=Paper=
{{Paper
|id=Vol-2307/DP2
|storemode=property
|title=Formula Concept Discovery and Recognition
|pdfUrl=https://ceur-ws.org/Vol-2307/DP2.pdf
|volume=Vol-2307
|authors=Philipp Scharpf
|dblpUrl=https://dblp.org/rec/conf/mkm/Scharpf18
}}
==Formula Concept Discovery and Recognition==
https://ceur-ws.org/Vol-2307/DP2.pdf
Formula Concept Discovery and Recognition
Philipp Scharpf
Dept. of Computer and Information Science
Konstanz, Germany
philipp.scharpf@uni-konstanz.de
Abstract
In my dissertation, I will develop a method to discover (define) and recognize (identify) formula concepts in
Wikipedia articles and STEM documents using Wikidata as a semantic knowledge-base. Both structural (syntax
tree) and semantic (identifier names) formula information will be considered. The approach is expected to
improve search engines, recommender systems, plagiarism and novelty detection and ontology learning.
Research Motivation
My research is motivated by 1) the need for Information Retrieval systems to match mathematical formulae when
assessing semantic content and similarity of STEM documents, and 2) the challenge that a given mathematical
formula concept usually appears in several variations or equivalent representations.
Figure 1: Various representations of the Klein-Gordon equation from arbitrarily selected sources (left)
and LaTeX document annotation system ”AnnotaTeX” (right).
Figure 1 (left) illustrates the observation that a selected formula concept - here the Klein-Gordon-equation from
quantum physics - appears in the literature in a variety of different but equivalent representations. In numerous
publications, it is at least slightly different from the rest. To make such a formula machine-interpretable, i.e.,
teach the computer how to understand its semantics, there needs to be a definition of the mathematical concept
that comprises as many variations as possible.
Research Plans
My research goals are 1) development and optimization of Formula Concept Discovery to elaborate a definition of
a formula concept by examining Wikipedia articles and STEM documents from the ArXiV, 2) seeding a scalable
large number to Wikidata with suitable annotation of the formula parts (identifiers and operators) and as a
preparation for 3) the development and optimization of Formula Concept Recognition techniques that match a
given formula to a Wikidata concept item, independently from its specific representation.
Databases
arXiv, Wikipedia, and Wikidata In my research, I will focus on examples from mathematics and physics
Copyright c by the paper’s authors. Copying permitted for private and academic purposes.
In: O. Hasan, D. Gallois-Wong (eds.): Proceedings of CICM 2018, Hagenberg, Austria, 13-08-2018, published at http://ceur-ws.org
retrieved from the arXiv repository of electronic preprints (http://arxiv.org/) and Wikipedia. I am striving to
develop a method that will be able to map, e.g., all of the formulae collected in figure 1 - in particular, linkable to
the Wikidata entry https://www.wikidata.org/wiki/Q868967. I chose the semantic knowledge-base Wikidata
because it is free, open and can be read and edited by humans and machines.
Research Method
Formula Feature Analysis
The first step in the formula feature analysis is tokenization, i.e., the decomposition into their components
(identifiers, operators, numbers, etc.) and Part-Of-Math-tagging: a formula consists of different terms, which
∂2 m2 c2
must be distinguished from each other. The Klein-Gordon equation c12 ∂t 2
2ψ − ∇ ψ + h̄2
ψ = 0 from quantum
2
1 ∂
physics, again used as an instructive example, contains a term c2 ∂t2 ψ with a double time derivative, one with
2 2
a double space derivative ∇2 ψ as well one with a constant prefactor mh̄2c ψ -the first term can then be further
decomposed into its characters (tokens), that is, the denominator c for the speed of light, the operator ∂t with
an exponent (number) 2 and the identifier ψ for the physical (quantum) wave function.
When analyzing the semantics of a formula, we are faced with the problem of identifier ambiguity, which re-
quires disambiguation with the help of the partial clarifications available in the text. A single identifier has a
theoretically unlimited number of possible meanings, e.g., E in physics often refers to both an energy and an
electric field, generally mathematically an expected value, etc. Thus, it is essential to improve the retrieval of
the semantics from the surrounding text.
Research Questions
The aim of Formula Concept Discovery (FCD) is to 1) retrieve a large number of formula examples from Wikipedia
articles and arXiv documents together with a mapping to formula concepts (Wikidata items), and 2) recover a
general definition of a formula concept using feature analysis and abstract mathematical formalization.
The aim of Formula Concept Recognition (FCR) is to identify formulae in arXiv documents or Wikipedia
articles as Wikidata formula concept items. Therefore, a measure of similarity that allows assigning a formula to
a mathematical concept (equation) if it exceeds a defined threshold needs to be defined. A first rough approach
could be a matching score = # recognized elements / # total elements. To successfully identify a single element,
for example, the Laplace operator ∇2 = ∆, it must be assigned to the corresponding concept in Wikidata,
at https://www.wikidata.org/wiki/Q203484, i.e. to QID Q203484. The aim is to motivate active users of
Wikidata to gradually build a hierarchical structure of the formula elements, assign elements to all available
formulae (property has part) and create new items for formulae concepts directly including the parts.
Evaluation Plans
I will compare and discuss 1) several possible Formula Concept Discovery methods (e.g., taking the first formula
from a Wikipedia article as defining formula of the concept, formula clustering, etc.), and 2) several possible
Formula Concept Recognition methods (e.g., simple TeX string search vs. parts identification, recognition by
identifier name, symbol and value, etc.).
Completed Research
In my first publication [SGPS+ 18], I significantly contributed to the creation of a Gold standard MathMLben
for the evaluation of the conversion between different mathematical formats (LaTeX vs. Computer Algebra
Systems). In my second publication [SSD+ 18], I presented the first math-aware QA system that can answer a
natural language question yielding a mathematical formula using Wikidata. My third recent publication [SSG18]
initiates my reasoning on a definition of a formula concept and its possible content representations in LaTeX,
MathML, and Wikidata.
Remaining Research
In my next publication, I will provide a thorough literature review on formula feature analysis. Together with M.
Schubotz and A. Greiner-Petter, I am planning to develop an annotation tool AnnotaTeX for LaTeX documents
that will facilitate the annotation process by recommending identifier names to the user. Figure 1 (right) shows
a proposed User Interface.
References
[SGPS+ 18] Moritz Schubotz, Andre Greiner-Petter, Philipp Scharpf, Norman Meuschke, Howard Cohl, and Bela
Gipp. Improving the representation and conversion of mathematical formulae by considering their
textual context. In Proceedings of the ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL),
Fort Worth, USA, Jun. 2018.
[SSD+ 18] Moritz Schubotz, Philipp Scharpf, Kaushal Dudhat, Yash Nagar, Felix Hamborg, and Bela Gipp.
Introducing mathqa - a math-aware question answering system. In Proceedings of the ACM/IEEE-
CS Joint Conference on Digital Libraries (JCDL), Workshop on Knowledge Discovery, Fort Worth,
USA, Jun. 2018.
[SSG18] Philipp Scharpf, Moritz Schubotz, and Bela Gipp. Representing mathematical formulae in content
mathml using wikidata. In Proceedings of the International ACM SIGIR Conference on Research
and Development in Information Retrieval (SIGIR), 2018.