How to Extract Unit of Measure in Scientific Documents?
Soumia Lilia Berrahou
1,2
, Patrice Buche
1,2
, Juliette Dibie-Barthelemy
3
and Mathieu Roche
1
1
LIRMM – CNRS – Univ. Montpellier 2, 34095 Montpellier Cedex 5, France
2
INRA – UMR IATE, INRIA – GraphIK 2, place Pierre Viala, 34060 Montpellier Cedex 2, France
3
INRA – M
´
et@risk & AgroParisTech, 75231 Paris Cedex 5, France
Keywords:
Information Retrieval, Unit of Measure Extraction, Ontological and Terminological Resource, Machine
Learning.
Abstract:
A large amount of quantitative data, related to experimental results, is reported in scientific documents in a
free form of text. Each quantitative result is characterized by a numerical value often followed by a unit of
measure. Extracting automatically quantitative data is a painstaking process because units suffer from different
ways of writing within documents. In our paper, we propose to focus on the extraction and identification of
the variant units, in order to enrich iteratively the terminological part of an Ontological and Terminological
Resource (OTR) and in the end to allow the extraction of quantitative data. Focusing on unit extraction
involves two main steps. Since we work on unstructured documents, units are completely drowned in textual
information. In the first step, our method aims at handling the crucial time-consuming process of unit location
using supervised learning methods. Once the units have been located in the text, the second step of our method
consists in extracting and identifying candidate units in order to enrich the OTR. The extracted candidates are
compared to units already known in the OTR using a new string distance measure to validate whether or not
they are relevant variants. We have made concluding experiments on our two-step method on a set of more
than 35000 sentences.
1 INTRODUCTION
Discovering and extracting information reported in
published scientific documents is a crucial stake in
several scientific domains in order to be able to reuse,
manage, exploit and analyze the information they
contain. Nevertheless, a large amount of data in any
scientific field is still published in today’s web in the
form of unstructured text rather than as structured se-
mantic information. Indeed, published scientific ar-
ticles use natural language combined with domain-
specific terminology that is extremely tedious to ex-
tract in the free form of text.
Linguistic methods for information extraction
(IE), such as the ones proposed in GATE (Cunning-
ham et al., 2002), are mostly rule-based techniques,
which usually use a combination of gazetteer lists and
pattern-matching rules in order to define if a candi-
date extracted term is valid or not. Those techniques
are then included in an overall architecture providing
various language processing tasks as building and an-
notating corpora. The system involves semantic anal-
ysis at the simplest level in the process with named
entities. Other systems integrate directly formal on-
tologies in the natural language processing (NLP)
environment, called ontology-based information ex-
traction (OBIE), and incorporate semantic knowledge
about terms in the context involving reasoning (May-
nard et al., 2008; Wimalasuriya and Dou, 2010).
Our work deals with unit recognition and extrac-
tion, widely known as a challenging issue of IE as
described in Related Work section below. Indeed,
most of techniques used in classical IE systems, based
on linguistic methods, cannot handle quantitative data
extraction issues, especially units of measure because
they rely on completely different syntactic defini-
tions. New methods and rules of information extrac-
tion must be provided and take into account those par-
ticularities such as simple or complex units, prefix of
multiple or submultiple of units, dimensions, consid-
ering the presence or absence of special characters,
superscript numbers and so on. Currently, none of
the existing NLP methods of the state-of-the-art con-
sider efficiently the extraction of units of measure.
However, it remains a paramount importance in many
quantitative research fields such as physics, chem-
istry, biology and food science, where the process of
quantitative data extraction in experimental results is
249
Berrahou S., Buche P., Dibie-Barthelemy J. and Roche M..
How to Extract Unit of Measure in Scientific Documents?.
DOI: 10.5220/0004666302490256
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval and the International Conference on Knowledge
Management and Information Sharing (SSTM-2013), pages 249-256
ISBN: 978-989-8565-75-4
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
difficult to overcome.
Our work aims at leveraging knowledge in food
science domain, which relies on an Ontological and
Terminological Resource (OTR). The concepts re-
lated to units of measure in this OTR have been de-
fined from the most extended ontology of units, On-
tology of units of Measure and related Concepts (OM)
(Rijgersberg et al., 2011), (Rijgersberg et al., 2013).
OM respects the International System of Units (SI)
(Thompson and Taylor, 2008), which organizes quan-
tities and units of measure. The definitions of base
and derived units of SI have been translated in a for-
mal representation in OM. Our work focuses on the
unit recognition and extraction issues in order to ex-
tract typographic variants, synonyms, abbreviations
and new units of measure, which are then used to
populate the OTR. This is a crucial step to be com-
pleted in order to build robust add-in semantic tag-
ging and extracting tools concerning units of measure
that could be integrated in OBIE or IE systems. These
tools facilitate domain expert participation in the task
of annotating textual corpus and data tables for gold
standard corpus construction. Moreover, the OTR
such populated is used as a basis of reasoning and to
make multi-criteria analysis in decision-making sys-
tems (Hignette et al., 2008).
We propose a two-step approach in order to per-
form unit recognition and extraction, which focuses
on the following contributions: (i) the first step of our
work demonstrates how supervised learning methods
can be successfully used to meet the challenge of re-
ducing the time-consuming process of unit location
in free form of text. (ii) the second step of our work
deals with unit extraction and identification using a
new string distance measure in order to overcome the
scaling issue.
The paper is structured as follows. In section 2, we
point out related work on unit recognition. In section
3, we describe our two-step approach, based on locat-
ing and extracting variant terms of units. In sections
4 and 5, each step is precisely detailed. In section 6
we deliver our results on real-world data and finally,
discuss and conclude on our work.
2 RELATED WORK
Recent work (Touhami et al., 2011), (Jessop et al.,
2011a) has revealed that most of extraction of exper-
imental data fails because units suffer from a large
variability of writing within scientific documents.
Despite works to produce proper formal standard
(Thompson and Taylor, 2008), (Rijgersberg et al.,
2013) in order to exchange and process quantitative
information, if a unit of measure is written differently
in the text, the process of identification fails even in
the smallest differences (e.g. mol written mole).
Other work in related fields as chemistry, focus
not only on the identification and annotation of chem-
ical entities (Jessop et al., 2011b) but also on the rela-
tionships linking these entities to each other (Hawizy
et al., 2011) using ChemicalTagger. ChemicalTagger
is the best performing state-of-the-art tool for text-
mining in chemistry, a phrase-based semantic natural
language processing tool for the language of chemi-
cal experiments. However, in (Jessop et al., 2011a),
the observation is similar when the authors note that
ChemicalTagger (Hawizy et al., 2011) fails in the pro-
cess of recognizing chemical names as reagent be-
cause of the variability of writing units in the text. It
perfectly matches the units of measure written in re-
spect to the SI (Thompson and Taylor, 2008) but fails
automatically when small changes occur in the text.
The authors present this issue as a challenge to be met
in future work.
In (Willems et al., 2012), in the same trend as we
aim at working, the authors present a method for an-
notating and extracting quantities and units. Their ap-
proach relies on the use of semantic markups avail-
able in Latex files. Our approach must address the
full-text unstructured documents.
In (Van Assem et al., 2010), the authors pro-
pose to overcome unit identification issue in data ta-
bles by using a string distance metric, called Jaro-
Winkler-TFIDF (Cohen et al., 2003). We have used
a new string distance measure, we adapted from
Damerau-Levenshtein distance (Damerau, 1964), in
order to tackle unit identification issue when units are
drowned in textual information. Moreover, unlike our
work, they do not need to consider unit location in or-
der to overcome scalability issue since they work on
structured data tables.
The current state-of-the-art methods of related do-
mains is not suitable enough to handle efficiently the
identification of variant terms of unit issue. Our ap-
proach aims at addressing this concern using a novel
method to extract unit information in a two-step ex-
traction process.
3 GLOBAL APPROACH
In this section, we present our two-step approach to
extract units in order to enrich a domain OTR. We
assume that the more we extract units from text, the
more we will succeed in the experimental data iden-
tification process. Indeed, variant terms of units en-
countered in scientific documents, due to the variabil-
KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
250
ity of writing, are seen as potential synonyms in the
terminological part of the OTR. Once referenced in
the OTR, they will increase the positive results in the
identification and extraction process of experimental
data. Thus, our work on unit identification task is seen
as a sub-step of an overall process of experimental re-
sult identification and extraction but remains a crucial
stake.
Before presenting our global approach, let us
briefly present the unit concepts of the considered
OTR. As shown in Figure 1, units are specifically or-
ganized around the concept Unit
concept and divided
into 4 classes of units (units being considered as in-
stances):
• Singular Unit such as day or meter;
• Unit Exponentiation such as Square Meter;
• Unit Division Or Multiplication such as
Milliliter Per Square Meter Per Day;
• Unit Multiple Or Submultiple such as Microme-
ter.
Figure 1: Excerpt of the OTR: Unit concept.
Each unit has several possible labels, called terms,
which are defined in the terminological part of the
OTR and denote unit instances, called units in the fol-
lowing, defined in the conceptual part of the OTR.
In Figure 2, we give an example in order to
clearly illustrate the main problem of unit recognition.
The terms amol.m
−1
.s
−1
.Pa
−1
and amol/m/s/Pa refer-
enced in the terminological part of the OTR denote the
unit AttoMole Per Meter Per Second Per Pascal,
instance of the Unit Division Or Multiplication con-
cept. Those terms are not sufficient enough to recog-
nize variant terms of units such as suggested in Fig-
ure 2. Our method aims at locating those variant terms
and extracting them to enrich the OTR with new terms
(e.g. amol x m
-1
x s
-1
x Pa
-1
where a multiplication
sign has been added).
The method must face several challenges: (i) the
question of locating units within the text, do we need
to skim all the text or can we predict the best informa-
tion location without any assumption. The question
is important to be asked because the major limit of
classical methods of string distance metrics concerns
Figure 2: Relevant variant terms of units.
the scalability issue. Currently, these metrics can not
be applied on a huge corpus. The first step is a ma-
jor concern in order to reduce significantly the search
space to the relevant sentences where we will be able
to apply string distance metrics (ii) the question of ex-
tracting variant terms of units, once the best location
has been identified, how do we extract variant terms
of units from that part of text; and (iii) once extracted,
the question of semantically identifying these variant
terms of units and find which unit instances they can
denote in the OTR. Our approach tackles these key
points and intends at discovering more variant terms
of units in order to enrich the terminological part of
OTR as an iterative process.
The objective of the first step is to reduce sig-
nificantly the size of search space and overcome the
painstaking process of unit location. As a matter of
fact, it is often assumed that experimental data are lo-
cated in specific paragraphs of the text, e.g. Exper-
iments or Results and discussion in (Hawizy et al.,
2011). We aim at locating it without any of these as-
sumptions. However, we are aware that going through
all documents involves a time-consuming process.
The method to reduce search space involves text min-
ing and supervised machine learning approaches. Our
aim in this first step is obtaining the most efficient
model from machine learning algorithms that could
be then applied in any corpus of a domain where ex-
tracting units is a major concern. In section 4, we give
more details on different experiments that have been
led with the purpose of achieving this best model.
The second step, detailed in section 5, works on
textual fragments obtained from the learning model
and focuses on terms of unit extraction. As soon as
the optimal location for discovering variant terms of
units is defined, we first highlight in this location the
ones to be extracted. After extracting potential candi-
dates, we compute a new string distance measure in
order to distinguish which of the known units in the
OTR can match these variant terms and so are poten-
tial synonyms to enrich the OTR.
HowtoExtractUnitofMeasureinScientificDocuments?
251
4 LOCATING UNITS
In the first step of the approach, our objective is to fo-
cus on locating relevant data concerning units of mea-
sure. In the work of (Hawizy et al., 2011), the authors
use a Regex-tagger based on regular expressions in or-
der to capture the sentences where quantitative data,
chemical entities and units appear. However, when
testing the tagger in (Jessop et al., 2011a), the process
of recognition fails because of typographic variations
of units in the text. This observation has motivated
our choices towards supervised learning methods in-
stead of using regular expressions. Hence, our pro-
cess consists of using supervised classifier to deter-
mine whether a part of text is classified as contain-
ing ”units” or ”non-units”. In order to obtain the best
model to be applied on the corpus, we work on sev-
eral training sets, each one corresponds to a specific
textual context. Thus, as soon as the process recog-
nizes a unit annotation in the sentence of the text, we
observe the context around it. We consider different
case studies to build these training sets for supervised
learning methods:
• the first case study takes into account only the sen-
tence where at least one unit appears and is recog-
nized according to the units defined in the OTR;
• the second and third case study take respectively
into account one sentence before and one sentence
after the one where at least one unit appears;
• the fourth and fifth case study take respectively
into account two sentences before and two sen-
tences after the one where at least one unit ap-
pears.
The following is a description of the transformation
process tasks and explains our choices in order to ob-
tain the model.
4.1 Preparing Data
Data preparation involves text processing and text
transformation tasks. Text processing sub-step con-
sists of:
• text segmentation in order to generate a set of sen-
tences, which deals with our issues, e.g. not going
to new sentence when units with point separators
are encountered;
• text cleanup, which removes punctuation and spe-
cial characters from text, except those involved in
units;
• text tokenization, which splits up a string of char-
acters into a set of tokens;
• text reducing, which prunes away tokens contain-
ing junk words according to a list of stop words;
• text tagging, which annotates the units in the text.
Text tagging relies on comparing the terms denot-
ing the referenced units in the OTR with tokens of
each sentence. As soon as the process recognizes a
unit or a part of unit, this one is tagged. Our main
purpose is obtaining a tagged data used to test our dif-
ferent case studies under learning algorithms.
4.2 Transforming Data
As soon as we have obtained our training data, we can
proceed with text transformation, which consists of
representing a text according to the words it contains
and their occurrences in the sentences. The repre-
sentative words (features) form the bag-of-words ac-
cording to a relevant threshold we can pre-select (e.g.
words appearing more than just once), and we com-
pute their occurrences in each sentence according to
three word weighting measures:
• Term Frequency (TF), which considers that the
word is more important if it appears several times
in sentences;
• Term Frequency-Inverse Document Frequency
(TF.IDF (Hiemstra, 2000)), which considers that
the word is more important if it appears in less
sentences;
• Okapi BM25 (Jones et al., 2000), which takes also
into consideration the sentences length where the
word appears to define its relevancy.
Sentences become vectors that constitute our
training corpus in order to define, with supervised
learning methods, whether we are in a ”unit” or ”non-
unit” context.
4.3 Learning Model
Our five case studies performed under those three
weight-based measures are run under several learning
algorithms. Our aim is to obtain exhaustive experi-
ments and results in order to conclude about the best
model in our issue. Various learning algorithms have
been tested in our experiments:
• Naive Bayes classifier and the Discriminative
Multinominal Naive Bayes (DMNB) classifier;
• J48 decision tree classifier ;
• Sequential Minimal Optimization (SMO), which
is a Support Vector Machines classifier (SVM).
The motivation of our choices relies on compar-
ing the behavior of those widely known learning algo-
rithms on a corpus containing many quantitative data.
KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
252
Naive Bayes (John and Langley, 1995) is competitive
for computational efficiency. Decision tree (Kohavi
and Quinlan, 2002) classifiers are known to obtain
good classification results but are less competitive in
speed of execution. SMO (Platt, 1999) is a discrim-
inative classifier known to efficiently behave on text
classification and, DMNB is a new text classification
algorithm (Su et al., 2008) which performs competi-
tively with discriminative classifiers in terms of accu-
racy, without losing the advantages of computational
efficiency of Naive Bayes classifiers.
The conclusion on the best model is taken after
paying a particular attention to several key points:
• Training time and best classification balance;
• Precision, Recall and F-measure results, we fo-
cus on the recall, which means we want to maxi-
mize the fraction of relevant sentences that are re-
trieved, without losing too much precision in the
results ;
• Confusion matrix, which compares the results
of the classifier under test with trusted external
judgements.
As we want to estimate how accurately the model
of each classifier will perform in practice, we use a
10-fold cross-validation: the original sample is ran-
domly partitioned into 10 equal size subsamples. One
subsample is used as the validation data for testing the
model while the other subsamples are used as training
data. This process is repeated 10 times with each of
the subsample used once as the validation data. The
averaged result produces the estimation of the model.
Using cross-validation is a crucial task in order to
avoid ”overfitting” effect of the model.
5 EXTRACTING UNITS
In the second step of the approach, we work on tex-
tual fragments obtained from the first step. We have
significantly reduced the search space and need now
to detect, extract and identify variant terms of units or
new units. It involves working on textual fragments
as follows:
• detecting terms of units inside textual fragments:
the tags used in the annotation process of the first
step are our entry point;
• extracting terms (or fragment of terms of units)
from the entry point: both sides of the tag are
scanned and the process stops when it encounters
a common word of the dictionary;
• identifying variant terms of units or new units us-
ing the new string distance measure.
The identifying process relies on a new string dis-
tance measure, adapted from Damerau-Levenshtein
(DL) distance (Damerau, 1964). The DL distance be-
tween two strings is defined as the minimum number
of edits needed to transform one string into another,
with the edit operations being insertion, deletion, or
substitution of a single character. The distance can
then be normalized by using (Maedche and Staab,
2002) approach:
SM
DL
(u1,u2) = max[0;
min(|u1|,|u2|) − DL(u1,u2)
min(|u1|,|u2|)
]
∈ [0;1]
The similarity measure is computed and the higher it
is, the closer the variant terms of units is to the units
of the OTR.
Example 1. Let us consider the similarity between
amol/(m.s.Pa) and amol/m.sec.Pa according to the
classical DL distance. 2 characters are removed ( ”(”
and ”)”), 2 characters are inserted (”e” and ”c”).
The D
c
(the distance between those units) is 4 and
the similarity distance normalized according to D
c
is
then:
SM
D
c
(amol/(m.s.Pa),amol/m.sec.Pa) = max[0;
13 − 4
13
]
SM
D
c
= 0.69
We adapted this distance by considering each
block of characters forming the terms of units rather
than the simple characters. Thus, in our previous ex-
ample, the process considers each block as amol, m,
s, Pa and compares each of them with each block of
characters from the units referenced in the OTR. The
edit distance adapted to the block becomes D
b
= 1
since only the block sec varies and
SM
D
b
(amol/(m.s.Pa),amol/m.sec.Pa) = max[0;
4 − 1
4
]
SM
D
b
= 0.75
The new string distance measure is more accurate
and more relevant to address units of measure simi-
larity.
6 EXPERIMENTS AND RESULTS
6.1 Data Description
Experiments have been led on real-world dataset com-
posed of 115 scientific documents in food science do-
main, which represents a set of more than 35 000 sen-
tences. The positive corpus, where at least one unit
appears, represents more than 5 000 sentences. We
HowtoExtractUnitofMeasureinScientificDocuments?
253
use a list of 211 terms denoting units of the OTR in
order to conduct unit identification. During experi-
ments, we can set the number of instances that will
constitute our training data. The results delivered in
our paper rely on a training set of 2 000 instances ran-
domly chosen from the corpus, and size balanced be-
tween the positive and negative instances. According
to the different case studies, the bag-of-words varies
from 3 000 to 4 800 features.
6.2 Learning the Best Location
6.2.1 Learning under Various Supervised
Algorithms
We have tested our five case studies. Only the 3 more
relevant cases are reported in Table 1. Each case is
represented by a symbol in order to make a readable
table:
• XP1: only the sentence where at least one unit
appears;
• XP2: two sentences after the one where at least
one unit appears;
• XP3: two sentences before the one where at least
one unit appears.
Table 1 restitutes the results with all weight-based
measures included and according to the case studies.
This first table helps us to determine which case is the
most relevant context.
Table 1: Results of ”Unit” instances: Precision (P), Recall
(R), F-measure (F) are given for each case study. Best recall
in bold considering P and F.
Dec. Tree J48 Naive Bayes DMNB SMO
P R F P R F P R F P R F
XP1 0.99 0.87 0.93 0.83 0.93 0.88 0.95 0.96 0.95 0.99 0.99 0.99
XP2 0.99 0.92 0.96 0.95 0.77 0.85 0.93 0.96 0.95 0.99 0.97 0.99
XP3 0.99 0.92 0.95 0.77 0.98 0.86 0.94 0.96 0.95 0.99 0.97 0.98
Table 2: Results of ”Unit” instances: Precision (P), Recall
(R), F-measure (F) are given for each weight-based mea-
sures and boolean matrix. Best results in bold.
Algo. Dec. Tree J48 Naive Bayes DMNB SMO
P R F P R F P R F P R F
Boolean 0.99 0.87 0.93 0.83 0.93 0.88 0.95 0.96 0.95 0.99 0.99 0.99
TF 0.99 0.86 0.92 0.69 0.85 0.76 0.95 0.96 0.95 0.84 0.90 0.87
TF.IDF 0.99 0.86 0.92 0.69 0.85 0.76 0.95 0.96 0.95 0.84 0.90 0.87
Okapi 0.99 0.86 0.92 0.69 0.86 0.76 0.95 0.96 0.95 0.77 0.88 0.82
We take into consideration several parameters in
order to conclude about the most relevant context: re-
call, overall classification, F-measure, stability of the
various results in the weight-based measures that are
presented in section 6.2.2 and confusion matrix re-
sults. We are particularly interested in recall since our
aim is obtaining the most relevant instances that are
retrieved considering ”unit” class but, without losing
too much precision in the results, which is described
by the F-measure. Firstly, we can say that Naive
Bayes returns F-measure rates ranging from 0.85 to
0.88. Decision tree (i.e. J48) returns better rates from
0.93 to 0.96. DMNB and SMO return better values
(0.95 to 0.99).
Secondly, we can observe that a larger context (i.e.
composed of two sentences – XP2 and XP3) does not
improve the results. We can conclude that consider-
ing the smallest context based on one sentence (i.e.
XP1) is enough for unit extraction. This allows us to
significantly reduce the search space while being in
an optimal context of discovery.
6.2.2 Learning under Various Weight-based
Measure
Table 2 restitutes the results on the XP1 context, pre-
viously underlined, according to the three weight-
based measures and the boolean matrix. This second
table allows us to conclude whether the learning algo-
rithm is constant on the weight-based measures.
As the previous analysis, we take into consid-
eration recall, overall classification and F-measure.
Firstly, we can say that, all weight-based measures
included, Naive Bayes returns rates declining from
86.9% (boolean matrix) to 73.4% (okapi measure),
decision tree J48 overall classification falls a bit
around 92 − 93%. SMO loses almost 20% with a
rate declining from 99.6% to 80.9%. DMNB stays
constant with 95.5% regardless of weight-based mea-
sures. The DMNB model stays constant on precision,
recall and F-measure and achieves a very good recall
at 0.96.
6.2.3 Discussion
In this first step, we have studied how supervised
learning methods can be successfully used in order
to optimize the process of unit location drowned in
unstructured scientific documents. For this, we have
compared several algorithms and weight-based mea-
sures before concluding that the best context for dis-
covering variant terms of units and even new units
(cf. next sub-section) is the sentence where at least
one unit appears. This result is constantly confirmed
regardless weight-based measures under DMNB al-
gorithm with a recall at 0.96 and a precision at
0.95. Thus, we have successfully overcome the time-
consuming process of location issue.
KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
254
Table 4: Identifying variant terms of units.
Variant term Reference SMDc SMDb
10e10 (cm3.m-1.sec-1.Pa-1) 10e10.cm3.m-1.sec-1.Pa-1 0.87 1
10e-14(cm3/m.s.Pa) 10e-14.cm3/(m.s.Pa) 0.89 1
10e-16cm3.cm/cm.cm2.s.Pa (10e-16cm3.cm)/(cm2.s.Pa) 0.76 0.8
10e18 (mol.m/Pa.sec.m2) 10e18.mol.m/(Pa.sec.m2) 0.87 1
amol.m-1.s-1.Pa-1 amol.s-1.m-1.Pa-1 0.88 0.75
amol/m.s.Pa amol/(m.s.Pa) 0.84 1
amol/m.sec.Pa amol/(m.s.Pa) 0.69 0.75
cm3.um/m2.d.kPa cm3.µm/(m2.d.kPa) 0.77 0.8
Table 3: Discovering new terms of units.
Class Nb Relevancy Terms of units
Singular Unit 9 55.5% mol, s, m, g, d, etc.
Unit Exponentiation 7 85.7% m
−2
, m
−1
, d
−1
,
cm
3
, s
−1
, m
2
, etc.
Unit Division or Multi. 39 43.6% mol.L
ˆ
-1
, mol/L,
mL/kg/h, etc.
Unit Multiple or Submulti. 4 50.0% amol, mL, etc.
Total 59 50.8% All previous units
Besides that, we have performed some extra ex-
periments in order to know what kind of features are
involved in the learning methods. The question asked
is: are the terms of units referenced in the OTR the
best features used in learning step? To answer this
question, we have removed from the bag-of-words
”unit” features and we have relaunched the learning
algorithm DMNB in the same context. The recall
and the precision drop slightly but remains quite hon-
orable with respectively 0.81 and 0.76. This result
shows that other features are involved in the learning
process. We have then conducted a feature classifi-
cation according to their different weight-based mea-
sures. The top k classification with k=10 shows that
there are not only units but features like nouns, verbs,
or adverbs describing a context where unit may ap-
pear (e.g. water, temperature, RH for relative humid-
ity, thickness, modified, solutions, samples, concen-
tration, min, permeability, respectively). We can say
that units and other types of terms are closely related
in the overall expression of an experimental data.
6.3 Extracting Variant Terms of Units
and Discovering New Ones
From the results obtained in step 1, we conclude that
the best location to work in, is the sentence where at
least one unit appears. We apply the process accord-
ing to different thresholds of the new similarity mea-
sure, adapted to the block and compare those experi-
ments to the results obtained with the classical string
distance measure as described in section 5.
In our experiments, we assume that a similarity
measure between 0.2 and 0.4 is favourable for discov-
ering new terms of units. Indeed, these values rep-
resent the case where at least one part of the blocks,
a subset of units is recognized using the new string
distance measure. Let us consider the following de-
rived unit g.m
-1
.d
-1
.Pa
-1
. This unit is referenced in
the OTR and used in the process of recognition. In
our approach, we consider 4 blocks g, m
-1
, d
-1
and
Pa
-1
. In the process of recognition, the base unit g,
found in the text and not referenced in the OTR yet, is
recognized compared to the derived unit with a sim-
ilarity measure equals 0.2. The g block is identified.
It is then relevant to propose this new base unit to be
introduced in the OTR. Considering this assumption,
we obtained 59 new possible units for the four classes
of the Unit concept, organized as described in Table
3.
The similarity measure between 0.5 and 1 is
favourable for identifying variant terms of units. In-
deed, the more the similarity measure is close to 1,
the more terms of units are the same. We obtained
17 variant terms of units, we have 8 relevant variant
terms that can be introduced in the OTR. The relevant
variant terms of units are described in Table 4.
In this second step, we have successfully detected
and extracted either relevant variant terms of units or
new units. We can potentially enrich the OTR of 38
terms of units or 18% of enrichment (the actual OTR
references 211 units). We have adapted a new string
distance measure in order to compare variant and new
terms of units to the ones referenced in the OTR. We
compare blocks of units rather than characters, which
proves to be more efficient in our issue. We have no-
ticed that this process is quite relevant to identify new
terms of units but is under our expectations concern-
ing the identification of variant terms of units.
In our process of identifying blocks, we currently
ignore all separators between those blocks, which are
very specific in the case of units (e.g. /, ., (, ), x).
Those characters have semantic meanings that need
to be added in the concern of extraction and identifi-
cation, in order to improve the results of the similarity
measure. We intend to carry out this work given these
promising results. Another future work is completing
the process with a new step in order to identify what
HowtoExtractUnitofMeasureinScientificDocuments?
255
kind of terms of units are discovered (e.g. a tempera-
ture unit, a relative humidity unit, a permeability unit)
and allow an automatic integration of variant or new
units directly in the OTR.
7 CONCLUSIONS AND FUTURE
WORK
The paper presents our overall two-step approach in
order to identify variant terms of units or new units in
unstructured scientific documents. We have success-
fully demonstrated how supervised learning methods
can help reduce the search space of terms of units
drowned in textual information. We have achieved
almost 86% of reducing. For this purpose, we have
used many algorithms and weight-based measures to
prove the relevancy of our approach. We have also
presented our first results to detect and extract vari-
ant terms and new units, using a new string distance
measure adapted to our unit identification issue. As
presented in the paper, this second step needs to be
deepened. Our approach, presented in this paper, ad-
dresses the entire issue: location, extraction and iden-
tification of units. Further work on semantical identi-
fication of variant terms and new units should be ad-
dressed in order to automatically populate the OTR.
REFERENCES
Cohen, W. W., Ravikumar, P., and Fienberg, S. E. (2003).
A comparison of string distance metrics for name-
matching tasks. In Proc. of the Workshop on Informa-
tion Integration on the Web (IIWeb-03), volume 47.
Cunningham, H., Maynard, D., Bontcheva, K., and Tablan,
V. (2002). GATE: A framework and graphical de-
velopment environment for robust NLP tools and ap-
plications. In Proceedings of the 40th Annual Meet-
ing of the Association for Computational Linguistics,
Philadelphia, PA, USA.
Damerau, F. J. (1964). A technique for computer detec-
tion and correction of spelling errors. Commun. ACM,
7(3):171–176.
Hawizy, L., Jessop, D., Adams, N., and Murray-Rust, P.
(2011). ChemicalTagger: a tool for semantic text-
mining in chemistry. Journal of cheminformatics,
3(1):17.
Hiemstra, D. (2000). A probabilistic justification for using
tf x idf term weighting in information retrieval. Int. J.
on Digital Libraries, 3(2):131–139.
Hignette, G., Buche, P., Couvert, O., Dibie-Barth
´
elemy, J.,
Doussot, D., Haemmerl
´
e, O., Mettler, E., and Soler,
L. (2008). Semantic annotation of Web data applied
to risk in food. International Journal of Food Micro-
biology, 128(1):174–180.
Jessop, D. M., Adams, S. E., and Murray-Rust, P. (2011a).
Mining chemical information from open patents.
Journal of cheminformatics, 3(1):40.
Jessop, D. M., Adams, S. E., Willighagen, E. L., Hawizy,
L., and Murray-Rust, P. (2011b). OSCAR4: a flexi-
ble architecture for chemical text-mining. Journal of
cheminformatics, 3(1):1–12.
John, G. H. and Langley, P. (1995). Estimating continuous
distributions in bayesian classifiers. In Proc. of the
conf. on Uncertainty in artificial intelligence, pages
338–345.
Jones, K. S., Walker, S., and Robertson, S. E. (2000). A
probabilistic model of information retrieval: develop-
ment and comparative experiments - part 1. Inf. Pro-
cess. Manage., 36(6):779–808.
Kohavi, R. and Quinlan, J. R. (2002). Data mining tasks and
methods: Classification: decision-tree discovery. In
Handbook of data mining and knowledge discovery,
pages 267–276. Oxford University Press, Inc.
Maedche, A. and Staab, S. (2002). Measuring similar-
ity between ontologies. In Knowledge Engineering
and Knowledge Management: Ontologies and the Se-
mantic Web, volume 2473 of LNCS, pages 251–263.
Springer.
Maynard, D., Li, Y., and Peters, W. (2008). Nlp techniques
for term extraction and ontology population. In Pro-
ceeding of the 2008 conference on Ontology Learning
and Population: Bridging the Gap between Text and
Knowledge, page 107127.
Platt, J. C. (1999). Fast training of support vector machines
using sequential minimal optimization. In Advances
in kernel methods, pages 185–208. MIT Press.
Rijgersberg, H., van Assem, M., and Top, J. (2013). Ontol-
ogy of units of measure and related concepts. Seman-
tic Web.
Rijgersberg, H., Wigham, M., and Top, J. (2011). How se-
mantics can improve engineering processes: A case of
units of measure and quantities. Advanced Engineer-
ing Informatics, 25(2):276–287.
Su, J., Zhang, H., Ling, C. X., and Matwin, S. (2008). Dis-
criminative parameter learning for bayesian networks.
In Proc. of the int. conf. on Machine learning, pages
1016–1023.
Thompson, A. and Taylor, B. N. (2008). Guide for the use
of the international system of units (SI).
Touhami, R., Buche, P., Dibie-Barth
´
elemy, J., and Ibanescu,
L. (2011). An ontological and terminological resource
for n-ary relation annotation in web data tables. On
the Move to Meaningful Internet Systems: OTM 2011,
pages 662–679.
Van Assem, M., Rijgersberg, H., Wigham, M., and Top, J.
(2010). Converting and annotating quantitative data
tables. The Semantic Web–ISWC 2010, pages 16–31.
Willems, D. J., Rijgersberg, H., and Top, J. (2012). Identify-
ing and extracting quantitative data in annotated text.
SWAIE.
Wimalasuriya, D. C. and Dou, D. (2010). Ontology-based
information extraction: An introduction and a survey
of current approaches. Journal of Information Sci-
ence, 36(3):306323.
KDIR2013-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
256
