Lexical and Morpho-syntactic Features in Word Embeddings

A Case Study of Nouns in Swedish

Ali Basirat and Marc Tang

Department of Linguistics and Philology, Uppsala University, Box 635, 75 126, Uppsala, Sweden

Keywords:

Neural Network, Nominal Classiﬁcation, Swedish, Word Embedding.

Abstract:

We apply real-valued word vectors combined with two different types of classiﬁers (linear discriminant analy-

sis and feed-forward neural network) to scrutinize whether basic nominal categories can be captured by simple

word embedding models. We also provide a linguistic analysis of the errors generated by the classiﬁers. The

targeted language is Swedish, in which we investigate three nominal aspects: uter/neuter, common/proper, and

count/mass. They represent respectively grammatical, semantic, and mixed types of nominal classiﬁcation

within languages. Our results show that word embeddings can capture typical grammatical and semantic fea-

tures such as uter/neuter and common/proper nouns. Nevertheless, the model encounters difﬁculties to identify

classes such as count/mass which not only combine both grammatical and semantic properties, but are also

subject to conversion and shift. Hence, we answer the call of the Special Session on Natural Language Process-

ing in Artiﬁcial Intelligence by approaching the topic of interfaces between morphology, lexicon, semantics,

and syntax via interdisciplinary methods combining machine learning of language and general linguistics.

1 INTRODUCTION

The continuous vector representation of words,

known as word embeddings or word vector, has been

widely used in different areas of natural language pro-

cessing. The studies on the distribution of word vec-

tors show that different types of word similarities are

captured by word vectors, e.g., the semantically sim-

ilar words are clustered together. However, it is not

still clear what types of linguistic information (e.g.,

semantic or syntactic word classes) are encoded in

these vectors. By way of illustration, the semantic

and syntactic information of different types of nom-

inal classes is mirrored through language via gram-

matical systems of nominal classiﬁcation. Thus, the

investigation of such linguistic structures in word vec-

tors is of high interest.

Nominal classiﬁcation refers to how a language

classify nouns of the lexicon. The most frequent

grammaticalized system of nominal classiﬁcation is

grammatical gender (Corbett, 1991; Seifart, 2010;

Corbett, 2013), e.g., all nouns in French are afﬁliated

to either masculine or feminine. Nouns may equiva-

lently be divided in categories according to different

semantic and/or syntactic criterion, i.e., count/mass

and common/proper (Delahunty and Garvey, 2010).

We study different types of nominal classiﬁcations

in Swedish with regard to the syntactic and seman-

tic information encoded to their word vectors. As a

research question, we aim at evaluating the perfor-

mance of word embeddings on the nominal classiﬁ-

cation task. The outputs of such investigation are not

only expected to serve as a baseline for future research

in terms of computational models of nominal classiﬁ-

cation, but also be compared with the ﬁndings of cog-

nitive and linguistic studies with regard to nominal

classiﬁcation (Aikhenvald, 2012; Kemmerer, 2017).

Our studies are on the basis of the accuracy of a classi-

ﬁer in predicting different classes of nouns from their

word vectors. To this end, we train a classiﬁer with

word vectors as input and nominal classes as output.

As a case study, we select three binary nomi-

nal features in Swedish: uter/neuter (i.e., grammati-

cal gender), count/mass, and proper/common nouns.

These distinctions represent three different types of

nominal classiﬁcation. First, grammatical gender is a

typically grammaticalized feature, which is reﬂected

in language via grammatical agreement with other el-

ements of phrase. For instance, in Swedish, the ar-

ticle and the adjective varies in terms of form de-

pending on the grammatical gender (uter or neuter) of

the following noun, c.f., ett stor-t

apple (a.SG.NEUT

big.SG.NEUT apple.SG.NEUT) ‘a big apple’ and en

stor-

0 h

ast (a.SG.UTER big.SG.UTER horse.SG.UTER)

Basirat, A. and Tang, M.

Lexical and Morpho-syntactic Features in Word Embeddings - A Case Study of Nouns in Swedish.

DOI: 10.5220/0006729606630674

In Proceedings of the 10th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2018) - Volume 2, pages 663-674

ISBN: 978-989-758-275-2

663

‘a big horse’. Every noun can only be associated to

one grammatical gender, regardless of register and

context. Hence, grammatical gender is considered as

a static nominal feature in Swedish.

Second, the distinction between common and

proper nouns (proper names) is considered as a static

semantic feature. Common nouns generally refer to

classes of things (e.g., printer, desk) while proper

nouns designate particular individual entities such as

Stockholm, Paris, among others (Delahunty and Gar-

vey, 2010, p.149). To be more precise, proper nouns

do not necessarily refer to only one speciﬁc individ-

ual. By way of illustration, the proper name Smith

may be given to different people (or even objects)

which do not share any property or quality in com-

mon. On the other hand, common nouns refers to a

set of entities which share speciﬁc properties. Thus,

common nouns may not be used in the same arbi-

trary way of proper nouns. As an example, naming a

desk as a fridge in English would not be semantically

valid, since the properties of a desk do not concord

with the intrinsic semantic content of the noun fridge

(e.g., a fridge must have a cold temperature compart-

ment). Following this logic, nouns are either com-

mon or proper and do not ﬂuctuate between the two

categories. Hence, it is deﬁned as semantically static.

Few exceptions of conversion are attested but they are

context speciﬁc (Gillon, 1999, p.58), thus we do not

consider them in the current paper.

Finally, the count/mass distinction is recognized

as an intermediary between syntactic and semantic

nominal features (Doetjes, 2012), as “the brain dif-

ferentiates between count and mass nouns not only

at the syntactic level but also at the semantic level”

(Chiarelli et al., 2011). With regard to semantics,

count nouns point to objects which represent a dis-

crete entity and may be counted, e.g., computer, book

in English. On the other hand, mass nouns (also

named non-count nouns) generally refer to objects

which are interpreted as a uncountable mass and are

not speciﬁed as how to individuate or divide them,

e.g., sand, milk in English. In terms of morpho-

syntax, the categories of count and mass nouns are

Static gender systems are not obligatorily found in every

gender language. By way of illustration, Pnar (Austro-

Asiatic) applies a grammatical gender system based on the

masculine/feminine/neuter distinction. However, a noun

in Pnar may have different genders depending on the prop-

erties attributed to it. For instance, the noun dei

n ‘tree’

can be marked by both masculine and feminine proclitics.

The masculine form u=dei

n refers to a standing, upright

tree, while ka=dei

n refers to a fallen tree (log) or wood.

Phenomena of gender variation are likewise attested in

languages such as Hindi, Nepali, among others (Nespital,

1990; Hall, 2002; Pokharel, 2010).

also differentiated, as mass nouns cannot occur in

the plural form even if the language has plural in-

ﬂection (among other criteria). However, it is con-

sidered difﬁcult to deﬁne absolute syntactic and se-

mantic parameters to entirely isolate count and mass

nouns. As an example, a noun may undergo con-

version between count and mass, e.g., pizza may be

used as a mass noun or or a count noun depending

on context, c.f., did you order pizza? and I ordered

a pizza (Gillon, 1999, p.51). Moreover, the pres-

ence/absence of grammatical number marking (i.e.,

plural) on the noun does not completely correlate

with the count/mass distinction (Corbett, 2000; Dryer,

2005). For instance, certain mass nouns may have

a count interpretation but be morphologically mass

nouns as they only occur in the singular form, e.g.,

luggage (Pelletier and Schubert, 1989).

As a summary, the count/mass distinction is more

versatile than the uter/neuter and proper/common

nouns classiﬁcation. Therefore, we estimate that the

classiﬁers should not have difﬁculties interpreting the

uter/neuter and proper/common features of the nouns,

since they provide transparent syntactic and semantic

clues. As for count/mass, we expect that the clas-

siﬁers will be able to distinguish between the two

classes, but with less accuracy than grammatical gen-

der and proper names.

The structure of this paper is as follow. Section 2

presents a literature review of the three nominal fea-

tures approached in our research question. Section 3

explains the structure of the selected computational

models. Section 4 lists the setting of our experiments.

Section 5 details how the classiﬁers performed with

regard to the three nominal features involved in our

study. In Section 6, we provide an explanation to the

performance of the classiﬁer, along with an error anal-

ysis. Finally, we conclude in Section 7.

2 LITERATURE REVIEW

In this section we describe the linguistic deﬁni-

tion and examples of grammatical gender, com-

mon/proper, and count/mass distinction. Moreover,

we provide an overview of the difﬁculties of classiﬁ-

cation observed from a linguistic approach.

2.1 Grammatical Gender

Nominal classiﬁcation, i.e., how languages classify

nouns of the lexicon, reﬂects cognitive and cultural

facets of the human mind (Aikhenvald, 2012; Contini-

Morava and Kilarski, 2013; Kemmerer, 2017). Gram-

matical genders are one of the most common systems

NLPinAI 2018 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

664

of nominal classiﬁcation (Corbett, 1991). They are

widely distributed in Africa, Europe, Australia, Ocea-

nia, while their presence is sporadicly attested in the

Paciﬁc, Asia and Americas (Aikhenvald, 2000, p.78).

The deﬁnition of grammatical gender involves

grammatical agreement between a noun and other

syntactic units which appear with the noun in a sen-

tence. In other words, all nouns of the lexicon are

assigned to certain classes. For instance, a language

has two genders if two classes of nouns can be differ-

entiated by their agreement marker (Senft, 2000). As

an example in French, frigo ’fridge’ is masculine and

chaise ’chair’ is feminine. Thus, grammatical gender

is reﬂected via agreement on the articles, adjectives,

and verbs, c.f., un grand frigo (one.MASC big.MASC

fridge.MASC) ‘A big fridge’ and une grande chaise

(one.FEM big.FEM chair.FEM) ‘A big chair’. Likewise

in Swedish, all nouns are afﬁliated to either uter or

neuter gender (Bohnacker, 2004, p.198), which is re-

ﬂected through grammatical agreement (see examples

in Section 1). Thus, the syntactic clues provided by

gender agreement are expected to be easily captured

by the word embedding model and mirrored in the

generated word vectors.

While the main functions of gender are relatively

transparent, e.g., to facilitate referent tracking in dis-

course (Dixon, 1986; Nichols, 1989; Contini-Morava

and Kilarski, 2013), gender assignment is considered

as much more opaque (Corbett, 1991, p.57). How-

ever, contradictory observation occur for Swedish.

While grammatical gender is commonly considered

as arbitrary (Andersson, 1992; Teleman et al., 1999),

several semantic principles are attested, i.e., nouns

referring to human and non-human animates tend to

be afﬁliated to the uter gender, while inanimates are

more likely to be neuter (Dahl, 2000). Moreover,

nouns pointing at concrete or countable entities are

generally uter while abstract or mass nouns are fa-

vored by the neuter gender (Fraurud, 2000).

2.2 Common/Proper Nouns

Nevertheless, not all classiﬁcation of nouns in the lex-

icon are interpreted via agreement marking. As an

example, the distinction between common and proper

nouns is based on semantics. Common nouns refer to

a general things which share certain properties, while

proper nouns name speciﬁc individual things (De-

lahunty and Garvey, 2010, p.149). By way of illustra-

tion for common nouns, one of the inherent properties

of a printer is ‘a device which can print materials’.

Proper nouns, on the other hand, do not carry such

intrinsic properties. For instance, any person, animal,

and object may be named after Simon. The differen-

tiation between common and proper nouns is realized

via capitalization in languages such as English. How-

ever, such marking is only represented in the writing

system, which is an additional invention of language.

Moreover, the rules of capitalization of proper nouns

are not consistent in a cross-language manner. En-

glish capitalizes names of the months (e.g., Decem-

ber) while Swedish does not. Furthermore, languages

do not apply the same parameters, e.g., in German, all

nouns are capitalized. Finally, the rule of capitalizing

sentence initial words in certain languages may also

interfere with the recognition of proper nouns solely

based on capitalization.

Similar observations are made in terms of syn-

tax, as no absolute criterion is valid to differentiate

between common and proper nouns. For instance,

in English, proper nouns generally appear as bare

forms, e.g., John, Apple. However, the deﬁnite ar-

ticle is preferred for certain types of proper nouns,

e.g., the United States. Moreover, the use of article

with proper noun is not universal cross-linguistically.

By way of illustration, country names are used with

the deﬁnite article in French, e.g., la France (the.FEM

France), while in English it is not required (e.g.,

France). Therefore, the classiﬁcation of common and

proper nouns is a relevant issue for machine transla-

tion (Lopez, 2008), sentiment analysis (Pang and Lee,

2008), topic tracking (Petrovic et al., 2010), web data

search (Baeza-Yates and Ribeiro-Neto, 2011), case

restoration (Baldwin et al., 2009), among others top-

ics (Preiss and Stevenson, 2013).

2.3 Count/Mass Distinction

The third main subdivision of nouns is the contrast

between count and mass. Count nouns are commonly

perceived as entities which can be individuated and

counted, while mass nouns incarnate things as a mass

whose parts are not considered as discrete units (De-

lahunty and Garvey, 2010, p.156). As an example, a

piece of a cake is still cake, however, a piece of a desk

is not a desk. The reﬂection of the cognitive princi-

ple of individuation have been studied with regard to

its connection to human cognition and how such cog-

nitive concept is mirrored through language (Quine,

1960; Chierchia, 1998; Chierchia, 2010; Doetjes,

2012). For instance, morphological marking of gram-

matical number has been attested as one of the main

linguistic realization of count/mass marking (Gillon,

1999; Borer, 2005). Hence, we expect that the word

embedding model is capable of retrieving the syntac-

tic information and encode it into the vectors.

Formally, count nouns may bear singular and

grammatical plural marking (c.f., the book is here and

Lexical and Morpho-syntactic Features in Word Embeddings - A Case Study of Nouns in Swedish

665

the books are here in English), while mass nouns may

solely occur in singular (c.f., the furniture is new, *the

furnitures are new). Moreover, only count nouns can

apply indeﬁnite article (c.f., a table and *a luggage),

among other syntactic criteria (Chierchia, 1998). Se-

mantically, count and mass nouns may be distin-

guished on the basis of cumulativity (Quine, 1960),

divisibility (Cheng, 1973), and speciﬁcity (Gillon,

1999, p.51-53). Mass nouns are unspeciﬁed as how

to be cumulated and divided, while count nouns are

speciﬁed for how to be cumulated or divided. By way

of illustration, coffee may be counted in terms of cup,

glass, barrel, brand, among other measure terms (Ki-

larski, 2014, p.9), with no intrinsic speciﬁcation. On

the opposite, count nouns such as book can inherently

be counted by the mean of cardinal numbers and can-

not be divided.

Nevertheless, even though both syntactic and se-

mantic criteria are available to distinguish count and

mass nouns, the fact that nouns may undergo con-

version (shift) and migrate to the other category rep-

resents a challenge for classiﬁcation (Gillon, 1999).

For instance, coffee in English may be used as a mass

noun when referring to a type of beverage, e.g., the

coffee is good. Nonetheless, the same word form can

equivalently be employed as a count noun when re-

ferring to ‘a cup of coffee’, e.g., I would like to have

two coffees. Such conversion is extremely productive

and common within languages of the world (Doetjes,

2012, p.14) and represents one of the major difﬁculty

of identifying count and mass nouns in terms of com-

putational linguistics (Katz and Zamparelli, 2012).

3 METHODOLOGY

We selected Swedish as language of analysis due to

the contradictory observation attested on grammatical

gender (see Section 2.1) and its availability of ﬁrst

hand data. For each nominal classiﬁcation task, we

train a classiﬁer on a set of word vectors which are

labeled with their corresponding nominal classes. The

process is divided into three steps: vector generation

(word embedding), data labeling, and classiﬁcation.

During the ﬁrst stage, word embedding, a corpus

of raw sentences with word segmentation is fed to a

word embedding model. The word embedding model

assigns a vector to each word in the corpus. This vec-

tors assignment is done in such a way that semantic

similarities of words is modeled by the correlation

between their corresponding vectors, i.e., the semanti-

cally similar words are assigned to the similar vectors.

Then, in the data labeling step, a dictionary is

used to associated a subset of word vectors with their

corresponding nominal classes. These are the word

vectors associated with some nouns in the dictionary

whose nominal classes could be extracted from the

dictionary. Later on, in Section 4, we elaborate how

the nominal classes are extracted from the dictionary.

This gives us a list of word vectors labeled with nom-

inal classes. This list is partitioned into train, and test

sets, to be used in the classiﬁcation step.

Finally, in the classiﬁcation, we train a classiﬁer

on the pairs of word vectors and nominal classes. The

classiﬁer takes word vectors as input and predict the

nominal classes in its output. The train data is used

to train the classiﬁer, and the test set is used to evalu-

ate the classiﬁcation model. Since, this research aims

to study the information provided by word vectors

for different nominal classiﬁcation tasks, we use sim-

ple classiﬁcation methods without performing com-

plicated tuning step. This is why we don’t keep a part

of data as development set to tune the classiﬁer.

The evaluation is on the basis of the performance

of word vectors on each of the nominal classiﬁcation

tasks, e.g., the higher the accuracy of the classiﬁcation

is, the more the information about the nominal classes

is encoded into the word vectors. In the next section,

we elaborate the tools, models, and data used in each

of these steps, word embedding, data labeling, and

classiﬁcation.

4 EXPERIMENTAL SETTINGS

Our word vectors are generated via RSV (Real-valued

Syntactic Word Vectors) model for word embedding

(Basirat and Nivre, 2017). RSV extracts a set of

word vectors from an unlabeled data in three major

steps: First, it builds a co-occurrence matrix whose

elements are the frequency of seeing words together.

The elements of this matrix are the frequency of see-

ing words in the domain of different context words.

The columns are associated with words and the rows

are associated with contexts. Each column forms a

high dimensional word vector that describes the word

with respect to its occurrence frequency in different

contexts. In the second step, the elements of the high

dimensional column vectors are transformed in such

a way that the distribution of the vectors is more close

to the Gaussian distribution with zero mean. Finally,

in the third step, it forms the low dimensional data

from the top K right singular vectors of the trans-

formed co-occurrence matrix. Within this process, the

RSV model has the following parameters:

• Context type: the context of a word may refer

to the preceding words (symmetric), following

words or include both directions (asymmetric).

NLPinAI 2018 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

666

• Context size: how many words does the system

count in the context. As an example, the most

popular setting is one preceding word.

• Dimensionality: the quantity of dimensions the

model may use to represent the word vectors. The

amount of dimensions is generally positively cor-

related to the accuracy, but negatively correlated

with the processing time and memory.

Within our experiments we set context type and con-

text size as the immediate preceding word, as pro-

posed by (Basirat and Nivre, 2017). The number of

dimensions is set to 50.

We include two types of classiﬁers in our exper-

iments: linear discriminant analysis, and neural net-

work. The purpose of this selection is to verify the

complexity of nominal properties and to see if the

nominal classes are linearly separable. Linear dis-

criminant analysis (LDA) is a linear generative clas-

siﬁer that ﬁts a Gaussian distributions on the training

data and predict the test data classes through the like-

lihood ratio of the data given the distributions. Neural

network (NN) is a non-linear discriminative classiﬁer

that make no assumption on the data distribution. It

searches for a boundary between the data points with

regard to their classes in such a way that the classi-

ﬁcation accuracy is minimized. LDA is more simple

in terms of structure and processing time but may be

less accurate depending on the complexity of the task.

While Neural Network is the more elaborate type of

classiﬁer but also costly in terms of processing time.

Our computational model is fed with two types

of data, which both originate from the Swedish

Language Bank (Spr

akbanken) at the University of

Gothenburg.

The word vectors are generated from a

corpus of Swedish raw sentences. This corpus is com-

piled by Spr

akbanken and involves data from Swedish

Wikipedia (available at Wikipedia Monolingual Cor-

pora, Swedish web news corpora (2001-2013) and

Swedish Wikipedia corpus). The OpenNLP sentence

splitter and tokenizer are used for normalizing the raw

corpus.

We replace all numbers with a special to-

ken NUMBER and convert uppercase letters to low-

ercase forms. Due to the high ratio of compound

nouns in Swedish (Carter et al., 1996; Ostling and

Wirn, 2013; Ullman and Nivre, 2014), we only in-

clude nouns which have more than 100 occurrences

in the corpus.

Second, the information on grammatical gender,

common/proper, and count/mass distinction is ex-

tracted from the SALDO (Swedish Associative The-

saurus version 2) dictionary. The categorization of

spraakbanken.gu.se/eng/resources/corpus

opennlp.apache.org

SALDO is rather “generous” and lists diverse prop-

erties of nouns (Borin et al., 2008, p.27). By way

of illustration, nouns which were assigned two gen-

ders according to speaker variation are afﬁliated to

neither uter nor neuter but rather to a third type vack-

lande. Moreover, some nouns are annotated as blank

if their gender was difﬁcult to interpret. The creation

of these categories was driven by speciﬁc pragmatic

and semantic classiﬁcation purposes. In our study,

we only incorporate the relevant categories. For in-

stance, with regard to grammatical gender, we solely

cover nouns annotated as uter and neuter since only

these two classes fulﬁll the conditions of grammati-

cal gender. Furthermore, the frequency and quantity

of peripheral nouns such as vacklande and blank is

not signiﬁcant compared to the whole dataset (1%).

Hence, we leave these patterns of variation for further

studies to explore. We also ﬁlter out those nouns that

are not in word vectors’ vocabulary set. The result of

our ﬁltering is shown in Table 1.

Table 1: Nominal features in Swedish based on SALDO.

Category Quantity Example

uter 13540 bok ‘book’

neuter 5518 hus ‘house’

count 16181 kontor ‘ofﬁce’

mass 3085 bagage ‘luggage’

comon 18549 bord ‘table’

proper 3142 Alyssa

The distinction of grammatical gender and com-

mon/proper nouns is rather straightforward, as the

information is annotated in SALDO. However, the

count/mass distinction is not a category transparently

speciﬁed in the dictionary. Thus, we follow the for-

mal syntactic deﬁnition and consider that only count

nouns may have plural inﬂection (Chierchia, 1998;

Chierchia, 2010).

5 RESULTS

The results of our experiments are evaluated accord-

ing to the Rand index (Rand, 1971) (accuracy) and

F-score (Ting, 2010). To obtain the general accuracy

of the model on the entire dataset, the Rand index is

calculated by dividing the total number of correctly

retrieved tokens by the total number of retrieved to-

kens. In the remaining parts of this paper, Rand in-

dex and accuracy are used interchangeably. As as

side note, we do not compare directly our results with

other studies as we are not aware of methodologies

and approaches being applied in a way similar to ours.

Due to the lack of balance between the investi-

Lexical and Morpho-syntactic Features in Word Embeddings - A Case Study of Nouns in Swedish

667

gated classes (e.g., 71% uter words vs 29% neuter

words), we use majority label prediction, called Zero

rule, as our baseline. In this case, our accuracy base-

line for each of the nominal classiﬁcations is equal to

relative size of the larger class, i.e., 71.0% for gram-

matical gender prediction, 84.0% for count/mass pre-

diction, and 85.5% for common/proper noun predic-

tion. Moreover, we expect to obtain adequate mea-

sures not only for the overall accuracy of the classi-

ﬁer, but also for the detailed performance on every

single class. For instance, in the classiﬁcation task of

uter/neuter, did one of the two classes represent more

difﬁculties for the classiﬁer. Hence, we generate from

the classiﬁer’s output the two values of precision and

recall. Precision evaluates how many tokens are cor-

rect among all the output of the classiﬁer, i.e., preci-

sion is equal to the total number of correctly retrieved

tokens divided by the total number of retrieved to-

kens. Recall quantiﬁes how many tokens are correctly

retrieved among all the expected correct output, i.e.,

recall is obtained by dividing the total number of cor-

rectly retrieved tokens with the total number of cor-

rect tokens in the dataset. The two measures evaluate

different facets of the output, thus they are merged

into the F-score, which is equal to the harmonic mean

of the precision and recall, 2

recall×precision

recall+precision

Furthermore, we also provide two ﬁgures for ev-

ery class of nouns we targeted. First, we display how

the noun classes are distributed in the distributional

semantic space formed by the word vectors. Second,

we show the histogram of the entropy of the neural

network’s output for each class of nouns. The en-

tropy measures the uncertainty involved in the neural

network’s output to identify the noun classes. High

values of the entropy can be interpreted as more un-

certainty in the classiﬁer’s outputs, which itself show

the weakness of the information provided by the input

word vectors with regard to the nominal classes. The

skewness of the histogram toward right or left shows

the certainty of the classiﬁer for a particular nominal

class. Once the histogram is skewed toward right, the

classiﬁer is uncertain about its outputs. However, the

left skewness means higher certainty in the output.

5.1 Grammatical Gender

The overall accuracy and processing time of the clas-

siﬁers is shown in Table 2. As expected, neural net-

work reaches higher accuracy (Rand index), as it was

able to identify correctly the grammatical gender of

93.6% of the nouns in the test set. Such performance

is conjointly higher than our baseline accuracy, which

is 71.0%. LDA could reach relatively high accuracy

as neural network but in signiﬁcantly shorter time.

Table 2: The performance of LDA and neural network (NN)

on the grammatical gender prediction.

LDA NN

accuracy 92.7% 93.6%

uter recall 94.8% 96.3%

uter precision 94.8% 95.0%

uter f-score 94.8% 95.6%

neuter recall 87.9% 87.4%

neuter precision 87.9% 90.1%

neuter f-score 87.9% 88.7%

This shows that the word vectors could encode

the information about the grammatical genders of the

nouns and they are almost linearly separable with re-

spect to the grammatical genders of the nouns. Such

statement is supported by the semantic space of neu-

ral network in Figure 1. The uter nouns (black) and

the neuter nouns (red) are forming two clusters which

only overlap at a small area. As expected, this in-

termediary zone is precisely where most of the errors

generated by the neural network (green and blue) are

located. Moreover, the classiﬁers had more difﬁcul-

ties identifying neuter nouns. For instance, neural

network could interpret uter nouns with higher f-score

(95.6%) compared to neuter nouns (88.7%).

Figure 1: tSNE representation of the word vectors with re-

gard to their grammatical genders associated predicted by

the neural network.

Figure 2 shows the histogram of the entropy of the

neural network’s outputs. The top left histogram, the

neuter nouns are classiﬁed as neuter, and the bottom

right histogram, the uter nouns are classiﬁed as uter,

are skewed toward left. This left skewness in the his-

tograms show that the classiﬁer predicts the correct

grammatical genders with high conﬁdence. This con-

ﬁrm that the information about the grammatical gen-

ders of the nouns are captured by the word vectors.

We also see that the histogram of the entropy of the

neural network’s outputs for the erroneous items, i.e.,

the top right and the bottom left graphs, are skewed

NLPinAI 2018 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

668

toward right. This displays the uncertainty involved in

the neural network’s outputs and indicates the lack of

information in the erroneous word vectors. Thus, the

analysis of the output’s entropy demonstrate that with

regard to grammatical gender, the neural network was

interpreting the grammatical gender of nouns with

high accuracy, with exception to some outliers for

which the entropy was unusually high. Further ex-

planation is provided in the following Section.

Figure 2: The histogram of the entropy of the neural net-

work’s outputs with regard to grammatical genders.

5.2 Common/Proper Nouns

Table 3) summarizes the classiﬁcation results ob-

tained from both LDA and the neural network to the

distinction between common and proper nouns. We

see that both classiﬁers outperform our baseline accu-

racy (85.5%).

Table 3: The performance of LDA and neural network (NN)

on the common noun versus proper noun distinction.

LDA NN

accuracy 93.4% 95.2%

common noun recall 96.8% 98.9%

common noun precision 95.5% 93.9%

common noun f-score 96.1% 96.3%

proper noun recall 73.4% 62.4%

proper noun precision 79

9% 91

proper noun f-score 76.5% 74.3%

In terms of accuracy, the neural network is more

accurate than LDA. However, most of this achieve-

ment is due to the unbalance data distribution which

biases the neural network toward the bigger popula-

tion. In term of f-scores, we see that LDA is more

accurate than the neural network in identifying the

proper nouns and it is as accurate as the neural net-

work in identifying the common nouns.

It is also worth noting that LDA is much more ef-

ﬁcient than neural network when the processing time

is of importance. The high accuracy obtained from

LDA show that the word vectors should be linearly

separable with regard to this nominal classiﬁcation.

We support this by the visualisation of the word vec-

tors with regard to their proper noun versus common

noun categories (see Figure 3).

Figure 3: tSNE representation of the word vectors associ-

ated with the proper noun and common noun categories by

the neural network.

Figure 4: The histogram of the entropy of the neural net-

work’s outputs with regard to the proper noun versus com-

mon noun classiﬁcation.

As shown, a certain part of proper nouns are

merged into the common nouns, thus, hard to iden-

tify. In this case, the recall of the proper nouns is

signiﬁcantly improved if we use quadratic discrim-

inant analysis (QDA) instead of the linear discrimi-

nant analysis. However, this improvement is together

with a dramatic decrease on the recall of the common

nouns. In our experiments with QDA, we obtained

the f-score of 93.3% on the common nouns and the f-

score of 71.4% on the proper nouns which are smaller

than LDA and the neural network.

As an additional evidence, the histogram of the

entropy of the neural network’s outputs in Figure 4

shows less conﬁdence in the outputs of the neural net-

Lexical and Morpho-syntactic Features in Word Embeddings - A Case Study of Nouns in Swedish

669

work when it faces with proper nouns in its inputs. In

comparison with Figure 2, we see higher amount of

uncertainty in the classiﬁer’s outputs which conﬁrms

the lack of information in the word vectors with re-

gard to the common/proper noun classes.

5.3 Count/Mass Distinction

As for the classes of count and mass nouns, the clas-

siﬁers did not perform as well as the two other tasks.

Table 4 summarizes the results obtained from the

count/mass distinction. As shown, the accuracy ob-

tained from both classiﬁers are below our baseline,

85.5%. As we mentioned before, our baseline uses

the Zero rule for classiﬁcation which simply predicts

the majority class, i.e., count noun in this case. So,

it will be completely unable to predict the minority

class of mass nouns.

Table 4: The performance of LDA and neural network (NN)

on the count noun versus mass noun distinction.

LDA NN

accuracy 74

8% 82

count noun recall 74.3% 99.0%

count noun precision 93.7% 83.2%

count noun f-score 82.9% 90.4%

mass noun recall 76.9% 7.2%

mass noun precision 39.1% 61.4%

mass noun f-score 51.8% 12.7%

The baseline’s recall for the mass nouns, however,

is 0.0% since all the mass nouns are classiﬁed as count

noun. This is similar to the results obtained from the

neural network, resulting in the high precision 99.0%

on the count nouns but very small precision on the

mass nouns. This basically shows that the neural net-

work’s prediction on the mass nouns is almost always

wrong, i.e., mass nouns are always classiﬁed as count

noun by the neural network. However we see that

the LDA’s recall on the mass nouns (76.9%) is sig-

niﬁcantly higher than the baseline’s recall (0.0%) and

the neural network’s recall (7.1%). This is because

of the generative nature of LDA that always give a

chance to all classes, regardless of their size, to ap-

pear in the prediction task. However, this might not

always work well specially when the data is not prop-

erly distributed with regard to the classes. In this case,

the classiﬁer will result in a small value of precision,

as we see for LDA’s precision on the mass nouns. In

general, the weak performance of the classiﬁers on

the count/mass classiﬁcation task shows that the word

vectors have almost no information about this fea-

ture of the nouns. This is because of the migration

of words between the two classes (see Section 2.3)

which allows a noun to be count or mass depending on

its syntactic environment. This dynamic behavior of

nouns results in a big overlap between the two classes

of word vectors associated with count and mass nouns

and makes the word vectors ineffective for this task.

The overlap between the two classes of word vectors

is seen in Figure 5. We see that mass nouns and the

count nouns appear everywhere in the semantic space.

Figure 5: tSNE representation of the word vectors associ-

ated with the mass nouns and count nouns by the neural

network.

Such observation is further supported by the his-

togram of the entropy of the neural network’s out-

puts in Figure 6. First of all, with regard to the tasks

of uter/neuter and common/proper classes, the neural

network is highly certain about the correctly classiﬁed

tokens (see Figure 2 and Figure 4).

Figure 6: The histogram of the entropy of the neural net-

work’s outputs with regard to the proper noun versus com-

mon noun classiﬁcation.

However, the neural network does not perform as

well on count/mass categorization. We see that the

classiﬁer is very uncertain about the identiﬁcation of

count versus mass nouns. This shows that the classi-

ﬁer could not ﬁnd any relevant information in the dis-

tribution of word vectors and its decisions rely highly

NLPinAI 2018 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

670

on chance. This uncertainty is seen is the four scenar-

ios shown in Figure 6. We observe that the amount

of uncertainty involved in the classiﬁer’s decisions on

the mass nouns that are correctly classiﬁes as mass

noun is higher than the count nouns that are correctly

classiﬁed as count noun. This because the likelihood

of migration (shift) from mass category to count cate-

gory is higher than the migration from the count cate-

gory to mass category (Gillon, 1999, p.57-58).

6 DISCUSSION

Our research question was to evaluate the perfor-

mance of word embedding on the nominal classiﬁ-

cation tasks, uter/neuter, proper/common, count/mass

noun distinctions. Through our preliminary analy-

sis, we deduce that word vectors are satisfactory for

uter/neuter and proper/common noun distinctions, but

they show weak performance on the count/mass nom-

inal classiﬁcation. Nonetheless, the classiﬁers still

encountered difﬁculties when categorizing the three

nominal classes in our experiment. Hence, we pro-

vide a preliminary error analysis of the erroneous out-

put of the classiﬁers. Due to ranking of performance

and space limitation, we only scrutinize the wrongdo-

ings of the neural network. As a preparatory analysis,

we do not provide a quantitative error analysis, we

rather describe qualitatively the main types of errors

attested. Further experiments and statistic measure-

ments are planned in future studies to investigate the

recurrent frequency of different types of errors gener-

ated by the classiﬁers.

Our analysis shows that most errors are due to pol-

ysemy, i.e., the coexistence of several meanings for a

unique word form. For instance, one frequently ob-

served type of error is due to participles which can be

used both as adjectives and nouns. By way of illustra-

tion with the task of uter/neuter distinction, the word

sovande ’sleeping’ may refer to the action of sleeping

as a neuter noun in (1).

(1) det

it.NEUT

go.PRS

att

bryta

break.INF

cirkeln

circle.DEF.UTER

aligt

bad.NEUT

sovande

sleeping

‘It is possible to break the circle of bad sleeping.’

However, it can also be interpreted as an adjective

in (2). Such situation results in incoherent context in-

formation when building the word vectors. Thus, the

classiﬁer equivalently encounters difﬁculties to iden-

tify the correct class of the noun.

(2) brasilien

Brazil

kan

can

inte

not

forts

atta

continue.INF

att

vara

be.INF

one.UTER

sovande

sleeping

atte

giant

‘Brazil can not continue to be a sleeping giant.’

(2) and (1) further demonstrates the source of con-

fusion for the word vectors via the different articles

leading the target word. In (1), neuter agreement

is reﬂected on the adjective ‘bad’. However, in (2),

sovande is preceded by the uter indeﬁnite article en

which actually refers to the next uter noun j

atte ‘gi-

ant’. Such situation are expected to occur frequently

as the majority of nouns in Swedish are afﬁliated to

the uter gender. Since the parameters of word em-

beddings were set to include the immediate preceding

word of the target, polysemous neuter nouns are more

likely to be erroneously identiﬁed as uter nouns. This

is exactly what we observed in Section 5.

Furthermore, polysemy may also occur with re-

gard to proper and common nouns. As an example in

(3), the word springer may refer to the name of a per-

son, e.g., Axel Springer. As a reminder, capitalization

effect was removed from our corpora. Thus, the word

embeddings could not simply retrieve the information

relevant to common and proper nouns by identifying

the starting letter of the encountered nouns.

(3) axel

Axel

springer

Springer

be.PRS

and

known

som

agare

owner

till

bild

Bild

som

which

be.PRS

tysklands

Germany.POSS

och

and

europas

Europe.POSS

orsta

biggest

tidning.

magazine

‘Axel Springer is known as the owner of Bild,

which is Germany’s and Europe’s largest maga-

zine.’

Nonetheless, the word form springer can also

carry the meaning of a verb, i.e., ‘run’, e.g. in (4).

Once more, the verb interpretation of springer is more

frequent than the reference proper noun. Thus, the

word vectors are disrupted when considering the syn-

tactic and semantic context of springer as a noun,

since it mistakenly involves the syntactic and seman-

tic contexts of springer as a verb.

(4) det

it.NEUT

be.PRS

andra

second

aret

year.DEF.NEUTER

som

which

jag

springer

run.PRS

here

‘That’s the second year I’m running here.’

Lexical and Morpho-syntactic Features in Word Embeddings - A Case Study of Nouns in Swedish

671

Moreover, (3) is in fact a double example of poly-

semy. As displayed in (5), the word bild is usually a

noun, i.e., ‘picture’. Yet, bild is employed as a proper

noun in (3) when referring to the name of a maga-

zine. Hence, such examples demonstrate the ease of

category shift for the same word form.

(5) en

one.UTER

gammal

old.UTER

bild

picture

barn

children

‘An old picture of children.’

Finally, polysemy likewise occurred in terms of

count/mass classiﬁcation. As shown in (6), tequila

can refer to tequila as a category of liquid, thus a

mass. In such situation, the noun is expected to oc-

cur without indeﬁnite articles or plural formation.

(6) samtliga

all

kan

can

oras

done

med

with

vodka

eller

tequila

‘All can be done with vodka or tequila.’

However, as explained in Section 2.3, mass nouns

regularly undergo conversion to count nouns. As dis-

played in (7), the noun tequila is considered as a count

noun since it does not refer to the liquid tequila, but

rather to a shot of tequila, which can be counted.

Hence, tequila as a count noun can be preceded by

the indeﬁnite article en.

(7) han

get

orja

begin.INF

med

with

att

ber

atta

tell.INF

hur

how

mycket

much

han

druckit.

drink.PERF

ett

one.NEUT

glas

glass

vin

wine

och

and

one.UTER

tequila.

tequila

‘He gets to start with telling how much he drank.

A glass of wine and a tequila.’

Conversion is extremely productive between the

count/mass classes of nouns (Gillon, 1999; Doetjes,

2012). Hence, it creates additional difﬁculties to the

word embedding models. Therefore, we may de-

duce that static syntactic and semantic nominal fea-

tures such as grammatical gender (uter/neuter) and the

common/proper nouns differentiation are more gen-

erally interpretable by word embedding. However,

cases of polysemy represent a challenge to word em-

bedding, which may require additional tuning to reach

high precision, c.f. the accuracy of the count/mass

distinction was lower since the count/mass category

is more versatile. Such results are in accordance with

ﬁndings from linguistic and psycholinguistic stud-

ies. The brain differentiates count and mass nouns in

terms of syntax and semantics. However, “one-to-one

mappings between mass-count syntax and semantics

is not supported by empirical ﬁndings” (Barner and

Snedeker, 2005; Chiarelli et al., 2011).

7 CONCLUSIONS

Through the application of word embedding (Basirat

and Nivre, 2017) with various classiﬁers such as lin-

ear discriminant analysis (LDA) and feed-forward

neural network (NN), we are able to demonstrate that

some types of nominal features can be captured by

word embedding models. We show that both gram-

matical and semantic properties of nouns may be

identiﬁed correctly through word embedding. How-

ever, we equivalently point out the importance of pol-

ysemy with regard to classiﬁcation task. By way of

illustration, the count/mass categorization represented

difﬁculties for the classiﬁers not only due to its com-

plex grammatical and semantic parameters, but also

for its high occurrences of conversion (shift) across

nouns. Such issues do incarnate complication for

word embeddings. Thus, the application of word em-

beddings is considered as adequate for static nomi-

nal features such as grammatical genders or semantic

features, but less appropriate for ﬂuctuating proper-

ties such as count/mass. The high accuracy of LDA

on the nominal classiﬁcation task shows that the word

vectors are almost linearly separable with regard to

the nominal classes.

The limitations of our study include a lack of di-

versity in terms of data. For instance, we only in-

cluded one language (Swedish) in our analysis. It

would be necessary to enlarge the sampling and run

the experiment on a phylogenetically weighted group

of languages. Furthermore, we solely selected one

model of word embeddings, i.e., Real-valued Syntac-

tic word Vectors. More existing architectures such as

GloVe and word2vec should be involved in our ex-

perimentation to further testify the pros and cons of

word embeddings. Finally, we identiﬁed the issue of

polysemy with regard to classiﬁcation tasks via linear

word embedding models. Thus, the next step would

be to undergo similar functions with different types of

vector generator, e.g., dependency parsing, and com-

pare their respective performance.

ACKNOWLEDGEMENTS

Our work on this paper was fully collaborative; the

order of the authors names is alphabetical and does

not reﬂect any asymmetry in contribution. We are

grateful for the fruitful discussion with the audience

NLPinAI 2018 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

672

of the Special Session on Natural Language Process-

ing in Artiﬁcial Intelligence at the 10th International

Conference on Agents and Artiﬁcial Intelligence in

Funchal, Madeira. We would like to express our grat-

itude to our colleagues Linnea

Oberg, Karin Koltay,

Rima Haddad, and Joseﬁn Lindgren for their com-

ments and support. We also appreciate the construc-

tive comments from the anonymous referees. We are

fully responsible for any remaining errors.

REFERENCES

Aikhenvald, A. Y. (2000). Classiﬁers: A Typology of Noun

Categorization Devices. Oxford University Press, Ox-

ford.

Aikhenvald, A. Y. (2012). Round women and long men:

Shape, size, and the meanings of gender in New

Guinea and beyond. Anthropological Linguistics,

54(1):33–86.

Andersson, A.-B. (1992). Second language learners’ ac-

quisition of grammatical gender in Swedish. PhD dis-

sertation, University of Gothenburg, Gothenburg.

Baeza-Yates, R. and Ribeiro-Neto, B. (2011). Modern in-

formation retrieval: The concepts and technology be-

hind search. Addison Wesley Longman Limited, Es-

sex.

Baldwin, T., Paul, M., and Joseph, A. (2009). Restoring

punctuation and casing in English text. Proceedings

of the 22nd Australian Joint Conference on Artiﬁcial

Intelligence (AI09), pages 547–556.

Barner, D. and Snedeker, J. (2005). Quantity judgments

and individuation: Evidence that mass nouns count.

Cognition, 97:41–66.

Basirat, A. and Nivre, J. (2017). Real-valued Syntactic

Word Vectors (RSV) for Greedy Neural Dependency

Parsing. In Proceedings of the 21st Nordic Confer-

ence on Computational Linguistics, NoDaLiDa, pages

21–28, Gothenburg. Linkping University Electronic

Press.

Bohnacker, U. (2004). Nominal phrases. In Josefsson, G.,

Platzack, C., and Hkansson, G., editors, The acquisi-

tion of Swedish grammar, pages 195–260. John Ben-

jamins, Amsterdam.

Borer, H. (2005). Structuring Sense, part I. Oxford Univer-

sity Press, Oxford.

Borin, L., Forsberg, M., and Lnngren, L. (2008). The hunt-

ing of the BLARK - SALDO, a freely available lexi-

cal database for Swedish language technology. Studia

Linguistica Upsaliensia, pages 21–32.

Carter, D., Kaja, J., Neumeyer, L., Rayner, M., Weng, F.,

and Wirn, M. (1996). Handling compound nouns in

a Swedish speech-understanding system. Proceed-

ings of the Fourth International Conference on Spoken

Language, 1:26–29.

Cheng, C.-Y. (1973). Response to Moravcsik. In Hintikka,

J., Moravczik, J., and Suppes, P., editors, Approaches

to natural language, pages 286–288. D.Reidel, Dor-

drecht.

Chiarelli, V., El Yagoubi, R., Mondini, S., Bisiacchi, P., and

Semenza, C. (2011). The Syntactic and Semantic Pro-

cessing of Mass and Count Nouns: An ERP Study.

PLoS ONE, 6(10):e25885.

Chierchia, G. (1998). Plurality of mass nouns and the notion

of semantic parameter. In Rothstein, S., editor, Events

and grammar, pages 53–104. Dordrecht, kluwer edi-

tion.

Chierchia, G. (2010). Mass nouns, vagueness and semantic

variation. Synthese, 174(1):99–149.

Contini-Morava, E. and Kilarski, M. (2013). Functions of

nominal classiﬁcation. Language Sciences, 40:263–

299.

Corbett, G. G. (1991). Gender. Cambridge University

Press, Cambridge.

Corbett, G. G. (2000). Number. Cambridge University

Press, Cambridge.

Corbett, G. G. (2013). Number of Genders. In Dryer,

M. S. and Haspelmath, M., editors, The World Atlas

of Language Structures Online. Max Planck Institute

for Evolutionary Anthropology, Leipzig.

Dahl, O. (2000). Elementary gender distinctions. In

Unterbeck, B. and Rissanen, M., editors, Gender in

grammar and cognition, pages 577–593. Mouton de

Gruyter, Berlin.

Delahunty, G. P. and Garvey, J. J. (2010). The English Lan-

guage: From Sound to Sense. West Lafayette, parlor

press edition.

Dixon, R. M. W. (1986). Noun class and noun classiﬁcation.

In Craig, C., editor, Noun classes and categorization,

pages 105–112. John Benjamins, Amsterdam.

Doetjes, J. (2012). Count/mass distinctions across lan-

guages. In Maienborn, C., Heusinger, K. v., and Port-

ner, P., editors, Semantics: an international handbook

of natural language meaning, part III, pages 2559–

2580. Mouton de Gruyter, Berlin.

Dryer, M. S. (2005). Coding of nominal plurality. In

Haspelmath, M., Dryer, M. S., Gil, D., and Comrie,

B., editors, The word atlas of language structures,

pages 138–141. Oxford University Press, Oxford.

Fraurud, K. (2000). Proper names and gender in Swedish.

In Unterbeck, B., Rissanen, M., Nevalainen, T., and

Saari, M., editors, Gender in Grammar and Cognition,

pages 167–220. Mouton de Gruyter, Berlin.

Gillon, B. S. (1999). The lexical semantics of English count

and mass nouns. In Viegas, E., editor, Breadth and

depth of semantic lexicons, pages 19–37. Springer,

Dordrecht.

Hall, K. (2002). Unnaturalgender in Hindi. In Gender

across Languages: The Linguistic Representation of

Women and Men, pages 133–162. John Benjamins,

Amsterdam.

Katz, G. and Zamparelli, R. (2012). Quantifying

count/mass elasticity. Proceedings of the 29th West

Coast Conference on Formal Linguistics, pages 371–

379.

Kemmerer, D. (2017). Categories of object concepts across

languages and brains: the relevance of nominal classi-

ﬁcation systems to cognitive neuroscience. Language,

Cognition and Neuroscience, 32(4):401–424.

Lexical and Morpho-syntactic Features in Word Embeddings - A Case Study of Nouns in Swedish

673

Kilarski, M. (2014). The Place of Classiﬁers in the

History of Linguistics. Historiographia Linguistica,

41(1):33–79.

Lopez, A. (2008). Statistical machine translation. ACM

Computing Surveys, 40(3):1–49.

Nespital, H. (1990). On the relation of Hindi to its re-

gional dialects. In Offredi, M., editor, Language ver-

sus dialect: Linguistic and literary essays on Hindi,

Tamil and Sarnami, pages 3–23. Manohar Publica-

tions, New Delhi.

Nichols, J. (1989). The origin of nominal classiﬁcation.

Proceedings of the ﬁfteenth annual meeting of the

Berkeley linguistics society, pages 409–420.

Ostling, R. and Wirn, M. (2013). Compounding in

a Swedish blog corpus. Acta Universitatis Stock-

holmiensis, pages 45–63.

Pang, B. and Lee, L. (2008). Opinion mining and sentiment

analysis. Foundations and Trends in Information Re-

trieval, 2(1-2):1–135.

Pelletier, F. J. and Schubert, L. K. (1989). Mass expressions.

In Gabbay, D. and Guenther, F., editors, Handbook of

philosophical logic. Volume IV: Topics in the philoso-

phy of language, pages 327–408. Reidel, Dordrecht.

Petrovic , S., Osborne, M., and Lavrenko, V. (2010).

Streaming ﬁrst story detection with application to

twitter. Human Language Technologies: The 2010

Annual Conference of the North American Chapter of

the Association for Computational Linguistics, pages

181–189.

Pokharel, M. P. (2010). Noun class agreement in Nepali.

Kobe papers in linguistics, 7:40–59.

Preiss, J. and Stevenson, M. (2013). Distinguishing com-

mon and proper nouns. Second joint conference on

lexical and computational semantics: Proceedings of

the main conference and the shared task: Semantic

textual similarity, 1:80–84.

Quine, W. v. O. (1960). Word and object. MIT Press, Cam-

bridge.

Rand, W. M. (1971). Objective Criteria for the Evaluation

of Clustering Methods. Journal of the American Sta-

tistical Association, 66(336):846–850.

Seifart, F. (2010). Nominal classiﬁcation. Language and

Linguistics Compass, 4(8):719–736.

Senft, G. (2000). Systems of nominal classiﬁcation. Cam-

bridge University Press, Cambridge.

Teleman, U., Hellberg, S., and Andersson, E. (1999).

Svenska Akademiens grammatik. Vol. 2: Ord.

[The Swedish Academy Grammar, Part 2: Words].

Norstedts, Stockholm.

Ting, K. M. (2010). Precision and Recall. In Sammut,

C. and Webb, G. I., editors, Encyclopedia of Machine

Learning, pages 781–781. Springer US, Boston, MA.

DOI: 10.1007/978-0-387-30164-8 652.

Ullman, E. and Nivre, J. (2014). Paraphrasing Swedish

Compound Nouns in Machine Translation. In MWE@

EACL, pages 99–103.

NLPinAI 2018 - Special Session on Natural Language Processing in Artiﬁcial Intelligence

674