A Case Study and Qualitative Analysis of Simple Cross-lingual Opinion

Mining

Gerhard Hagerer

1 a

, Wing Sheung Leung

, Qiaoxi Liu

, Hannah Danner

2 b

and Georg Groh

1 c

Social Computing Research Group, Department of Informatics, Technical University of Munich, Germany

Chair of Marketing and Consumer Research, TUM School of Management, Technical University of Munich, Germany

Keywords:

Opinion Mining, Topic Modeling, Sentiment Analysis, Cross-lingual, Multi-lingual, Market Research.

Abstract:

User-generated content from social media is produced in many languages, making it technically challenging

to compare the discussed themes from one domain across different cultures and regions. It is relevant for

domains in a globalized world, such as market research, where people from two nations and markets might

have different requirements for a product. We propose a simple, modern, and effective method for building

a single topic model with sentiment analysis capable of covering multiple languages simultanteously, based

on a pre-trained state-of-the-art deep neural network for natural language understanding. To demonstrate its

feasibility, we apply the model to newspaper articles and user comments of a speciﬁc domain, i.e., organic food

products and related consumption behavior. The themes match across languages. Additionally, we obtain an

high proportion of stable and domain-relevant topics, a meaningful relation between topics and their respective

textual contents, and an interpretable representation for social media documents. Marketing can potentially

beneﬁt from our method, since it provides an easy-to-use means of addressing speciﬁc customer interests from

different market regions around the globe. For reproducibility, we provide the code, data, and results of our

study

https://github.com/apairofbigwings/cross-lingual-opinion-mining

1 INTRODUCTION

Topic modeling on social media texts is difﬁcult,

since lack of data as well as spelling and grammati-

cal errors can make the approach unfeasible. Dealing

with multiple languages at the same time adds more

complexity to the problem which oftentimes makes

the approach unusable for domain experts. Thus,

we propose a cross-lingually pre-trained deep neural

network as a black box with very little textual pre-

processing necessary before embedding the texts and

forming their clustering and topic distributions.

For our method, we leverage current research

regarding multi-lingual topic modeling, see Section

2. We provide an extensive description of a simple

method to support domain experts from speciﬁc so-

cial media domains in its application in Section 3.

We qualitatively demonstrate our topic model, its fea-

sibility, and its cross-lingual semantic characteristics

https://orcid.org/0000-0002-2292-0399

https://orcid.org/0000-0001-8387-0818

https://orcid.org/0000-0002-5942-2297

on English and German newspaper and social media

texts in Section 4. We aim at inspiring pragmatic

ideas to explore the potential for comparative, inter-

cultural market research and agenda setting studies.

Unsolved problems and future potential are given in

Section 5.

2 RELATED WORK

Topic modeling is meant to learn thematic structure

from text corpora. With probabilistic topic model-

ing methods, such as latent semantic indexing (LSI)

(Deerwester et al., 1990) or latent Dirichlet allocation

(LDA) (Blei et al., 2003), researchers try to extend the

capabilities of topic modeling for application from a

single language to multiple languages. Using multi-

lingual dictionaries and translated corpora is an intu-

itive way to tackle cross-lingual topic modeling prob-

lems (Zhang et al., 2010; Vuli

c et al., 2013). Further

examples exist for topic modeling with either dictio-

naries or translation text collections (Guti

errez et al.,

2016; Boyd-Graber and Blei, 2012; Jagarlamudi and

Hagerer, G., Leung, W., Liu, Q., Danner, H. and Groh, G.

A Case Study and Qualitative Analysis of Simple Cross-lingual Opinion Mining.

DOI: 10.5220/0010649500003064

In Proceedings of the 13th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2021) - Volume 1: KDIR, pages 17-26

ISBN: 978-989-758-533-3; ISSN: 2184-3228

Figure 1: Plain text is ﬁrst tokenized into sentences and passed to topic modeling and sentiment analysis. Topic modeling

involves (1) converting sentences of both languages into embeddings with XLING, (2) clustering all embeddings with K-

means and (3) deriving a topic label of each cluster. Sentiment analysis is performed using Textblob. Topic and sentiment

scores are aggregated for the analysis.

Daum

e, 2010). However, this puts dependence on the

availability of dictionary or good quality of transla-

tions. Signiﬁcant manual labor and veriﬁcation are

required to prevent deteriorating noise.

Recently, methods converting words to vectors

according to their semantics are widely adopted

(Mikolov et al., 2013). Several studies showed text

embeddings improve topic coherence (Bianchi et al.,

2020; Srivastava and Sutton, 2017). Regarding multi-

linguality, embeddings in word level and sentence

level enable text in different languages to be projected

to the same vector space (Cer et al., 2018) such that

semantically similar texts are clustered together in-

dependently of their languages. This favors studies

on multi-lingual topic modeling without relying on

dictionaries and translation (Xie et al., 2020; Chang

et al., 2018). Although providing highly coherent top-

ics, a recreation of word spaces is required when new

text corpora are introduced. In our scenario, these

limitations are not present.

Regarding the application of topic modeling, var-

ious social media corpora are studied by domain ex-

perts (Tsur et al., 2015; Ko et al., 2018). They cov-

ered on different domains, such as politics, market-

ing, and public health. Regarding media agenda set-

ting, (Field et al., 2018) studied on how much degree

a Russian newspaper related to economic downturn.

They also ”introduced embedding-based methods for

cross-lingually projecting English frame to Russian”

based on Media Frames Corpus. In contrast, we pro-

pose a straightforward topic modeling method with-

out ﬁne-tuning but only clustering necessary on a so-

cial media corpus. This enables further investigation

on media agenda setting cross-lingually and cross-

culturally.

3 TOPIC MODELING METHOD

Figure 1 shows the overall workﬂow of our topic mod-

eling approach. We aim to conduct simple, cross-

lingual topic modeling on user-generated content with

no translation, dictionary, and parallel corpus re-

quired for aligning the semantic meanings across lan-

guages. Our approach solely depends on clustering

sentence embeddings for topic modeling. Ready-

made sentence representations simplify the approach,

since these suppress too frequent, meaningless, and

unimportant words automatically without the need to

model that part explicitly (Kim et al., 2017).

3.1 Preprocessing

The raw texts of articles and comments are ﬁrst tok-

enized into sentences with Natural Language Toolkit

(NLTK). Then, URLs, specially for those enclosed

with HTML <a> tag, are replaced with string ’url’.

After that, sentences with character length smaller

than 15 are omitted to minimize noise, since they ap-

pear inscrutable and they are only 6.6% out of all sen-

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

Figure 2: AIC plot indicates k = 15 is the global minimum.

tences which is a small portion. After preprocessing,

there are 127,464 English sentences and 200,627 Ger-

man sentences, i.e., total 328,091.

3.2 Cross-lingual Embeddings

In the following paragraph, we provide an explana-

tion of the pre-trained XLING model, which we use

for the present work, based on the words of the au-

thors (Chidambaram et al., 2018). XLING calculates

”sentence embeddings that map text written in differ-

ent languages, but with similar meanings, to nearby

embedding space representations”. Similarity is cal-

culated mathematically as dot product between two

sentence embeddings. In order to train the model, the

authors ”employ four unique task types for each lan-

guage pair in order to learn a function g”, i.e., the

eventual sentence-to-vector model. The architecture

is based on a Transformer neural network (Vaswani

et al., 2017) tailored for modeling multiple languages

at once. The tasks on which the model is eventu-

ally trained are ”(i) conversational response predic-

tion, (ii) quick thought, (iii) a natural language in-

ference, and (iv) translation ranking as the bridge

task”. The data for training ”is composed of Red-

dit, Wikipedia, Stanford Natural Language Inference

(SNLI), and web mined translation pairs”.

3.3 Sentence Clustering

K-means clustering algorithm is implemented on both

English and German sentence embeddings at the same

time. Since XLING provides semantically aligned

sentence embeddings of both languages, this joint

clustering step helps to establish one topic model for

two disjunct datasets irrespective of their language.

Clustering is established for a varying number of

clustersk, ranging from 1 to 30. Elbow method is ﬁrst

used for choosing the optimal k but the inertia (sum

of squared distances of samples to their closest clus-

ter center) of increasing k decreases rapidly at the be-

ginning and then gently without a signiﬁcant elbow

point. Therefore, Akaike Information Criterion (AIC)

is adopted and k = 15 is chosen as optimal value as it is

the global minimum, see Figure 2. In Section 4.2, fur-

ther discussion on topic coherence is conducted prov-

ing the fact that k = 15 resulted semantically coherent

topics.

3.4 Topic Labeling

To be able to derive a meaningful topic label for each

sentence cluster, the respective top words of each

cluster are required. In order to get the top word list,

the clarity score is adopted (Cronen-Townsend et al.,

2002). According to (Angelidis and Lapata, 2018), it

ranks terms with respect to their importance of each

cluster c and language l, such that

score

l,c

(w) = t

l,c

(w)log

l,c

(w)

, (1)

where t

l,c

(w) and t

(w) are the l1-normalized tf-idf

scores of the word w in the sentences within cluster c

and in all sentences, respectively, for a certain lan-

guage.

Additionally, stopword removal from the top word

lists is also a concern when calculating the clarity

score. Generally, stopwords are the most frequent

words in the documents and sometimes they are too

dominant such that they interfere with the result from

clarity scoring. Thus, we remove domain-speciﬁc

high frequency words for each language from corre-

sponding topic top word lists.

Topics are labeled manually based on the English

and German top word lists. The results are shown in

Table 1 and will be discussed further in Section 4.2

evaluating topic coherence across languages.

3.5 Sentiment Analysis

In addition to topic modeling, we conduct sentiment

analysis to investigate the feasibility and meaning

of cross-lingual topic-related sentiments in articles

and respective comment sections. We make use of

Textblob

and Textblob-de

to assign each of the

English and German pre-processed sentences a po-

larity score. The polarity assignment is ﬁrst pro-

posed by (Pang and Lee, 2004) and reimplemented by

https://github.com/sloria/textblob

https://github.com/markuskiller/textblob-de

A Case Study and Qualitative Analysis of Simple Cross-lingual Opinion Mining

Table 1: Top words for all meaningfull topics with k = 15 of English and German data.

Topic English top words German top words

Environment pesticide, plant, soil, use, crop, fertilize,

pesticide, garden, herbicide, grow

pﬂanze, pestizid, dunger, boden, gulle, garten,

anbau, gemuse, tomate, feld

Retailers store, whole, shop, groceries, supermar-

ket, local, market, amazon, price, online,

discount

aldi, supermarkt, lidl, kauf, laden, lebensmittel,

cent, einkauf, wochenmarkt

GMO

& organic

gmo, label, gmos, monsanto, product,

certiﬁcate, usda, genetic, product

produkt, bioprodukt, lebensmittel, gesund, kon-

ventionell, biodiesel, herstellung, enthalt, mon-

santo, pestizid

Food products

& taste

taste, milk, sugar, cook, eat, fresh, ﬂavor,

fruit, potato, sweet

kase, schmeckt, gurke, essen, analogkase,

schmeckt, tomate, milch, geschmack, kochen

Food safety chemical, cancer, body, acid, effect,

cause, toxic, toxin, glyphosat, disease

dioxin, gift, grenzwert, ddt, menge, giftig, tox-

isch, substanz, chemisch, antibiotika

Research science, study, scientiﬁc, research, gene,

scientist, genetic, human, stanford, na-

ture

gentechnik, natur, mensch, wissenschaft,

lebenserwartung, genetisch, studie, men-

schlich, planet

Health

& nutrition

eat, diet, healthy, nutritious, health, fat,

calory, obesity, junk

lebensmittel, essen, ernahrung, gesund,

nahrungsmittel, lebensmittel, nahrung, fett,

billig

Politics

& policy

govern, public, politic, corporate, regu-

lation, law, obama, vote

politik, skandal, verantwortung, bundestag,

schaltet, bestraft, strafe, kontrolle, kriminell

Animals

& meat

meat, chicken, anim, cow, beef, egg, fed,

raise, pig, grass

tier, ﬂeisch, eier, huhn, schwein, futter, kuh,

verunsichert, vergiftet, deutsch

Farming farm, farmer, agriculture, land, sustain,

crop, yield, acre, grow, local

landwirtschaft, landwirt, bau, ﬂache, okolo-

gisch, nachhaltig, konventionell, landbau, pro-

duktion, ertrag

Prices & proﬁt price, consume, market, company, proﬁt,

product, cost, amazon, money

verbrauch, preis, produkt, billig, qualitat,

kunde, kauf, geld, unternehmen, kosten

(De Smedt and Daelemans, 2012). Since the subjec-

tivity assignment is not well-developed in Textblob-

de, we ﬁlter out sentences with polarity equals to 0 for

both English and German sentences in order to derive

comparable results.

3.6 Topic and Sentiment Distributions

After assigning a labeled cluster, i.e., a topic, and a

sentiment score to each sentence of the corpus, we

derive the corresponding distributions.

For topic distributions, all sentences from a doc-

ument are counted per topic. The distribution is then

normalized to be comparable. For sentiment distri-

butions, all sentences from a document are grouped

per topic. Topic-wise sentiment distribution is derived

based on the sentence-wise polarity scores and the re-

spective median and quartiles. A document in that

regard is either an article or all of its comments, i.e.,

its comment section.

4 TOPIC COHERENCE

In this section, we evaluate the feasibility and se-

mantic coherence of our cross-lingual topic modeling

qualitatively. Instead of providing quantitative coher-

ence scores, we aim at a detailed, qualitative analy-

sis of textual examples. We depict representative sen-

tences and words of each topic in subsection 4.2. We

investigate to what extent these are semantically co-

herent, also across languages. We expose the ratio

of coherent and incoherent topics and how it devel-

ops with increasing number of topics in in subsection

4.3. Eventually, we show the distribution of topics in

selected newspaper articles and their respective com-

ment sections to relate the discussed content with our

actual topic model on English and German texts.

4.1 Data

The collection of the data used in this study is de-

scribed in another publication (Danner et al., 2021) as

follows. For the analysis we downloaded ”news arti-

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

k = 5 [23.72%]

k = 10 [30.26%]

k = 15 [29.26%]

k = 20 [23.79%]

Figure 3: Topic distributions with increasing number of topics k. The percentage is the amount of sentences in garbage topics.

cles and reader comments of two major news outlets

representative of the German and the United States

(US) context”, i.e., spiegel.de and nytimes.com. The

creation dates of the texts are ”spanning from January

2007 to February 2020”. ”Articles and related com-

ments on the issue of organic food were identiﬁed us-

ing the search terms organic food and organic farm-

ing and the German equivalents. For topic modeling,

we utilized ”534 articles and 41,320 comments from

the US for the years 2007 to 2020, and 568 articles

and 63,379 comments from Germany for the years

2007 to 2017 and the year 2020”.

4.2 Multi-linguality of Topics

In this section, we evaluate semantic coherence of

our cross-lingual topic modeling by depicting the rep-

resentative sentences and words for each topic and

showing the semantic relation. Table 1 shows the ﬁrst

10 English and German words having the highest clar-

ity scores (see Section 3.4) in each cluster for k = 15.

Table 2 shows the ﬁrst 3 English and German sen-

tences whose embeddings have the largest cosine sim-

ilarity to their corresponding cluster centroids. Both

top words and top sentences indicate that the clusters

are grouped reasonably in terms of semantics. For ex-

ample, this is the case for the topic Environment (pes-

ticides & fertilizers) which is indeed related to use of

pesticides in planting. Even though this also appears

to be the case for the sentences in GMO & organic on

the ﬁrst glance, those are actually about organic food

and how aspects such as GMO and pesticides relates

to the food itself. This and the other representative

top words and sentences indicate that clustering on

cross-lingual sentence embeddings yield semantically

coherent topics.

According to our analysis, top sentences from

garbage clusters are always short in length with

slightly more than 15 characters. Together with top

words (Table 1), these hardly contribute to the organic

food domain and corresponding entities. Thus, it is

feasible in our case to ignore them.

4.3 Amount of Meaningful Topics

Besides providing coherent cross-lingual topics, our

method performs well to distinguish usable from un-

usable topics, and it provides a constantly high num-

ber of relevant topics independently of the number of

overall topics. Figure 3 is a Sankey diagram showing

A Case Study and Qualitative Analysis of Simple Cross-lingual Opinion Mining

Figure 4: Topic and sentiment distribution for Grocer.

the ﬂow of topic assignments for all English and Ger-

man sentences with increasing number of clusters k.

Topic modeling is performed for each k with all pre-

processed sentences independently. It can be seen that

more speciﬁc topics descend from general but related

topics as indicated by the colors.

For instance, GMO & organic, Food safety, En-

vironment (pesticides & fertilizers) and Farming &

agricultural policy & food security for k = 15 are

derived from Organic vs. conventional farming for

k = 5. Organic vs. conventional farming in k = 5 gen-

erally focuses on advantages brought by organic farm-

ing when comparing to conventional farming, such as

reducing persistent toxin chemicals from entering to

environment, food, and bodies; thus, bio-products are

recommended. For k = 15, the children topics are

more speciﬁc. For example, GMO & organic shows

the aims for having organic food, i.e., avoidance of

GMO and poisoning with pesticides. Moreover, Food

safety in k = 15 is further split into Food safety and

Environmental pollution.

To see how the topics relate to their actual sen-

tences, we try to observe the top sentences of each

topic, i.e., those sentences of which the embeddings

are closest to the centroid. Both English and Ger-

man sentences are similar and share strong seman-

tic similarity. The food safety topic focuses on the

toxicity issue of dioxin and other chemicals towards

consumers. Environmental pollution, which is further

splitted from it, for k = 20 indeed tells contamina-

tion of water resources by chemicals. This shows that

ﬁne-grained topics and the way they develop with in-

creasing k have a meaningful relation to ancestor and

sister topics.

Sentences without contribution to the organic food

domain always remain in garbage clusters in a way

that the proportion of usable and unusable clusters

does not ﬂuctuate. Thus, the topic model maintains

its coherence independent to the number of topics and

the despite the fact that k-means is not deterministic

in its clustering. This property is helpful, since the

number of topics can be chosen as high as necessary

to provide a sufﬁcient level of detail for the domain of

interest. Moreover, this highlights the meaningfulness

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

Figure 5: Topic and sentiment distribution for

Oko-Test.

and robustness of the given sentence representations

being able to separate noise from informative data in

an unsupervised fashion.

4.4 Validation of Opinion Distributions

In this section, two real text examples are given to

evaluate our method qualitatively. The ﬁrst one is an

article from New York Times, titled ’Major Grocer

to Label Foods With Gene-Modiﬁed Content’

, here-

after referred to as Grocer. It reported that the ﬁrst

retailer in the United States announced to label all of

its genetically modiﬁed food sold in its stores. Ad-

vocating and opposing stakeholders stated their argu-

ments regarding different aspects. The second ex-

ample is from Der Spiegel, titled ’”

Oko-Test” und

Co. – Welche Lebensmittelsiegel wirklich taugen’

below denoted as

Oko-Test. It reported that number

https://www.nytimes.com/2013/03/09/business/

grocery-chain-to-require-labels-for-genetically-modiﬁed-food.

html

https://tinyurl.com/spiegel-lebensmittelsiegel

of food claims, certiﬁcations, and seals in Germany

were growing as organic labeling was a good promo-

tional strategy indicating high food quality. However,

consumers knew little about the details even when

the tests for each label were transparent and well-

documented. Based on these two summaries, it would

be expected that topics related to supermarkets, retail-

ers, and GMO labels are shown to be present in those

articles. The Grocer article, however, expresses con-

cerns about the consumption of genetically modiﬁed

food, whereas

Oko-Test discusses organic food label-

ing issues from various point of views, among others

fair trading and organic ﬁshing.

Topic Distribution. Figures 4 and 5 show the dis-

tribution of topics in the overall article sentences. It

can be seen that the two topics Retailers and GMO &

organic are mentioned the most in both articles, sup-

porting our hypothesis. The comment section of the

Grocer article corresponds to the article itself such

that most of its sentences also talk about GMO &

organic and the second most for Retailers. How-

A Case Study and Qualitative Analysis of Simple Cross-lingual Opinion Mining

ever, the commenters of the

Oko-Test articles com-

mented more about GMO & organic followed with

Consumer prices & proﬁt and Food products & taste.

Even though the dominating topics in German differ

between article and comments, it can be stated that

the topic distribution overall still refers to the actual

topics of the given texts and domains. At the same

time, differences in the distribution not only between

article and comments but also between languages and

thus cultures are directly visible, providing a means

for clear comparability in several respects.

Sentiment Distribution. Figures 4 and 5 also show

the sentiment distribution. Generally, the sentiment

of the Grocer article spreads out less than that of the

Oko-Test article. It is observed that, in topic GMO &

organic, comments score sentiment polarity ranging

between 0.50 to −0.70 in Grocer and between 1 and

−0.85 in

Oko-Test. This means sentences from Gro-

cer show weaker sentiment compared to those from

Oko-Test. The actual texts indicate that sentiment on

our German data indeed has more variance than on

English. Thus, the proposed multi-lingual sentiment

analysis, Textblob and Textblob-de, appears to repre-

sent the data adequately in the given use case. How-

ever, it cannot be excluded that the sentiment distribu-

tion could be affected given the fact that two different

but methodically similar frameworks are used. Differ-

ent biases and variances could be caused by different

models which have differences in the sentiment dic-

tionary size and the subjectivity of human-assigned

sentiment scores based on different cultures. Further

studies should examine this problem for more robust,

domain-independent multi-lingual sentiment predic-

tion.

5 CONCLUSION

This case study shows that our technically simple ap-

proach successfully generates an high proportion of

relevant and coherent topics for our domain, i.e., or-

ganic food products and related consumption behav-

ior based on English and German social media texts.

Moreover, the topics display the text contents cor-

rectly and support a domain expert in the content

analysis of social media texts written in multiple lan-

guages.

However, the presented paper did not provide

quantitative measurements of topic coherences and

comparisons with the state-of-the-art. For mono-

language topic modeling, it would be LDA (Blei et al.,

2003); for advanced cross-lingual topic modeling, it

could be attention-based aspect extraction (He et al.,

2017) utilizing aligned multi-lingual word vectors

(Conneau et al., 2017). Several multi-lingual datasets

would need to be included for a representative com-

parison. Since pre-trained models trained on exter-

nal data are used for the proposed method, it might

be relevant for coherence score calculation to include

intrinsic coherence scoring methods based on train

test splits, such as, UMass coherence score (Mimno

et al., 2011), and explore extrinsic methods calculated

on external validation corpora, e.g., Wikipedia (R

oder

et al., 2015).

Regarding multi-lingual sentiment analysis, the

difference in the sentiment analysis frameworks for

different languages must be considered. For example,

since two independent but similar sentiment analysis

models are applied for English and German, the senti-

ment distribution could be affected. Therefore, future

studies on developing and evaluating comparable sen-

timent models should be conducted.

REFERENCES

Angelidis, S. and Lapata, M. (2018). Summarizing

opinions: Aspect extraction meets sentiment predic-

tion and they are both weakly supervised. ArXiv,

abs/1808.08858.

Bianchi, F., Terragni, S., and Hovy, D. (2020). Pre-training

is a hot topic: Contextualized document embeddings

improve topic coherence.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent

dirichlet allocation. the Journal of machine Learning

research, 3:993–1022.

Boyd-Graber, J. and Blei, D. (2012). Multilingual topic

models for unaligned text.

Cer, D., Yang, Y., yi Kong, S., Hua, N., Limtiaco, N., John,

R. S., Constant, N., Guajardo-Cespedes, M., Yuan,

S., Tar, C., Sung, Y.-H., Strope, B., and Kurzweil, R.

(2018). Universal sentence encoder.

Chang, C.-H., Hwang, S.-Y., and Xui, T.-H. (2018). In-

corporating word embedding into cross-lingual topic

modeling. 2018 IEEE International Congress on Big

Data (BigData Congress), pages 17–24.

Chidambaram, M., Yang, Y., Cer, D., Yuan, S., Sung, Y.-H.,

Strope, B., and Kurzweil, R. (2018). Learning cross-

lingual sentence representations via a multi-task dual-

encoder model. arXiv preprint arXiv:1810.12836.

Conneau, A., Lample, G., Ranzato, M., Denoyer, L., and

egou, H. (2017). Word translation without parallel

data. arXiv preprint arXiv:1710.04087.

Cronen-Townsend, S., Zhou, Y., and Croft, W. B. (2002).

Predicting query performance.

Danner, H., Hagerer, G., Pan, Y., and Groh, G. (2021). The

news media and its audience: Agenda-setting on or-

ganic food in the united states and germany.

De Smedt, T. and Daelemans, W. (2012). Pattern for

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval

python. The Journal of Machine Learning Research,

13(1):2063–2067.

Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer,

T. K., and Harshman, R. (1990). Indexing by latent

semantic analysis. Journal of the American Society

for Information Science, 41(6):391–407.

Field, A., Kliger, D., Wintner, S., Pan, J., Jurafsky, D., and

Tsvetkov, Y. (2018). Framing and agenda-setting in

Russian news: a computational analysis of intricate

political strategies. In Proceedings of the 2018 Con-

ference on Empirical Methods in Natural Language

Processing, pages 3570–3580, Brussels, Belgium. As-

sociation for Computational Linguistics.

Guti

errez, E., Shutova, E., Lichtenstein, P., de Melo, G.,

and Gilardi, L. (2016). Detecting cross-cultural dif-

ferences using a multilingual topic model. Transac-

tions of the Association for Computational Linguis-

tics, 4:47–60.

He, R., Lee, W. S., Ng, H. T., and Dahlmeier, D. (2017).

An unsupervised neural attention model for aspect ex-

traction. In Proceedings of the 55th Annual Meet-

ing of the Association for Computational Linguistics

(Volume 1: Long Papers), pages 388–397, Vancouver,

Canada. Association for Computational Linguistics.

Jagarlamudi, J. and Daum

e, H. (2010). Extracting multi-

lingual topics from unaligned comparable corpora. In

Gurrin, C., He, Y., Kazai, G., Kruschwitz, U., Little,

S., Roelleke, T., R

uger, S., and van Rijsbergen, K., ed-

itors, Advances in Information Retrieval, pages 444–

456, Berlin, Heidelberg. Springer Berlin Heidelberg.

Kim, H. K., Kim, H., and Cho, S. (2017). Bag-of-

concepts: Comprehending document representation

through clustering words in distributed representation.

Neurocomputing, 266:336–352.

Ko, N., Jeong, B., Choi, S., and Yoon, J. (2018). Identify-

ing product opportunities using social media mining:

Application of topic modeling and chance discovery

theory. IEEE Access, 6:1680–1693.

Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).

Efﬁcient estimation of word representations in vector

space.

Mimno, D., Wallach, H., Talley, E., Leenders, M., and Mc-

Callum, A. (2011). Optimizing semantic coherence in

topic models. In Proceedings of the 2011 Conference

on Empirical Methods in Natural Language Process-

ing, pages 262–272, Edinburgh, Scotland, UK. Asso-

ciation for Computational Linguistics.

Pang, B. and Lee, L. (2004). A sentimental education:

Sentiment analysis using subjectivity summarization

based on minimum cuts. arXiv preprint cs/0409058.

oder, M., Both, A., and Hinneburg, A. (2015). Exploring

the space of topic coherence measures. In Proceedings

of the eighth ACM international conference on Web

search and data mining, pages 399–408.

Srivastava, A. and Sutton, C. (2017). Autoencoding varia-

tional inference for topic models.

Tsur, O., Calacci, D., and Lazer, D. (2015). A frame of

mind: Using statistical models for detection of fram-

ing and agenda setting campaigns. In Proceedings of

the 53rd Annual Meeting of the Association for Com-

putational Linguistics and the 7th International Joint

Conference on Natural Language Processing (Volume

1: Long Papers), pages 1629–1638, Beijing, China.

Association for Computational Linguistics.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L., and Polosukhin, I.

(2017). Attention is all you need. arXiv preprint

arXiv:1706.03762.

Vuli

c, I., De Smet, W., and Moens, M.-F. (2013). Cross-

language information retrieval models based on latent

topic models trained with document-aligned compara-

ble corpora. Information Retrieval, 16.

Xie, Q., Zhang, X., Ding, Y., and Song, M. (2020).

Monolingual and multilingual topic analysis using

lda and bert embeddings. Journal of Informetrics,

14(3):101055.

Zhang, D., Mei, Q., and Zhai, C. (2010). Cross-lingual

latent topic extraction. In Proceedings of the 48th

Annual Meeting of the Association for Computational

Linguistics, pages 1128–1137, Uppsala, Sweden. As-

sociation for Computational Linguistics.

A Case Study and Qualitative Analysis of Simple Cross-lingual Opinion Mining

APPENDIX

Table 2: Top sentences of meaningful topics from the whole dataset for k = 15 in English and German.

Topic Top 3 sentences for English and German

Environ-

ment

(pesti-

cides,

fertilizers)

Usually, the plant which uses conventional farming will produce the residue of the pesticides. – Some pesticides used in conventional farming, however, may reduce the level

of resveratrol in plants. – Also, there is the question of naturally occurring pesticides produced by the plant itself.

Viele Biopﬂanzen werden zwar nicht mit Pestiziden behandelt, Ihnen wird jedoch sehr viel mehr Wachstumsﬂ

ache zugestanden. – Nicht nur Biobauern benutzen G

ulle, und

Herbizide und Pestizide werden vor allem in der konventionellen Landwirtschaft eingesetzt. – Zum einen bauen sich die Pestizide und Herbizide relativ schnell ab, nicht zu

verwechseln mit

Uberd

ungung durch G

ulle oder Belastung mit Schwermetallen.

Retailers

Whole Foods also sells a lot of high quality grocery items that aren’t available elsewhere in a lot of places. – Whole Foods executives, however, say their supermarkets can be

high quality, organic and natural but also inexpensive. – Larger competitors like Safeway and Kroger have vastly expanded their store-brand offerings of natural and organic

products, and they are often cheaper than those at Whole Foods.

Auch bei ALDI und CO lassen sich hochwertige Lebensmittel erwerben. – Wobei ich feststellen muss, dass andere Superm

arkte - zumindest die, die ich frequentiere - auch

Wert darauf legen, das gewisse Produkte aus der Region stammen, auch wenn sie konventionell hergestellt wurden. – Der Trend zur Feinkost beschert dem Handel vor allem

in den Großst

adten steigende Ums

atze, wo Bio-L

aden hip sind und die kaufkr

aftigen Kunden beim Einkaufen nicht auf jeden Cent schauen.

GMO &

organic

In a nutshell, though, organic means the product meets a number of requirements, such as no GMOs, no non-organic pesticides, etc. – When consumers buy organic, they are

guaranteed little more than food that is (in theory at least) produced without synthetic chemicals or G.M.O.’s (genetically modiﬁed organisms), and with some attention (again,

in theory) to the health of the soil. – Organic food includes products that are grown without the use of synthetic fertilisers, sludge, irradiation, GMOs, or drugs, which already

shows how much better it is for health.

Jeder weiß doch, daß der Vorteil von bio nicht in der erh

ohten Aufnahme von N

ahrstoffen gegen

uber konventionellen Produkten liegt, sondern in der Vermeidung, sich

mit Pestiziden zu vergiften. – Bio-Lebensmittel genießen einen guten Ruf, weil sie wesentlich weniger Schadstoffe enthalten als konventionell hergestellte Lebensmittel. –

Bioprodukte sind kaum ges

under als konventionelle Lebensmittel

Food

products

& taste

It tastes totally different from your normal vegetables. – It can also be mixed in with the other foods (milk and fruit, oats/rice cooked in milk). – The increased ﬂavor is the

result of the food containing more micronutrients.

Die meisten frischen Zutaten m

ussen etwas aufbereitet werden, damit sie gegessen werden k

onnen. – Es braucht wirklich nur Mehl, Wasser und etwas Salz, keinerlei andere

Inhaltsstoffe, Punkt. – Nichts geht

uber frisch zubereitete Speisen aus gesunden Zutaten.

Food

safety

Dioxins are extremely toxic chemicals, and their bioaccumulation in the food chain may potentially lead to dangerous levels of exposure. – Many of the toxins found in

non-organic foods are toxins that have a cumulative effect on our bodies. – As proved by various researchers, these chemicals have harmful effects not only on the consumers

but also on the environment and farmers.

Tatsache ist die Dosen um die es bei Nahrungsmittelkontaminationen durch Dioxine geht sind derart gering dass sie auf keinem Wege zu eine signiﬁkanten Gesundheitsgefahr

uhren. – Dioxine sind UBIQUIT

AR und entstehen in nicht unerheblichen Mengen durch nat

urliche Vorg

ange. – Es bedarf erheblicher Dosen um gef

ahrliche Effekte von

Dioxin und Lindan nachzweisen.

Research

Science is not always applied in benign ways, even when we know as much - growth hormones and indiscriminate use of antibiotics in livestock, for example. – The bottom

line is that genetically modiﬁed organisms have not been proven in any way to be safe, and most of the studies are actually leaning the other direction, which is why many of

the world’s countries have banned genetically engineered foods. – However, evolution and adaptation, especially to unknown and unnatural substances, takes many generations

for humans to achieve.

Zu wenig verstehen die Wissenschaftler noch von

okologischen und evolution

aren Prozessen. – Hinzu kommt dass die Natur st

andig neue genetisch ver

anderte Organismen

hervorbringt. – Es geht nicht um die Behauptung von Gentechnik sondern um die Behauptung ihrer Resultate.

Health &

nutrition

While the government wants us to eat healthy, it is very true that organic foods are outrageously priced for the small amount of food we recieve. – The health issue with foods

lies in our collective wisdom that insists on making foods as cheap as possible. – Of course there are health beneﬁts to eating “organic” food.

ur die Bev

olkerungsschichten die auf g

unstige Lebensmittel angewiesen sind es komplizierter sich gesund zu ern

ahren. – Im Vordergrund unserer Lebensmittelwirtschaft

steht eben der Proﬁt und nicht die gesunde Ern

ahrung. – Es ist sowieso viel ges

under Lebensmittel zu essen, die einen m

oglichst geringen Verarbeitungsgrad aufweisen.

Politics

& policy

& com-

pliance

There are more that ”government regulators” involved. – We full well know that the industry in all of its glory takes precedence over the concerns or welfare of the people of

this country. – This is all being decided in PRIVATE, There is no involvement by the political or judicial processes that normally make laws in this country.

Lobbyismus m

usste als Straftatbestand angesehen werden und

ahnlich schwerwiegend behandelt werden wie Landesverrat. – Das ist das Ergebnis der Lobbyarbeit und unsere

Volksvertreter verabschieden solche Strafrahmen nicht versehentlich, sondern ganz bewußt. – Das und

ahnliches,

andert nichts an der kriminellen Energie der durch die Politik

und Gesetzgebung Vorschub geleistet wird.

Animal

welfare

& meat

con-

sumption

Organically raised animals used for meat must be given organic feed and be free of steroids, growth hormone and antibiotics. – When it comes to meat, again organic is

the better option as animals are often treated cruelly and inhumanely to increase production. – Manure produced by organically raised animals wreaks less havoc on the

environment, but the meat may still wreak havoc on arteries.

Diejenigen die noch Fleisch essen nehmen Bio weil diese Tiere etwas weniger gequ

alt werden als wie im konventionellen Bereich. – Im Falle von Fleisch geht das nicht anders,

als dass man Tiere qu

alt und dazu mit Dingen f

uttert, die man kaum noch als Futter bezeichnen kann. – Rinder fressen nat

urlicherweise kein Zucht-Getreide wegen des hohen

arke und Fettgehalts.

Farming

& agri-

cultural

policy &

food

security

Regardless of capacity the modern crops still need more and more land to feed the more and more people, even if the inefﬁciencies and failures of corporate agriculture are

overcome. – The main problem in organic farming is the availability of adequate organic sources of nutrients (crop residues, composts, manures) to supply crops with all the

required nutrients and to maintain soil health. – Without this, the organic farming industry can’t sustain economically as most of the Organic food that is produced is bought

by the big packaged food companies.

Daß der D

unger f

ur Bio-

Acker n

amlich von Nutztieren hergestellt wird und mehr Bedarf daran auch mehr Nutztiere zur Folge hat, geh

ort nicht zu den Notwendigkeiten, mit

denen die Bio-Lobby hausieren geht. – Wegen der F

orderung von Biogasanlagen und dem dadurch entstandenen Bedarf etwa an Mais sei Ackerland inzwischen vielerorts so

teuer, dass die Bio-Bauern nicht mehr konkurrieren k

onnten. – Hinzu kommt ein Trend, der immer mehr Landwirte zu Energiewirten werden l

asst: Nachwachsende Rohstoffe

sind gefragt wie nie.

Consu-

mer

prices &

proﬁt

Acquisitions such as this takes away consumers’ prerogative on where to spend our hard-earned dollars. – The consumer gains nothing from this. – But ﬁnal cost to consumer

is based on supply and demand.

Beim Verbraucher bleibt so gut wie kein Preisvorteil. – Das man f

ur ”bessere” Erzeugnisse mehr zahlen muss, liegt auf der Hand. – Nur m

ussen diese Billigartikel erst mal

produziert werden, bevor der Verbraucher zugreifen kann.

KDIR 2021 - 13th International Conference on Knowledge Discovery and Information Retrieval