Prompt-Time Ontology-Driven Symbolic Knowledge Capture

with Large Language Models

Tolga Çöplü, Arto Bendiken, Andrii Skomorokhov, Eduard Bateiko, Stephen Cobb
Haltia, Inc.
{tolga, arto, andriy, eduard, steve}@haltia.ai
arXiv:2405.14012v1 [cs.AI] 22 May 2024

Abstract
In applications such as personal assistants, large language models (LLMs) must consider the user’s
personal information and preferences. However, LLMs lack the inherent ability to learn from
user interactions. This paper explores capturing personal information from user prompts using
ontology and knowledge-graph approaches. We use a subset of the KNOW ontology, which models
personal information, to train the language model on these concepts. We then evaluate the success of
knowledge capture using a specially constructed dataset. Our code and datasets are publicly available
at https://github.com/HaltiaAI/paper-PTODSKC

Keywords ontology-driven symbolic knowledge capture · KNOW ontology · symbolic representation · knowledge
graphs · large language models · fine-tuning

1 Introduction
Currently, many generative artificial intelligence (AI) applications, particularly personal assistants, strive to offer users
personalized experiences. To achieve this, AI applications must learn personal information and preferences from
user interactions (knowledge capture) and use this learned knowledge in future conversations (knowledge utilization).
Implementing this fundamental personal AI approach depends on addressing several complex sub-problems, such as
discerning which user prompt information is personal, extracting it, determining whether the extracted information is
duplicate, and figuring out its association with other personal data.
These challenges have been the focus of extensive research within the AI field for many years. However, the emergence
of neuro-symbolic approaches through the collaboration between large language models (LLMs) and symbolic AI
has provided researchers with new perspectives [1, 2, 3, 4]. LLMs’ capabilities in natural language processing can be
integrated with the representational and factual reasoning abilities of knowledge graphs, enhanced by the structure,
rules, and inference mechanisms offered by an ontology. For targeted personal AI applications, this ontology approach
presents several benefits:
• Ontology schemas enable language models to determine which personal information will be captured and how
it will be associated with other captured knowledge.
• Ontology rules can help identify inconsistencies in the captured knowledge, allowing for validation before
storage.
• Ontology relationships allow the extraction of implicit information from captured knowledge, effectively
enabling automatic inference that expands the knowledge graph.
• A robust, personalized knowledge graph forms a reliable foundation for facilitating personalized interactions
with the application through language models.
In this paper, we address a specific aspect of the AI personalization challenge by focusing on prompt-time, ontology-
driven symbolic knowledge capture using language models. We explore the extraction from user prompts of subject-
predicate-object triples1 that conform to a specified ontology. We have investigated various methods to enable the
1
https://www.w3.org/TR/rdf12-concepts/
underlying language model to comprehend a pre-defined ontology, ensuring effective symbolic knowledge capture. By
utilizing a specially designed dataset, we evaluate the effectiveness of these methods, emphasizing their strengths and
identifying potential areas for improvement.
The structure of this paper is as follows: Section 2 discusses in-context learning and fine-tuning approaches for
ontology-driven symbolic knowledge capture and focuses on the details of the fine-tuning approach. Section 3 describes
the experimental setup by presenting the development framework, the language model selection, and the ontology and
dataset creation process. Section 4 outlines our performance evaluation framework and the test results. Finally, Section
5 concludes the paper and suggests future directions.

2 Ontology-Driven Symbolic Knowledge Capture

In the literature, language models have demonstrated their capability to transform unstructured text into a knowledge
graph [5, 6, 7, 8, 9]. However, the process of populating a knowledge graph from user prompts in alignment with a
pre-defined ontology has been explored only marginally [10, 11, 12, 13, 14]. Research typically centers on in-context
learning, which heavily relies on prompt engineering. A significant limitation of this approach is the requirement to
incorporate the entire custom ontology into the prompt. This necessity not only slows down the knowledge capture
process, because of the high token overhead but also restricts the use of larger ontologies due to the constraint on
context-window length. Given these constraints, in-context learning methods do not provide a scalable solution for
ontology-driven symbolic knowledge capture.
An alternative approach involves training a language model with a pre-defined ontology, so that the model internalizes
it. There are two strategies to consider: pre-training the LLM on the ontology or fine-tuning it. This paper does not
explore pre-training due to its extensive data, computational resources, energy, and time requirements. Additionally,
pre-training does not offer a flexible response to ongoing changes or expansions in the ontology. Therefore, this paper
will focus on fine-tuning as a method to train language models on personal ontologies, highlighting advantages in
feasibility and maintainability

2.1 Ontology-Driven Knowledge Capture with Fine-Tuning

Fine-tuning is a process whereby a pre-trained language model is further trained on a specific dataset to tailor its
capabilities to a particular task. In our study, the language model is expected to learn the classes, object properties, and
data properties defined in an ontology, and to use them to populate a knowledge graph from user prompts. The first step
involves preparing a fine-tuning dataset, which includes user prompts, system prompts, and expected model responses
for each concept in the ontology. This dataset is used to fine-tune the language model, which is then evaluated by testing
it with new prompts to assess the effectiveness of the knowledge capture operation. We define a system prompt for this
task with the requirement of maintaining the model’s generality across other tasks.
The following points highlight the key aspects of ontology fine-tuning:
• The training dataset’s coverage and diversity are vital for successful fine-tuning. These characteristics greatly
influence the LLM’s ability to effectively capture knowledge. Details about the dataset and how it is constructed
are discussed in Section 3.4.
• The training dataset must include a variety of examples for each element of the predefined ontology. This
approach avoids scalability issues typically associated with in-context learning and ensures comprehensive
learning coverage.
• If the LLM encounters a user prompt that is not relevant to the predefined ontology concepts, it should not
attempt to capture knowledge. Therefore, the dataset should also contain sufficient out-of-context samples to
enable the LLM to distinguish between relevant and irrelevant information for capture.

3 Experimental Setup
This section explores the components of our experimental setup.

3.1 Development Framework

The methods suggested in this paper have been implemented using the Apple MLX framework ([15]). MLX is a
specialized array framework designed for machine learning applications, akin to NumPy, PyTorch, or JAX, with the
distinction of being exclusive to Apple silicon.

2
Ontology fine-tuning has been conducted using the parameter-efficient QLoRA approach ([16]) on our custom dataset,
comprising randomly selected, non-overlapping sets of training, validation, and test samples.

3.2 Language Model

The methods we have developed here do not have a structural dependency on a particular underlying foundation model.
The key factors guiding our language model selection were its proven effectiveness across diverse domains in community
benchmarks and its prevalence in the field. Owing to its performance in the Hugging Face Open LLM Leaderboard
([17]) and its robust ecosystem, the Mistral-7B-Instruct-v0.2 ([18]), based on the Llama 2 ([19]) architecture, was
selected for our research. We ran all examples, tests, and benchmarks on the MLX 4-bit quantized version of this model
to be able to run the tests on personal laptops.

3.3 Predefined Ontology

Our study utilized the Knowledge Navigator Ontology for the World (KNOW)2 for representing personal information.
KNOW is the first ontology crafted specifically to capture everyday knowledge, thereby enhancing language models
in practical real-world scenarios. The domain of KNOW encompasses human life, addressing both daily concerns
and significant life milestones. The initial scope of KNOW’s modeled concepts is currently limited to established
human universals that are not subject to revision: spacetime (places, events) and social elements (people, groups,
organizations).
Due to the requirement that each element in the ontology be associated with a diverse set of prompt and response
samples within the training dataset, our research focuses on a specific subset of the KNOW ontology. This subset
concentrates on core family relationships with four ontology classes, eleven object properties, and one data property. A
visual depiction of this subset is presented in Figure.1.

Figure 1: A visual representation of the ontology design used in this paper.

3.4 Dataset

For a language model to effectively learn a predefined ontology and use it to perform knowledge extraction and capture,
a robust and diverse training dataset is essential. Our paper focuses on a subset of the KNOW ontology that includes
the concepts of ’person’, ’name’, ’sex’, ’child’, ’father’, ’mother’, ’sibling’, ’sister’, ’brother’, ’spouse’, ’partner’ and
’knows’. We created 143 manually crafted user prompts along with their respective ontology responses for training and
tests. Additionally, to manage inputs that fall outside these ontology concepts, we included 32 generic user prompts in
the dataset. The composition of this dataset, which consists of 175 user prompts, is illustrated in Figure.2. Concepts not
associated with the ontology are labeled as the ’none’ legend in the figure. As each sample prompt typically contains
multiple modeled concepts, the chart shows a total number of concept occurrences greater than the number of prompts.
The Turtle format was chosen for serializing the ontology population in our research because of its straightforward
structure, readability, and prevalent use in existing pre-training datasets for LLMs.
2
https://know.dev

3
 Occurrences of Ontology Concepts

Person
Name
Sex
Child
Father
Mother
Sibling
Sister
Brother
Spouse
Partner
Knows
None

0 100 200 300

Number of Occurrences

Figure 2: Occurrences of ontology concepts in the prepared dataset.

4 Performance Evaluation

Our research focuses on fine-tuning a language model with predefined ontology concepts and capturing knowledge from
user prompts that fits the ontology. This section will detail the performance evaluations associated with these efforts.

Occurrences of Ontology Concepts

Person
Name
Sex
Child
Father
Mother
Sibling
Sister
Brother
Spouse
Partner
Knows
None

0 20 40 60

Number of Occurrences

Figure 3: Occurrences of ontology concepts in the test dataset.

Initially, we investigated how many training samples each ontology concept required to effectively teach the core family
ontology to the selected Mistral-7B-Instruct-v0.2 model. From the dataset described in Section3.4, 41 random samples
were selected and reserved for method evaluation. The distribution of ontology concepts within this test set is shown in
Figure.3.
From the remaining 134 samples, we created three different training datasets. In these three datasets, we have ensured
the inclusion of 2, 4, and 8 sample prompts for the ontology concepts ’child’, ’father’, ’mother’, ’sibling’, ’sister’,
’brother’, ’spouse’, ’partner’, and ’knows’. Subsequent evaluations using the test set measured the precision, recall, and
f1-scores for each fine-tuning session. During these evaluations, the generated prompt responses were processed triple
by triple and compared against the ground truth established for the test set. The findings are displayed in Figure.4.

4
 Performance vs. Samples per Ontology Concept
Precision Recall F1-Score

1.00

0.75

0.50

0.25

0.00
2 samples/concept 4 samples/concept 8 samples/concept entire training dataset

Figure 4: Ontology population performance (precision, recall, and f1-score) for various fine-tuning dataset sizes.

The tests were conducted using the default QLoRA hyperparameters specified in the MLX framework. To ensure
consistent test results, each training session was set to run for 18 epochs.

• Layer keys to apply: ["self_attn.q_proj", "self_attn.v_proj"]

• Rank: 8
• Alpha: 16
• Scale: 10
• Optimizer: Adam
• Learning rate: 1x10−5
• Number of layers to fine-tune: 16
• Minibatch size: 4

Performance vs. Epochs of Training

Precision Recall F1-Score

1.00

0.75

0.50

0.25

0.00
9 training epochs 18 training epochs 27 training epochs 36 training epochs

Figure 5: Ontology population performance (precision, recall, and f1-score) of 8 samples per ontology concept for
various epochs.

As illustrated in Figure.4, our study, which encompasses twelve key ontology concepts, demonstrates that providing
eight diverse examples for each concept yields acceptable success rates. Although our training and test datasets are not

5
sufficiently diverse or large enough to generalize the results to real user scenarios, the high success achieved with a
small number of training samples is promising for the feasibility of the proposed approach.
In the subsequent phase, we explored the optimal number of training epochs required to achieve maximum performance
for the training set. For this analysis, we continued using the default MLX QLoRA hyperparameters with the 8 samples
per ontology concept, but trained the QLoRA adapter over various epoch lengths. We then conducted evaluations on the
test set using each trained adapter, and the findings are presented in Figure.5.
As depicted in Figure.5, the success rate of the ontology population increases with longer training. However, considering
the resource usage and energy consumption, we observe that 18 epochs are sufficient for fine-tuning.

5 Conclusion
In this paper, we first explored prompt-driven, ontology-based symbolic knowledge capture and its importance in the
generative AI domain. We then discussed the ontology approach and how to teach ontology concepts to the language
model through in-context learning and training. The language model was fine-tuned using a custom dataset focused on
core family relationships, and we evaluated the model’s ability to learn personal ontology concepts.
Our findings indicate that fine-tuning is particularly effective for teaching ontology concepts to a language model for
prompt-time knowledge capture. In our future work, we aim to integrate the generated knowledge graph with the
language model for knowledge utilization, combining the strengths of the neural and symbolic AI approaches.
Please refer to the paper’s corresponding GitHub repository at https://github.com/HaltiaAI/paper-PTODSKC

References
[1] Amit Sheth, Kaushik Roy, and Manas Gaur. Neurosymbolic AI – Why, What, and How, May 2023.
arXiv:2305.00813 [cs].
[2] Lauren Nicole DeLong, Ramon Fernández Mir, and Jacques D. Fleuriot. Neurosymbolic AI for Reasoning over
Knowledge Graphs: A Survey, May 2024. arXiv:2302.07200 [cs, stat].
[3] Fajar J. Ekaputra, Majlinda Llugiqi, Marta Sabou, Andreas Ekelhart, Heiko Paulheim, Anna Breit, Artem Revenko,
Laura Waltersdorfer, Kheir Eddine Farfar, and Sören Auer. Describing and Organizing Semantic Web and Machine
Learning Systems in the SWeMLS-KG, March 2023. arXiv:2303.15113 [cs].
[4] Lars-Peter Meyer, Claus Stadler, Johannes Frey, Norman Radtke, Kurt Junghanns, Roy Meissner, Gordian Dziwis,
Kirill Bulert, and Michael Martin. LLM-assisted Knowledge Graph Engineering: Experiments with ChatGPT,
July 2023. arXiv:2307.06917 [cs].
[5] Qipeng Guo, Zhijing Jin, Xipeng Qiu, Weinan Zhang, David Wipf, and Zheng Zhang. CycleGT: Unsupervised
Graph-to-Text and Text-to-Graph Generation via Cycle Training, December 2020. arXiv:2006.04702 [cs].
[6] Xiaoya Li, Fan Yin, Zijun Sun, Xiayu Li, Arianna Yuan, Duo Chai, Mingxin Zhou, and Jiwei Li. Entity-Relation
Extraction as Multi-Turn Question Answering, September 2019. arXiv:1905.05529 [cs].
[7] Yi Xu, Luoyi Fu, Zhouhan Lin, Jiexing Qi, and Xinbing Wang. INFINITY: A Simple Yet Effective Unsupervised
Framework for Graph-Text Mutual Conversion, September 2022. arXiv:2209.10754 [cs].
[8] Raghu Anantharangachar, Srinivasan Ramani, and S. Rajagopalan. Ontology Guided Information Extraction
from Unstructured Text. International journal of Web & Semantic Technology, 4(1):19–36, January 2013.
arXiv:1302.1335 [cs].
[9] Tolga Çöplü, Arto Bendiken, Andrii Skomorokhov, Eduard Bateiko, Stephen Cobb, and Joshua J. Bouw. Prompt-
Time Symbolic Knowledge Capture with Large Language Models, February 2024. arXiv:2402.00414 [cs].
[10] Hamed Babaei Giglou, Jennifer D’Souza, and Sören Auer. LLMs4OL: Large Language Models for Ontology
Learning, August 2023. arXiv:2307.16648 [cs, math].
[11] Patricia Mateiu and Adrian Groza. Ontology engineering with Large Language Models, July 2023.
arXiv:2307.16699 [cs].
[12] Keming Lu, Xiaoman Pan, Kaiqiang Song, Hongming Zhang, Dong Yu, and Jianshu Chen. PIVOINE: Instruction
Tuning for Open-world Information Extraction, May 2023. arXiv:2305.14898 [cs].
[13] Nandana Mihindukulasooriya, Sanju Tiwari, Carlos F. Enguix, and Kusum Lata. Text2KGBench: A Benchmark
for Ontology-Driven Knowledge Graph Generation from Text, August 2023. arXiv:2308.02357 [cs].

6
[14] Maurice Funk, Simon Hosemann, Jean Christoph Jung, and Carsten Lutz. Towards Ontology Construction with
Language Models, September 2023. arXiv:2309.09898 [cs].
[15] Awni Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. MLX: Efficient and flexible machine
learning on Apple silicon. https://github.com/ml-explore, 2023.
[16] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient Finetuning of Quantized
LLMs, May 2023. arXiv:2305.14314 [cs].
[17] Edward Beeching, Clementine Fourrier, Nathan Habib, Sheon Han, Nathan Lambert, Nazneen Rajani, Omar
Sanseviero, Lewis Tunstall, and Thomas Wolf. Open LLM Leaderboard. https://huggingface.co/spaces/
HuggingFaceH4/open_llm_leaderboard, 2023.
[18] Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las
Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne
Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed.
Mistral 7B, October 2023. arXiv:2310.06825 [cs].
[19] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov,
Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya
Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao,
Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas,
Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux,
Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar
Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan
Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor,
Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie
Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2:
Open foundation and fine-tuned chat models, 2023.