Search

With the advances of the Natural Language Processing area, corpora are resources that have had a prominent place. More than subsidizing linguistic studies, they constitute the basis for training Machine Learning models and developing cutting-edge computational applications. In particular, there is a great need for annotated corpora, but their production requires another essential resource, the annotation manual, which instantiates the annotation model of interest for the language in question and outlines the annotation decisions that should be adopted. In this paper, we explore issues related to the development of manuals for the annotation of Brazilian Portuguese corpora according to the Universal Dependencies model, widely adopted in the field. We discuss the evolution of NLP and the use of corpora, the fundamental issues, resources and tools related to syntactic representation, the Universal Dependencies model, and the main decisions made in the instantiation of UD guidelines in Brazilian Portuguese. For practical and didactic reasons, we divided the manual into two parts: the PoS Tag Annotation Manual (morphosyntactic annotation) and the Dependency Relations Annotation Manual. Both resulted from the process reported in this paper and are available for free access on the POeTiSA project's Web site. Com o avanço da área de Processamento de Linguagem Natural (PLN), corpora são recursos que têm tido um lugar de destaque. Mais do que subsidiar estudos linguísticos, eles constituem as bases para o treinamento de modelos de Aprendizagem de Máquina e para o desenvolvimento de aplicações computacionais de ponta. Particularmente, há grande necessidade de corpora anotados, porém sua geração requer outro recurso essencial, o manual de anotação, que instancia o modelo de anotação de interesse para a língua em questão e delineia as decisões de anotação que devem ser adotadas. Neste artigo, exploramos questões relacionadas ao desenvolvimento de manuais para a anotação de corpus em português brasileiro segundo o modelo internacional Universal Dependencies, amplamente adotado na área. Partimos da discussão da evolução do PLN e o uso de corpora, passamos pelas questões, recursos e ferramentas fundamentais relacionados à representação sintática, discutimos o modelo Universal Dependencies e apresentamos as principais decisões tomadas na instanciação de suas diretrizes no português brasileiro. Por questões práticas e de didática, dividimos o manual em duas partes: o Manual de Anotação de PoS tags (anotação morfossintática) e o Manual de Anotação Relações de Dependência. Ambos foram resultado do processo relatado neste artigo e estão disponíveis para livre acesso no site do projeto POeTiSA na Web.

В статье представлена методология извлечения морфологических признаков технических систем в виде компонентов устройства и связей между ними. Объектом анализа для извлечения данных выступает главный пункт формулы изобретения в текстах русскоязычных патентов. Информация о компонентах устройства является наиболее фундаментальной и важной и может использоваться во множестве задач анализа патентного массива, а поиск эффективных подходов по извлечению такой информации все еще продолжается. В настоящем исследовании областью применения указанных данных рассматривается направление автоматизированного изобретательства. Целью работы являлся анализ качества извлечения данных по деревьям зависимостей для русского языка. Деревья зависимостей являются результатом работы систем синтаксической разметки естественного языка. Для сравнения были выбраны следующие синтаксические анализаторы: UdPipe, Stanza, DeepPavlov и spaCy. Выходные данные представлены в виде семантических структур SAO (Subject-Action-Object). Дана оценка качества извлечения данных с помощью метрик точности (precision), полноты (recall) и F1-меры. Для этого вручную было размечено 20 патентных формул с 252 структурами SAO. При текущих методологических ограничениях из тестовой выборки в лучшем случае удалось извлечь 79 % связок SAO в терминах метрики recall при нестрогой оценке данных, т. е. без учета полноты именных групп субъекта и объекта. Значение F1-меры по инструментам несколько ниже и находится в пределах от 48 % до 66 % в зависимости от типа оценки. Сделаны общие выводы по текущему уровню работы синтаксических анализаторов в рамках исследуемой области применения. Материалы статьи представляют практическую ценность при проработке эффективных подходов извлечения структурированных данных из русскоязычного патентного массива., The article presents a methodology for extracting morphological features of technical systems in the form of device components and connections between them. The main section of Russian patents claims is chosen as the subject of the study for data extraction. Information about device components is the most fundamental and important part. It can be used in many tasks of computer-aided patent analysis, while the search for effective approaches to extracting such information is still in progress. In the present inquiry, computer-aided development of inventions is considered as a range of applications for such data. The aim of the study was to explore the quality of data extraction using dependency tree analysis for Russian language. The dependency tree is the result of markup by natural language processing tools. Several parsers were chosen for the comparison: UdPipe, Stanza, DeepPavlov and spaCy. The output data are presented in the form of semantic SAO (Subject-Action-Object) structures. The quality of data extraction has been evaluated using precision, recall and F1 metrics. For this purpose, 20 patent claims with 252 SAO structures were manually marked. Under the current methodological constraints, we were able to extract from the dataset 79 % of the SAO structures at best according to the recall metric with a non-strict data evaluation, i.e. without accounting for the completeness of noun groups. The value of F1-measure is lower and ranges from 48 % to 66 % depending on the evaluation type. Conclusions are drawn about the current level of the syntactic analyzer performance within the field of application under review. The results can be useful for developing efficient approaches to extracting structured data from Russian patent arrays., МОДЕЛИРОВАНИЕ, ОПТИМИЗАЦИЯ И ИНФОРМАЦИОННЫЕ ТЕХНОЛОГИИ, Выпуск 4 (39) 2022, Pages 6-7

In this paper, we aim to introduce the dependency annotation process of the largest and the only cross-linguistic Turkish dependency treebank which was translated from the original Penn Treebank corpus. Within the scope of this project, 16.400 sentences have been morphologically and semantically annotated, and the dependency relations were manually carried out by a team of linguists. It is hoped that this project will serve as a base for a successful dependency parser and a system which can automatically perform the bi-directional conversion between constituency and dependency trees. Publisher's Version

This report describes a quantitative analysis performed to compare two different methods on the task of pronoun resolution for Swedish. The first method, an implementation of Mitkov’s algorithm, is a heuristic-based method — meaning that the resolution is determined by a number of manually engineered rules regarding both syntactic and semantic information. The second method is data-driven — a Support Vector Machine (SVM) using dependency trees and word embeddings as features. Both methods are evaluated on an annotated corpus of Swedish news articles which was created as a part of this thesis. SVM-based methods significantly outperformed the implementation of Mitkov’s algorithm. The best performing SVM model relies on tree kernels applied to dependency trees. The model achieved an F1-score of 0.76 for the positive class and 0.9 for the negative class, where positives are pairs of pronoun and noun phrase that corefer, and negatives are pairs that do not corefer. Rapporten beskriver en kvantitativ analys som genomförts för att jämföra två olika metoder för automatisk pronomenbestämning på svenska. Den första metoden, en implementation av Mitkovs algoritm, är en heuristisk metod vilket innebär att pronomenbestämningen görs med ett antal manuellt utformade regler som avser att fånga både syntaktisk och semantisk information. Den andra metoden är datadriven, en stödvektormaskin (SVM) som använder dependensträd och ordvektorer som särdrag. Båda metoderna utvärderades med hjälp av en annoterad datamängd bestående av svenska nyhetsartiklar som skapats som en del av denna avhandling. Den datadrivna metoden överträffade Mitkovs algoritm. Den SVM-modell som ger bäst resultat bygger på trädkärnor som tillämpas på dependensträd. Modellen uppnådde ett F1-värde på 0.76 för den positiva klassen och 0.9 för den negativa klassen, där de positiva datapunkterna utgörs av ett par av pronomen och nominalfras som korefererar, och de negativa datapunkterna utgörs av par som inte korefererar.

This thesis studies the connections between parsing friendly representations and interlingua grammars developed for multilingual language generation. Parsing friendly representations refer to dependency tree representations that can be used for robust, accurate and scalable analysis of natural language text. Shared multilingual abstractions are central to both these representations. Universal Dependencies (UD) is a framework to develop cross-lingual representations, using dependency trees for multlingual representations. Similarly, Grammatical Framework (GF) is a framework for interlingual grammars, used to derive abstract syntax trees (ASTs) corresponding to sentences. The first half of this thesis explores the connections between the representations behind these two multilingual abstractions. The first study presents a conversion method from abstract syntax trees (ASTs) to dependency trees and present the mapping between the two abstractions – GF and UD – by applying the conversion from ASTs to UD. Experiments show that there is a lot of similarity behind these two abstractions and our method is used to bootstrap parallel UD treebanks for 31 languages. In the second study, we study the inverse problem i.e. converting UD trees to ASTs. This is motivated with the goal of helping GF-based interlingual translation by using dependency parsers as a robust front end instead of the parser used in GF. The second half of this thesis focuses on the topic of data augmentation for parsing – specifically using grammar-based backends for aiding in dependency parsing. We propose a generic method to generate synthetic UD treebanks using interlingua grammars and the methods developed in the first half. Results show that these synthetic treebanks are an alternative to develop parsing models, especially for under-resourced languages without much resources. This study is followed up by another study on out-of-vocabulary words (OOVs) – a more focused problem in parsing. OOVs pose an interesting problem in parser development and the method we present in this paper is a generic simplification that can act as a drop-in replacement for any symbolic parser. Our idea of replacing unknown words with known, similar words results in small but significant improvements in experiments using two parsers and for a range of 7 languages.

One of the main goals of service managers is to determine the strengths and weaknesses of the service delivery and to improve the service experience continually. Online feedback provided by customers offers service managers a valuable source of information to help achieve these goals. Despite the usefulness of reading such comments, service managers do not usually have sufficient resources to read all feedback thoroughly. Deriving concisely and focused managerially useful insights from online user content without wasting resources calls for an automated extraction technique. In this paper, we examine three different techniques for mining the service aspects from the same online data. The used techniques require a different degree of processing and provide a different level of information for service managers. The first technique retrieves nouns from the reviews, weights them and assesses the nature of an aspect, whether it is positive or negative, according to an overall user rating. The second technique counts co-occurrence statistics of nouns and adjectives and assesses them in the same way as the first technique. The third technique differs from the others in a deepness of text processing and its representation in a machine. A method parses grammatical dependencies between words and analyses the resulted trees by patterns. The third technique does not require rating scales to assess the nature of a service aspect. The grammatical dependencies involve sentence negation. So, the nature of an aspect can be extracted from the text directly. The data source for this paper is a major online review site that collects and presents customers’ opinions on different hotel and travel related services. The results show that the grammatical dependencies technique provides the most meaningful output but requires a careful extraction pattern setting. The outputs of the other two techniques are not ready for direct use to inform service managers.

One of the main goals of service managers is to determine the strengths and weaknesses of the service delivery and to improve the service experience continually. Online feedback provided by customers offers service managers a valuable source of information to help achieve these goals. Despite the usefulness of reading such comments, service managers do not usually have sufficient resources to read all feedback thoroughly. Deriving concisely and focused managerially useful insights from online user content without wasting resources calls for an automated extraction technique. In this paper, we examine three different techniques for mining the service aspects from the same online data. The used techniques require a different degree of processing and provide a different level of information for service managers. The first technique retrieves nouns from the reviews, weights them and assesses the nature of an aspect, whether it is positive or negative, according to an overall user rating. The second technique counts co-occurrence statistics of nouns and adjectives and assesses them in the same way as the first technique. The third technique differs from the others in a deepness of text processing and its representation in a machine. A method parses grammatical dependencies between words and analyses the resulted trees by patterns. The third technique does not require rating scales to assess the nature of a service aspect. The grammatical dependencies involve sentence negation. So, the nature of an aspect can be extracted from the text directly. The data source for this paper is a major online review site that collects and presents customers’ opinions on different hotel and travel related services. The results show that the grammatical dependencies technique provides the most meaningful output but requires a careful extraction pattern setting. The outputs of the other two techniques are not ready for direct use to inform service managers.

This article describes the generation of aggregated and elliptic sentences, using Dependency Trees connected by rhetorical relations as input. The system we have developed can generate both hypotactic and paratactic constructions with appropriate cue words, and various forms of ellipsis such as Gapping and Conjunction Reduction. We contend that Dependency Trees connected by rhetorical relations are excellent input for a generation system that has to generate ellipsis, and we propose a taxonomy of the most common Dutch cue words, grouped according to the kind of discourse relations they signal. Finally, we argue that syntactic aggregation should be performed in the Surface Realizer of a language generation system, because it requires access to language-specific syntactic information.