Bifid: Parallel corpus alignment at the document, sentence and vocabulary levels

Bifid is a program for parallel corpora alignment:

Web demo: http://www.bifidalign.com/

State of this project on January 17, 2021:

DSELE: a dictionary of Spanish verbs with 'se'

Web demo: http://www.tecling.com/dsele

DSELE is the proposed model of a Dictionary verbs with "se" for ELE. A sample of verbs is available for online consultation. It is based on an analysis of 13,754 total corpus concordances, which results in the analysis of 273 usage patterns (CPA patterns). Such patterns are also available online. In the model, concordances are used as examples, and patterns are transformed into definitions. The verbal entry of the proposed model gives visibility to the grammatical information and not just semantics, and takes into account the degree of information to which the student has access. Therefore, an initial index incorporates a central column which is largely semantic information and a column to the right where the most grammatical information is situated.

Related publications:

+ Renau, I. (2014). Entre el léxico y la gramática: diccionarios de ELE para el aprendizaje de los verbos pronominales. XXV Congreso Internacional de la ASELE. La enseñanza de ELE centrada en el alumno. Madrid, 2014.

Contact: irene.renau at gmail.com

EMaD:

Introducción

Emad es una herramienta que permite encontrar y clasificar elementos lingüísticos que, en un texto dado, funcionan como marcadores discursivos. Su implementación computacional se basa en los resultados (provisorios) de un proyecto de investigación en curso en lingüística computacional que tiene por objetivo la creación de un detector y clasificador automático de marcadores discursivos (Robledo, Nazar y Renau, 2017; Robledo y Nazar, 2018). La particularidad de este proyecto es que en él se recurre a una metodología totalmente inductiva, basada en datos obtenidos de grandes corpus textuales para la extracción y la clasificación de elemententos lingüísticos que pueden funcionar como marcadores del discurso.

Demo Web: http://www.tecling.com/cgi-bin/emad

La interfaz está en línea desde agosto 2018 y funciona solamente en castellano.

Metodología de clasificación

La metodología de extracción y clasificación de marcadores del discurso está basada en datos de corpus paralelos alineados a nivel de oración. La clasificación que aquí se propone se origina a partir de un proceso de agrupación por clusters jerárquicos (Rokach y Maimon, 2005; Dixon, 2003). El clustering es algoritmo estadístico que agrupa automáticamente los elementos, en virtud de la similitud entre sus componentes. El objetivo del clustering es, por lo tanto, identificar particiones en un conjunto no estructurado de objetos descritos según ciertos atributos (clasificación no supervisada). Esta identificación se basa solo en estos atributos y no requiere de ningún dato anotado.

Corpus de referencia de la clasificación

En esta investigación se ha recurrido a los datos paralelos bilingües proporcionados por el OPUS Corpus (Tiedemann, 2012, 2016). En este caso, se seleccionaron los textos paralelos español-inglés e inglés-español, fundamentalmente, debido a la disponibilidad de los datos: 1,1 mil millones de tokens en total para estos bitextos. El formato elegido es el tmx (memoria de traducción), alineado a nivel de oración. El proyecto OPUS Corpus (Tiedemann, 2012, 2016) proporciona conjuntos de datos paralelos que cubren varios dominios y están disponibles online de forma pública y gratuita (http://opus.nlpl.eu/).

Procesamiento de los datos: experimentos en lengua española

El procesamiento de los datos se inicia con la descarga del corpus paralelo español-inglés de la web del Opus Corpus y termina con la formación de los clusters o conjuntos de elementos aglomerados a partir de procedimiento de clustering jerárquico aglomerativo. Estos clusters preliminares darán origen a las categorías de marcadores discursivos que se proponen en este trabajo. Una vez descargados los archivos en formato tmx, se procede a su lectura automática implementando instrucciones de búsqueda a través de expresiones regulares y se extraen todos los segmentos textuales que ocurren entre signos de puntuación en cada lengua. Una vez aplicados una serie de filtros para seleccionar los segmentos textuales candidatos a marcadores discursivos (número de subcorpus en que ocurre una expresión, eliminación de nombres comunes), se cuenta con dos listados independientes de candidatos a marcadores discursivos en ambas lenguas. Se procede a alinear las unidades léxicas de estos listados en ambas lenguas a partir del cálculo del grado de asociación estadística en el corpus para obtener un conjunto de marcadores en inglés correspondientes a cada marcador discursivo en español de la lista. Por ejemplo, para el marcador en español "no obstante" se obtuvo un conjunto de marcadores correspondientes en inglés, ordenados según el grado de asociación estadística entre las ocurrencias de ellos en el corpus paralelo: however, nevertheless, though, nonetheless, entre otros.

Luego, de manera análoga, se procede a realizar el mismo procedimiento pero partiendo de los marcadores en inglés. De manera que, por cada marcador en inglés se obtiene un conjunto de marcadores correspondientes en español, ordenados según el grado de asociación estadística de sus coocurrencias en el corpus paralelo. El resultado es la obtención de datos en lengua española ordenados en una tabla que contiene, para cada marcador en español, un listado de otros marcadores en la misma lengua que cumplen una función similar en el corpus. A partir de estos datos se generan un total de {número de marcadores en español} vectores binarios compuestos por {número de marcadores en español} atributos, donde cada uno de ellos se rellenará con un 1 o un 0, dependiendo de si la coordenada del vector corresponde a o no a un marcador que está en la lista de atributos del marcador representado por ese vector. Estos datos binarios se procesan luego con un software estadístico para generar aglomeraciones de marcadores discursivo agrupados en virtud de su similitud y de la diferencia con los demás elementos. Para esto se usa el coeficiente de Jaccard como medida de similitud entre los vectores binarios y el algoritmo de Ward o el método de la media para obtener las aglomeraciones o clusters. Este método permite seleccionar el número de clusters finales que, para este elemento fue de 100. Finalmente, estos clusters son evaluados manualmente y etiquetados según nombres descriptivos extraídos de la literatura sobre marcadores discursivos en español, por ejemplo, "contraargumentativos", "causales", "consecutivos", entre otros.

En una etapa posterior, se pretende examinar el fenómeno de la polifuncionalidad de los marcadores discursivos en el corpus de estudio. Con ello, se espera obtener la adscripción de un marcador del discurso a una o más categorías, según si el elemento cumple o no más de una función en el corpus.

Publicaciones relacionadas:

+ Robledo, H.; Nazar, R. (2018). "Una clasificación automatizada de marcadores discursivos", Procesamiento del Lenguaje Natural, n. 61, pp 109-116.

Concepts relacionados:

Contact: rogelio.nazar at gmail.com

Estilector:

This proposal is aimed at improving academic writing skills of students through the creation, development and implementation of a web tool that assists in detecting these problems of style that can be found in drafting academic work. It offers additional explanations, bibliographic support and online resources. The tool is not intended to correct grammatical or spelling errors, but those problems such as repeating words close in the text, poor vocabulary, the use of colloquialisms, the unequal structure of paragraphs, and so on. All these issues cannot be detected by programs such as Word, and yet they are critical to academic achievement. Our proposal is not to create a merely "corrector", but a teaching tool that fosters independent learning because the student can work on these aspects independently of the work of the classroom, albeit also complementary. The idea is that the tool will help students improve their writing during the process of performing the task. In addition, the program also encourages autonomy in the sense that it suggests solutions to the student, but does not correct the text, so that it is the student who ultimately decides whether or not the suggested changes apply.

Web demo: http://www.estilector.com/

Contact: rogelio (dot) nazar (at) gmail.com

GeNom:

GeNom: automatic detection of the gender of proper names is a project we have been granted on June 20, 2017, funded by the Technology Prototypes track of the Innovation and Entrepreneurship 2017 Competition (Vicerrectoría de Investigación y Estudios Avanzados - Pontificia Universidad Católica de Valparaíso). The result is offered as a web service for batch processing of information for terminography or lexicography projects or for mailing purposes.

Abstract: This software is designed to automatically determine the gender of a list of names based on their co-occurrence with words and abbreviations in a large corpus. GeNom is different from other forms of automatic name gender recognition software because it is based on natural language processing and does not rely on already compiled lists of first names, systems that get quickly outdated and cannot analyze previously unseen names. GeNom uses corpora to address the problem, because it offers the possibility of obtaining real and up-to-date name-gender links and performs better than machine learning methods: 93% precision and 88% recall on a database of ca. 10,000 mixed names. This software can be used to conduct large scale studies about gender, as gender bias for example, or for a variety of other NLP tasks, such as information extraction, machine translation, anaphora resolution and others. It is designed to work with Spanish names, as it works with a Spanish corpus, but it will be able to process names in other languages as well, provided that they use the same alphabet.

Web demo: http://www.tecling.com/genom

The interface is at the moment only in Spanish.

Contact: rogelio.nazar (imagine the 'at' symbol here) gmail.com

Jaguar:

Jaguar is a tool for corpus exploitation. This software can analyze textual corpora from a user or from the web and it is currently available as a web application as well as a Perl module. The functions that are available at this moment are: vocabulary analysis of corpora, concordance extractions, n-gram sorting and measures of association, distribution and similarity.

Jaguar is essentially a Perl module instantiated as a web application. A web application has the advantage of being executable in any platform without installation procedures. However, with the module users are capable of building their own sequence of procedures, taking the output of a process to be the input of another process. The web interface has the limitation that only one procedure can be executed at a time, meaning that the output of a process has to be manually fed as input for the next process.

Since July 2016, this project is funded by the Innovation and Entrepeneurship 2016 Program of Pontificia Universidad Católica de Valparaíso, within the "Technological Prototyes" track.

The project is a full renovation and extension of the old "Jaguar Project" carried out at Universitat Pompeu Fabra in Barcelona from 2006 to 2012. The title of the current project is: "Jaguar: an open-source prototype for quantitative corpus analysis"

The results of this project will be officialy presented in January 2017 at the university headquarters, in Av. Brasil #2950, Valparaíso, Chile.

We are also planning to offer an introductory Workshop on the use of this tool in the summer of 2017, maybe in Valparaíso, maybe in Santiago, or maybe in both places. Drop a line if interested.

Web demo: http://www.tecling.com/jaguar

Related publications:

+ Nazar, R.; Vivaldi, J.; Cabré, MT. (2008). A Suite to Compile and Analyze an LSP Corpus. Proceedings of LREC 2008 (The 6th edition of the Language Resources and Evaluation Conference) Marrakech (Morocco), May 28-30, 2008.

We are preparing a new paper to describe the new version of the program.

Related concepts: corpus exploitation, concordances, n-grams, measures of association, distribution and similarity

Contact: rogelio.nazar at gmail.com

KIND (aka The Taxonomy Project)

We designed a statistically-based taxonomy induction algorithm consisting of a combination of different strategies not involving explicit linguistic knowledge. Being all quantitative, the strategies we present are however of different nature. Some of them are based on the computation of distributional similarity coefficients which identify pairs of sibling words or co-hyponyms, while others are based on asymmetric co-occurrence and identify pairs of parent-child words or hypernym-hyponym relations. A decision making process is then applied to combine the results of the previous steps, and finally connect lexical units to a basic structure containing the most general categories of the language. We evaluate the quality of the taxonomy both manually and also using Spanish Wordnet as a gold-standard. We estimate an average of 89.07% precision and 25.49% recall considering only the results which the algorithm presents with high degree of certainty, or 77.86% precision and 33.72% recall considering all results.

Website: http://www.tecling.com/kind

Funding: This research is supported by a grant from the Chilean Government: Conicyt-Fondecyt 11140686, “Inducción automática de taxonomías de sustantivos generales y especializados a partir de corpus textuales desde el enfoque de la lingüística cuantitativa” (Automatic taxonomy induction from corpora for terminology and general vocabulary using quantitative measures). Lead researcher: Rogelio Nazar.

Related publications:

+ Nazar, R.; Balvet, A.; Ferraro, G.; Marín, R.; Renau, I. (2020). "Pruning and repopulating a lexical taxonomy: experiments in Spanish, English and French". Journal of Intelligent Systems, vol. 30 num. 1, pp. 376-394.

+ Nazar, R.; Obreque, J.; Renau, I. (2020). "Tarántula –> araña –> animal : asignación de hiperónimos de segundo nivel basada en métodos de similitud distribucional". Procesamiento del Lenguaje Natural, núm 64, pp. 29-36.

+ Nazar, R.; Soto, R.; Urrejola, K. (2017). Detección automática de nombres eventivos no deverbales en castellano: un enfoque cuantitativo basado en corpus. Revista Linguamatica, vol. 9, num. 2, pp. 21-31.

+ Nazar, R.; Arriagada, P. (2017). POL: un nuevo sistema para la detección y clasificación de nombres propios. Procesamiento del Lenguaje Natural, n. 58, pp. 13-20.

+ Nazar, R., Renau, I. (2016). “A Quantitative analysis of the semantics of verb-argument structures”. In S. Torner and E. Bernal (eds.), Collocations and other lexical combinations in Spanish. Theoretical and Applied approaches. New York: Routledge, pp. 92-109.

+ Nazar, R. (2016). Distributional analysis applied to terminology extraction: example in the domain of psychiatry in Spanish. Terminology: International Journal of Theoretical and Applied Issues in Specialized Communication, 22(2):142-170.

+ Nazar, R.; Renau, I. (2016). A taxonomy of Spanish nouns, a statistical algorithm to generate it and its implementation in open source code. Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC'16). European Language Resources Association (ELRA), May 2016.

+ Nazar, R.; Renau, I. (2016). Automatic extraction of lexico-semantic patterns from corpora. Proceedings of the XVII EURALEX International Congress: Lexicography and Linguistic Diversity. Tinatin Margalitadze and George Meladze (eds). Tbilisi, Gergia: Ivane Javakhishvili Tbilisi State University, pp. 823-830.

+ Nazar, R.; Renau, I. (2015). Agrupación semántica de sustantivos basada en similitud distribucional: implicaciones lexicográficas. In María Pilar Garcés Gómez (ed.): "Lingüística y diccionarios" (Anexos Revista de Lexicografía, vol. 2: 272-295). Universidade da Coruña.

+ Nazar, R.; Renau, I. (2015). Ontology Population Using Corpus Statistics. Proceedings of the Joint Ontology Workshops 2015 co-located with the 24th International Joint Conference on Artificial Intelligence (IJCAI 2015). Buenos Aires, Argentina, July 25-27, 2015.

Related concepts: corpus statistics, distributional semantics, Spanish, taxonomy induction

KWiCo:

This project is part (or a "spin-off") of the Perl module Jaguar, which is currently ongoing with funding from the Innovation and Entrepeneurship 2016 Program of Pontificia Universidad Católica de Valparaíso, within the "Technological Prototyes" track.

KWiCo is a corpus indexing algorithm. It takes a corpus as input and produces a table with an index of the corpus, thus significantly reducing the time needed to retrieve concordances, especially when the corpus is very large.

Web demo: http://www.tecling.com/kwico

Source code and documentation: http://www.tecling.com/index.php?l=kwico_source Comments within the same scripts are at the moment only in Spanish but we are working in their translation to English.

NEVEN

We present a study in the field of the automatic detection of non-deverbal eventive nouns, which are those nouns that designate events but have not experienced a process of derivation from verbs, such as fiesta (‘party’) or cóctel (‘cocktail’) and, for this reason, do not present the typical morphological features of deverbal nouns, such as -ci´on, -miento, and are therefore more difficult to detect. In the present research we continue and extend the work initiated by Resnik (2010), who offers a number of cues for the detection of this type of lexical unit. We apply Resnik’s ideas and we also add new ones, among them, the inductive analysis of the words that tend to co-occur with eventive nouns in corpora, in order to use them as predictors of this condition. Furthermore, we simplify the classification algorithm considerably, and we apply the experiments to a larger corpus, the EsTenTen (Kilgarriff & Renau, 2013), comprising more than 9 billion running words. Finally, we present the first results of the automatic extraction of eventive nouns from the corpus, among which we find plenty non-deverbal nouns.

Web demo: http://www.tecling.com/neven

Source code: http://www.tecling.com/neven/neven.pl

Usage:

perl neven.pl input.txt > result.htm

Beforehand, you need the contexts of occurrence of a word extracted from the corpus. But you will need to edit the script in order to set the right path to the folder where the contexts are stored. These concordances are stored in a file bearing the same name of the word's lemma. You can obtain these concordances from any corpus using our free corpus concordancer Kwico.
Comments in the script are at the moment only in Spanish.

Pending Work: Users interested only in non-deverbal eventive nouns will need a few changes in the script que filter out those nouns having deverbal morphology (e.g. -ción, -miento). What is interesting about this program is that it completely ignores such morphological cues. The morphology filter is a safe and simple method and will be ready soon.

Funding: This research is supported by a grant from the Chilean Government: Conicyt-Fondecyt 11140686, “Inducción automática de taxonomías de sustantivos generales y especializados a partir de corpus textuales desde el enfoque de la lingüística cuantitativa” (Automatic taxonomy induction from corpora for terminology and general vocabulary using quantitative measures). Lead researcher: Rogelio Nazar.

Related publications: Nazar, R.; Soto, R.; Urrejola, K. (2017). Detección automática de nombres eventivos no deverbales en castellano: un enfoque cuantitativo basado en corpus. Revista Linguamatica, vol. 9, num. 2, pp. 21-31.

Related concepts: computacional lexicography, inductive corpus analysis, non-deverbal eventive nouns

Questions or comments? Feel free to drop a line.

Termout:

Termout.org is the first implementation of a new method for terminology extraction based on distributional analysis. The intuition behind the algorithm is that single or multi-word lexical units that refer to specialised concepts will show a characteristic co-occurrence pattern, described as a tendency to appear in the same contexts with other conceptually related terms. E.g. the term fluoxetine will systematically appear in the same sentences with other related terms such as depression, serotonin reuptake inhibitor, obsessive–compulsive disorder and others. Of course, terms will co-occur with general vocabulary units as well, but not with a characteristic pattern as when a conceptual relation holds. Experimental evaluation of this method was conducted in a corpus of psychiatry journals from Spain and Latin America, and concluded that the results are significantly better than other methods.

Web demo: http://www.termout.org/

A new version of the web interface has been published online today 8 September, 2018.

Related publications:

+ Nazar, R. (2016). Distributional analysis applied to terminology extraction: example in the domain of psychiatry in Spanish. Terminology: International Journal of Theoretical and Applied Issues in Specialized Communication, 22(2):142-170.

Related concepts: co-occurrence, distributional semantics, terminology extraction, topic signatures, text-mining

Contact: rogelio.nazar at gmail.com

POL

The purpose of this research is to develop a methodology for the detection and categorisation of named entities or proper names (PPNN), in the categories of geographical place, person and organisation. The hypothesis is that the context of occurrence of the entity –a context window of n words before the target– as well as the components of the PN itself may provide good estimators of the type of PN. To that end, we developed a supervised categorisation algorithm, with a training phase in which the system receives a corpus already annotated by another NERC system. In the case of these experiments, such system was the open-source suite of language analysers FreeLing, annotating the corpus of the Spanish Wikipedia. During this training phase, the system learns to associate the category of entity with words of the context as well as those from the PN itself. We evaluate results with the CONLL- 2002 and also with a corpus of geopolitics from the journal Le Monde Diplomatique in its Spanish edition, and compare the results with some well-known NERC systems for Spanish.

Web demo: http://www.tecling.com/pol

Source code: http://www.tecling.com/pol/source/sourcePol.zip

It contains:

• config.pm: Configuration file. The user needs to adjust its values before execution.
• poltrain.pl: Script used for training.
• pol.pl: Script used for the actual processing of new data.
• convertmodel.pl: Script used to convert the model produced by poltrain.pl to the model that pol.pl needs to work.

Corpus and models: experiments have only been conducted in Spanish for the moment. Models for new languages will be added in the future. If you would like to help, you are welcome.

• WikipediaFreeling.zip (2,6Gb !!!). This is the training corpus, a Spanish Wikipedia tagged with Freeling.
• Model.zip: An example of model produced after training and conversion, ready to be used with pol.pl.

These models were created with a x86_64 HP Proliant machine with GenuineIntel CPU 1064.000 MHz running Linux (Ubuntu 14.04). If you have a different kind of machine (e.g., a Desktop pc on Windows), then you will probably need to create the models again by using poltrain.pl.

Funding: This research is part of the Fondecyt Project 11140686: “Inducción automática de taxonomías de sustantivos generales y especializados a partir de corpus textuales desde el enfoque de la lingüística cuantitativa” (Automatic taxonomy induction from corpora for terminology and general vocabulary using quantitative measures). Lead researcher: Rogelio Nazar.

Related publications:

+ Nazar, R.; Arriagada, P. (2017). POL: un nuevo sistema para la detección y clasificación de nombres propios. Procesamiento del Lenguaje Natural, n. 58, pp. 13-20.

Related concepts: Named entities, proper names, text linguistics

Poppins:

Poppins a very simple and yet effective algorithm for document categorization. Text categorization has became a very popular issue in computational linguistics and it has developed to great complexity, motivating a large amount of literature. Document categorization can be used in many scenarios. For instance, an experiment on authorship attribution can be seen as a text categorization problem. That is to say, each author represents a category and the documents are the elements to be classified. This system can be used as a general purpose document classifier, for example by content instead of authorship, because it only reproduces the criterion that it learned during the training phase. This program is language independent because it uses purely mathematical knowledge: an n-gram model of texts. It works in a very simple way and is therefore easy to modify. In spite of its simplicity, this program is capable of classifying documents by author obtaining more than 90% of accuracy.

Web demo: http://poppinsweb.com/

Document related with this project:

• Nazar, R & Sánchez Pol, M. (2006). "An Extremely Simple Authorship Attribution System", (PDF),
  Proceedings of the Second European IAFL Conference on Forensic Linguistics / Language and the Law, Barcelona 2006.
  
  Contact: rogelio.nazar at gmail.com
  
  

Verbario:

Verbario is our first attempt to extract lexical patterns using corpus statistics. A pattern is a structure that combines syntactic and semantic features and is linked to a conventional meaning of a word. This means, for example, that the verb to die does not have intrinsic meanings, but potential meanings which are activated by the context: in ‘His mother died when he was five’, the meaning of the verb differs from ‘His mother is dying to meet you’, due to collocational restrictions and syntactic differences. With the automatic analysis of thousands of concordances per verb, we can make a first approach to the problem of detecting these structures in corpora, a very time-consuming task for lexicographers. The average precision is around 50%. The next step to increase precision is adding a dependency parser to the system and make adjustments to the automatic taxonomy we have created for semantic labeling.

Web demo: http://www.verbario.com/

Funding: This research is supported by a grant from the Chilean Government: Conicyt-Fondecyt 11140704, “Detección automática del significado de los verbos del castellano por medio de patrones sintáctico-semánticos extraídos con estadística de corpus” (Automatic Extraction of patterns of use of Spanish verbs using corpus statistics). Lead researcher: Irene Renau.

Related publications:

+ Nazar, R.; Renau, I. (forthcoming). A Quantitative Analysis of the Semantics Of Verb-Argument Structures. In S. Torner and E. Bernal (eds.) "Collocations and other lexical combinations in Spanish. Theoretical and Applied approaches", Routledge.

+ Nazar, R.; Renau, I. (forthcoming). Automatic extraction of lexico-semantic patterns from corpora. Proceedings of EURALEX 2016. Tbilisi, Georgia.

Related concepts: computational lexicography; lexical patterns; Spanish verbs; taxonomy

Contact: irene.renau at gmail.com