:https://doi.org/10.
1007/978- 3- 319- 22363- 6_9 .
[16] L. Cadorel, A. Blanchi, A. G. Tettamanzi, Geospatial knowledge in housing advertisements:
Capturing and extracting spatial information from text, in: Proceedings of the 11th
Knowledge Capture Conference, K-CAP ’21, ACM Press, New York, NY, United States,
2021, pp. 41–48. doi:10.1145/3460210.3493547 .
[17] P. Mooney, W. Cui, B. Guan, L. Juhász, Towards understanding the geospatial skills of
chatgpt: Taking a geographic information systems (gis) exam, in: Proceedings of the 6th
ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery,
GeoAI ’23, ACM Press, New York, NY, 2023, pp. 85–94. doi:10.1145/3615886.3627745 .
[18] N. Ramrakhiyani, V. Varma, G. K. Palshikar, Extracting orientation relations between
geo-political entities from their wikipedia text, in: Proceedings of the First International
Workshop on Geographic Information Extraction from Texts, GeoExT ’23, CEUR Workshop
Proceedings, 2023, pp. 44–50.
[19] P. Rayson, A. Reinhold, J. Butler, C. Donaldson, I. N. Gregory, J. Taylor, A deeply annotated
testbed for geographical text analysis: The corpus of lake district writing, in: Proceedings
of the 1st ACM SIGSPATIAL Workshop on Geospatial Humanities, GeoHumanities ’17,
ACM Press, New York, NY, 2017, pp. 9–15. doi:10.1145/3149858.3149865 .
[20] S. Sinclair, G. Rockwell, Voyant tools, 2016. URL: http://voyant-tools.org/.
[21] Azure AI Services, Azure openai service models (GPT-4), 2023. URL: https://learn.microsoft.
com/en-us/azure/aiservices/openai.
[22] P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, G. Neubig, Pre-train, prompt, and predict: A
systematic survey of prompting methods in natural language processing, ACM Comput.
Surv. 55 (2023). doi:10.1145/3560815 .
[23] VERBI Software, Maxqda 2022. software, 2021. URL: maxqda.com.
[24] M. Bastian, S. Heymann, M. Jacomy, Gephi: an open source software for exploring and
manipulating networks, in: Proceedings of the International AAAI Conference on Weblogs
and Social Media, ICWSM, AAAI, 2009, pp. 361–362. doi:10.1609/icwsm.v3i1.13937 .
[25] Space Time Narratives Project, Understanding imprecise space and time in narratives
through qualitative representations, reasoning, and visualisation, 2022. URL: https:
//spacetimenarratives.github.io/.