An Innovative Approach to Represent Tacit Knowledge of Fishing with
Knowledge Graphs
Tanjila Kanij
1 a
, Shafia Husna
1
, Afzal Azeem Chowdhary
1 b
, Misita Anwar
1 c
,
Md. Khalid Hossain
2 d
and John Grundy
2 e
1
Swinburne University of Technology, Melbourne, Australia
2
Monash University, Melbourne, Australia
Keywords:
Tacit Knowledge, Knowledge Graph, Fisherfolk, Large Language Models.
Abstract:
Fisherfolk communities in developing nations face marginalization due to low literacy levels and socio-
economic challenges, leading to reduced interest in fishing and the loss of vital, undocumented tacit knowl-
edge. To address this, we conducted focus group discussions and interviews in Bangladesh and Indonesia,
extracting knowledge components from the conversational data to develop knowledge graphs. These graphs
visually represent facts, attributes, and relationships, facilitating knowledge extraction. We compared man-
ual and automated graph development using Large Language Models (LLMs), demonstrating their potential
to systematically identify, preserve, and share critical fishing knowledge. Future work involves testing this
framework with fisherfolk to preserve and disseminate this essential knowledge.
1 INTRODUCTION
From our experience with a large-scale ICT4D (In-
formation and Communication for Development)
project, we worked closely with fishing communities
in two developing nations to empower marginalized
fisherfolk through digital technologies. During the
user-centred design of prototype solutions, a critical
challenge emerged: the loss of vital fishing knowl-
edge. Due to low literacy and socio-economic bar-
riers, fisherfolk acquire tacit knowledge through ob-
servation and mentorship of experienced “seniors”.
Being unexpressed and often semi-conscious, Tacit
knowledge is rarely documented (Polanyi, 2009).
However, declining interest in the profession among
younger generations threatens this knowledge, lead-
ing to a scarcity of skilled fisherfolk. To address this,
it is essential to identify, systematically preserve, and
share this invaluable tacit knowledge within the com-
munity.
While previous research proposed user-focused
a
https://orcid.org/0000-0002-5293-1718
b
https://orcid.org/0000-0003-2722-6424
c
https://orcid.org/0000-0003-3979-1746
d
https://orcid.org/0000-0001-5040-8619
e
https://orcid.org/0000-0003-4928-7076
applications to facilitate knowledge sharing (Kanij
et al., 2023), these do not address the systematic
preservation of tacit knowledge. We systematically
explore methods to identify and preserve tacit knowl-
edge to address this gap.
After reviewing various approaches, we adopted
“Knowledge Graphs” for two key reasons: (1) their
ability to graphically represent heterogeneous facts,
attributes, and relationships, and (2) the ease of ex-
tracting knowledge from the graphs. Developing
knowledge graphs for tacit knowledge posed com-
putational challenges. Initially, we manually created
knowledge graphs to evaluate their feasibility. Then,
we leveraged large language models (LLMs) to au-
tomate the extraction of text-based tacit knowledge
components and develop these graphs. This article
details both approaches and discusses insights from a
pilot study, highlighting the potential for large-scale
deployment to systematically identify, preserve, and
share vital fishing knowledge.
The rest of the paper is organised as follows: Sec-
tion 2 presents theories of tacit knowledge and knowl-
edge graphs and information on fisherfolk communi-
ties in Bangladesh and Indonesia, Section 3 presents
the details of user studies, Section 4 describes the idea
of developing knowledge graphs based on the find-
ings, Section 5 and Section 6 illustrate the manual and
500
Kanij, T., Husna, S., Chowdhary, A. A., Anwar, M., Hossain, M. K. and Grundy, J.
An Innovative Approach to Represent Tacit Knowledge of Fishing with Knowledge Graphs.
DOI: 10.5220/0013282400003928
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2025), pages 500-507
ISBN: 978-989-758-742-9; ISSN: 2184-4895
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

automated process of knowledge graphs generation,
respectively, Section 7 discusses the implications and
challenges and finally Section 8 concludes the article.
2 BACKGROUND
2.1 Tacit Knowledge
The concept of tacit knowledge, introduced by
Michael Polanyi in The Tacit Dimension, highlights
that ”we can know more than we can tell” (Polanyi,
2009). Polanyi illustrated this with examples like
face recognition and bike riding, showing that tacit
knowledge resides in unconscious understanding and
is challenging to externalize. Philosophers and cog-
nitive psychologists differentiate tacit (unconscious)
from explicit (conscious) knowledge. Pathirage et al.
describe tacit knowledge as rooted in personal expe-
riences, shaped by factors like attitudes, emotions,
and perspectives (Pathirage et al., 2008). Sanderson
emphasizes physical skills gained through practice
and implicit learning (Sanderson, 2001), while Clarke
highlights its evolving nature through interaction and
experience (Clarke, 2010).
Nonaka and Takeuchi define tacit knowledge as
personal, context-specific, and hard to formalize, con-
trasting it with explicit knowledge, which is codi-
fied and easily communicated (Nonaka and Takeuchi,
1995). Leonard and Sensiper position these as op-
posite ends of the knowledge spectrum (Leonard and
Sensiper, 1998). Knowledge management literature
debates two views: (1) knowledge exists in distinct
structures (unconscious and conscious) that are not
mutually convertible, and (2) tacit and explicit knowl-
edge are two states that can be converted (Grandinetti,
2014). Nonaka et al. support the externalization of
tacit knowledge into explicit forms but note that this
process is complex due to its integration with dynamic
human processes (Nonaka, 2009).
To address these challenges, an Entity-
Relationship Management (ERM) system has
been proposed to codify tacit knowledge into explicit
forms and manage it in a digital repository (Chen and
Nunes, 2019). Unlike traditional databases reliant
on Boolean logic, the ERM model is envisioned
as a flexible, heterogeneous repository capable of
accommodating the complexities of tacit knowledge.
2.2 Knowledge Graph (KG)
Knowledge Graphs (KGs) organize and link infor-
mation in a graph-like structure, where nodes repre-
sent entities (e.g., people, places, or concepts) and
edges define their relationships; as semantic net-
works, they integrate diverse information sources to
represent knowledge on specific topics (Fensel et al.,
2020). Knowledge Management Systems structure
and connect information, simplifying navigation and
discovery within organizations (Dalkir, 2013).
KGs capture complex relationships and semantic
information, enabling the querying of interconnected
data, supporting knowledge discovery, decision-
making, and efficient information retrieval. Conse-
quently, they are widely adopted in AI, data inte-
gration, NLP, recommendation systems, and general
knowledge management (Ladeinde et al., 2023).
2.3 Fisherfolk
2.3.1 Fisherfolk in Bangladesh
Due to its geographical location, Bangladesh is pros-
perous in rivers and wetlands, making fishing a cru-
cial livelihood for a significant portion of its popu-
lation. Mahmud et al. (Mahmud et al., 2015) esti-
mate that 10% of the population depends on fishing,
while Tran et al. (Tran et al., 2023) report fisheries
support 12% of the 170 million residents in full-time
and part-time roles. Livelihood experiences and chal-
lenges vary by region and water source (rivers, seas,
or wetlands) (Kabir et al., 2012), with common chal-
lenges including poverty, low living standards, lim-
ited credit, lack of knowledge, disasters, and diseases
(Kabir et al., 2012; Das et al., 2015; Mahmud et al.,
2015; Hossain et al., 2009).
Recent studies emphasize the importance of pre-
serving and sharing tacit knowledge among fisher-
folk. Kanij et al. (Kanij et al., 2023) designed a
mobile application to facilitate knowledge exchange
among boat captains, aligning with Shuva et al.’s find-
ings that fisherfolk primarily rely on friends and fam-
ily for information (Shuva, 2017). Both studies stress
the importance of transforming the ”information ser-
vice culture” to promote equitable information access,
addressing barriers such as poor mobile networks, ra-
dio signal limitations, illiteracy, and poverty. Further-
more, Miah and Islam (Miah and Islam, 2020) high-
lighted that inequitable benefits, weak regulations,
and power imbalances also shape these practices.
2.3.2 Fisherfolk in Indonesia
Like Bangladesh, Indonesian fisherfolk communities
face significant socio-economic and environmental
challenges, including poverty, income instability, lim-
ited access to essential services, and environmental
degradation (Prasetyo et al., 2023). Despite efforts
by government and NGOs, these initiatives often fail
An Innovative Approach to Represent Tacit Knowledge of Fishing with Knowledge Graphs
501

to address entrenched structural issues (Supriati and
Umar, 2020; Prasetyo et al., 2023).
Local wisdom and traditional knowledge are vital
in sustaining livelihoods and social cohesion, as seen
in Bunaken Island communities (Supriati and Umar,
2020). Practices such as sustainable fishing methods,
cultural rituals, and communal activities promote en-
vironmental protection and unity, embodying values
of cooperation and respect for nature.
3 DATA COLLECTION
To explore knowledge management practices and
identify critical tacit fishing knowledge, we con-
ducted eight focus group discussions each in
Bangladesh (Barguna and Chandpur districts) and
Indonesia (South Galesong, Mangarabombang, and
North Galesong districts). The participants, involved
in both marine and inland fisheries, were recruited
through trusted local partner organizations.
Separate focus groups were held for men and
women, employing a storytelling approach that be-
gan with, “Please describe a usual day in your life.”
Discussions covered various topics, including demo-
graphics, information practices, decision-making, re-
sponsibilities, career aspirations, training, challenges,
and environmental issues. Although not directly in-
volved in fishing, women fisherfolk also shared their
contributions. Focus group had 8–12 participants.
In addition, we interviewed 20 key stakeholders
(referred to as Key informants (KI)) working closely
with fisherfolk, including local government officials,
fish business representatives, NGO workers, and fish-
erfolk’s association leaders to explore stakeholder en-
gagement with fisherfolk and their perceptions of
livelihood challenges, expectations, and tacit knowl-
edge practices. All focus groups were in person,
and the interviews were both face-to-face and online,
in local languages, and audio-recorded with partici-
pants’ consent. Using Thematic Analysis (Nowell
et al., 2017), we developed over 200 codes grouped
into 12 themes to find insight from the data.
Our findings highlight that fisherfolk’s tacit
knowledge management involves practical and cul-
tural elements deeply embedded in generational wis-
dom and integral to daily life. It encompasses eco-
logical understanding, economic considerations, and
socio-cultural practices; the key features are:
• The information/knowledge and their respective
sources are mainly heterogeneous.
• The relationships between knowledge compo-
nents are complex and do not always conform to
a specific pattern (semantic ambiguity).
4 REPRESENTATION OF TACIT
KNOWLEDGE OF FISHING
WITH KNOWLEDGE GRAPHS
Our research shows that fisherfolks in Bangladesh and
Indonesia rely heavily on their fishing tacit knowledge
for . The tacit knowledge of fishing must be docu-
mented and preserved systematically to prevent the
loss of this important knowledge over time. We devel-
oped KGs to systematically preserve tacit knowledge
and facilitate sharing to achieve this objective.
KGs enhance accessibility and streamline the
sharing and preservation of tacit fishing knowledge,
promoting their adoption among fisherfolk in both
countries. Their selection followed an evaluation of
paradigms like knowledge bases, ontologies, and se-
mantic networks. Unlike knowledge bases, which
focus on structured data storage, KGs utilize graph
structures to effectively capture and represent experi-
ential knowledge.
While ontologies enable semantic modelling and
link concepts, they lack the flexibility and granularity
of KGs. By integrating varied data types and repre-
senting complex, domain-specific relationships, KGs
proved to be the most suitable approach for preserv-
ing and sharing the nuanced knowledge of fisherfolk.
5 DEVELOPMENT OF
KNOWLEDGE GRAPHS
A manual approach was initially adopted to ensure the
effectiveness and accuracy of the knowledge graphs.
Focus group discussions and interview transcripts
were used as inputs to extract knowledge components.
These responses typically contained long sentences,
so the information was first organized by separating
sentences based on their themes.
Sentence Separation
Subject Identification
Structuring Facts
Identify Relationships
Representing Relationships
Development of Cypher Query
Figure 1: KG Development Process (Manual).
The subjects of these sentences were identified to cre-
ate graph nodes, and their relationships to themes
were structured and encoded in Cypher. The queries
were executed in Neo4j (NEuler) to construct and
display the knowledge graph, leveraging Neo4j’s ef-
ficiency in extracting knowledge components with
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
502

Cypher. The process is shown in Figure 1.
Demonstration with an Example. The following
example demonstrates the manual process of develop-
ing knowledge graphs using a focus group discussion
and interview transcript.
The transcript from an Indonesian focus group
reads as follows: ”We all use Jabba (fish trap) to catch
fish. Jabba we set at the bottom of the lake. So those
of us who use Jabba as our main means of catching
fish must be clever and endure diving to the bottom
of the lake. We don’t have a specific time when we
should install Jabba. It just depends on the feeling
and the proper time slot available.”
Table 1: Sentence Separation & Subject Identification.
Sentence Separation Subject Identification
S1: We all use Jabba (fish trap) to catch fish. We [Sub: Fisherfolk]
S2: Jabba we set at the bottom of the lake. Jabba [Sub: Jabba]
S3: Those of us who use Jabba as our main means
of catching fish must be clever and endure diving
to the bottom of the lake.
Those of us [Sub: Fisherfolk]
S4: We don’t have a specific time when we should
install Jabba.
We [Sub: Fisherfolk]
S5: It depends on the feeling and proper time slot
available.
It [Sub: Jabba installation]
Based on the method outlined in Section 5 and
shown in Figure 1, the following steps were followed:
1. The interview response was split into five sen-
tences. 2. The subjects of each sentence were iden-
tified by analyzing grammatical structure and context
to ensure accurate understanding. This step is shown
in Table 1. 3. We extracted five key facts from the
sentences and subjects. To simplify, we assumed that
all fisherfolk in the example used “Jabba” for fishing.
4. Next, Five relationships were identified from the
facts extracted. 5. The relationships were formatted
for easy conversion into Cypher queries. Facts were
enclosed in curly braces, with relationships following
the same format, starting with a colon (:), as illus-
trated in Table 2. 6. Lastly, Nine Cypher queries
were created for node generation and six for estab-
lishing relationships, executed on the Neo4j platform
to produce the knowledge graph shown in Figure 2.
Creating Nodes:
CREATE (fishermen:Identity {name:‘Fishermen’, description:‘Catches
fish using Jabba’ })
CREATE (fish:Identity {name:‘Fish’})
CREATE (jabba:Tool {name:‘Jabba’, description:‘Fish trap’ })
CREATE (depths:Position {name:‘Depths’, description:‘Bottom of the
lake’ })
CREATE (lake:Location {name:‘Lake’})
CREATE (wit:Attribute {name:‘Wit skills’})
Table 2: Fact & Relationship Identification and Representa-
tion.
Fact Identifica-
tion
Relationship Identification Relationship Repre-
sentation
Jabba is a fish
trap.
Jabba ≡ Fishtrap Jabba ≡ Fishtrap
Fishermen use
Jabba to catch
fish.
Relationship: Fishermen use Jabba
nets to catch fish from the lake.
Fishermen (identity)
[:use] Jabba (tool) to
catch fish.
Jabba is set at
the bottom of the
lake.
Relationship: Deepest part of the lake
location where Jabba is positioned for
use.
Jabba(tool) [:is set at]
the bottom (position)
of the lake(location).
Fishermen re-
quire wit and
diving endurance
to set up the
Jabba.
Relationship: Deploying Jabba re-
quires fishermen to demonstrate wit
and diving endurance to reach and in-
stall it at the lake bottom.
Fishermen (identity)
[:require] wit (at-
tribute) and diving
endurance (attribute)
to set up Jabba (tool).
Jabba installation
depends on the
feeling and avail-
ability of proper
time slot.
Relationship: The decision to install
Jabba underwater is influenced by sub-
jective factors, intuition, and practical
considerations, such as the availability
of a suitable time slot.
Jabba (tool) [:installa-
tion depends on] the
intuition (attribute)
and availability of
proper time slot (at-
tribute).
CREATE (diving:Attribute {name:‘Diving skills’})
CREATE (intuitiveTiming:Attribute {name:‘Intuition’})
CREATE (availability:Attribute {name:‘Timeslot availability’})
Creating Relationships:
CREATE (fishermen)-[ : USE]→(jabba)-[ : TO CATCH]→(fish)
CREATE (jabba)-[ : IS SET AT]→(depths)-[ : OF]→(lake)
CREATE (fishermen)-[ : REQUIRE]→(wit)-[ : TO SET UP]→(jabba)
CREATE (fishermen)-[ : REQUIRE]→(diving)-[ : TO SET UP]→(jabba)
CREATE (jabba)-[ : SET UP DEPENDS ON]→(intuitiveTiming)
CREATE (jabba)-[ : SET UP DEPENDS ON]→(availability)
Fishermen Wit Skills
Jabba
Driving
Skills
Intuition
Fish
Timeslot
Available
Lake
Depths
REQUIRE
Figure 2: Knowledge Graph Snippet.
Feedback on KG. We developed separate KGs for
each country, with the Indonesian knowledge graph
presented to five experts and academics from Indone-
sian universities who have experience working with
fishermen in the region. All experts welcomed the
idea of visualizing knowledge in this format but ex-
pressed concerns about the complexity of the graphs
as more knowledge components are added.
This manual feasibility study evaluated knowl-
edge graphs for preserving tacit knowledge in fish-
erfolk communities. The successful creation of two
graphs and positive expert feedback confirm its via-
bility.
An Innovative Approach to Represent Tacit Knowledge of Fishing with Knowledge Graphs
503

6 AUTOMATED DEVELOPMENT
OF KNOWLEDGE GRAPHS
Building on the successful creation of KGs and pos-
itive feedback, we streamlined the process by em-
ploying computational methods to extract facts and
relationships from text, automating the generation of
Cypher queries.
This section explains the automated extraction of tacit
knowledge from translated focus group discussions
and interview transcripts (in English) to create knowl-
edge graphs. Using a six-step framework (see Fig-
ure 3), the process mirrors the manual approach, with
the best algorithms validated against manual results.
Each step is detailed in the following subsections.
Demonstration with an Example: We will demon-
strate the automated process of developing knowledge
graphs from the focus group discussion and interview
transcripts. We will consider the previous example
transcript from the manual KG part Section 5.
Preprocessing & Text
Segmentation
Fact Identification
Entity Recognition
Relation Identification
Database Integration &
Knowledge Graph Generation
System Integration
Figure 3: Automated Knowledge Graph Methodology.
A. Preprocessing and Text Segmentation: This
stage was vital in preparing transcription data for
knowledge graph creation. It automatically identi-
fies and extracts question-answer (Q&A) pairs and fil-
ters sentences containing key knowledge components
while removing irrelevant information.
Initial Approach: SpaCy retained conversational parts
but lacked precise filtering and segmentation, often
missing relevant tacit fishing knowledge. NLTK im-
proved segmentation using its Text-Filling Algorithm
but struggled with accuracy and filtering irrelevant
content. Both identified five and four Q&A pairs in
the example above but failed to filter sensitive infor-
mation consistently.
Final Approach: Mistral 7B (Jiang et al., 2023),
trained without tacit fishing knowledge, significantly
improved text segmentation by accurately pairing rel-
evant questions and answers. This ensured precise
filtering, retaining only sentences with valuable tacit
fishing knowledge. For example, it identified the fol-
lowing Q&A pair: “Question”: “Why do you use
Jabba?”, “Answer”: “We use Jabba (fish trap) to
catch fish. We set Jabba at the bottom of the lake.”
The automation reduced manual effort and enabled
faster, more efficient processing of large datasets.
B. Summarisation and Fact Identification: This
involves identifying essential facts from Q&A pairs to
create summarised sentences that exclude pronouns.
Initial Approach: Various ML techniques were ex-
plored for extracting meaningful facts from text,
including libraries like spaCy & NLTK, LLMs
such as Bard, GPT, Llama 2, Facebook/bart-
large-cnn, and summarization models like
Falconsai/text summarization & IlyaGusev/m-
bart ru sum gazeta. Testing on 20 segmented
paragraphs revealed Facebook/bart-large-cnn as
the most accurate, though it introduced pronouns,
ambiguous sentences, and omitted crucial details.
For example: Facebook/bart-large-cnn: “We all use
Jabba (fish trap) to catch fish. Jabba we set at the
bottom of the lake. We don’t have a specific time
when we should install Jabba.”.
Llama 2 initially produced accurate outputs without
omitting key facts but struggled with hallucination.
Final Approach: Mistral 7B was selected over Llama
2 for its superior efficiency and accuracy in context
retention and fact identification. Prompt engineering
enhanced information clarity by removing pronouns
and refining context while ensuring gender-inclusive
language by referring to individuals as fisherfolk.
Using the demo example, Mistral 7B summarised the
sentences as follows: Sentence 1: “Fisherfolk use
Jabba as a fishing tool.” Sentence 2: “Jabba is set at
the lake bottom for catching fish.”
C. Named Entity Recognition (NER): This step fo-
cuses on categorisation accuracy and efficiency. It
extracts key concepts from the input data and cate-
gorises them into predefined labels.
Initial Approach: Initially, spaCy struggled with en-
tity extraction, especially using native terms from
Bangladesh and Indonesia. BERT performed poorly
on unstructured data and native words. Doccano’s
manual labelling was inefficient, leading to the adop-
tion of the OpenAI API for auto-labelling, which im-
proved consistency. However, fine-tuning BERT with
limited training samples still yielded low accuracy.
Both approaches misidentified Jabba as PERSON.
Final Approach: Mistral 7B effectively captured key
facts from summarised text and accurately identified
entities, categorizing them into 16 predefined labels
(Table 3). Context-based categorization structured
the knowledge for graph creation. Automation of
labelling expedited processing and significantly re-
duced manual effort for large datasets.
Using the demo example, the following entities were
found in each sentence:
Sentence 1: “Fisherfolk use Jabba as a fishing
tool.” Entities:[{“entity”: “Jabba”, “label”: “TOOL”
},{“entity”: “Fisherfolk”,“label”: “GROUP”}]; Sentence
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
504

Table 3: Description of the 16 Entity Categories.
Entity Name Description
Identity (ID) Names of individuals or entities involved in fishing practices
(e.g., ”Fisherman Ali”)
Location
(LOC)
Geographical locations relevant to fishing activities (e.g., ”Bay of
Bengal”)
Tool (TOOL) Equipment and tools (e.g., ”Fishing Net”)
Food (FOOD) Types of food or bait (e.g., ”Shrimp”)
Season (SSN) Seasonal information affecting fishing practices (e.g., ”Monsoon
season”)
Attribute
(ATT)
Characteristics or qualities related to fishing (e.g., ”Strong cur-
rents”)
Product As-
pect (PROD)
Specific aspects of fishing products (e.g., ”Fish size”)
Infosource
(INFO)
Sources of information such as local knowledge or expert advice
(e.g., ”Local guide”)
Group (GRP) Groups or communities involved in fishing (e.g., ”Fishing coop-
erative”)
Factor (FAC) Factors influencing fishing practices (e.g., ”Water temperature”)
Date (DATE) Specific dates relevant to fishing activities (e.g., ”March 2023”)
Time (TIME) Time-related information (e.g., ”Early morning”)
Countable
(CNT)
Quantitative data such as counts or measurements (e.g., ”10
nets”)
Concept
(CON)
Abstract concepts or ideas related to fishing (e.g., ”Sustainabil-
ity”)
Activity
(ACT)
Specific activities or actions related to fishing (e.g., ”Casting
nets”)
Cost (COST) Financial aspects or costs associated with fishing (e.g., ”Cost of
the boat”)
2: “Jabba is set at the lake bottom for catching fish.”
Entities: [{“entity”: “Jabba”,“label”: “ID”},{“entity”:
“lake bottom”,“label”: “LOC” },{“entity”: “for catching
fish”,“label”: “ACT” }, {“entity”: “fish”,“label”: “FISH” }]
D. Relationship Identification- This identifies rela-
tionships between key entities in a text, represented as
edges in Cypher queries for the Neo4j KG.
Initial Approach: The BERT model, fine-tuned with
OpenNRE struggled with low accuracy due to lim-
ited training data and predefined relations, resulting
in generic labels. TinyLlama was used to identify en-
tity relationships but often produced hallucinated, un-
structured outputs. Efforts to format results in JSON
increased inaccuracies, highlighting the need for a
more robust LLM.
Final Approach: Mistral 7B optimized with Unsloth
(Han and Han, 2023), enhanced accuracy in identify-
ing relationships between entities. Prompt engineer-
ing, as used with TinyLlama, structured outputs in
JSON format for seamless processing and integration.
Mistral 7B, extracted three relations from sen-
tences, {“node 1”: “Fisherfolk”, “node 2”: “Jabba”,
“relation”: “FISHING TOOL” }; {“node 1”: “Jabba”,
“node 2”: “lake bottom”, “relation”: “PLACED AT”
}; {“node 1”: “Jabba”, “node 2”: “fish”, “relation”:
“CATCHES” }, and two unique entities as a part of the
relations: {“entity”: “fisherfolk”,“label”: “UNKNOWN”
}; {“entity”: “jabba”, “label”: “UNKNOWN”}].
While Mistral 7B showcased substantial advance-
ments over others, future iterations will require fur-
ther refinement through prompt engineering and the
reduction of hallucinations.
E. Database Integration & KG Generation- After
extracting nodes, relationships, and labels, the next
step was to generate knowledge graphs by integrating
the Neo4j database, executing Cypher queries, and
visualizing the results. A cloud-based solution was
implemented to optimize data ingestion, query execu-
tion, and graph visualization. Below are the key steps:
Database Integration & Cypher Queries: Neo4j Au-
raDB was selected for its scalability, reliability, and
ability to efficiently handle large datasets and com-
plex queries. Cypher, a declarative language tailored
for graph databases, was used to query and update the
database. Two primary Cypher queries were devel-
oped: one created nodes with specified labels and en-
tities, ensuring no duplication. The relationships were
created only if they did not already exist.
Jabba
Lake_
bottom
Fisher
folk
Fish
CATCHES
Figure 4: Automated Knowledge Graph for Demo Exam-
ple.
KG Generation: The knowledge graph data was re-
trieved and visualised once the database was popu-
lated with nodes and relationships. Python’s Plotly
and NetworkX libraries were used for creating inter-
active and visually appealing graphs. Nodes, labels,
and relationships from AuraDB were processed to
generate an interactive graph with distinct colours for
node labels, annotated relationships, and individual
node visualizations for detailed insights. NetworkX
managed the graph structure, while Plotly rendered
the visuals. The demo example is shown in Figure 4.
F. System Integration - Google Colab was chosen
for its intuitive interface and seamless integration, en-
abling efficient execution. The pipeline optimized
data preprocessing, entity extraction, relation identi-
fication, and query generation, ensuring modularity
and scalability. Validation with Mistral 7B confirmed
correct formatting, data integrity, and JSON compli-
ance, allowing conversion into Python dictionaries for
efficient handling.
7 DISCUSSION
This pilot study created knowledge graphs from con-
versational data with fisherfolk in Bangladesh and In-
donesia. Feedback on the manually developed graphs
underscored their value in preserving orally transmit-
ted tacit knowledge. These KGs provide an innova-
tive way to document such knowledge by capturing
complex relationships and contextual nuances. Inte-
grating automated approaches with LLMs like Mis-
An Innovative Approach to Represent Tacit Knowledge of Fishing with Knowledge Graphs
505

tral 7B shows promise for scalability and efficiency,
enabling systematic preservation on a scale unattain-
able through traditional methods.
7.1 Technical Contributions
The successful implementation of this framework
serves as the ”back-end” module for a complete plat-
form aimed at preserving and sharing tacit knowl-
edge. Fisherfolk would communicate knowledge via
a user interface in the proposed platform, while the
back-end would extract knowledge components and
construct knowledge graphs. Future plans include de-
veloping a front-end interface and integrating it with
the back-end module to assess the feasibility of the
entire process, as visualised in Figure 5.
KGs effectively preserve tacit knowledge, en-
abling easy extraction via Cypher queries. Future
plans include an interface for fisherfolk to ask ques-
tions, automatically converting them into Cypher
queries for retrieving specific knowledge.
Front-end Back-end
Pre-processing KG
Segmentation
Fact Identification
Cypher Query
Figure 5: Proposed tacit knowledge management platform.
Currently, no LLMs, including spaCy and Mistral
7B, address the specific needs of the fishing occupa-
tion. Developing specialized LLMs could help fisher-
folk share knowledge efficiently and process queries
to retrieve information from knowledge graphs via
Cypher, improving usability and accessibility.
7.2 Societal & Broader Implications
This work’s societal impact extends beyond preserv-
ing knowledge and addressing key socio-economic
challenges fisherfolk face. Systematizing tacit knowl-
edge enables intergenerational transfer, with knowl-
edge graphs helping younger fisherfolk acquire tra-
ditional skills and bridging generational gaps amid
declining interest in the profession. Moreover,
the framework democratizes access to critical fish-
ing knowledge, reducing dependence on hierarchi-
cal knowledge-sharing networks and advancing dig-
ital inclusion by integrating traditional practices.
The knowledge graphs provide actionable insights
for policymakers and NGOs, aiding in designing tar-
geted interventions. For example, sustainable fishing
practices encoded in the graphs can inform environ-
mental conservation efforts while supporting liveli-
hood sustainability. The model’s adaptability also
makes it a blueprint for preserving and sharing tacit
knowledge in other marginalized communities, such
as those relying on indigenous agriculture or forestry.
This approach is crucial for preserving cultural
heritage and promoting sustainable practices facing
socio-economic and environmental changes. It also
advances digital equity by demonstrating how tech-
nologies like LLMs can serve marginalized commu-
nities, fostering inclusive technological progress.
7.3 Challenges
Focus group discussions and interviews were con-
ducted in the local language, transcribed, and trans-
lated into English, a process that can alter meanings.
To mitigate this, multiple researchers collaborated to
cross-check the translations.
The accuracy of the knowledge graphs generated
using automated algorithms on the Google Colab plat-
form was manually evaluated. However, the evalua-
tion was limited due to the small dataset from focus
groups and interviews, with plans to expand data col-
lection. While Mistral 7B outperformed its peers, it
occasionally produced hallucinated information, re-
sulting in undocumented entities and relationships.
Future efforts will focus on reducing such halluci-
nations through refined prompt engineering and en-
hanced model training with more labelled data.
8 CONCLUSION
This study introduces an innovative approach to pre-
serving tacit fishing knowledge for fisherfolks in
Bangladesh and Indonesia. Initially, knowledge was
manually identified from collected data, structured
into entities and relationships, and visualized as
knowledge graphs using Cypher queries on the Neo4j
platform. Encouraged by positive feedback on the
manually developed graphs, we experimented with
various LLMs to automate the process, following sim-
ilar steps and verifying outcomes at each stage.
Despite challenges encountered during devel-
opment, the automatically generated knowledge
graphs demonstrate significant potential for lever-
aging LLMs to help fisherfolk preserve their tacit
knowledge. This approach can also benefit other
marginalized communities in similar contexts. Fu-
ture plans include trialling the framework with fish-
erfolk in Bangladesh and Indonesia, addressing iden-
tified challenges, and refining the methodology.
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
506

ACKNOWLEDGEMENTS
Kanij and Grundy are supported by ARC Laureate
Fellowship FL190100035. Research funding is also
provided by the Empowerment Charitable Trust and
the Whyte Fund. We thank all the participants and
local collaborators.
REFERENCES
Chen, H. and Nunes, M. B. (2019). Retaining professional
tacit knowledge and evidence of experience through
electronic records management. IADIS International
Journal on Computer Science & Information Systems,
14(2).
Clarke, T. (2010). The Development of a Tacit Knowl-
edge Spectrum based on the Interrelationships be-
tween Tacit and Explicit Knowledge. PhD thesis,
Cardiff Metropolitan University.
Dalkir, K. (2013). Knowledge Management in Theory and
Practice. Routledge.
Das, M., Ray, S., Kumar, U., Begum, S., and Tarafdar,
S. (2015). Livelihood assessment of the fishermen
community in the South West region of Bangladesh.
Journal of Experimental Biology and Agricultural Sci-
ences, 3:353–361.
Fensel, D., S¸ims¸ek, U., Angele, K., Huaman, E., K
¨
arle, E.,
Panasiuk, O., Toma, I., Umbrich, J., and Wahler, A.
(2020). Introduction: What Is a Knowledge Graph?,
chapter 1, pages 1–10. Springer International Publish-
ing, Cham.
Grandinetti, R. (2014). The explicit dimension: What we
could not learn from polanyi. The Learning Organi-
zation, 21(5):333–346.
Han, D. and Han, M. (2023). Unsloth AI: Making AI ac-
cessible to everyone.
Hossain, I., Chamhuri, S., Mokhtar, M., Madan, M., Dey,
M., and Jaafar, A. (2009). Socio-economic condition
of fishermen in seasonal floodplain beels in Rajshahi
District, Bangladesh. Research Journal of Social Sci-
ences, 4:74–81.
Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C.,
Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel,
G., Lample, G., Saulnier, L., et al. (2023). Mistral 7b.
Kabir, K., Adhikary, R., Hossain, M. B., and Minar, M. H.
(2012). Livelihood Status of Fishermen of the Old
Brahmaputra River, Bangladesh. World Applied Sci-
ences Journal, 16:869–873.
Kanij, T., Anwar, M., Oliver, G., and Hossain, M. K.
(2023). Developing software for low socio-economic
end users: Lessons learned from a case study of fisher-
folk communities in bangladesh. In 2023 IEEE/ACM
45th International Conference on Software Engineer-
ing: Software Engineering in Society (ICSE-SEIS),
pages 96–107. IEEE.
Ladeinde, A., Arora, C., Khalajzadeh, H., Kanij, T., and
Grundy, J. (2023). Extracting queryable knowledge
graphs from user stories: An empirical evaluation. In
International Conference on Evaluation of Novel Ap-
proaches to Software Engineering 2023, pages 684–
692. Scitepress.
Leonard, D. and Sensiper, S. (1998). The role of tacit
knowledge in group innovation. California manage-
ment review, 40(3):112–132.
Mahmud, S., Ali, M. L., and Ali, M. M. (2015). Present
scenario on livelihood status of the fishermen in the
Paira River, Southern Bangladesh: Constraints and
Recommendation. International Journal of Fisheries
and Aquatic Studies, 2(4):23–30.
Miah, M. and Islam, M. (2020). Defending fishing space:
Conflicts and marginalization of small-scale fisheries
in the Saint Martin’s Island of Bangladesh. In Blue
Justice For Small-Scale Fisheries: A Global Scan,
pages 1–3. TBTI Global Publication Series, St. John’s,
NL, Canada.
Nonaka, I. (2009). The knowledge-creating company. In
The economic impact of knowledge, pages 175–187.
Routledge.
Nonaka, I. and Takeuchi, H. (1995). The Knowledge-
Creating Company: How Japanese Companies Cre-
ate the Dynamics of Innovation. Oxford University
Press.
Nowell, L. S., Norris, J. M., White, D. E., and Moules, N. J.
(2017). Thematic analysis: Striving to meet the trust-
worthiness criteria. International Journal of Qualita-
tive Methods, 16(1):1–13.
Pathirage, C., Amaratunga, D., and Haigh, R. (2008). Tacit
knowledge generation and utilisation in the construc-
tion industry: from process perspective. In Proceed-
ings of the RICS Construction and Building Research
Conference (COBRA 2008).
Polanyi, M. (2009). The tacit dimension. Routledge.
Prasetyo, T., Widodo, P., Saragih, H., Suwarno, P., and Said,
B. (2023). Poverty Reduction For Coastal Communi-
ties In Indonesia Through Community Empowerment
Training. International Journal Of Humanities Edu-
cation and Social Sciences (IJHESS), 2.
Sanderson, M. (2001). Records management and the cap-
ture of tacit knowledge. Records Management Jour-
nal, 11:7–17.
Shuva, N. Z. (2017). The Information Practices of the Fish-
ermen in the Bay of Bengal, Bangladesh. Open Infor-
mation Science, 1(1):21–40.
Supriati, A. and Umar, M. (2020). The role of coastal com-
munities to preserve local wisdom (a study in com-
munity of bunaken island, indonesia). In 2nd Annual
Civic Education Conference (ACEC 2019).
Tran, N., Rodriguez, U.-P., Chan, C. Y., Aung, Y. M., Chu,
L., Islam, A. H. M. S., Barman, B. K., and Phillips,
M. J. (2023). Future scenarios of fish supply and de-
mand for food and nutrition security in Bangladesh:
An analysis with the AsiaFish model. Aquaculture,
568:739288.
An Innovative Approach to Represent Tacit Knowledge of Fishing with Knowledge Graphs
507