Improving text generation through introducing coherence metrics
Text-based interaction using mobile devices is now ubiquitous, its main outlets being social networks, messengers, email conversations, virtual assistants, accessibility applications, etc. Its status implies the need to facilitate text input by the user and to devise ways to provide verbal feedback....
Gespeichert in:
| Veröffentlicht in: | Кибернетика и системный анализ |
|---|---|
| Datum: | 2020 |
| Hauptverfasser: | , , , , |
| Format: | Artikel |
| Sprache: | Englisch |
| Veröffentlicht: |
Інститут кібернетики ім. В.М. Глушкова НАН України
2020
|
| Schlagworte: | |
| Online Zugang: | https://nasplib.isofts.kiev.ua/handle/123456789/190337 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Назва журналу: | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| Zitieren: | Improving text generation through introducing coherence metrics / O.O. Marchenko, O.S. Radyvonenko, T.S. Ignatova, P.V. Titarchuk, D.V. Zhelezniakov // Кибернетика и системный анализ. — 2020. — Т. 56, № 1. — С. 16–25. — Бібліогр.: 22 назв. — англ. |
Institution
Digital Library of Periodicals of National Academy of Sciences of Ukraine| _version_ | 1860088558173290496 |
|---|---|
| author | Marchenko, O.O. Radyvonenko, O.S. Ignatova, T.S. Titarchuk, P.V. Zhelezniakov, D.V. |
| author_facet | Marchenko, O.O. Radyvonenko, O.S. Ignatova, T.S. Titarchuk, P.V. Zhelezniakov, D.V. |
| citation_txt | Improving text generation through introducing coherence metrics / O.O. Marchenko, O.S. Radyvonenko, T.S. Ignatova, P.V. Titarchuk, D.V. Zhelezniakov // Кибернетика и системный анализ. — 2020. — Т. 56, № 1. — С. 16–25. — Бібліогр.: 22 назв. — англ. |
| collection | DSpace DC |
| container_title | Кибернетика и системный анализ |
| description | Text-based interaction using mobile devices is now ubiquitous, its main outlets being social networks, messengers, email conversations, virtual assistants, accessibility applications, etc. Its status implies the need to facilitate text input by the user and to devise ways to provide verbal feedback. In this paper, we discuss a method of unique text generation for mobile devices and its evaluation methodology as a solution for both stated challenges. We consider the opportunities given by the use of context (location, weather, scheduled events, etc.), the limitations in terms of computational resources and data usage, and the inherent subjectivity of creative task assessment given the number variety of possibly acceptable outputs. The comparison with other text generation approaches shows that the use of coherence metrics helps to achieve higher quality in terms of human perception. The Spearman correlation between the values of the proposed coherence metric and the human assessment of text readability is 0.86, which indicates the high quality of the metrics and the effectiveness of the method as a whole.
Взаємодія, що ґрунтується на тексті з використанням мобільних пристроїв, стала повсюдною, її основними джерелами є соціальні мережі, месенджери, електронні листи, віртуальні помічники, застосунки для забезпечення доступності тощо. Це передбачає потребу у створенні систем полегшення введення тексту користувачем та розробленні способів підтримки вербального зворотного зв’язку. Описано метод генерації унікального тексту для мобільних пристроїв та методологію його оцінювання як розв’язки обох зазначених вище задач. Розглянуто можливості, надані використанням контексту (місцезнаходження, погода, заплановані події тощо), обмеження обчислювальних ресурсів та використання даних, а також притаманну суб’єктивність оцінювання творчої задачі з урахуванням різноманіття можливих прийнятних результатів. Порівняння з іншими методами генерації текстів свідчить про те, що використання метрик зв’язності дає змогу досягти більш високого рівня якості з погляду сприйняття людиною. Кореляція Спірмена між значеннями пропонованої метрики та оцінкою читабельності тексту людиною становить 0.86, що свідчить про високу якість метрики та ефективність методу в цілому.
Взаимодействие на основе текста с использованием мобильных устройств стало повсеместным, его основными источниками являются социальные сети, мессенджеры, электронные письма, виртуальные помощники, приложения для обеспечения доступности и т.д. Это подразумевает необходимость создания систем облегчения ввода текста пользователем и разработки способов поддержки вербальной обратной связи. В этой статье мы обсуждаем метод генерации уникального текста для мобильных устройств и методологию его оценки в качестве решения обеих заявленных проблем. Мы рассматриваем возможности, предоставляемые использованием контекста (местоположение, погода, запланированные события и т.д.), ограничения вычислительных ресурсов и использования данных, а также присущую субъективность оценки творческой задачи с учетом разнообразия возможных приемлемых результатов. Сравнение с другими методами генерации текстов показывает, что использование метрик связности помогает достичь более высокого качества с точки зрения человеческого восприятия. Корреляция Спирмена между значениями предлагаемой метрики связности и человеческой оценкой читабельности текста составляет 0.86, что свидетельствует о высоком качестве метрики и эффективности метода в целом.
|
| first_indexed | 2025-12-07T17:21:44Z |
| format | Article |
| fulltext |
UDC 681.3
O.O. MARCHENKO, O.S. RADYVONENKO, T.S. IGNATOVA,
P.V. TITARCHUK, D.V. ZHELEZNIAKOV
IMPROVING TEXT GENERATION THROUGH INTRODUCING
COHERENCE METRICS
Abstract. Text-based interaction using mobile devices is now ubiquitous, its
main outlets being social networks, messengers, email conversations, virtual
assistants, accessibility applications, etc. Its status implies the need to facilitate
text input by the user and to devise ways to provide verbal feedback. In this
paper, we discuss a method of unique text generation for mobile devices and its
evaluation methodology as a solution for both stated challenges. We consider the
opportunities given by the use of context (location, weather, scheduled events,
etc.), the limitations in terms of computational resources and data usage, and the
inherent subjectivity of creative task assessment given the number variety of
possibly acceptable outputs. The comparison with other text generation
approaches shows that the use of coherence metrics helps to achieve higher
quality in terms of human perception. The Spearman correlation between the
values of the proposed coherence metric and the human assessment of text
readability is 0.86, which indicates the high quality of the metrics and the
effectiveness of the method as a whole.
Keywords: natural language processing, automatic natural language text
generation, coherency, coherence metrics.
INTRODUCTION
In this paper, we consider the task of automatic generation of short coherent text com-
prising several sentences based on a query or a given set of input keywords. A short
chunk of text is given as an input, and we expect a two or three sentence long unique
text as a result. The output text should be related to the input query without resorting
to directly repeating or paraphrasing it, as can be seen in examples below.
On query “Attended, rock concert” the system should generate “We attended the
concert together. I went to a rock and roll concert for the first time in my life”.
On query “Cat, window” we expect to read something like “The cat practically lives
in front of the window. She loves watching birds in summer.”
Promising results have been achieved in generating sentences of short and even
medium length. There are various effective approaches for measuring text coherency,
allowing to assess and control the level of coherency. However, generating longer
coherent texts is still an open problem.
Based on the most effective methods of measuring the coherence of texts, we try
to develop principles, approaches, and methods for the formation of coherent texts
from individual sentences. But while structural methods as explicit ones can be
researched and analyzed from within, identifying features and developing selection
heuristics, dissecting neural network systems in the similar vein is almost impossible
due to their “black box” nature and interpretability issues.
The main purpose of this paper is to research and develop effective methods for
constructing coherent texts on request without using lexical-semantic knowledge bases
specially prepared for this task. Given the above, we reformulate the task:
Problem. There is a set of correct sentences S S S S N� 1 2, , ,� .
It is necessary to form a coherent text consisting of 2 or 3 sentences relevant to
a query Q by choosing from the S a set of the most appropriate instances (with possible
subsequent modification of these sentences at the syntax level, where necessary).
16 ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1
© O.O. Marchenko, O.S. Radyvonenko, T.S. Ignatova, P.V. Titarchuk, D.V. Zhelezniakov, 2020
RELATED WORK
Text Generation. The most promising approach to implementing such systems is
currently the use of neural networks, for example, convolutional neural networks,
recurrent networks, recursive networks, and their various combinations for learning
word sequences from the sentences in the train data set. After training, a neural
network has to be able to generate sentences that form a unique coherent text.
Neural networks have been widely used to solve a number of problems of
computational linguistics where it is necessary to find correspondences between pieces
of text written by people, according to some sets of features, such as authorship
attribution [1] and paraphrase identification [2]. Neural networks are also able to
adequately generate very short messages for interactive systems [3, 4]. GPT-2 [5] model
allows generating texts of significant length, but the size of the model makes its use very
limited. At the moment, there are still difficulties with generating a single coherent text.
The challenges can be generally demonstrated using the example of two well-known
approaches: “skip thoughts” [6] and Sequence to Sequence Recurrent Neural
Network [7]. When the “skip thoughts” model is used, a chain of semantic connections
between word meanings is broken beyond the sequence of length n, where word n �1 is
dissonant with the previous sequence (both between sentences and within one sentence,
if it is longer than n). Recurrent neural networks tend to loop, and often generate
sentences that are almost completely identical to the query and to each other by semantic
meaning and lexical content, which cannot be considered a cohesive and coherent text.
Problems such as the curse of dimensionality make it hard to use exclusively neural
network approaches to solve the problem of generating complete coherent texts. Hybrid
systems that employ the principles of structural algorithmic natural language processing along
with neural networks make it possible to use the advantages of neural network technologies,
compensating their weak points by optimizing data, heuristics, special metrics, etc.
Structural methods for generating texts were shown to be useful in a number of
works [8–10]. Despite the well-known advantages, they have two serious
interconnected drawbacks. First, these systems often use large lexical-semantic
databases that in addition to linguistic models contain a description of large semantic
structures, e.g. ontologies. The databases are built mainly by hand and, as a rule, are
specialized for a narrow subject area. Consequently, these systems are hard and
expensive to scale for new subject areas due to the annotation cost.
For neural networks, as well as for structural methods, forming a training set is also
a serious challenge, especially considering that after training or building a knowledge
base there has to be enough material to form a coherent text on request.
Coherence Metrics. The solution of the task is complicated by the lack of a
single accepted effective metric for assessing the quality of the generated texts. There
are works in which BLEU [11] is used as such metric, while originally having been
developed to assess the quality of machine translation. It is well able to identify
paraphrase, plagiarism, and other similarities between texts [12]. It can be used to
measure the relevance of the generated text to the request. However, the aim of this
research is to create texts that are not identical to the request. Since BLEU will
produce the highest values for a set of verbatim paraphrases of the input request, its use
in tasks similar to the one at hand can be limited.
Taking into account our goal of creating coherent texts, we have to consider
coherence models and metrics. Various approaches to control and measure coherence
were proposed in a number of works.
Foltz, Kintsch, and Landauer’s [13] model measures coherence as a function of
semantic relatedness between adjacent sentences. Semantic relatedness is computed
automatically using Latent Semantic Analysis from raw texts without employing
syntactic or any other annotation. Barzilay and Lapata [14] propose an approach that
captures local coherence by modeling patterns of entity distribution in discourse using
an entity-grid model.
ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1 17
Recent deep learning coherence works [15, 16] adopt recursive and recurrent neural
networks for computing semantic vectors for sentences. Coherence models that use
recursive neural networks often depend severely on external resources, e.g. syntactic
parsers, to construct their recursion structure. Coherence models that rely purely on
recurrent neural networks process the words within a text sequentially. However, in such
models, long-distance dependencies between words cannot be captured effectively due to
the limits of the memorization capability of recurrent networks.
Basile et al. [17] describe a relatedness model built using FrameNet, which
formally describes semantic structures of predicate-argument relations of English
clauses. This allows the authors to consider the relatedness of two clauses as a sum of
two estimates: the relatedness of the predicates of these two clauses and the relatedness
of the corresponding arguments of these two predicates.
Mesgar and Strube [18] propose a neural model of local text coherency using
recurrent networks and a convolutional network and demonstrate state-of-the-art
results for two tasks: readability assessment and essay scoring.
METHOD
S is a set of sentences relevant to a query consisting of a set of keywords. A set
of sentences S is formed as a result of a search in the corpus of texts. The search
is performed using a search index built over the text corpus using the search plat-
form SOLR (http://lucene.apache.org/solr/).
The search index is implemented as an inverted index. Each word from the corpus
provided with a list with sublists:
w d n s n s n sk k: (( , (( , )( , )... ( , ))),1 11 11 12 12 1 1 �
� , ( (( , )( , )... ( , )))) ,d n s n s n sr r r r r rt rt1 1 2 2 (1)
where di is the index of the document containing this word w; nij is the index of
the word w in the text di ; and sij is the index of the sentence containing the given
word w in document di .
As a result of a search by request, sentences that contain at least two of the k
words from the query are selected from the corpus of texts, provided that the paragraph
of each of the returned sentences (or its entire text, if it is sufficiently short — three
paragraphs or less) contains all query
keywords or their synonyms. The next
task is to assemble a coherent text of three
sentences from the set S . We emphasize
that all sentences must be taken from
different texts.
To combine sentences into text,
a hierarchical agglomerative clustering
algorithm is used. At the first stage of the
algorithm, each sentence is a cluster.
Then the nearest pair of clusters is
selected and merged into one (see Fig. 1).
The metric for finding the nearest
sentences is calculated in a similar way
to [17] using the formula for calculating
relatedness:
FRel S S FRelPred S S FRelArgs S S( , ) ( , ) ( ) ( ,1 2 1 2 1 21� � � � �� � ), (2)
where FRelPred is the relatedness of predicates:
FRelPred S S
Cp p
Cp Cp
( , ) log
| |
| | | |
1 2 2
1 2
1 2
� , (3)
18 ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1
Coherence
metrics
S11 S21
S12 S22
S13 S23
Candidate
for K1
Candidate
for K2
K1 K2
Fig. 1. Process of sentence combination to create
a unique coherent text with respect to the proposed
metrics (K1 and K2 are subsets of keywords from
the query)
where Cp1 and Cp2 are subsets sentences from the corpus that have common predi-
cate p1 with the first sentence S1 and p2 with the second sentence S 2 , respec-
tively. Cp p1 2 is a subset of contexts– the adjacent sentences in the text corpus,
where p1 and p2 are the main verbs-predicates of the first and the second sen-
tences respectively.
FRelPred is similar to the PMI (pointwise mutual information) metric. It can be
calculated dynamically using a special inverted index, constructed after preprocessing
the sentences in the text corpus with a part-of-speech tagger. The resulting special
index contains only verbs that are predicates of the sentences in the corpus.
It should be mentioned that order matters in (3). Due to language nature,
FRelPred is a quasimetric that does not satisfy the condition of symmetry. Changing
the order of sentences in the text will likely change the meaning or affect its coherence.
Relatedness for predicate arguments is calculated as follows:
FRelArgs S S( , )1 2 �
�
�
�
�
��
�
1
2
1 1
1
1
2
| |
max ( , )
arg
arg
s
wpsim N N
N arg s
N s
i j
i
j
| |
max ( , ) ,
arg
arg
arg
s
wpsim N N
N s
N s
i j
i
j
2
2
1
�
�
(4)
where arg s1 and arg s2 are sets of noun arguments of verb predicates p1 and p2 ,
respectively, wpsim is the Wu-Palmer similarity measure for two words [19], � is
the coefficient of balancing between the two components of (2). Basile suggests the
optimal value � � 0.5 obtained experimentally [17].
Wu and Palmer calculate relatedness of two concepts-words by considering
depths of the two corresponding synsets in the WordNet taxonomies, along with the
depth of the LCS (Least Common Subsumer, or closest common ancestor):
wpsim S S
depth LCS S S
depth S depth S
( , )
( ( , ))
( ) (
1 2
1 2
1 2
2
�
�
� )
. (5)
Fig. 2 demonstrates the process of sentence combination according to (2). Solid
outline and double outline boxes identify nouns, connections between which are found
using (4). Relations between verbs predicates depicted in dotted outline boxes
calculated with (3).
First, we select and concatenate such sentences S1 and S 2 , forming two-sentence
clusters, for which the following condition is satisfied:
FRel S S R( , )1 2 � , (6)
where R is a threshold found empirically. Sentences can go into several separate
clusters at once.
At the next stage of the algorithm, when singular sentences join clusters of two
sentences, we try to put them either as the first or the last sentence constructed texts. In
this case, condition (6) for the new sentence with the adjacent one should be satisfied.
ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1 19
walked forest
One day I walk woods walk woods .went on a with my family in the . I took my dog for a in the
Fig. 2. The example of search
When two clusters sharing a sentence are formed — e.g., (S S1 2, ) and (S S2 3, ) — they
are merged together to form cluster (S S S1 2 3, , ). The coherence metric of the
constructed text is calculated by the formula:
Coherence T
FRel S S
n
i i
i
n
( )
( , )
�
�
�
�
�
1
1
1
1
. (7)
We accept the text corresponding to the maximum value of Coherence T( ),
provided that
Coherence T P( ) � , (8)
where P is a threshold. Its optimal value also figured out in the experiment when
calculating Coherence T( ) for paragraphs from the original texts of the corpus.
The algorithm can potentially continue the assembly of texts of any length, but for
the current research, we have limited our scope to shorter texts of no more than
three sentences.
The flaw of the method is the need to enumerate all possible variants of the
sentences found by the search engine of the system from the corpus of texts. There are
hundreds and even thousands of sentences that are relevant to a query, which makes
a complete search through possible pairs and triples of sentences almost impossible
when operating in real time. Randomly reducing the set of sentences can lead to a lack
of pairs or triples of sentences satisfying the conditions of coherence (6) and (8). It is
necessary to effectively filter the initial set of sentences found by the system’s search
engine, leaving a subset of ones most relevant to query. We propose to use the
word2vec vector model to construct a vector metric for proximity of sentences to
a given query. In the word2vec model, each word corresponds to a certain vector.
The vector of a sentence will be calculated as a weighted sum of vectors of all
meaningful words in this sentence. The words for which the TF-IDF metric exceeds
a certain threshold level Th are considered meaningful. The weight of each vector in
the sum is proportional to TF-IDF of its word. Query vector is calculated in a similar
way. We use cosine as proximity metric. Having a query vector and vectors of N
sentences selected from the text corpus, it is possible to select from the N sentences k
of them that are most relevant to query for linear time. We can also reduce this search
time to logarithmic time O N( )log . For this purpose, we have to make a special
pre-processing of the corpus.
We propose to implement hierarchical agglomerative clustering for all sentences
in the text corpus.
Step 1. All sentences in the corpus are separate single clusters.
Step 2. While we don’t have one global merged cluster do
— Select the nearest pairs of clusters according to the cosine measure;
— Merge the nearest pair of clusters;
— Recalculate centroids in new clusters.
After merging all the clusters into one global cluster, we will have a binary search
tree of sets of sentences.
Having received the query vector, we go to the root of the search tree, and begin
to compare the centroid vectors of the left and right son of the given node with the
query vector by means of the cosine measure. After measurement we go in the
direction of the best option. Having done this several times, we will find the desired
subset of sentences Q that satisfy the criterion of proximity to query vector with a
given threshold level or for the required number of sentences. After that, text assembly
works only with found sentences that are included in subset Q. The text corpus
together with the search index for the system is an implicit knowledge base that stores
information about the realities of the surrounding world and subject areas described in
the texts. The building of such a knowledge base is fully automatic, as is its
20 ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1
replenishment by including additional
texts to the corpus and automatically
updating the search index.
In Fig. 3, the general pipeline of
text creation based on the proposed
coherence metrics is shown.
EXPERIMENTS
We study the effectiveness of the met-
ric by training a text generation model
and evaluating it against two other ap-
proaches.
Dataset . We conduct the
experiments on the ROCStories
corpus [20]. It consists of 98,162 short
commonsense stories containing on
average 50 words each. Our system constructs a search index for this corpus and uses
the index to form coherent texts. The index used for calculating (3) was built from a
large collection of blog texts.
Baselines. To evaluate the effectiveness of the proposed approach, we compare
our method against such representative baselines as the skip-thoughts model [6]
specially trained on the above-mentioned corpus and the plan-and-write hierarchical
generation system presented in [21] that actually also uses the ROCStories corpus.
Skip-thoughts is an encoder-decoder model. Decoder consists of two parts: one
predicts the previous sentence for the current one, and another — the next one. As
an encoder we use bidirectional RNN with GRU activations (64 units each). As
decoders, we use two RNNs with GRU activations with 128 units. As an output, we get
the vector of probabilities for all words from the dictionary. Dictionary size: 11 000
words. For word embeddings in encoder and decoder models, we use a pre-trained
fasttext (https://fasttext.cc) model for simple English. The model was trained during
500 steps (16000 randomly chosen samples were given at each step).
Model Training.Training of the proposed model is conducted with calculating the
optimal values for parameters R and P (conditions (6) and (8) correspondingly). We
partition all text corpus into 3 equal parts. The first part is left as is. The second part has
been mixed to form incoherent text set. In this case, in constructed texts, every sentence
is selected from a distinct text. In the third part of the corpus, only half of the texts were
mixed in the same way as mentioned above and they are marked as incoherent.
Several machine learning models were used to figure out the optimal values for R
and P. The first and the second part of corpus were used as a training set and the third
as a test set.
The best results of classification on test set were obtained by a support vector
classifier with parameters R � 52.81 and P � 59.16.
After the training, the three systems receive a query for a future story as an input
and generate their texts. Typical outputs of each system are shown in Table 1.
Evaluation. For the expert evaluation study, we offered 50 fluent English speakers
to assess texts generated by the three approaches without disclosing which system
generated which passage. Every expert received a questionnaire with texts generated
according to 20 queries. The queries were taken from the pool of image captions in the
COCO dataset [22], making sure that all the words in the chosen caption are present in
the ROCStories dataset. For each text, three parameters were evaluated on a scale from 1
to 10: Readability (coherency and consistency), Likability (interestingness) and
Appropriateness (relevance to the query). Averaged estimates are shown in Table 2.
ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1 21
Index
Plot (keywords)
User input
Sentence candidate
selection
Text generated based
on the coherence metric
Show result
Sentences
Candidates of text
Fig. 3. The approach of text generation using the
proposed coherence metric
For Readability assessment, the coefficient of variation of human expert estimations
on the given set of generated texts was between 6% and 14% (as shown in Table 3) that
demonstrated acceptable quality of coherency assessment. For Appropriateness
assessment, the coefficient of variation was between 5% and 11%. For Likability
assessment, the coefficient of variation was between 16% and 23% due to the extra
subjectivity of that specific part of estimating.
We propose to deploy (2) as the metrics of coherency for obtained texts. We can
calculate its values for all three sets of texts but there is a problem due to the essence of
our method. It selects such combinations of sentences that already satisfy the found
optimal values of parameters R and P. Still, we present the results of such measuring in
Table 4. During the experiments, it was found that the Spearman correlation between the
values of Coherence T( ) and the human assesses of Readability T( ) was 0.86, which
indicates the quality of the proposed metrics and the effectiveness of the method as
a whole.
Results and Discussion. As can be seen from Table 2, the proposed method
overcomes competitors in all assessments. The results obtained during the expert
assessment may be hard to reproduce since the evaluation fully relies on human judgment,
but unfortunately for this specific case there are no better way to asses such subjective
qualities of text as coherency, appropriateness and especially likability without
participation of people.
22 ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1
T a b l e 1. Examples of generated stories
Input query A man riding skis down a snow covered slope
Skip-thoughts
1. He landed on the ground. He landed on the ground. He ran to the hospital.
2. He ran to the hospital. The doctor diagnosed him. The doctor diagnosed him.
Plan-and-write
1. It was a cold winter day. John decided to shovel the snow. He put on his
boots and went outside. He decided to go for a walk. He was covered in snow.
2. Tim was driving down the road. He saw a car in the driveway. He whipped
it up and took it home. He put it in his truck. He decided to keep it.
Proposed
solution
1. This morning,I woke up and saw the ground in snow. The streets were
covered in snow.
2. He fell down a snowy hill. When he got up he was covered in snow. He
saw that his walkway was covered with deep snow.
Input query
Several large commercial airplanes stationed at an airport with service vehicles
nearby.
Skip-thoughts
1. In the end, the flight was delayed. The flight was delayed. The plane was
delayed.
2. The plane was delayed. The flight was delayed. The plane was
delayed.
Plan-and-write
1. I was on my way to the airport. I got my luggage. I was waiting in line.
I pulled over. I realized I missed the flight.
2. Tom was on his way to the airport. He was on his way to his flight. He
couldn’t find his windows. He called the cops. They had to call the police.
Proposed
solution
1. A large family went to the airport. It was their first time flying a commer-
cial plane of a large airline.
2. The plane shook as it rose into the air. The weather was changing when
the plane was high in the air.
T a b l e 2. Results of human estimation
System Readability Likability Appropriateness Overall
Skip-thoughts 3.714 3.281 3.433 3.476
Plan-and-write 5.161 4.212 5.369 4.914
Our system 6.751 6.281 6.004 6.345
There are no widely available AI systems
that read and understand natural language texts
better than humans. Comparison with other
existing approaches to measuring the coherency
of texts is an interesting but technically difficult
to implement experiment, and discrepancies
between different metrics still lead to the need
for human experts to choose a more correct
option. In our experiment, good coefficient of
variation of human expert estimations ensures the reliability of obtained estimation.
For the second part of the experiment, one can note that it is not quite correct to
evaluate baselines and the proposed method on FRel and Coherence, since the two
metrics are locally defined in this work, and the proposed method is specifically
designed to optimize on them while the compared baselines are not. FRel and
Coherence for baselines are shown not to demonstrate the advantage of the approach,
but to measure the correlation of calculated values of our metrics and estimations
obtained during expert assessment. And the obtained high correlation demonstrates
good quality of the proposed metrics.
One of the flaws of the metrics is its excessively high evaluation of texts with repeated
almost or completely identical sentences generated by neural networks. This influence of
repetitions on the metrics can be compensated by imposing a penalty for repeats, which are
successfully identified by the BLEU metrics: If the text contains sentences S i and S j , such
that BLEU S Si j( , ) � 0.8, then Coherence T Coherence T Penalty( ) ( )� � .
Additionally, it should be mentioned that the proposed approach is applicable for
usage in standalone on-device running for mobile devices. The average text generation
time per one request is ~ 2.30 sec, which is comparable with the analyzed NN-based
solutions. But the size of the necessary components is much lower: for our solution the
ROCStories corpus size is 27 MB and the SOLR index size is ~ 500 MB, while in case
of NN-based generation the combined size of required components is about 2.7 GB.
CONCLUSIONS
This paper proposes the use of coherence metrics during text generation. We demon-
strate the advantages of controlling coherence of the output by implementing a sys-
tem for assembling coherent texts based on a search engine with special search in-
dexes that uses the proposed coherence metrics and conducting an expert study.
The search system selects from the text corpus a set of sentences that relevant to
the request, after that the system composed a coherent story from them. Coherence
metrics that serve as the main criterion for combining sentences into text can be
calculated dynamically using data from the built indices. The experimental results
confirm the reliability of the proposed metrics and the effectiveness of the method as
a whole. Compared to NN-based approaches, the proposed method is more applicable
for use on mobile devices due to lower storage and computation power requirements.
Future work includes conducting a more extensive expert study involving comparison
of auto-generated texts with human-written ones; studying the behavior of the proposed
metrics in longer texts; and integrating text coherence control into NN-based approaches.
ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1 23
T a b l e 3. Coefficient of variation (CV) in %, calculated for expert assessments of
Readability of stories generated for 20 queries ( )Q by three systems: encoder-de-
coder ( )S1 , plan-and-write ( )S2 , and our system ( )S3
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 Q13 Q14 Q15 Q16 Q17 Q18 Q19 Q20
S1 7.1 9.8 10.2 8.8 8.6 12.2 7.9 13.7 11.9 12.7 13.5 11.9 12.4 10.5 13.6 12.8 11.1 12.3 11.1 10.8
S2 9.8 13.8 10.7 11.2 12.2 10.3 9.1 12.9 8.3 10.1 12.4 13.2 9.9 13.9 8.4 8.9 7.3 8.8 11.9 9.5
S3 11.0 8.5 9.4 12.9 11.7 9.8 11.3 13.1 10.4 13.2 8.6 11.7 8.7 10.9 11.7 11.8 6.4 9.8 9.1 10.0
T a b l e 4. Estimation of intro-
duced metrics
System FRel Coherence
Skip-thoughts 27.12 34.82
Plan-and-write 39.65 45.91
Our system 60.18 64.15
REFERENCES
1. Ruder S., Ghaffari P., Breslin J.G. Character-level and multi-channel convolutional neural networks
for large-scale authorship attribution. 2016. URL: http://arxiv.org/abs/1609.06686.
2. Agarwal B., Ramampiaro H., Langseth H., Ruocco M. A deep network model for paraphrase detection
in short text messages. Information Processing and Management. 2018. Vol. 54, N 6. P. 922–937.
3. Asghar N., Poupart P., Hoey J., Jiang X., Mou L. Affective neural response generation. Advances in
Information Retrival. Proc. 40th European Conf. on IR Research, ECIR 2018 (March 26–29, 2018,
Grenoble, France). P. 154–166.
4. Chen Y.-N., Celikyilmaz A., Hakkani-Tur D. Deep learning for dialogue systems. Proc.
27th Int. Conf. on Computational Linguistics: Tutorial Abstracts. 2018. P. 25–31. URL: https://
www.aclweb.org/anthology/C18-3006.pdf.
5. Radford A., Wu J., Child R., Luan D., Amodei D., Sutskever I. Language models are unsupervised
multitask learners. OpenAI Blog. URL: https://openai.com/blog/better-language-models/ (Feb 14, 2019).
6. Kiros R., Zhu Y., Salakhutdinov R.R., Zemel R., Urtasun R., Torralba A., Fidler S. Skip-thought vectors.
Proc. NIPS 2015. (December 7–12, 2015, Montreal, Quebec, Canada). Vol. 2. P. 3294–3302.
7. Jain P., Agrawal P., Mishra A., Sukhwani M., Laha A., Sankaranarayanan K. Story generation from
sequence of independent short descriptions. 2017. URL: http:// arxiv.org/abs/1707.05501.
8. McIntyre N., Lapata M. Learning to tell tales: a data-driven approach to story generation. Proc. 47th
Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP (2–7 August 2009, Suntec,
Singapore). 2009. P. 217–225.
9. Rishes E., Lukib S.M., Elson D.K., Walker M.A. Generating different story tellings from semantic
representations of narrative. Proc. 6th ICIDS 2013. (November 6–9, Istanbul, Turkey). P. 192–204.
10. Lepp��anen L., Munezero M., Granroth-Wilding M., Toivonen H. Data-driven news generation for
automated journalism. Proc. 10th INLG 2017. (September, 2017, Santiago de Compostela, Spain).
P. 188–197.
11. Papineni K., Roukos S., Ward T., Zhu W.-J. BLEU: A method for automatic evaluation of machine
translation. Proc. 40th Annual Meeting on Association for Computational Linguistics. (July, 2002,
Philadelphia, Pennsylvania, USA). 2002. P. 311–318.
12. Ji Y., Eisenstein J. Discriminative improvements to distributional sentence similarity. Proc. 2013
Conference on Empirical Methods in Natural Language Processing. (October, 2013, Seattle,
Washington, USA). P. 891–896.
13. Foltz P. W., Kintsch W., Landauer T.K. The measurement of textual coherence with latent semantic
analysis. Discourse Processes. 1998. Vol. 25, N 2–3. P. 285–307.
14. Barzilay R., Lapata M. Modeling local coherence: an entity-based approach. Computational
Linguistics. 2008. Vol. 34, N 1. P. 1–34.
15. Li J., Hovy E. A model of coherence based on distributed sentence representation. Proc. EMNLP
2014. (October 25–29, 2014, Doha, Qatar). P. 2039–2048.
16. Li J., Jurafsky D. Neural net models of open-domain discourse coherence. Proc. EMNLP 2017.
(September, 2017, Copenhagen, Denmark). P. 198–209.
17. Basile V., Condori R.L., Cabrio E. Measuring frame instance relatedness. Proc. 7th Joint Conference
on Lexical and Computational Semantics (*SEM) (June 5–6, 2018, New Orleans). 2018. P. 245–254.
18. Mesgar M., Strube M. A neural local coherence model for text quality assessment. Proc. 2018
Conference on Empirical Methods in Natural Language Processing (October 31 – November 4,
2018, Brussels, Belgium). P. 4328–4339.
19. Wu Z., Palmer M. Verbs semantics and lexical selection. Proc. 32nd Annual Meeting on Association
for Computational Linguistics (June 27–30, 1994, Las Cruces, New Mexico). 1994. P. 133–138.
20. Mostafazadeh N., Chambers N., He X., Parikh D., Batra D., Vanderwende L., Kohli P., Allen J. A corpus
and cloze evaluation for deeper understanding of commonsense stories. Proc. NAACL-HLT 2016.
(June 12–17, 2016, San Diego, California). P. 839–849.
21. Yao L., Peng N., Weisahedel R., Kbight K., Zhao D., Yan R. Plan-And-Write: Towards better
automatic storytelling. 2018. URL: https://arxiv.org/abs/1811.05701.
22. Lin T.-Y, Maire M., Belongie S., Hays J., Peroba P., Ramanan D., Dollar P., Zitnicl L. Microsoft
COCO: Common objects in context. Proc. ECCV 2014. (September 6–12, 2014, Zurich,
Switzerland). P. 740–75.
Íàä³éøëà äî ðåäàêö³¿ 15.08.2019
24 ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1
Î.Î. Ìàð÷åíêî, Î.Ñ. Ðàäèâîíåíêî, Ò.Ñ. ²ãíàòîâà,
Ï.Â. Òèòàð÷óê, Ä.Â. Æåëåçíÿêîâ
ÏÎÊÐÀÙÅÍÍß ßÊÎÑÒ² ÃÅÍÅÐÓÂÀÍÍß ÒÅÊÑÒÓ ÇÀ ÄÎÏÎÌÎÃÎÞ Ì²ÐÈ ÇÂ’ßÇÍÎÑÒ²
Àíîòàö³ÿ. Âçàºìîä³ÿ, ùî ´ðóíòóºòüñÿ íà òåêñò³ ç âèêîðèñòàííÿì ìîá³ëüíèõ
ïðèñòðî¿â, ñòàëà ïîâñþäíîþ, ¿¿ îñíîâíèìè äæåðåëàìè º ñîö³àëüí³ ìåðåæ³,
ìåñåíäæåðè, åëåêòðîíí³ ëèñòè, â³ðòóàëüí³ ïîì³÷íèêè, çàñòîñóíêè äëÿ çàáåç-
ïå÷åííÿ äîñòóïíîñò³ òîùî. Öå ïåðåäáà÷ຠïîòðåáó ó ñòâîðåíí³ ñèñòåì ïî-
ëåãøåííÿ ââåäåííÿ òåêñòó êîðèñòóâà÷åì òà ðîçðîáëåíí³ ñïîñîá³â ï³äòðèìêè
âåðáàëüíîãî çâîðîòíîãî çâ’ÿçêó. Îïèñàíî ìåòîä ãåíåðàö³¿ óí³êàëüíîãî òåê-
ñòó äëÿ ìîá³ëüíèõ ïðèñòðî¿â òà ìåòîäîëîã³þ éîãî îö³íþâàííÿ ÿê ðîçâ’ÿçêè
îáîõ çàçíà÷åíèõ âèùå çàäà÷. Ðîçãëÿíóòî ìîæëèâîñò³, íàäàí³ âèêîðèñòàííÿì
êîíòåêñòó (ì³ñöåçíàõîäæåííÿ, ïîãîäà, çàïëàíîâàí³ ïî䳿 òîùî), îáìåæåííÿ
îá÷èñëþâàëüíèõ ðåñóðñ³â òà âèêîðèñòàííÿ äàíèõ, à òàêîæ ïðèòàìàííó
ñóá’ºêòèâí³ñòü îö³íþâàííÿ òâîð÷î¿ çàäà÷³ ç óðàõóâàííÿì ð³çíîìàí³òòÿ ìîæ-
ëèâèõ ïðèéíÿòíèõ ðåçóëüòàò³â. Ïîð³âíÿííÿ ç ³íøèìè ìåòîäàìè ãåíåðàö³¿
òåêñò³â ñâ³ä÷èòü ïðî òå, ùî âèêîðèñòàííÿ ìåòðèê çâ’ÿçíîñò³ äຠçìîãó äî-
ñÿãòè á³ëüø âèñîêîãî ð³âíÿ ÿêîñò³ ç ïîãëÿäó ñïðèéíÿòòÿ ëþäèíîþ. Êîðå-
ëÿö³ÿ Ñï³ðìåíà ì³æ çíà÷åííÿìè ïðîïîíîâàíî¿ ìåòðèêè òà îö³íêîþ ÷èòà-
áåëüíîñò³ òåêñòó ëþäèíîþ ñòàíîâèòü 0.86, ùî ñâ³ä÷èòü ïðî âèñîêó ÿê³ñòü
ìåòðèêè òà åôåêòèâí³ñòü ìåòîäó â ö³ëîìó.
Êëþ÷îâ³ ñëîâà: êîìï’þòåðíà ë³íãâ³ñòèêà, àâòîìàòè÷íå ãåíåðóâàííÿ ïðèðîä-
íîìîâíèõ òåêñò³â, çâ’ÿçí³ñòü òåêñò³â, ìåòðèêè çâ’ÿçíîñò³ òåêñò³â.
À.À. Ìàð÷åíêî, Î.Ñ. Ðàäèâîíåíêî, Ò.Ñ. Èãíàòîâà,
Ï.Â. Òèòàð÷óê, Ä.Â. Æåëåçíÿêîâ
ÓËÓ×ØÅÍÈÅ ÊÀ×ÅÑÒÂÀ ÃÅÍÅÐÀÖÈÈ ÒÅÊÑÒÀ Ñ ÏÎÌÎÙÜÞ ÌÅÐÛ ÑÂßÇÍÎÑÒÈ
Àííîòàöèÿ. Âçàèìîäåéñòâèå íà îñíîâå òåêñòà ñ èñïîëüçîâàíèåì ìîáèëüíûõ
óñòðîéñòâ ñòàëî ïîâñåìåñòíûì, åãî îñíîâíûìè èñòî÷íèêàìè ÿâëÿþòñÿ ñîöè-
àëüíûå ñåòè, ìåññåíäæåðû, ýëåêòðîííûå ïèñüìà, âèðòóàëüíûå ïîìîùíèêè,
ïðèëîæåíèÿ äëÿ îáåñïå÷åíèÿ äîñòóïíîñòè è ò.ä. Ýòî ïîäðàçóìåâàåò íåîáõî-
äèìîñòü ñîçäàíèÿ ñèñòåì îáëåã÷åíèÿ ââîäà òåêñòà ïîëüçîâàòåëåì è ðàçðà-
áîòêè ñïîñîáîâ ïîääåðæêè âåðáàëüíîé îáðàòíîé ñâÿçè. Â ýòîé ñòàòüå ìû
îáñóæäàåì ìåòîä ãåíåðàöèè óíèêàëüíîãî òåêñòà äëÿ ìîáèëüíûõ óñòðîéñòâ è
ìåòîäîëîãèþ åãî îöåíêè â êà÷åñòâå ðåøåíèÿ îáåèõ çàÿâëåííûõ ïðîáëåì.
Ìû ðàññìàòðèâàåì âîçìîæíîñòè, ïðåäîñòàâëÿåìûå èñïîëüçîâàíèåì êîíòåê-
ñòà (ìåñòîïîëîæåíèå, ïîãîäà, çàïëàíèðîâàííûå ñîáûòèÿ è ò.ä.), îãðàíè÷åíèÿ
âû÷èñëèòåëüíûõ ðåñóðñîâ è èñïîëüçîâàíèÿ äàííûõ, à òàêæå ïðèñóùóþ
ñóáúåêòèâíîñòü îöåíêè òâîð÷åñêîé çàäà÷è ñ ó÷åòîì ðàçíîîáðàçèÿ âîçìîæ-
íûõ ïðèåìëåìûõ ðåçóëüòàòîâ. Ñðàâíåíèå ñ äðóãèìè ìåòîäàìè ãåíåðàöèè
òåêñòîâ ïîêàçûâàåò, ÷òî èñïîëüçîâàíèå ìåòðèê ñâÿçíîñòè ïîìîãàåò äîñòè÷ü
áîëåå âûñîêîãî êà÷åñòâà ñ òî÷êè çðåíèÿ ÷åëîâå÷åñêîãî âîñïðèÿòèÿ. Êîððå-
ëÿöèÿ Ñïèðìåíà ìåæäó çíà÷åíèÿìè ïðåäëàãàåìîé ìåòðèêè ñâÿçíîñòè è ÷å-
ëîâå÷åñêîé îöåíêîé ÷èòàáåëüíîñòè òåêñòà ñîñòàâëÿåò 0.86, ÷òî ñâèäåòåëü-
ñòâóåò î âûñîêîì êà÷åñòâå ìåòðèêè è ýôôåêòèâíîñòè ìåòîäà â öåëîì.
Êëþ÷åâûå ñëîâà: êîìïüþòåðíàÿ ëèíãâèñòèêà, àâòîìàòè÷åñêàÿ ãåíåðàöèÿ åñ-
òåñòâåííî-ÿçûêîâûõ òåêñòîâ, ñâÿçíîñòü òåêñòîâ, ìåòðèêè ñâÿçíîñòè òåêñòîâ.
Marchenko Oleksandr,
Dr. Sc. (Phys.-Math.), Professor, Taras Shevchenko National University of Kyiv, Ukraine,
e-mail: rozenkrans17@gmail.com.
Radyvonenko Olga,
Ph.D in Technical Science, Associate Professor, Head of Lab, Samsung R&D Institute Ukraine (SRK),
Kyiv, Ukraine, e-mail: o.radyvonenk@samsung.com.
Ignatova Tetiana,
Engineer, Samsung R&D Institute Ukraine (SRK), Kyiv, Ukraine, e-mail: te.ignatova@samsung.com.
Titarchuk Pavlo,
Engineer, Samsung R&D Institute Ukraine (SRK), Kyiv, Ukraine, e-mail: p.tytarchuk@samsung.com.
Zhelezniakov Dmytro,
Staff Engineer, Samsung R&D Institute Ukraine (SRK), Kyiv, Ukraine, e-mail: d.zheleznyak@samsung.com.
ISSN 1019-5262. Êèáåðíåòèêà è ñèñòåìíûé àíàëèç, 2020, òîì 56, ¹ 1 25
|
| id | nasplib_isofts_kiev_ua-123456789-190337 |
| institution | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| issn | 1019-5262 |
| language | English |
| last_indexed | 2025-12-07T17:21:44Z |
| publishDate | 2020 |
| publisher | Інститут кібернетики ім. В.М. Глушкова НАН України |
| record_format | dspace |
| spelling | Marchenko, O.O. Radyvonenko, O.S. Ignatova, T.S. Titarchuk, P.V. Zhelezniakov, D.V. 2023-05-31T12:46:32Z 2023-05-31T12:46:32Z 2020 Improving text generation through introducing coherence metrics / O.O. Marchenko, O.S. Radyvonenko, T.S. Ignatova, P.V. Titarchuk, D.V. Zhelezniakov // Кибернетика и системный анализ. — 2020. — Т. 56, № 1. — С. 16–25. — Бібліогр.: 22 назв. — англ. 1019-5262 https://nasplib.isofts.kiev.ua/handle/123456789/190337 681.3 Text-based interaction using mobile devices is now ubiquitous, its main outlets being social networks, messengers, email conversations, virtual assistants, accessibility applications, etc. Its status implies the need to facilitate text input by the user and to devise ways to provide verbal feedback. In this paper, we discuss a method of unique text generation for mobile devices and its evaluation methodology as a solution for both stated challenges. We consider the opportunities given by the use of context (location, weather, scheduled events, etc.), the limitations in terms of computational resources and data usage, and the inherent subjectivity of creative task assessment given the number variety of possibly acceptable outputs. The comparison with other text generation approaches shows that the use of coherence metrics helps to achieve higher quality in terms of human perception. The Spearman correlation between the values of the proposed coherence metric and the human assessment of text readability is 0.86, which indicates the high quality of the metrics and the effectiveness of the method as a whole. Взаємодія, що ґрунтується на тексті з використанням мобільних пристроїв, стала повсюдною, її основними джерелами є соціальні мережі, месенджери, електронні листи, віртуальні помічники, застосунки для забезпечення доступності тощо. Це передбачає потребу у створенні систем полегшення введення тексту користувачем та розробленні способів підтримки вербального зворотного зв’язку. Описано метод генерації унікального тексту для мобільних пристроїв та методологію його оцінювання як розв’язки обох зазначених вище задач. Розглянуто можливості, надані використанням контексту (місцезнаходження, погода, заплановані події тощо), обмеження обчислювальних ресурсів та використання даних, а також притаманну суб’єктивність оцінювання творчої задачі з урахуванням різноманіття можливих прийнятних результатів. Порівняння з іншими методами генерації текстів свідчить про те, що використання метрик зв’язності дає змогу досягти більш високого рівня якості з погляду сприйняття людиною. Кореляція Спірмена між значеннями пропонованої метрики та оцінкою читабельності тексту людиною становить 0.86, що свідчить про високу якість метрики та ефективність методу в цілому. Взаимодействие на основе текста с использованием мобильных устройств стало повсеместным, его основными источниками являются социальные сети, мессенджеры, электронные письма, виртуальные помощники, приложения для обеспечения доступности и т.д. Это подразумевает необходимость создания систем облегчения ввода текста пользователем и разработки способов поддержки вербальной обратной связи. В этой статье мы обсуждаем метод генерации уникального текста для мобильных устройств и методологию его оценки в качестве решения обеих заявленных проблем. Мы рассматриваем возможности, предоставляемые использованием контекста (местоположение, погода, запланированные события и т.д.), ограничения вычислительных ресурсов и использования данных, а также присущую субъективность оценки творческой задачи с учетом разнообразия возможных приемлемых результатов. Сравнение с другими методами генерации текстов показывает, что использование метрик связности помогает достичь более высокого качества с точки зрения человеческого восприятия. Корреляция Спирмена между значениями предлагаемой метрики связности и человеческой оценкой читабельности текста составляет 0.86, что свидетельствует о высоком качестве метрики и эффективности метода в целом. en Інститут кібернетики ім. В.М. Глушкова НАН України Кибернетика и системный анализ Кібернетика Improving text generation through introducing coherence metrics Покращення якості генерування тексту за допомогою міри зв’язності Улучшение качества генерации текста с помощью меры связности Article published earlier |
| spellingShingle | Improving text generation through introducing coherence metrics Marchenko, O.O. Radyvonenko, O.S. Ignatova, T.S. Titarchuk, P.V. Zhelezniakov, D.V. Кібернетика |
| title | Improving text generation through introducing coherence metrics |
| title_alt | Покращення якості генерування тексту за допомогою міри зв’язності Улучшение качества генерации текста с помощью меры связности |
| title_full | Improving text generation through introducing coherence metrics |
| title_fullStr | Improving text generation through introducing coherence metrics |
| title_full_unstemmed | Improving text generation through introducing coherence metrics |
| title_short | Improving text generation through introducing coherence metrics |
| title_sort | improving text generation through introducing coherence metrics |
| topic | Кібернетика |
| topic_facet | Кібернетика |
| url | https://nasplib.isofts.kiev.ua/handle/123456789/190337 |
| work_keys_str_mv | AT marchenkooo improvingtextgenerationthroughintroducingcoherencemetrics AT radyvonenkoos improvingtextgenerationthroughintroducingcoherencemetrics AT ignatovats improvingtextgenerationthroughintroducingcoherencemetrics AT titarchukpv improvingtextgenerationthroughintroducingcoherencemetrics AT zhelezniakovdv improvingtextgenerationthroughintroducingcoherencemetrics AT marchenkooo pokraŝennââkostígeneruvannâtekstuzadopomogoûmírizvâzností AT radyvonenkoos pokraŝennââkostígeneruvannâtekstuzadopomogoûmírizvâzností AT ignatovats pokraŝennââkostígeneruvannâtekstuzadopomogoûmírizvâzností AT titarchukpv pokraŝennââkostígeneruvannâtekstuzadopomogoûmírizvâzností AT zhelezniakovdv pokraŝennââkostígeneruvannâtekstuzadopomogoûmírizvâzností AT marchenkooo ulučšeniekačestvageneraciitekstaspomoŝʹûmerysvâznosti AT radyvonenkoos ulučšeniekačestvageneraciitekstaspomoŝʹûmerysvâznosti AT ignatovats ulučšeniekačestvageneraciitekstaspomoŝʹûmerysvâznosti AT titarchukpv ulučšeniekačestvageneraciitekstaspomoŝʹûmerysvâznosti AT zhelezniakovdv ulučšeniekačestvageneraciitekstaspomoŝʹûmerysvâznosti |