Search

Despite different applications, transformer-based language models, like BERT and GPT, learn about language by predicting missing parts of text. BERT is pretrained in Masked Language Modelling and GPT generates text from a given sequence. We explore such models for answering cloze questions in Portuguese, following different approaches. When options are not considered, the largest BERT model, trained exclusively for Portuguese, is the most accurate. But when selecting the best option, top performance is achieved by computing the most probable sentence, and GPT-2 fine-tuned for Portuguese beats BERT.

This paper describes how we tackled the development of Amaia, a conversational agent for Portuguese entrepreneurs. After introducing the domain corpus used as Amaia&rsquo, s Knowledge Base (KB), we make an extensive comparison of approaches for automatically matching user requests with Frequently Asked Questions (FAQs) in the KB, covering Information Retrieval (IR), approaches based on static and contextual word embeddings, and a model of Semantic Textual Similarity (STS) trained for Portuguese, which achieved the best performance. We further describe how we decreased the model&rsquo, s complexity and improved scalability, with minimal impact on performance. In the end, Amaia combines an IR library and an STS model with reduced features. Towards a more human-like behavior, Amaia can also answer out-of-domain questions, based on a second corpus integrated in the KB. Such interactions are identified with a text classifier, also described in the paper.

Poetry generation is a specific kind of natural language generation where several sources of knowledge are typically exploited to handle features on different levels, such as syntax, semantics, form or aesthetics. But although this task has been addressed by several researchers, and targeted different languages, all known systems have focused on a limited purpose and a single language. This article describes the effort of adapting the same architecture to generate poetry in three different languages – Portuguese, Spanish and English. An existing architecture is first described and complemented with the adaptations required for each language, including the linguistic resources used for handling morphology, syntax, semantics and metric scansion. An automatic evaluation was designed in such a way that it would be applicable to the target languages. It covered three relevant aspects of the generated poems, namely: the presence of poetic features, the variation of the linguistic structure and the semantic connection to a given topic. The automatic measures applied for the second and third aspect can be seen as novel in the evaluation of poetry. Overall, poems were successfully generated in the three languages addressed. Despite minor differences in different languages or seed words, poems revealed to have a regular metre, frequent rhymes, to exhibit an interesting degree of variation, and to be semantically-associated with the initially given seeds.

An overwhelming quantity of messages is posted in social networks every minute. To make the utilization of these platforms more productive, it is imperative to filter out information that is irrelevant to the general audience, such as private messages, personal opinions or well-known facts. This work is focused on the automatic classification of public social text according to its potential relevance, from a journalistic point of view, hopefully improving the overall experience of using a social network. Our experiments were based on a set of posts with several criteria, including the journalistic relevance, assessed by human judges. To predict the latter, we rely exclusively on linguistic features, extracted by Natural Language Processing tools, regardless the author of the message and its profile information. In our first approach, different classifiers and feature engineering methods were used to predict relevance directly from the selected features. In a second approach, relevance was predicted indirectly, based on an ensemble of classifiers for other key criteria when defining relevance—controversy, interestingness, meaningfulness, novelty, reliability and scope—also in the dataset. The first approach achieved a F 1-score of 0.76 and an Area under the ROC curve (AUC) of 0.63. But the best results were achieved by the second approach, with the best learned model achieving a F 1-score of 0.84 with an AUC of 0.78. This confirmed that journalistic relevance can indeed be predicted by the combination of the selected criteria, and that linguistic features can be exploited to classify the latter.

Word2vec-like word embeddings are known for keeping linguistic regularities and thus good for solving analogies. Following this, we explore such embeddings for Portuguese in the discovery of lexical-semantic relations, which can be used for augmenting lexical-semantic knowledge bases. In this exploratory approach, we tested different methods for discovering relations of different types and confirm that word embeddings can be used, at least, for suggesting new candidate relations.

This paper offers a brief overview on the ASSIN 2, an evaluation shared task collocated with STIL 2019. ASSIN 2 covered two different but related tasks: Recognizing Textual Entailment (RTE), also known as Natural Language Inference (NLI), and Semantic Textual Similarity (STS). The ASSIN 2 collection was made of pairs of sentences annotated with human judgments for NLI and STS. Participating teams could take part in any of the tasks or both: nine teams participated in the STS task and eight in the NLI task.

Computational creativity (CC) is a multidisciplinary research field, studying how to engineer software that exhibits behavior that would reasonably be deemed creative. This paper shows how composition of software solutions in this field can effectively be supported through a CC infrastructure that supports user-friendly development of CC software components and workflows, their sharing, execution, and reuse. The infrastructure allows CC researchers to build workflows that can be executed online and be easily reused by others through the workflow web address. Moreover, it enables the building of procedures composed of software developed by different researchers from different laboratories, leading to novel ways of software composition for computational purposes that were not expected in advance. This capability is illustrated on a workflow that implements a Concept Generator prototype based on the Conceptual Blending framework. The prototype consists of a composition of modules made available as web services, and is explored and tested through experiments involving blending of texts from different domains, blending of images, and poetry generation.

New forms of communication, like emojis, are frequent today in social media. Having in mind their strong connection with expressed emotions, we exploit emojis towards the creation of a model for emotion recognition in Portuguese. We gather short texts from Twitter and follow a traditional text classification task, where emojis are used as labels. After the process of feature engineering, two types of Naive Bayes and SVM classifiers are trained: one for classifying emotion, based on related emojis; another for predicting emojis. Interesting but debatable results were obtained on the former task, while the latter revealed to be more challenging, mainly due to emoji similarity. Yet, this also suggests that we can rely on them as an alternative to manually labelling emotions.

Within the domain of Artificial Intelligence, humor has been a research topic for some time, but the automatic recognition of its verbal expression has never been tackled for Portuguese. This work aims to change this scenario. We describe a set of experiments towards the development of computational models that recognize humor written in Portuguese, based on content and humor-specific features extracted. Interesting results, with F1-scores up to 0.93, are achieved when classifiers for this purpose are trained and tested on texts with a similar style (question-answers or news headlines). Yet, when the testing examples are of a different style, results are poor, which suggests that much more has to be done towards effective humor recognition.

Automatic recognition of named entities from clinical text lightens the work of health professionals by helping in the interpretation and easing tasks such as the population of databases with patient health information. In this study, we evaluated the performance of Conditional Random Fields, a sequence labelling model, for extracting entities from neurology clinical texts written in Portuguese. More than achieving F1-scores of about 73% or 80%, respectively for a relaxed or strict evaluation, the more discriminant features in this task were also analyzed.

Having in mind that different languages might present different challenges, this paper presents the following contributions to the area of Information Extraction from clinical text, targeting the Portuguese language: a collection of 281 clinical texts in this language, with manually-annotated named entities; word embeddings trained in a larger collection of similar texts; results of using BiLSTM-CRF neural networks for named entity recognition on the annotated collection, including a comparison of using in-domain or out-of-domain word embeddings in this task. Although learned with much less data, performance is higher when using in-domain embeddings. When tested in 20 independent clinical texts, this model achieved better results than a model using larger out-of-domain embeddings.

Ontologising is the task of associating terms, in text, with an ontological representation of their meaning, in an ontology. In this article, we revisit algorithms that have previously been used to ontologise the arguments of semantic relations in a relationless thesaurus, resulting in a wordnet. For increased flexibility, the algorithms do not use the extraction context when selecting the most adequate synsets for each term argument. Instead, they exploit a term-based lexical network which can be established by knowledge extracted automatically, or obtained from the resource the relations are being ontologised to. On the latter idea, we made several experiments to conclude that the algorithms can be used both for wordnet creation and for their enrichment. Besides describing the algorithms with some detail, the aforementioned experiments, which target both English and Portuguese, and their results are reported and discussed.

This paper describes the creation of PT-LKB, new Portuguese word embeddings learned from a large lexical-semantic knowledge base (LKB), using the node2vec method. Resulting embeddings combine the strengths of word vector representations and, even with lower dimensions, achieve high scores in genuine similarity, which so far were obtained by exploiting the graph structure of LKBs.

A wordnet is an important tool for developing natural language processing applications for a language. However, most wordnets are handcrafted by experts, which limits their growth. In this article, we propose an automatic approach to create wordnets by exploiting textual resources, dubbed ECO. After extracting semantic relation instances, identified by discriminating textual patterns, ECO discovers synonymy clusters, used as synsets, and attaches the remaining relations to suitable synsets. Besides introducing each step of ECO, we report on how it was implemented to create Onto.PT, a public lexical ontology for Portuguese. Onto.PT is the result of the automatic exploitation of Portuguese dictionaries and thesauri, and it aims to minimise the main limitations of existing Portuguese lexical knowledge bases.

This paper presents several approaches for computing word similarity in Portuguese and is motivated by the recent availability of state-of-the-art distributional models of Portuguese words, which add to several lexical knowledge bases (LKBs) for this language, available for a longer time. The previous resources were exploited to answer word similarity tests, also recently available for Portuguese. We conclude that there are several valid approaches for this task, but not one that outperforms all the others in every single test. For instance, distributional models seem to capture relatedness better, but LKBs are better suited for computing genuine similarity.

There is much research on Semantic Textual Similarity (STS) in English, specially since its inclusion in the SemEval evaluations. For other languages, it is not as common, mostly due to the unavailability of benchmarks. Recently, the ASSIN shared task targeted STS in Portuguese and released training and test collections. This paper describes an incremental approach to ASSIN, where the computed similarity is gradually improved by exploiting different features (e.g., token overlap, semantic relations, chunks, and negation) and approaches. The best reported results, obtained with a supervised approach, would get second place overall in ASSIN.

This paper presents MemeGera, a prototype tool that generates image-based memes from Portuguese news headlines. All is done automatically, with the help of computational linguistic resources, uncovered here with the rules for selecting images and adapting the text.

This paper revisits PoeTryMe, a poetry generation platform, and presents its most recent instantiation for producing poetry inspired by trends in the Twitter social network. The presented system searches for tweets that mention a given topic, extracts the most frequent words in those tweets, and uses them as seeds for the generation of new poems. The set of seeds might still be expanded with semantically-relevant words. Generation is performed by the classic PoeTryMe system, based on a semantic network and a grammar, with a previously used generate&test strategy. Illustrative results are presented using different seed expansion settings. They show that the produced poems use semantically-coherent lines with words that, at the time of generation, were associated with the topic. Resulting poems are not really about the topic, but they are a way of expressing, poetically, what the system knows about the semantic domain set by the topic.

There are several lexical resources available for the computational processing of Portuguese, organised differently and created by different people with different approaches and limitations. This paper presents the first experiments towards the exploitation of seven of those resources in the automatic creation of a large wordnet, where numerical scores are assigned to the inclusion of words in synsets and to the connection of synsets by semantic relations. Experiments confirm that a large wordnet can indeed be created and, to some extent, computed scores can be used as a confidence measure, which will enable the users to select only a portion of the resource, depending on the needs of their application on quantity and quality of lexical-semantic knowledge.

Identifying similar and related words is not only key in natural language understanding but also a suitable task for assessing the quality of computational resources that organise words and meanings of a language, compiled by different means. This paper, which aims to be a reference for those interested in computing word similarity in Portuguese, presents several approaches for this task and is motivated by the recent availability of state-of-the-art distributional models of Portuguese words, which add to several lexical knowledge bases (LKBs) for this language, available for a longer time. The previous resources were exploited to answer word similarity tests, which also became recently available for Portuguese. We conclude that there are several valid approaches for this task, but not one that outperforms all the others in every single test. Distributional models seem to capture relatedness better, while LKBs are better suited for computing genuine similarity, but, in general, better results are obtained when knowledge from different sources is combined.

In the last decade, several lexical-semantic knowledge bases (LKBs) were developed for Portuguese, by different teams and following different approaches. Most of them are open and freely available for the community. Those LKBs are briefly analysed here, with a focus on size, structure, and overlapping contents. However, we go further and exploit all of the analysed LKBs in the creation of new LKBs, based on the redundant contents. Both original and redundancy-based LKBs are then compared, indirectly, based on the performance of automatic procedures that exploit them for solving four different semantic analysis tasks. In addition to conclusions on the performance of the original LKBs, results show that, instead of selecting a single LKB to use, it is generally worth combining the contents of all the open Portuguese LKBs, towards better results.

ThisThis work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ paper describes the second version of the ASAP system 1 and its participation in the SemEval-2015, task 2a on Semantic Textual Similarity (STS). Our approach is based on computing the WordNet semantic relatedness and similarity of phrases from distinct sentences. We also apply topic modeling to get topic distributions over a set of sentences as well as some linguistic heuristics. In a special addition for this task, we retrieve named entities and compound nouns from DBPedia. All these features are used to feed a regression algorithm that learns the STS function.

Although typically used in classic topic models, surface words cannot represent meaning on their own. Consequently, redundancy is common in those topics, which may, for instance, include synonyms. To face this problem, we present SemLDA, an extended topic model that incorporates semantics from an external lexical-semantic knowledge base. SemLDA is introduced and explained in detail, pointing out where semantics is included both in the pre-pocessing and generative phase of topic distributions. As a result, instead of topics as distributions over words, we obtain distributions over concepts, each represented by a set of synonymous words. In order to evaluate SemLDA, we applied preliminary qualitative tests automatically against a state-of-the-art classical topic model. The results were promising and confirm our intuition towards the benefits of incorporating general semantics in a topic model.

PoeTryMe is a platform for the automatic generation of poetry, with a versatile architecture that provides a high level of customisation. The user can define features that go from the poem configuration and the line templates, to the initial seed words that will define a generation domain, and also the generation strategy. In this chapter, we introduce PoeTryMe’s architecture and describe how we used it to generate poetry in Portuguese, using natural language processing resources for this language as well as patterns that denote semantic relations in human-created poetry. After presenting the resources used with PoeTryMe, the problem of poetry generation is tackled incrementally, as our decisions are explained and illustrated, step-by-step. In the end, the objective features of the poems generated by the implemented strategies are compared, while the best-scoring poems are shown.

A wordnet is an important tool for developing natural language processing applications for a language, but the manual creation of such a resource limits its development. This dissertation studied the automatic construction of Onto.PT, a large Portuguese wordnet, aiming to minimise the main limitations of existing Portuguese wordnets. On this context, we propose ECO, an approach for creating wordnets automatically from text – relation instances are extracted, synonymy clusters (synsets) are discovered, and the remaining relations are then attached to suitable synsets. This document also reports on the contents of Onto.PT, its comparison to other wordnets, and its evaluation.

This paper presents the approach of the CISUC-KIS team to the SemEval 2014 task on Sentiment Analysis in Twitter, more precisely subtask B - Message Polarity Classification. We followed a machine learning approach where a SVM classifier was trained from a large and diverse set of features that included lexical, syntactic, sentiment and semantic-based aspects. This led to very interesting results which, in different datasets, put us always in the top-7 scores, including second position in the LiveJournal2014 dataset.

Onto.PT is a new public wordnet-like lexical ontology for Portuguese. To make its creation faster and to deal with information sparsity, Onto.PT is created automatically and integrates information extracted from existing lexical resources. After introducing Onto.PT's construction approach, this paper gives a closer look on this resource and provides information on its availability.

This paper presents work on the automatic creation of a polarity lexicon based on a lexical-semantic network. During this work, we noticed that the language registers of a relation should be considered in polarity propagation. After analysing the possible registers and performing some experiments, our intuition was confirmed --- there are registers that should invert the transmitted polarity (irony), while others always transmit negative polarity (pejorative, disparaging).

Polarity lexicons are lists of words (or meanings) where each entry is labelled as positive, negative or neutral. These lists are not available for different languages and specific domains. This work proposes and evaluates a new algorithm to classify words as positive, negative or neutral, relying on a small seed set of words, a common dictionary and a propagation algorithm. We evaluate the positive and negative polarity propagation of words, as well as the neutral polarity. The propagation is evaluated with different settings and lexical resources.

Regarding that information in broad-coverage knowledge bases, such as thesauri, is usually incomplete, merging information from different sources is a good option to amplify coverage. We propose a method for the enrichment of a thesaurus with information acquired automatically from dictionaries: pairs of synonyms are assigned to candidate synsets and, the pairs whose elements are not in the thesaurus are clustered to identify new synsets. This method was used in the enrichment of a Brazilian Portuguese thesaurus with synonyms from a European Portuguese dictionary, and resulted in a larger and broader thesaurus with new words and new concepts. The assignments and the obtained synsets were manually evaluated and yielded correction scores higher than 71% and 85% respectively.

This paper presents SAHARA, an online service for the evaluation platform of Second HAREM. SAHARA allows a fast evaluation of any NER system that conforms with HAREM guidelines, making it easier to perform post-hoc evaluations and keep track of the overall performance of NER systems.

This paper presents PAPEL, a lexical resource for Portuguese, consisting of relations between terms, extracted by (semi) automatic means from a general dictionary. After a short overview of the building process, a quantitative overview is given together with some examples. Evaluation is then presented and discussed: for synonymy, we used a public thesaurus, Tep, for the other relations, we queried Portuguese corpora through the AC/DC interface.

This paper describes a project aimed at creating a lexical ontology extracted (semi) automatically from a large Portuguese general dictionary. Although using machine readable dictionaries to extract semantic information is not new, we believe this is the first attempt for the Portuguese language. The paper describes a (to be) freely available resource, dubbed PAPEL, explaining the process used and the tools developed, and illustrating it with one specific relation: Causation.

This research addresses the role of the lyrics in the context of music emotion variation detection. To accomplish this task we create a system to detect the predominant emotion expressed by each sentence (verse) of the lyrics. The system employs Russell’s emotion model and contains 4 sets of emotions associated to each quadrant. To detect the predominant emotion in each verse, we propose a novel keyword-based approach, which receives a sentence (verse) and classifies it in the appropriate quadrant. To tune the system parameters, we created a 129-sentence training dataset from 68 songs. To validate our system, we created a separate ground-truth containing 239 sentences (verses) from 44 songs annotated manually with an average of 7 annotations per sentence. The system attains 67.4% F-Measure score.