LIVING-LANG: Living digital entities by human language technologies

Abstract

This project pursues the dynamic modeling at a spatial-temporal level of digital entities in social media for predicting their behavior. Firstly, digital entities are modelled by identifying the characteristics of individuals through their language and footprint on the network. Then, the extraction of relationships between digital entities is one of the nuclear challenges of the project. The proposal pursues this objective on a semantic level, structuring the information into representations of knowledge suitable for logical processing. Considering the heterogeneous nature of the sources to be dealt with, filtering of information is fundamental, using metrics and quality criteria. This spatial-temporal characterization, together with screening processes, will allow us to study high-performance predictive strategies in the evolution of digital entities. This project is coordinated by the SINAI and GPLSI research groups.

Keywords

• eol>Natural Language Processing
• Sentiment Analysis
• Emotion Mining
• Sentiment Enrichment

-

1. Introduction nario and often resulting in the deterioration of our behaviours of digital entities. These predictions can cognitive abilities [9] [ 10 ]. This enormous amount be used for the early detection of problems assoof information and digital connectivity entails the ciated with violence, mental health problems such development of a technology capable of modelling as suicides, inappropriate behaviours and other sethe new paradigm, as well as determining the rela- curity and health risks. Therefore, for example, a tionships that arise, their evolution in time and the change in pattern of the type of language used in the ability to interfere with or predict their behaviour communication between two people can help detect in the future. the start of practices such as sexual harassment,

Our previous project set out to identify digital en- when language moves from a suggestive, captivattities, considered to be any entity in the real world ing or friendly language to that of a coercive or (people, companies, organisations, tourist attrac- threatening nature. As shown in Figure 1, the relations. . . ) with presence in the digital world and tionships between entities are dynamic and change from which we can obtain a complete profile from with time as do their properties. By identifying their activity in such an environment. This pro- these variations and their patterns based on human ifle is generated by processing unstructured content language, we can prepare these networks for the (web pages, articles, comments. . . ) using human future by creating predictive models of peoples’ belanguage technologies. However, the present “digi- haviour (risk detection, prevention of cyberbullying, tal” situation requires us to go a step further and terrorist warnings, etc.). attempt to answer the following questions: a) How can we ensure the social contextualization of these entities, and model situations that change from day 2. Objetives to day? b) How can we deduce new semantic relationships between entities? c) How can we guarantee The project started in 2018 and will be completed in that captured knowledge is real and contrasted by 2022, and it involves a number of specific challenges multiple sources? d) How can we guarantee the and objectives of the overall project in the field of coexistence of knowledge in the long term? NLP research, which are detailed below:

This project aims to take this several steps fur- OBJ1. Generation of the human language models ther. In this way and thanks to these characteristics, used by digital entities through recognition of their we can establish relationships between the entities primary characteristics (linguistic, cognitive, social, from a social and human perspective, improve the cultural and emotional) and independent of the comprehension of the content exchanged, create new domains and scenarios in which they act. knowledge in the analysis of these relational struc- OBJ2. Use of the knowledge generated by digital tures and eventually, characterise and predict these entities and discovery of the semantic relationships networks between entities on a human language between them. All available sources of information level by using temporal dimension, behaviours or (unstructured, structured and open linked data), phenomena. extraction mechanisms, identity enrichment, and

This ability to understand language, model it other inference mechanisms will be taken into acand analyse its changes in time will allow us to count. This will enable the integration of informaface new challenges in the digital society in which tion related to an identity, determining the roles and we live. By measuring the veracity and credibil- properties associated to a space-time framework. It ity of the relationships extracted, we can confront also enables the definition of relationships between phenomena such as fake news, defined as a delib- identities using dynamic aspects such as context, erate distortion of a reality with the objective of temporary nature or importance. creating and shaping public opinion and influencing OBJ3. Use of knowledge of relationships to desocial attitudes. Thanks to this project, tasks such termine the coherence, quality and contrast of the as fact-checking, the automatic detection of ideo- semantic relationships extracted. For this, we will logical or confirmation bias, and the detection of use veracity assessment techniques, emotion analyclickbaits can be handled automatically, as well as sis and subjectivity, as well as the detection of bias other post-truth problems that are dificult to detect in the information to guarantee and contrast the and treat because of their “viral” content. Further- information that arises from the relationship. more, language modelling and new knowledge about OBJ4. Prediction of future behaviour of digital these dynamic relationships and their evolution over entities by discovering potential future semantic time will allow us, through the application of di- relationships between them, through the analysis verse techniques, to identify new characteristics and of pre-existing networks and based on previously make inferences that provide predictions of future detected relationships.

In summary, this project contributes to the Span- [ 11, 12 ], studying the characteristics of the diferent ish national Plan for the Promotion of Human Lan- scenarios in order to model the language in each guage Technologies, which has aimed to promote the of them. Resources associated with the diferent development of natural language processing since scenarios and domains defined have been created 2015. and used to train machine learning systems.

To achieve the above global objective and the Results regarding OBJ2: The project has worked specific objectives of the global project, the coordi- on various techniques for knowledge extraction in nation of two complementary sub-projects is pro- the diferent domains and scenarios defined, as well posed, whose specific objectives will cover the global as on the organisation of workshops such as eHealthobjectives proposed, and whose reunification will KD 2020 to model human language in health docprovide the added value sought by the coordination. uments in Spanish [13]. In addition, knowledge discovery techniques are being applied to the health domain [14, 15]. In addition, work has been done 3. Results and conclusions on the discovery of temporal information to enrich the entities by automatically extracting timelines This section describes the most significant results from the documents and generating summaries from of the project. these timelines [16].

Results regarding OBJ1: In this project, the do- Results regarding OBJ3: In relation to this objecmains to be worked on are mainly health and edu- tive, a systematic study of the state of the art in cation, as well as the following scenarios: fake news, this matter has been carried out[17] and, based on knowledge extraction, violence and hate speech this study, work has been done to determine both the veracity of the news and its parts and to study [6] Y. Eshet, Thinking in the digital era: A revised the detection of satire, achieving an architecture model for digital literacy, Issues in informing capable of determining 74% accuracy [18]. Within science and information technology 9 (2012) this task, progress has been made in the detection 267–276. of incongruent headlines as well as in fact-checking [7] D. Salyer, Reading the web: Internet guided tasks, as part of the disinformation detection archi- reading with young children, The Reading tecture [19]. In addition, work has been done on Teacher 69 (2015) 35–39. emotion detection [20] [21] and negation [22]. [8] N. K. Hayles, How we read: Close, hyper,

Results regarding OBJ4: Regarding this objec- machine, ADE 150 (2010) 62–79. tive, the project focused on the discovery of virality [9] M. Bauerlein, The Dumbest Generation—How patterns, applying opinion mining techniques that the Digital Age Stupefies Young Americans enable us to structure the information based on and Jeopardizes Our Future, Jeremy P. Tarcher the polarity of the messages and the emotions they / Penguin, New York, 2008. contain [23]. After transforming the information [ 10 ] N. Carr, The shallows: What the Internet is from an unstructured textual representation to a doing to our brains, WW Norton & Company, structured one, association rules mining were used, 2020. concluding that messages with a high-negative po- [ 11 ] F. M. P. del Arco, M. D. Molina-González, larity and a very high emotional charge, especially L. A. Ureña-López, M. T. Martín-Valdivia, emotions that have intensified with the COVID-19 Comparing pre-trained language models for pandemic, such as fear, sadness, anger and surprise spanish hate speech detection, Expert Syst. are more likely to go viral in social media. Appl. 166 (2021) 114120. URL: https://doi.

All publications related to the project can be org/10.1016/j.eswa.2020.114120. doi:10.1016/ found on the project website1. j.eswa.2020.114120. [12] F. M. P. del Arco, M. D. Molina-González,

L. A. Ureña-López, M. T. Martín-Valdivia, 4. Acknowledgments Detecting misogyny and xenophobia in spanish tweets using language technologies, ACM This research work is funded by MCIN/AEI/ Trans. Internet Techn. 20 (2020) 12:1–12:19. 10.13039/501100011033 and, as appropriate, by URL: https://doi.org/10.1145/3369869. doi:10. “ERDF A way of making Europe”, by the “European Union” or by the “European Union NextGenera- [13] 1A1.45P/i3a3d6-M98o6r9fis., Y. Gutiérrez, Y. AlmeidationEU/PRTR” through the grant LIVING-LANG Cruz, R. Muñoz, A computational ecosysProject (RTI2018-094653-B-C21 / C22). It is a tem to support ehealth knowledge discovery coordinated project with SINAI and GPLSI as par- technologies in spanish, J. Biomed. Informatticipating research groups. It is also funded by Gen- ics 109 (2020) 103517. URL: https://doi.org/ eralitat Valenciana through the project NL4DISMIS: 10.1016/j.jbi.2020.103517. doi:10.1016/j.jbi. Natural Language Technologies for dealing with disand misinformation (CIPROM/2021/21). [14] 2P0.20L.1ó0p3e5z1-Ú7.beda, M. C. Díaz-Galiano, T. Martín-Noguerol, A. Luna, L. A. U.

References López, M. T. Martín-Valdivia, Automatic medical protocol classification using machine [1] M. Tomasello, A natural history of human learning approaches, Comput. Methods thinking, Harvard University Press, 2014. Programs Biomed. 200 (2021) 105939. URL: [2] G. W. Grace, The linguistic construction of https://doi.org/10.1016/j.cmpb.2021.105939.

reality, Routledge, 2016. doi:10.1016/j.cmpb.2021.105939. [3] R. Rommetveit, Words, Meaning, and Mes- [15] P. López-Úbeda, M. C. Díaz-Galiano, sages: Theory and Experiments in Psycholin- T. Martín-Noguerol, A. Luna, L. A. U. guistics, Academic Press, 2014. López, M. T. Martín-Valdivia, COVID-19 [4] H. Wallach, Computational social science, detection in radiological text reports inte

Comput. Soc. Sci. 307 (2016). grating entity recognition, Comput. Biol. [5] M. Arnd-Caddigan, Sherry turkle: Alone to- Medicine 127 (2020) 104066. URL: https: gether: Why we expect more from technology //doi.org/10.1016/j.compbiomed.2020.104066. and less from each other, 2015. doi:10.1016/j.compbiomed.2020.104066.

[16] C. Barros, E. Lloret, E. Saquete, B. Navarro1https://livinglang.gplsi.es/ Colorado, Natsum: Narrative abstractive

References

Manag. 56 ( 2019 ) 1775 - 1793 . URL:

article/pii/S0306457318305922. doi:https:

//doi.org/10.1016/j.ipm. 2019 . 02 .010. [17] E. Saquete , D. Tomás , P. Moreda , P. Martínez-

141 ( 2020 ). URL: https://doi.org/10.1016/j.

eswa. 2019 . 112943 . doi: 10 .1016/j.eswa. 2019 .

112943. [18] A. Bonet-Jover , A. Piad-Morfis , E. Saquete,

news detection , Expert Syst. Appl. 169

( 2021 ) 114340 . URL: https://doi.org/10.1016/j.

eswa. 2020 . 114340 . doi: 10 .1016/j.eswa. 2020 .

114340. [19] R. Sepúlveda-Torres , M. E. Vicente , E. Sa-

mant. 71 ( 2021 ) 100660 . URL: https://doi.org/

10.1016/j.websem. 2021 . 100660 . doi: 10 .1016/j.

websem. 2021 . 100660 . [20] L. Canales , C. Strapparava , E. Boldrini,

put. 11 ( 2020 ) 335 - 347 . URL: https://doi.org/

10.1109/TAFFC. 2017 . 2764470 . doi: 10 .1109/

TAFFC. 2017 . 2764470 . [21] L. Canales , W. Daelemans , E. Boldrini,

access ( 2019 ) 1 - 1 . doi: 10 .1109/TAFFC. 2019 .

2927564. [22] S. M. Jiménez-Zafra , R. Morante , M. T.

Comput. Linguistics 46 ( 2020 ) 1 - 52 . URL:

https://doi.org/10.1162/coli_a_00371. doi:10.

1162/coli\_a\_ 00371 . [23] E. Saquete , J. Zubcof , Y. Gutiérrez ,

Syst. Appl . 197 ( 2022 ) 116676 . URL: https:

//doi.org/10.1016/j.eswa. 2022 . 116676 . doi:10.

1016/j.eswa. 2022 . 116676 .