In this paper, we present NegesAPI, an API for negation detection in Spanish texts using Machine Learning techniques, as well as its Web Application. We describe its architecture, the development technologies used and the integrated negation detection system. It is a powerful tool that can assist in diferent Natural Language Processing tasks.
ple if a patient “no tiene cáncer” (does not have cancer) and the entity “cáncer” is recognised, but the presence of Negation is a grammatical phenomenon present in all negation is not detected, it changes the meaning of the languages that can change the polarity of a sentence. It sentence and the diagnosis of the patient. expresses the negative form of a sentence or statement There are various approaches to handle negation in [1]. In Spanish there exist diferent types of negation de- NLP, such as rule-based approaches [2, 3], and traditional pending on the elements used to negate: syntactic nega- machine learning [4, 5] and deep learning approaches tion (no, nunca, nadie,...), lexical negation (impensable, [6, 7]. Rule-based approaches use a set of pre-defined ilegible, imposible,...) and morphological negation (en la rules to detect negation cues and modify the meaning of vida). Also, negation can be found in multiple ways, in- the sentence accordingly. Traditional machine learning cluding the use of simple markers (one negation cue: No and deep learning approaches, on the other hand, use sé cocinar), continuous markers (two or more consecu- annotated data to train models that can automatically tive negation cues: Casi no llego a la presentación) and identify negation cues and their scope in a sentence. discontinuous markers (two or more non-consecutively In this paper we describe the architecture of NegesAPI, negation cues: No tengo nada de dinero para la comida). an API to detect negation cues and scopes in Spanish
Therefore, negation can be very challenging to detect texts, that uses the machine learning approach of Jiménezand process. In fact, the computational treatment of nega- Zafra et al. [5] and that is trained in the SFU Review SPtion has not been solved due to its complexity. In Natural NEG corpus [8], which consists of 9,455 sentences from Language Processing (NLP) negation is a phenomenon reviews on 8 domains (cars, hotels, washing machines, of special interest, since it afects the polarity of texts, books, mobile phones, music, computers and movies). such as opinions, where the opinion of a product can be This API can analyze any text in Spanish and annotate drastically altered if negation is present. It also has an its negation automatically, to assist in other NLP tasks. impact on information retrieval systems, where search- The rest of the paper is organized as follows: In Section ing for “Películas que no sean de fantasía” (Non-fantasy 2 we provide a description of the API and its architecture. iflms) is not the same as searching for “Películas que sean In Section 3 we describe the components in the backde fantasía” (Fantasy films). It is also of special relevance end. The front-end components are set out in Section for entity recognition in biomedicine, where for exam- 4. In Section 5 we present the negation system used by NegesAPI. Finally, Section 6 summarizes the conclusions and future work.
• Data model: NegesAPI uses a relational database system that stores the user list and information 3.1. Request Management about them, in order to implement a login/register system. It should be noted that it does not The general request management architecture is prestore any information about the processed and sented in Figure 1 . First, when any request is sent, it is analyzed texts. It is built in MariaDB as it is open received by NGINX, serving as reverse proxy. Then, NGsource and it is the default database system in INX sends the request to an specified port of the machine, most Linux distributions. where Gunicorn is listening. Gunicorn is a Python WSGI
HTTP Server for UNIX, and it can replicate instances of • Views: The views contains all the information our flask application, named as workers. The requests are that is presented in the client side of the web ap- forwarded to an specific instance of the Flask application, plication, as well as the design and functionalities. where the request is resolved.
The information is written in HTML, the design is defined with CSS and minor functionalities executed in the client are implemented in JavaScript. 3.1.1. NGINX web server • Controller and routing: As routing for the web ap- NGINX is a web server and reverse proxy that is fast, replication we use Flask, a powerful Python frame- liable, and secure. It can manage HTTP connections and work with which we have created the diferent trafic by using various web acceleration techniques like endpoints. Flask uses a routes file, where data load balancing, SSL termination, connection and request validation is implemented and forms are created policing, static content ofload, and content caching. NGor retrieved. This routes file serves as controller, INX can also act as a secure application gateway that communicating with all other components when- passes trafic from users to applications using built-in ever it is necessary. The API is also written in interfaces.
Flask, but it has been previously generated using The reason to use NGINX as an application gateway Swagger Documentation. is that it provides an all-in-one solution for HTTP con• Negation system: The negation system is imple- nection management, load balancing, content caching, mented in the API /analyze route. It uses the Ma- and trafic security. This makes it easier to manage and chine Learning approach of Jiménez-Zafra et al. secure the application backend, improve scalability and [5], with an updated version of the corpus SFU performance, and build highly available applications by Review SP-NEG [1]. clustering application instances behind NGINX. • Request management: Flask already implements a WSGI server, the toolset of Werkzeug. Although 3.1.2. Gunicorn WSGI server it is more than enough for a development environment, it falls short in production, and it is necessary to apply several technologies and tools for a better user experience and system scalability. To achieve that, NegesAPI uses the Gunicorn WSGI HTTP server to replicate our Flask application and NGINX as reverse proxy and application gateway.
source Python Web Server Gateway Interface (WSGI) HTTP server. It allows Python web applications to be served by communicating with web servers like NGINX or Apache. Gunicorn is designed to be a lightweight and simple HTTP server that can handle multiple requests concurrently. Gunicorn is commonly used in combination with popular web frameworks like Flask, Django, Pyramid, and others to serve web applications with high performance and reliability.
We will use Gunicorn to replicate our Flask application, improving our API reliability by providing redundancy in case of server failures. Also, Gunicorn can improve the performance of a web application by using multiple worker processes to handle requests concurrently. This allows the application to handle a larger number of requests without slowing down or crashing. Scalability of the system must be taken into account, with Gunicorn we can easily adjust the number of workers, allowing more instances of the application to run concurrently.
The back-end is composed by all the server-side code, frameworks, tools and database management systems that are running in the server where the web application is allocated. It receives requests from the front-end, processes the data, and sends responses back to the frontend. database system that was created by the original developers of MySQL. It is designed to be high-performing, stable, and open-source, and is included as the default database system in most Linux distributions and many cloud services. To be able to communicate between the database and the Flask application, we will create a component called Database Model.
The Database Model is a class written in Python that implements all the necessary functions to communicate with the database. It builds the SQL statements needed to create, read, update and delete any element in the database. This way the connection and communication with the database is abstracted from the other components, leaving a cleaner and more maintainable code.
points and user management, it will be in charge of loading the models and the negation system, as well as connecting the database with the application and displaying the diferent views that will be available in the web.
Regarding the security aspect of our application, with Flask we can make use of diferent libraries to implement forms, data validators, session data, login, logout and CSRF protection: • Flask-Login: Flask-Login is a Python library for
Flask web applications that enables user authentication and authorization. This library helps in managing user sessions, securing user data, and restricting access to specific parts of the appli- 3.4. API security cation for authenticated users only. Flask-Login allows defining a User model that represents the To ensure the API security, we need to implement an users in the application and implementing login authentication system in the web application and link and registration views to handle user authenti- it to the API requests, so that every call is identified. In cation. It also enables protecting views and end- each call to an endpoint where you need identification, a points by requiring users to log in before access- JSON Web Token must be added in the request header. ing them. • Flask-WTF: Flask-WTF is a Python library that 3.4.1. JSON Web Tokens simplifies the process of creating and validating A JSON Web Token (JWT) is a secure method of transmitweb forms in Flask applications. With Flask-WTF, ting information a JSON object between parties in web forms can be created using Python classes and applications. It is a compact and self-contained format rendered in HTML templates. It includes various consisting of three parts: header, payload, and signature. form fields such as text fields, checkboxes, radio The header includes information about the type of token buttons, and dropdown menus, among others, and algorithm used to sign it, while the payload contains and supports form validation that helps prevent the claims or information that the token carries, such as common errors like missing required fields and user ID or role. The signature is a cryptographic signainvalid data types. Flask-WTF’s built-in security ture that verifies the token’s integrity and ensures it has features protect against CSRF attacks and enable not been tampered with. The stateless nature of JWTs secure file uploads. means that NegesAPI do not need to track user sessions or login credentials and can rely on JWTs to validate a
NegesAPI implements JWTs as an API Key, provided As database management system we will use MariaDB. in the /profile page of each user, with an expiration date MariaDB Server is a widely used open source relational of 24 hours and the option to regenerate a new one. This can easily be implemented with the jwt Python module, and starting each authorized request with the header “Authorization: Bearer userJwt”, being userJwt the user’s token. and debugging, and create client code in various programming languages. Swagger uses a machine-readable format called OpenAPI Specification, which describes the structure of the API, including endpoints, parameters, request/response formats, and authentication mechanisms.
The code executed in the client-side, normally interpreted by web navigators, is called front-end. It is the responsible to provide an interface with which the user can interact. NegesAPI ofers a simple and intuitive interface, with OpenAPI 3.0.0 documentation standards. In addition, it includes a test environment for the negation analyzer, a user profile tab and a tutorial on how to use the tool.
use test environment to try the diferent endpoints, the mandatory data, formats and response codes, without writing any code or command. This page is perfect to try the diferent output formats and it also provides a cURL example of how to create this request.
This documentation page has been generated using a Swagger tool, named Swagger Editor. Swagger is an open-source set of tools that provides a standardized way of documenting RESTful APIs. It comes with tools that can generate API documentation, perform API testing
such as name, email, password and API key. Here we can display, copy and regenerate our API key, change the user’s password and delete the account.
The API Key is the JWT generated in the back-end, necessary to make API requests that needs identification.
It also adds code snippets, output examples and descriptions of the diferent analyzer options, negation annotations and output formats.
match the latest improvements in FreeLing 4.2, since we use this tool as a feature extraction method. This corpus used a deprecated tag set generated with FreeLing 4.0 [9, 10], which needed to be updated in order to make the negation system work.
The negation detection system models the negation processing as a sequence labelling task, for that reason, it makes use of a CRF algorithm [11], as it has shown its efectiveness in this task [ 12, 13, 14]. This is because CRF makes predictions based on the elements in the sequence, not only in the current one, and negation cues and scopes behaves the same way. The results of its efectiveness using 10-fold cross validation are: i) Negation cues detection (Precision=92.70, Recall=82.09, F1-score=87.07), ii) Scope identification (Precision=90.77, Recall=63.64, F1score=74.79).
More information on the corpus and the system development can be found in [15].
NegesAPI is a powerful API that can assist diferent tasks in NLP and can be easily implemented in multiples scenarios, providing a fast and precise prediction of the negation cues and scopes in any Spanish text. In future works, we can improve NegesAPI so that it will add the negation cues in the scope, prefixing the negator to the negation word. It is also important that NegesAPI can be used in any type of algorithm and technology, therefore a future work would be to adapt the output suitable for Deep Learning algorithms. In addition, we plan to test its accuracy in diferent types of texts (reviews, posts, tweets, clinical texts, etc.). NegesAPI can provide a lot of relevant information to information retrieval systems, to detect negation in the searches that can be made, for better efectiveness. This knowledge could be used, for example, to distinguish "Películas que sean de fantasía" from "Películas que no sean de fantasía", or any other search with negation.
CONSENSO (PID2021-122263OB-C21), Project MODERATES (TED2021-130145B-I00) and Project SocialTox (PDC2022-133146-C21) funded by MCIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR, Project PRECOM (SUBV-00016) funded by the Ministry of Consumer Afairs of the Spanish Government, Project FedDAP (PID2020-116118GA-I00) supported by MICINN/AEI/10.13039/501100011033 and WeLee project (1380939, FEDER Andalucía 2014-2020) funded by the Andalusian Regional Government. Salud María Jiménez-Zafra has been partially supported by a grant from Fondo Social Europeo and the Administration of the Junta de Andalucía (DOC_01073). [6] F. Fancellu, A. Lopez, B. Webber, Neural networks for negation scope detection, in: Proceedings of the 54th annual meeting of the Association for Computational Linguistics (volume 1: long papers), 2016, pp. 495–504. [7] O. S. Pabón, O. Montenegro, M. Torrente, A. R.
González, M. Provencio, E. Menasalvas, Negation and uncertainty detection in clinical texts written in spanish: a deep learning-based approach, PeerJ
Computer Science 8 (2022) e913. [8] S. M. Jiménez-Zafra, M. Taulé, M. T. Martín-Valdivia,
L. A. Urena-López, M. A. Martí, Sfu review spneg: a spanish corpus annotated with negation for sentiment analysis. a typology of negation patterns, Language Resources and Evaluation 52 (2018) 533– 569. [9] L. Padró, E. Stanilovsky, Freeling 3.0: Towards
wider multilinguality, in: LREC2012, 2012. [10] M. Marimon, L. Padró, J. Turmo Borras, Coreference resolution in freeling 4.0, in: LREC 2018: 11th International Conference on Language, Resources and Evaluation: Miyazaki, Japan: May 7-12, 2018: proceedings book, 2018, pp. 376–381. [11] J. Laferty, A. McCallum, F. C. Pereira, Conditional random fields: Probabilistic models for segmenting and labeling sequence data (2001). [1] S. M. Jiménez-Zafra, R. Morante, M. T. Martín- [12] I. Councill, R. McDonald, L. Velikovich, What’s Valdivia, L. A. U. Lopez, Corpora annotated with great and what’s not: learning to classify the scope negation: An overview, Computational Linguistics of negation for improved sentiment analysis (2010). 46 (2020) 1–52. [13] H. Loharja, L. Padró, J. Turmo Borras, Negation [2] J. C. de Albornoz, L. Plaza, A. Díaz, M. Ballesteros, cues detection using crf on spanish product review Ucm-i: A rule-based syntactic approach for resolv- texts, in: NEGES 2018: Workshop on Negation ing the scope of negation, in: * SEM 2012: The First in Spanish: Seville, Spain: September 19-21, 2018: Joint Conference on Lexical and Computational proceedings book, 2018, pp. 49–54. Semantics–Volume 1: Proceedings of the main con- [14] L. Domınguez-Mas, F. Ronzano, L. I. Furlong, Superference and the shared task, and Volume 2: Pro- vised learning approaches to detect negation cues ceedings of the Sixth International Workshop on in spanish reviews, in: Proceedings of the Iberian Semantic Evaluation (SemEval 2012), 2012, pp. 282– Languages Evaluation Forum (IberLEF 2019), CEUR 287. Workshop Proceedings, Bilbao, Spain. CEURWS, [3] Y. Peng, X. Wang, L. Lu, M. Bagheri, R. Summers, 2019.
Z. Lu, Negbio: a high-performance tool for nega- [15] J. Valle-Aguilera, S. M. Jiménez-Zafra, M. T. Martíntion and uncertainty detection in radiology reports, Valdivia, L. A. Ureña-López, NegesAPI: An API for AMIA Summits on Translational Science Proceed- Negation Cue Detection and Scope Identification ings 2018 (2018) 188. in Spanish, Procesamiento del Lenguaje Natural 71 [4] N. P. Cruz, M. Taboada, R. Mitkov, A machine- (2023).
learning approach to negation and speculation detection for sentiment analysis, Journal of the Association for Information Science and Technology 67 (2016) 2118–2136. [5] S. M. Jiménez-Zafra, R. Morante, E. Blanco, M.-T.
Martín-Valdivia, L. A. U. López, Detecting negation cues and scopes in spanish, in: Proceedings of the Twelfth Language Resources and Evaluation Conference, 2020, pp. 6902–6911.