Hierarchical Ontology Graph for Solving Semantic Issues in Decision

Support Systems

Hua Guo and Kecheng Liu

Informatics Research Centre, University of Reading, Reading, U.K.

Keywords: Ontology Graph, NLP, Knowledge Graph, Decision Support Systems, Semantic Composition,

Neural-symbolic Integration.

Abstract: In the context of the development of AI algorithms in natural language processing, tremendous progress has

been made in knowledge abstraction and semantic reasoning. However, for answering the questions with

complex logic, AI system is still in an early stage. Hierarchical ontology graph is proposed to establish

analysis threads for the complex question in order to facilitate AI system to further support in business

decision making. The study of selecting the appropriate corpora is intended to improve the data asset

management of enterprises.

1 INTRODUCTION

Strategic questions are often formulated as openly as

possible in order to stimulate more considerations

involving different perspectives of business

operations. Those questions are not easy to answer

and usually demand a great deal of effort of analysis

before they can be addressed adequately. In order to

achieve the goal to help managers identify the hidden

impact factors of decision making, the enormous

academic explorations over query understanding

(Moldovan et al., 1999), information retrieval and

process (Harman,1993) or over heterogeneous data

sources (Kumar et al., 2014) have been ongoing for

years. Along with the solutions of key issues for

knowledge engineering, e.g.: semantic parsing (Cai

and Yates, 2013), knowledge graph (Kwiatkowski et

al., 2013), have been proposed in recent years, the

open-domain question answering system is becoming

one of the most important applications in AI-NLP

arenas. Most state-of-the-art searches adopts the

bottom-up approach, which firstly classifies the

question into a few classes (Moldovan et al.,1999; Li

and Roth, 2002), secondly retrieve the relative

information (Stoyanchev et al, 2008), and finally

extract the answer from the relative documents

(Ravichandran and Hovy, 2002). The purpose of the

question processing stage is to narrow down the

search scope and it focuses on question classification

(Allam and Haggag, 2012). For the open-domain QA

system, which relies on universal ontology and

information, the bottom-up approach handles well.

However, the closed-domain questions, like the

strategic questions related to the business operation,

usually consist of multiple aspects and contain

complex judgment logic. It is often the case that two

of the keywords in completely unrelated fields are

logically linked when answering a particular question.

This really is a challenge for AI algorithms to abstract

the tacit relationship from the limited corpus. So, how

can these AI techniques help in answering strategic

questions? In this paper, a top-down approach of

hierarchical ontology graph is proposed, which is

starting from the question analysis. This approach

will embed the logic of business operation and the

collaborative relationships of departments in the

organization into the procedure of decomposing user

queries into sub-questions. In other words, this

approach focuses on question reformulation (Allam

and Haggag, 2012) to understand the query in an

enterprise context and transform the complex logic

question into a few simple logical questions to

empower the search engine. From a practical point of

view, the hierarchical ontology graph is a visualized

procedure of decision-making, and the threads of

analysing the strategic questions could be mapped to

the entities on the hierarchical ontology graph.

Hierarchical ontology graph, rooted in the

knowledge graph (Singhal, 2012), illustrates complex

question from the relevant aspects, breaks one general

question into several domain-level sub-questions, and

Guo, H. and Liu, K.

Hierarchical Ontology Graph for Solving Semantic Issues in Decision Support Systems.

DOI: 10.5220/0007769904830487

In Proceedings of the 21st International Conference on Enterprise Information Systems (ICEIS 2019), pages 483-487

ISBN: 978-989-758-372-8

483

gradually decomposes the domain-level questions

layer by layer until the sub-questions can be answered

by the existing databases/documents. The reasoning

procedure will display on a hierarchical ontology

graph.

2 STATEMENT OF POSITION

A thorough understanding of the company

operational details is a prerequisite for decision

making. However, those details dispersed among

organizations and isolated within domains. The

routine approach of an executive making decision is

to call a meeting involving the heads of all

departments (domain experts), who have the ability to

interpret or decompose the strategic question into

sub-questions in domain-level, and their subordinates

can decompose these sub-questions into queries and

bridge these queries with the existing database or

documents. With the help of frontline staff or data

analysts, these queries will be answered, and bringing

together the answers to these queries, the

subordinates can answer sub-questions to their

department head. As these sub-questions are

answered individually and then aggregated, finally

the strategic question could be answered. This is the

normal procedure of an executive makes a strategic

decision, which is not only inefficient but also

significantly impacted by the personal experience of

domain experts, reflected in the ability to interpret the

question. Even more, interest disputes are also likely

to arise between departments due to the knowledge

barriers.

Hierarchical ontology graph can speed up this

process and can bridge the knowledge islands. It can

break down complex problem layers into simple

questions that can be answered by existing data

sources, and it can also provide an enterprise-wide

knowledge graph that effectively eliminates

knowledge barriers between departments. A shared

and reusable knowledge graph is an effective tool to

construct an agile organization in the quickly

evolving business competitive environment. It will

help on rapidly engaging in multidirectional

communication and complex collaboration.

3 PROPOSED SOLUTION

The proposed hierarchical ontology graph consists of

four different levels, shown as in Figure 1. The

application ontology is constructed by the data of

department-level, which theoretically docks with the

data warehouse, indicating the analysis results and

reports of the current business performance. Task

ontology is a department-level knowledge graph

abstracted from the internal corpora, which refer to

the documents of enterprise processes and operational

logics. Considering the number of relevant internal

documents is comparatively limited, it is highly

recommended to annotate these documents as much

as possible in order to apply the full supervision or

semi supervision learning algorithms. Different to

task ontology, domain ontology is built on the

industry professional corpora, which may or may not

be well trained. In order to get the enterprise-level

knowledge graph, the outcomes got from training the

semantically-rich annotated corpora in department

level can be used to do semi supervision training. The

semi-automatic tools, like ONION (Mitra et al.,

2000), are also recommended to bridge department-

level ontologies in the procedure of creating domain

ontologies. The top layer in the figure is a top-level

ontology, which includes some tacit knowledge.

These inexplicit factors could be abstracted from the

business activities based on the strategic management

theory and may vary in industries.

Figure 1: Hierarchical ontology graph.

Considering that the number of the corpus in a

specific domain is usually very limited in practice,

deep learning algorithms, which is based on a large-

scale corpus, often find the difficulty to obtain high-

quality training results. The latest research direction

of neural-symbolic integration network (Garcez et al.,

2008) is trying to combine the strengths of connective

and symbolic paradigms to enhance the ability of

machine learning and reasoning. The framework of

Object-oriented Neural Programming (OONP) (Lu et

al., 2017) is proposed for semantically parsing

documents in specific domains, which leverage the

advantages of reasoning feature of the symbolic

network to construct object-oriented ontology in the

process of text comprehension. The OONP

framework provides another approach to constructing

ICEIS 2019 - 21st International Conference on Enterprise Information Systems

484

the hierarchical ontology graph, furthermore, the

design of carry-on memory (Lu et al., 2017) model

can effectively store and reuse the prior knowledge,

which can be used in dealing with the existing

business logic in enterprise ontology graph. Evans

and Grefenstette (2018) proposed a logic

programming method based on Inductive Logical

Programming (ILP) to reason on symbolic domains,

which could effectively support the reasoning

function of a hierarchical ontology graph.

The following example will explain how a

hierarchical ontology graph support decision making

on a specific strategic question.

Figure 2: Decomposing a complex question over

hierarchical ontology graph.

Suppose the CEO of an e-commerce company is

considering adding a new category of middle-aged

clothing. His question is “Whether should we add a

category of middle-aged clothing?”. Cost and income

might be the first two key factors to be taken into

consideration because all the business activities

revolve around the purpose of making a profit. So,

primarily this business-related question type can be

decoded into a ‘quantitative’ question in terms of cost

and potential income. Furthermore, profitability is not

the only indicator of business decision-making, but

also the need to take into account the company's long-

term strategic plan. Feasibility is another dimension

to evaluate the possibility of success of the project,

and it is listed as a separating factor in the top-level

ontology as well.

At the second level of the ontology graph, the four

aspects sub-questions have been allocated to the

different domains. They are HR department, R&D

department, purchasing department, and marketing

department. In practice, entities at this level are often

closely related to the organizational structure.

At the third level of the ontology graph, the same

indicator may appear in different domains, such as

“supplier selecting” appears in both “cost” and

“feasibility”. The same indicator can map to multiple

domains and also one concept can be interpreted into

different meanings. In this case, “cost” has different

implications in HR, purchasing, R&D and marketing

departments. There are 1-to-N and N-to-1 tree-like

relationships and even can emerge N-to-N network

relationships in some cases. At this point, at the third

level, a complex strategic issue has been broken down

into a variety of queries belonging to various

departments.

The fourth level of the ontology graph does not

show in Figure 2, which are the answers to the queries

in the third level and got from the data analysis of the

existing database.

To build an enterprise-level knowledge graph,

selecting appropriate corpora is one of the most

important tasks besides selecting algorithms or

models. Corpus selection will be one of the future

research directions of this research.

The above description presents a decomposition

process of how hierarchical ontology graph performs

on complex questions, and AI algorithms will help on

constructing the hierarchical ontology graph.

Artificial intelligence is changing the business

landscape, especially with the development of weak

supervision learning and self-learning neural in NLP.

Usually, the result is highly correlated with the

quality of the corpus. Comparing with the general

linguistic corpus, the corpus in a business context has

clear boundaries and are comparatively simple in

terms of less ambiguity, polysemy, and vagueness

issues. The workload of getting a semantically-rich

annotated corpora is manageable, which is a crucial

impact factor of the computing result.

For the corpus with rich semantic annotations, the

full supervision training models can be applied, e.g.:

DeepCoder (Balog, 2016), NPI (Reed & Freitas,

2015) and Seq2Tree (Dong & Lapata, 2016). To train

the small volume of databases, the end-to-end models

like Neural Programmer (Neelakantan et al, 2015)

and Neural Turing Machines (Graves et al, 2014) can

be adopted. For those context-sensitive processing,

LSTM (Hochreiter & Schmidhuber 1997), Seq2Seq

(Sutskever et al, 2014), named entity recognition

(Lample et al, 2016), and reading Comprehension

models (Yu et al., 2018) can be considered. Those

NLP related algorithms provide the feasible methods

for concepts extraction from text and from semi-

structured tables (Pasupat & Liang 2015).

The complex question defined in this paper

referrers to a question that contains multiple

independent variables, rather than the complex syntax

logic of the sentence. These independent variables are

often difficult to obtain from ready-made documents,

and they exist in the minds of domain experts in the

Hierarchical Ontology Graph for Solving Semantic Issues in Decision Support Systems

485

form of knowledge or experience. The human

knowledge composed in different forms, including

tree-like relationships, two-dimensional grid

relationships, single dimensional sequential

relationships and directed grid relationships (Kemp

and Tenenbaum, 2009). The classical TransE (Bordes

et al., 2013) model and its derivations are not strong

enough to present these cognitive models from a

mathematical approach. This is the reason why a

symbolic network will be considered to do

relationship reasoning and neural networks will focus

on learning and information extraction. The

hierarchical ontology graph proposes an approach to

building the close-domain question and answering

system by leveraging the prior experiences to support

decision making.

4 SUGGESTED COURSES OF

ACTION

The hierarchical ontology graph is proposed to solve

semantic issues through injecting business operation

logic and the experiences of domain experts to

support executives to make strategic decisions. The

procedure of constructing an enterprise-level

ontology graph is also the process of establishing the

organizational knowledge graph. A unified

knowledge graph can not only help on decision

making but also be the basis for efficient business

operation. Further research will include the following

aspects:

1. Enterprise Semantic Model: constructing the

abductive reasoning model for decision support

2. Algorithm: selecting appropriate algorithms to

match the requirements for semantic analysis

3. Corpus acquirements: working out which types of

documents in an enterprise can be trained as

corpus

4. Tacit knowledge transfer: visualizing the tacit

enterprise experience in a hierarchical ontology

graph.

REFERENCES

Allam, A.M.N. and Haggag, M.H., 2012. The question

answering systems: A survey. International Journal of

Research and Reviews in Information Sciences

(IJRRIS), 2(3).

Balog, M., Gaunt, A.L., Brockschmidt, M., Nowozin, S.

and Tarlow, D., 2016. Deepcoder: Learning to write

programs. arXiv preprint arXiv:1611.01989.

Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J. and

Yakhnenko, O., 2013. Translating embeddings for

modeling multi-relational data. In Advances in neural

information processing systems (pp. 2787-2795).

Cai, Q. and Yates, A., 2013. Large-scale semantic parsing

via schema matching and lexicon extension. In

Proceedings of the 51st Annual Meeting of the

Association for Computational Linguistics (Volume 1:

Long Papers) (Vol. 1, pp. 423-433).

D. K. Harman. Overview of the first Text REtrieval

Conference (TREC-1). In Proceedings of the First Text

REtrieval Conference (TREC-1), pages 1–20. NIST

Special Publication 500-207, March 1993.

Dong, L. and Lapata, M., 2016. Language to logical form

with neural attention. arXiv preprint arXiv:1601.01280.

Evans, R. and Grefenstette, E., 2018. Learning explanatory

rules from noisy data. Journal of Artificial Intelligence

Research, 61, pp.1-64

Garcez, A.S.A., Lamb, L.C. and Gabbay, D.M., 2008.

Neural-symbolic cognitive reasoning. Springer Science

& Business Media.

Graves, A., Wayne, G. and Danihelka, I., 2014. Neural

turing machines. arXiv preprint arXiv:1410.5401.

Hochreiter, S. and Schmidhuber, J., 1997. Long short-term

memory. Neural computation, 9(8), pp.1735-1780.

Kemp, C. and Tenenbaum, J.B., 2009. Structured statistical

models of inductive reasoning. Psychological review,

116(1), p.20.

Kwiatkowski, T., Choi, E., Artzi, Y. and Zettlemoyer, L.,

2013. Scaling semantic parsers with on-the-fly

ontology matching. In Proceedings of the 2013

conference on empirical methods in natural language

processing (pp. 1545-1556).

Lample, G., Ballesteros, M., Subramanian, S., Kawakami,

K. and Dyer, C., 2016. Neural architectures for named

entity recognition. arXiv preprint arXiv:1603.01360.

Li, X. and Roth, D., 2002, August. Learning question

classifiers. In Proceedings of the 19th international

conference on Computational Linguistics-Volume 1

(pp. 1-7). Association for Computational Linguistics.

Liu Zhiyuan, Sun Maosong, Lin Yankai, et al., 2016.

Knowledge Representation Learning: A Review[J].

Journal of Computer Research and Development,

53(2): 247-261.

Lu, Z., Liu, X., Cui, H., Yan, Y. and Zheng, D., 2017.

Object-oriented neural programming (oonp) for

document understanding. arXiv preprint

arXiv:1709.08853.

Mitra, P., Wiederhold, G. and Kersten, M., 2000, March. A

graph-oriented model for articulation of ontology

interdependencies. In International Conference on

Extending Database Technology (pp. 86-100).

Springer, Berlin, Heidelberg.

Moldovan, D., Harabagiu, S., Pasca, M., Harabagiu, A.,

Mihalcea, R., Girju, R., Goodrum, R. and Rus, V.,

1999. Lasso: A tool for surfing the answer net.

Neelakantan, A., Le, Q.V. and Sutskever, I., 2015. Neural

programmer: Inducing latent programs with gradient

descent. arXiv preprint arXiv:1511.04834.

ICEIS 2019 - 21st International Conference on Enterprise Information Systems

486

Panupong Pasupat and Percy Liang. 2015. Compositional

semantic parsing on semi-structured tables. In

Proceedings of the 53rd Annual Meeting of the

Association for Computational Linguistics and the 7th

International Joint Conference on Natural Language

Processing (Volume 1: Long Papers), pages 1470–

1480.

Ravichandran, D. and Hovy, E., 2002, July. Learning

surface text patterns for a question answering system.

In Proceedings of the 40th annual meeting on

association for computational linguistics (pp. 41-47).

Association for Computational Linguistics.

Reed, S. and De Freitas, N., 2015. Neural programmer-

interpreters. arXiv preprint arXiv:1511.06279.

Singhal, A ‘‘Introducing the knowledge graph: Things, not

strings,’’ May 2012. [Online]. Available:

https://www.blog.google/products/search/introducing-

knowledge-graph-things-not/.

Suresh kumar, G., Zayaraz, G., 2014. Concept relation

extraction using Naıve Bayes classifier for ontology-

based question answering systems. J. King Saud Univ.

Sutskever, I., Vinyals, O., and Le, Q. V. (2014). Sequence

to sequence learning with neural networks. Technical

report. arXiv:1409.3215 [cs.CL] Google. NIPS’2014.

Stoyanchev, S., Song, Y.C. and Lahti, W., 2008, August.

Exact phrases in information retrieval for question

answering. In Coling 2008: Proceedings of the 2nd

workshop on Information Retrieval for Question

Answering (pp. 9-16). Association for Computational

Linguistics.

Yih, S.W.T., Chang, M.W., He, X. and Gao, J., 2015.

Semantic parsing via staged query graph generation:

Question answering with knowledge base.

Yu, Adams, Wei, David Dohan, Quoc Le, Thang Luong,

Rui Zhao, and Kai Chen. 2018. Fast and accurate

reading comprehension by combining self-attention and

convolution. In International Conference on Learning

Representations.

Hierarchical Ontology Graph for Solving Semantic Issues in Decision Support Systems

487