=Paper=
{{Paper
|id=Vol-2693/paper3
|storemode=property
|title=A New Approach for Extracting the Conceptual Schema of Texts Based on the Linguistic Thematic Progression Theory
|pdfUrl=https://ceur-ws.org/Vol-2693/paper3.pdf
|volume=Vol-2693
|authors=Elena del Olmo,Ana María Fernández-Pampillón
|dblpUrl=https://dblp.org/rec/conf/ecai/SuarezF20
}}
==A New Approach for Extracting the Conceptual Schema of Texts Based on the Linguistic Thematic Progression Theory==
https://ceur-ws.org/Vol-2693/paper3.pdf
Proceedings of the Workshop on Hybrid Intelligence for Natural Language Processing Tasks HI4NLP (co-located at ECAI-2020)
Santiago de Compostela, August 29, 2020, published at http://ceur-ws.org
A new approach for extracting the conceptual schema of
texts based on the linguistic Thematic Progression theory
Elena del Olmo1 and Ana María Fernández-Pampillón2
Abstract.3 The purpose of this article is to present a new approach approaches theoretically improve the consistency issue, they
for the discovery and labelling of the implicit conceptual schema of introduce a new complexity layer: a natural language generator
texts through the application of the Thematic Progression theory. module. Despite this greater complexity, nowadays text
The underlying conceptual schema is the core component for the summarization research is progressively shifting towards abstractive
generation of summaries that are genuinely consistent with the
approaches (Lin et al. 2019).
semantics of the text.
Traditionally, abstractive summarization techniques have been
1. INTRODUCTION classified into structure-based, intended to populate predefined
information structures out of the relevant sentences of the texts, and
Automatic Summary Generation was first proposed in the late semantic-based, involving a wide variety of knowledge
1950s. Outstanding examples of this early stage are Luhn (1958), representation techniques. Regarding the former, depending on the
whose method is based on sentence extraction relying on its words structural schema chosen, it is possible to identify, (i), tree-based
weightings, inferred from TF-IDF metrics, or Edmundson (1969), models (Kikuchi at. al. 2014), which perform different strategies for
who proposed novel sentence weighting metrics, such as the syntactic parsing analysis in order to codify paraphrasing
presence of words from a predefined list, the presence of the words information mainly by linking and reducing the syntactic sentence
of the title of the document or its positioning at the beginning of trees of the text, (ii), template-oriented models (Elhadad et al. 2015,
documents and paragraphs. These are paradigmatic examples of the Wu et al. 2018, Wang et al. 2019), which rely on extraction rules led
first extractive summarization techniques: techniques based on the by linguistic patterns matching sequences of tokens to be mapped
verbatim extraction of the most relevant parts of a text. The into predefined templates, (iii), ontology-based models (Nguyen
generated text summary was, thus, a collection of sentences 2009, Baralis et al. 2013), which are highly domain-dependent and
considered relevant but, often, semantically inconsistent because of include a hierarchical classifier mapping concepts into the nodes of
the overall weakness in coherence (the text does not make overall an ontology, and, (iv), rule-based models (Genest et al. 2011), based
sense) and cohesion (the sentences are connected incorrectly). The on extraction rules operating on categories and features
summary generated was consequently a poorly connected text with representative of the content of the text. Regarding the latter,
no global meaning, presumably due to the assumption of semantic-based techniques for abstractive summarization, there are
independence of the extracted sentences (Lloret et al. 2012). interesting approaches based on the concept of information item
(Gatt et al. 2009), the smallest units with internal coherence, in the
Currently, five main approaches to extractive techniques can be format of subject-verb-object triplets obtained through semantic role
distinguished: (i), statistical approaches (Luhn 1958, McCargar labeling, disambiguation, coreference resolution and the
2004, Galley 2006), based on different strategies for term counting, formalization of predication. Besides, there are approaches based on
(ii), topic-based approaches (Edmundson 1969, Harabagiu et al. discourse information (Gerani et al. 2014, Goyal et al. 2016),
2005), which assume that several topics are implicit in a text and predicate-argument tuples (Li 2015, Zhang et al. 2016) and
attempt to formally represent those topics, (iii), graph-based semantic graphs (Liu et al. 2019).
approaches (Erkan et al. 2004, Giannakopoulos et al. 2008), based
on the representation of the linguistic elements in texts judged to be The aforementioned tendency towards abstractive approaches in
relevant as nodes connected by arcs, (iv), discourse-based recent years is framed at a stage when the new Deep Learning
approaches (Marcu 2000, Cristea et al. 2005, da Cunha et al. 2007), models have proved to be particularly promising for using vector
whose target is to capture the discursive relations within texts, and, spaces as a way to address the shortcomings of discrete symbols as
(v), machine learning approaches (Aliguliyev 2010, Hannah et al. the input for Natural Language Processing tasks, such as tokens or
2014), intended to reduce the text summarization task to a lemmas, which cannot represent the underlying semantics of the
classification task by assigning a relevance value to each sentence. concepts involved. This new paradigm has provided techniques for
both extractive and abstractive summarization, such as the clustering
Although historically less addressed in the literature, abstractive of sentence and document embeddings, or the generation of correct
models try to address the lack of coherence and cohesion in the sentences given a sentence embedding and a language model.
summaries produced, using some source of semantic internal Remarkable examples are the contributions of Templeton et al.
representation of the text (which can be merely the output of an (2018), who compare different methods of sentence embeddings
extractive process) to generate the ultimate summary, composed of computing their cosine similarity, or Miller et al. (2019), who
sentences not necessarily included in the original text. Although this
1
General Linguistics department, Complutense University of Madrid, Spain, email: elenadelolmo@ucm.es
2
General Linguistics department, Complutense University of Madrid, Spain, email: apampi@ucm.es
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
23
proposed k-means clustering to identify sentences closest to the 3.1. Thematic theory
centroid for summary selection.
The thematic theory is framed within the optics of linguistic analysis
In addition to the distinction between extractive and abstractive corresponding to the informational layer. The uses and applications
approaches, there is a crucial challenge in automatic summarization that the authors have been giving to terms such as theme, focus, topic
which affects them both: the subjectivity of the accuracy scoring of and notions such as new information or emphasis have been
summaries. This implies a new difficulty in the creation of objective overwhelmingly numerous (Gutiérrez 2000). In accordance with the
gold datasets composed of correct summaries. In this context, Thematic theory, in descriptive and narrative texts, which are the
unsupervised summary models, such as the one proposed in this ones most typically to summarize, known information, or theme, is
paper, which does not require training labelled data, has become consensually described to be positioned at the beginning of
particularly relevant. Among the unsupervised approaches we can sentences. By contrast, the phrases containing the informative
highlight, (i), approaches which are extensions of word embedding contribution of the sentence, also known as rheme, tend to be located
techniques, such as the n-grams embeddings (Mikolov et al. 2013), further to the right, ahead in the time of enunciation. This description
or doc2vec (Le et al. 2014), (ii), the skip-thought vectors (Kiros et is consistent with how the acquisition of new knowledge is described
al. 2015), (iii), the Word Mover’s Distance model (Kusner et al. at the neurological level, through networking the known with the
2015), (iv), quick-thought vectors (Logeswaran et al. 2018), and, (v), novel or by altering pre-existing relationships (McCulloch et al.
models based on contextual embeddings obtained from 1943).
transformers, such as SBERT (Reimers et al. 2019).
In order to clarify how we will use these concepts, we present here
This paper addresses one of the weaknesses of extractive models a series of examples adapted from Gutiérrez (2000: 18) and their
discussed in the previous section, i. e. the lack of coherence in the corresponding answers:
summaries produced, especially when there are insufficient
1. Who joined Luis this morning?
linguistic datasets in a language for applying machine or Deep
Ludwig was joined this morning by Peter.
Learning methods. In this respect, the solution we propose identifies
implicit conceptual schemas from texts using the morpho-syntactic 2. When did Pedro join Luis?
knowledge currently provided by NL analyzers. Pedro joined Luis this morning.
3. Who joined Pedro this morning?
The paper is organized as follows: in section 2 we define the This morning Peter joined Ludwig.
hypothesis and objectives of the research work. In section 3 we
present a review of the linguistic theories on which we base our That these statements are different is a standard judgment for any
solution: the Thematic theory and Thematic Progression theory. In native speaker. Although they share the same representative
section 4 we present our solution: the application of both theories function, i. e. their syntactic and semantic relations do not differ,
for the identification of the text conceptual schema. In section 5 we they show different informative functions. Therefore, in spite of
study the feasibility of our solution for the automatic extraction of transmitting the same information about the world, they do not
thematic progression in Spanish, a language with few linguistic inform in the same way. Accordingly, the underlying assumption of
datasets for text summarization. Finally, in section 6 we draw the our proposal is that the thematic status of a phrase (like who, when
conclusions of this work and present our future research lines. and who in the examples above, respectively) is relevant in terms of
the prevalence of the concept involved for the summarization of a
document. Their clustering along a document, taking into account
2. HYPOTHESIS AND OBJECTIVES the thematic progressions patterns found, as further explained in the
Our hypothesis is that applying the Thematic Theory and the next section, is expected to reveal the conceptual schema of the text.
Thematic Progression Theory to annotate the discourse features
theme, rheme and their coreferences will allow us to extract thematic
progression schemas, which represent the implicit conceptual
3.2. Thematic Progression theory
schemas of texts. Daneš (1974: 114) presents the thematic progression as the choice
and arrangement of discourse themes, their concatenation, hierarchy
Therefore, our aim is obtaining an internal representation of the text and relation with the topics of texts. Accordingly, he argues that it is
informational structure as a formal representation for text possible to infer the informational schema of a text from its theme-
summarization. The advantage of this solution is that it can be rheme organization. It is considered that there are three main
applied to any language regardless of whether or not there are typologies of thematic progressions: (i), linear progression, in which
enough training data for the implementation of machine learning and the rheme of one sentence is the theme of the subsequent sentence,
Deep Learning techniques. In our work we will use Spanish as the (ii), constant progression, in which a theme is repeated over several
language to study the feasibility of the solution. We also hope to sentences, and, (iii), derived progression, in which several topics are
contribute to the generation of summaries in Spanish, a task derived from the same inferred hypertheme. Apart from these three
currently performed with moderate efficiency due to the limited basic types, Daneš (1974: 120) also proposed that the combination
availability of linguistic resources. of them can lead to thematic progressions of higher levels of
abstraction, such as, (iv), the split rheme progression, which consists
of the existence of a complex rheme, whose hyponyms and
3. REVIEW OF THEMATIC AND meronyms are themes of the subsequent sentences. Finally, he
THEMATIC PROGRESSION concludes (1974: 122) that the study of the thematic organization of
24
a text could be useful for numerous practical applications, among In short, it is possible to locate the discourse elements theme and
which outstands information retrieval, given the performance rheme using syntactic knowledge. To the extent that syntactic
achieved nowadays by the tools available for the automatic text analysis is a task that can be considered well solved in NLP, it seems
analysis. feasible to be able to automatically locate the theme and rheme in
every sentence of a text. The next natural step in order to obtain the
thematic progression schema, i. e. the conceptual schema of a text,
4. THEMATIC PROGRESSION AS A is to connect each theme and rheme of the sentences of the text in
MODEL FOR SEMANTIC TEXT the ultimate thematic path.
REPRESENTATION
The usefulness of the thematic or rhematic roles of concepts along 5. A FIRST STUDY OF THE FEASIBILITY
texts for automatic text summarization arises from two main facts. OF THEMATIC PROGRESSION
On the one hand, the theoretical validation of the concept of thematic
THEORY IN SPANISH TEXTS
progression enjoys consensus among researchers as a relevant
description for the semantic structure of texts. On the other hand, Aiming to conduct a first study to verify the applicability of the
although it has been traditionally examined through the optics of the Thematic Progression theory for the extraction of the underlying
Pragmatics layer, the thematic or rhematic status of a concept is conceptual schema of a text, we carried out an exploratory corpus
actually embodied in the surface syntactic layer, which is prone to survey with Spanish descriptive texts. We analyzed the mean ratio
be represented in an easy-to-compute form. and the ratio per text of preceding subjects, since they are the
prototypically unmarked themes. The examined corpus is AnCora
Concerning the correlation between the theme of a sentence and its Surface Syntax Dependencies3 (AnCora GLiCom-UPF 1.1),
syntactic structure, which is crucial for its automatic annotation, published in 2014 at Pompeu Fabra University, which contains
Halliday (1985) proposed an interesting categorization based on the 17,376 sentences manually annotated with the lemmas, the PoS tags
concept of linguistic markedness. Thus, in SVO languages, such as plus other morphological features and both dependency heads and
English or Spanish, for declarative sentences there are unmarked relations for every token. The analysis was based on a symbolic rule-
themes, prototypically the syntactic subjects preceding principal based grammar expressed as sub-tree extraction operations from the
verbs, and marked themes, such as circumstantial attachments, dependency tree of the sentences. In order to ensure the generality
complements, or sentences with predicate construction. Examples of the grammar, the elements obtained through rules for sampling
for the former are the first and second sentence of the examples preceding subjects were compared with the corresponding results in
provided above with Ludwig and Pedro as unmarked themes a second version of the corpus, resulting from its automatic
respectively, whilst the third sample sentence is an example for the annotation with the Freeling analyzer. The thematic progression
latter, with this morning as theme. Thematic equative sentences, analyzer scripts used to process the grammar and to generate the
such as What I want is a proper cup of coffee, would be excluded outputs of the corpus rendering are publicly accessible from github4.
from this categorization. For interrogative sentences, the unmarked
themes are definite verbs for yes-no questions, such as did deliver in Basically, the grammar and the analysis algorithm to extract the
Did the president deliver a speech, and interrogative pronouns and thematic progression schema of the text were designed in three
similar phrases for non-polar questions, such as where in Where is consecutive steps: (i), first, the automatic identification and labelling
Berlin located?, whilst marked themes are circumstantial adjuncts. of themes for every sentence; (ii), second the subsequent
identification of their rheme; and, (iii), finally, the identification of
Besides, a constituent which is not the theme of a sentence may concepts corresponding to the same theme or rheme in a text. Each
appear occasionally in a prototypical theme position. This step was carried out using a symbolic syntactic-semantic rule-based
phenomenon has been referred to by several names, such as grammar expressed as sub-tree extraction operations. For the
focussing (Campos et al. 1990), focus preposition (Ward 1985) or grammar definition, an approach based on the transformation of
thematization (Chomsky 1972). Examples of this type of dependency trees has been applied. Thus, for the simplest scenario
informational structure are It was Pedro who lied to me. A number of finding unmarked subjects in SVO languages, such as Spanish,
of authors (e.g. Gutiérrez 2000: 34) have argued that the intent of two categories of rules were defined: (i), matching rules of child
this particular information schemas is to gain the attention of the dependencies from a selected head token, consisting of the
interlocutor to overcome their presumed predisposition to receive identification of a dependency relation as the name of a relation of
information that is at some point contrary to that which is intended arity two with a first argument as the key and the value expected for
to be communicated, or simply to emphasize the importance of a the selected parent and a second argument with the options for
certain aspect in the informational process. This nuance of matching, being ALL if all children nodes from the head of a scope
enunciative modality would undoubtedly be applicable for the should be matched or ONE if only the immediate child should be
weighing of the relevant concepts for a proper summary, especially matched. For example, the SUB(deprel:ROOT, ALL) rule would
since the syntactic structures involved are relatively easy to match match subtrees consisting of all child nodes of the SUB children of
with rules only out of the tokens positions, the dependency relation a token tagged with a ROOT dependency relation (as shown in figure
tags and the dependency heads. 1, obtained from the Freeling 4.1 demo5); and, (ii), matching rules
of head dependencies from a selected child token, consisting also of
3
http://clic.ub.edu/corpus/es/ancora-descarregues
4
https://github.com/eelenadelolmo/HI4NLP/tree/master
5
http://nlp.lsi.upc.edu/freeling/demo/demo.php
25
the identificator of a dependency relation as the name of a relation through careful examination. We found that the vast majority of
of arity two, whose arguments are the same as for the child mismatched annotations involve some type of coordinated or
dependency rules but in the opposite order. The second type of rules juxtaposed clauses. These syntactic structures are analyzed by
can apply, for example, for sentence compression when several Freeling with a highly fluctuating dependency structure, which is
propositions are involved. quite different from the analysis in GLiCom-UPF 1.1. This high
variability in the syntactic tree accounts for the vast majority of both
the undermatched and overmatched sentences ratios.
A qualitative analysis of unmatched sentences has also been
conducted, revealing a strong presence of thematization as the most
relevant finding. This sentence pattern has been referred to by the
Thematic theory as a relevant discourse feature, indicating a break
in the information flow of the text, as further discussed above. A
promising conclusion of the analysis of the patterns found is their
suitability for being implemented in the rule formalism designed.
Besides, as a synthesis of the findings obtained from the qualitative
analysis of the manually annotated version, there is a strong presence
Figure 1. Input sentence and output (subtree) for a rule of the first type. of subordinate clauses in the corpus, which implies the necessity of
more complex rules to select the most informative proposition in the
The analyzer accepts various corpus formats as input and transforms sentence. The observed patterns have been categorized into three
them into the universal CoNLL-U format, where the additional main categories (the most informative clauses appear in bold and the
theme and rheme features are added for every token in the main selected theme is underlined):
proposition of sentences. We found that, with our first version of the
rules, theme and rheme annotation was correct in roughly half of the 1. Sentences whose root clause is the most relevant (e. g. Since
cases, as shown in table 1. Through a careful manual review of the pharmacists work with a high profit margin, the business
data, these results have enabled us to confirm a significant opportunity is huge).
correlation between the syntactic-semantic and discourse layers 2. Sentences whose root clause is not the most relevant. (e. g. The
outlined by the Thematic theory, and, consequently, the feasibility main factor is that electricity consumption during the summer
of automating identification of, at least, half of the themes. is now not much lower than it used to be).
Table 1. Ratio of preceding subjects in AnCora corpus.
3. Sentences whose root clause is not the most relevant but
provides a crucial modality feature for information retrieval (e.
GLiCom-UPF 1.1 Freeling version g. Investigators are convinced that someone deliberately cut
that rubber).
Sentences with preceding 50.4 % 46.2 %
subjects as themes
6. CONCLUSIONS AND FUTURE WORK
With the syntactically annotated GLiCom-UPF 1.1 version of the As observed in the study conducted, this first approach to rule-based
AnCora corpus we seek to provide objective metrics, not prone to theme annotation seems to claim the theoretically hypothesized
major annotation errors, as the corpus annotation has been manually correlation between the syntactic-semantic and discourse layers
reviewed. The qualitative analysis of this data is intended to assess required by our proposal. However, the qualitative analysis of both
whether or not the preceding subjects match the main themes of matched and unmatched sentences revealed the need for more
sentences in order to ultimately detect the underlying thematic complex tree rewriting rules to achieve a more accurate theme
progression template of texts. By contrast, with the Freeling version selection in order to obtain thematic progression schemas from texts.
of the corpus we aim to assess the accuracy in applying the rule for
the extraction of unmarked themes with no dependence on manually Regarding subordination, i. e. sentences with several propositions
annotated data, because Freeling is the best option for syntax with different syntactic status, we are working on two feasible
annotation in Spanish at this stage. According to this objective, we options for sentence compression: (i), the choice of the most relevant
have generated two different files, one for each version of the proposition for every sentence, and, (ii), the choice of the ordered
corpus, with the suspected overmatched and undermatched subset of its n more relevant clauses. In addition, this study shows
sentences, as shown in table 2. the necessity to implement an algorithm to infer the modality from
the main verb. We also found that the rules should be refined in order
Table 2. Ratio of suspected annotation errors. to capture the various ways in which coordinated and juxtaposed
Freeling version clauses could be analyzed, given the high variability observed in the
automatic syntactic annotation. Finally, the study revealed that it
Suspected overmatches 1283 (7.2 %) will be necessary to design lexicon-based rules to capture lexical
semantic generalizations. With all this, in principle, it seems
Suspected undermatches 3550 (20.1 %) possible to apply this new linguistic approach for the extraction of
implicit conceptual schemas from texts.
As the figures suggest, a pervasive tendency for undermatching
sentences with preceding subjects by Freeling has been confirmed
26
REFERENCES 24. M.J. Kusner, Y. Sun, N.I. Kolkin and K.Q. Weinberger. ‘From Word
Embeddings To Document Distances’, Proceedings of the 32nd
International Conference on International Conference on Machine
1. A. Gatt and E. Reiter, ‘SimpleNLG: A realisation engine for practical Learning, 37, 957–966, (2015).
applications’, Proceedings of the 12th European Workshop on Natural 25. N. Elhadad, K. McKeown, D. Kaufman, and D. Jordan, ‘Facilitating
Language Generation, 90-93, (2009). physicians’ access to information via tailored text summarization’,
2. A. Templeton and J. Kalita. ‘Exploring Sentence Vector Spaces Proceedings of the AMIA Annual Symposium, 226-300, (2005).
through Automatic Summarization’, Proceedings of the 17th IEEE 26. N. Chomsky, Studies on Semantics in Generative Grammar, Mouton,
International Conference on Machine Learning and Applications, La Haya, 1972
(2018). 27. N. Goyal and J. Eisenstein. ‘A Joint Model of Rhetorical Discourse
3. C.Q. Nguyen and T.T. Phan. ‘An ontology-based approach for key Structure and Summarization’, Proceedings of the Workshop on
phrase extraction’, Proceedings of the ACL-IJCNL, 181-184, (2009). Structured Prediction for NLP, 25-34 , (2016).
4. D. Cristea, O. Postolache and I. Pistol. ‘Summarisation through 28. N. Reimers and I. Gurevych. ‘Sentence-BERT: Sentence Embeddings
discourse structure’, Lecture Notes in Computer Science, 3406, 632- using Siamese BERT-Networks’, Proceedings of the 2019 Conference
644, (2005). on Empirical Methods in Natural Language Processing and 9th
5. D. Marcu. The Theory and Practice of Discourse and Summarization, International Joint Conference on Natural Language Processing,
The MIT Press, Cambridge, 2000. (2019).
6. D. Miller. ‘Leveraging BERT for Extractive Text Summarization on 29. P.E. Genest and G. Lapalme. ‘Framework for Abstractive
Lectures’, Proceedings of arXiv, (2019). Summarization using Text-to-Text Generation’, Proceedings of the
7. E. Baralis, L. Cagliero, S. Jabeen, A. Fiori and S. Shah. ‘Multi- 49th Annual Meeting of the Association for Computational Linguistics,
document summarization based on the Yago ontology’, Expert 64-73, (2011).
Systems with Applications, 9, 6976-6984, (2013). 30. P. Wu,, Q. Zhou, Z. Lei, W. Qiu and X. Li, ‘Template Oriented Text
8. E. Lloret and M. Palomar. ‘Text summarization in progress: a literature Summarization via Knowledge Graph’, International Conference on
review’, Artificial Intelligence Review, 37, 1-41, (2012). Audio, Language and Image Processing (ICALIP), 79-83, (2018).
9. F. Daneš, Papers on functional sentence perspective, Mouton, The 31. Q. Le and T. Mikilov. ‘Distributed Representations of Sentences and
Hague,1974. Documents’, Proceedings of the 31st International Conference on
10. G. Erkan and D.R. Radev. ‘LexRank: Graph-based lexical centrality Machine Learning, (2014).
as salience in text summarization’, Journal of Artificial Intelligence 32. R. Kiros, Y. Zhu, R. Salakhutdinov, R.S. Zemel, A. Torralba, R.
Research, 22, 457-459, (2004). Urtasun and S. Fidler. ‘Skip-Thought Vectors’, Proceedings of the
11. G. Giannakopoulos, V. Karkaletsis and G.A. Vouros. ‘Testing the Use 28th International Conference on Neural Information Processing
of N-gram Graphs in Summarization Sub-tasks’, Proceedings of the Systems, 2, 3294–3302, (2015).
Text Analytics Conference, (2008). 33. R.M. Aliguliyev. ‘Clustering techniques and discrete particle swarm
12. G. Ward, The Semantics and Pragmatics of Preposing, Universidad de optimization algorithm for multi-document summarization’,
Pennsylvania, 1985. Computational Intelligence, 26, 420-448, (2010).
13. H. Campos and M. Zampini. ‘Focalization Strategies in Spanish’, 34. S. Gerani, Y. Mehdad, G. Carenini, R.T. Ng and B. Neja. ‘Abstractive
Probus, 2, 47-64, 1990. summarization of product reviews using discourse structure’,
14. H. Lin and V. Ng. ‘Abstractive Summarization: A Survey of the State Proceeding of the Conference on Empirical Methods in NLP, 1602-
of the Art’, Proceedings of the Thirty-Third AAAI Conference on 1613, (2014).
Artificial Intelligence, 9816-9822, (2019). 35. S. Gutiérrez, Temas, remas, focos, tópicos y comentarios, Arco Libros,
15. H.P Edmundson. ‘New methods in automatic extracting’, Journal of Madrid, 2000.
the ACM, 16(2), 264–285, (1969). 36. S. Harabagiu and F. Lacatusu. ‘Topic themes for multi-document
16. H.P. Luhn. ‘ The Automatic Creation of Literature Abstracts’, IBM summarization’, Proceedings of the 28th Annual International ACM
Journal of research and development, 2(2), 159-165, (1958). SIGIR, (2005).
17. I. da Cunha, S. Fernández, P.V. Morales, J. Vivaldi, E. San Juan and 37. T. Mikolov, I. Sutskever, K. Chen, G. Corrado and J. Dean.
J.M. Torres-Moreno. ‘A new hybrid summarizer based on vector space ‘Distributed Representations of Words and Phrases and their
model, statistical physics and linguistics’, Lecture Notes in Computer Compositionality’, Advances in neural information processing
Science, 4827, 872-882, (2007). systems, 26, (2013).
18. J. Zhang, Y. Zhou and C. Zong. ‘Abstractive Cross-Language 38. V. McCargar. ‘Statistical approaches to automatic text
Summarization via Translation Model Enhanced Predicate Argument summarization’, Bulletin of the American Society for Information
Structure Fusing’, IEEE/ACM, 24, (2016). Science and Technolog, 30, (2004).
19. K. Wang, X. Quan and R. Wang. ‘BiSET: Bi-directional Selective 39. W. Li. ‘Abstractive multi-document summarization with semantic
Encoding with Template for Abstractive Summarization’, The 57th information extraction’, Proceedings of the Conference on Empirical
Annual Meeting of the Association for Computational Linguistics, Methods in NLP, 1908-1913, (2015).
(2019). 40. W.S. McCulloch and W. Pitts. ‘A logical calculus of the ideas
20. L. Logeswaran and H. Lee. ‘An efficient framework for learning immanent in nervous activity’, Bulletin of Mathematical Biophysics,
sentence representations’, Proceedings of the 6th International 5, 115-133, (1943).
Conference on Learning Representations, (2018). 41. Y. Kikuchi, T. Hirao, H. Takamura, M. Okumura and M. Nagata.
21. M.A.K. Halliday, An Introduction to Functional Grammar, Arnold, ‘Single Document Summarization based on Nested Tree Structure’,
London, 1985. Proceedings of the 52nd Annual Meeting of the Association for
22. M.E. Hannah and S. Mukherjee. ‘A classification-based Computational Linguistics, 315–320, (2014).
summarisation model for summarising text documents’, International 42. Y. Liu, I. Titov and M. Lapata. ‘Single Document Summarization as
Journal of Information and Communication Technology, 6, 292-308, Tree Induction’, Proceedings of NAACL-HLT, 1745-1755, (2019).
(2014).
23. M. Galley. ‘Skip-chain Conditional Random Field for ranking meeting
utterances by importance’, Proceedings of the Conference on
Empirical Methods in Natural Language Processing, (2006).
27