Search

One of the long-standing challenges in lexical semantics consists in learning representations of words which reflect their semantic properties. The remarkable success of word embeddings for this purpose suggests that high-quality representations can be obtained by summarizing the sentence contexts of word mentions. In this paper, we propose a method for learning word representations that follows this basic strategy, but differs from standard word embeddings in two important ways. First, we take advantage of contextualized language models (CLMs) rather than bags of word vectors to encode contexts. Second, rather than learning a word vector directly, we use a topic model to partition the contexts in which words appear, and then learn different topic-specific vectors for each word. Finally, we use a task-specific supervision signal to make a soft selection of the resulting vectors. We show that this simple strategy leads to high-quality word vectors, which are more predictive of semantic properties than word embeddings and existing CLM-based strategies.

Comunicació presentada a: EACL 2021 celebrat del 19 a 23 d'abril de 2021 en línia. Lexical collocations are idiosyncratic combinations of two syntactically bound lexical items (e.g., “heavy rain”, “take a step” or “undergo surgery”). Understanding their degree of compositionality and idiosyncrasy, as well their underlying semantics, is crucial for language learners, lexicographers and downstream NLP applications alike. In this paper we analyse a suite of language models for collocation understanding. We first construct a dataset of apparitions of lexical collocations in context, categorized into 16 representative semantic categories. Then, we perform two experiments: (1) unsupervised collocate retrieval, and (2) supervised collocation classification in context. We find that most models perform well in distinguishing light verb constructions, especially if the collocation’s first argument acts as a subject, but often fail to distinguish, first, different syntactic structures within the same semantic category, and second, finer-grained categories which restrict the set of correct collocates. This work was partially supported by the European Commission via its H2020 Program under the contract number 870930.

Analogies play a central role in human commonsense reasoning. The ability to recognize analogies such as "eye is to seeing what ear is to hearing", sometimes referred to as analogical proportions, shape how we structure knowledge and understand language. Surprisingly, however, the task of identifying such analogies has not yet received much attention in the language model era. In this paper, we analyze the capabilities of transformer-based language models on this unsupervised task, using benchmarks obtained from educational settings, as well as more commonly used datasets. We find that off-the-shelf language models can identify analogies to a certain extent, but struggle with abstract and complex relations, and results are highly sensitive to model architecture and hyperparameters. Overall the best results were obtained with GPT-2 and RoBERTa, while configurations using BERT were not able to outperform word embedding models. Our results raise important questions for future work about how, and to what extent, pre-trained language models capture knowledge about abstract semantic relations., Comment: Accepted by ACL 2021 main conference

Pre-trained language models such as ClinicalBERT have achieved impressive results on tasks such as medical Natural Language Inference. At first glance, this may suggest that these models are able to perform medical reasoning tasks, such as mapping symptoms to diseases. However, we find that standard benchmarks such as MedNLI contain relatively few examples that require such forms of reasoning. To better understand the medical reasoning capabilities of existing language models, in this paper we introduce DisKnE, a new benchmark for Disease Knowledge Evaluation. To construct this benchmark, we annotated each positive MedNLI example with the types of medical reasoning that are needed. We then created negative examples by corrupting these positive examples in an adversarial way. Furthermore, we define training-test splits per disease, ensuring that no knowledge about test diseases can be learned from the training data, and we canonicalize the formulation of the hypotheses to avoid the presence of artefacts. This leads to a number of binary classification problems, one for each type of reasoning and each disease. When analysing pre-trained models for the clinical/biomedical domain on the proposed benchmark, we find that their performance drops considerably., Accepted by ACL 2021 Findings

In this paper, we introduce a new annotated dataset which is aimed at supporting the development of NLP models to identify and categorize language that is patronizing or condescending towards vulnerable communities (e.g. refugees, homeless people, poor families). While the prevalence of such language in the general media has long been shown to have harmful effects, it differs from other types of harmful language, in that it is generally used unconsciously and with good intentions. We furthermore believe that the often subtle nature of patronizing and condescending language (PCL) presents an interesting technical challenge for the NLP community. Our analysis of the proposed dataset shows that identifying PCL is hard for standard NLP models, with language models such as BERT achieving the best results.

The experimental landscape in natural language processing for social media is too fragmented. Each year, new shared tasks and datasets are proposed, ranging from classics like sentiment analysis to irony detection or emoji prediction. Therefore, it is unclear what the current state of the art is, as there is no standardized evaluation protocol, neither a strong set of baselines trained on such domain-specific data. In this paper, we propose a new evaluation framework (TweetEval) consisting of seven heterogeneous Twitter-specific classification tasks. We also provide a strong set of baselines as starting point, and compare different language modeling pre-training strategies. Our initial experiments show the effectiveness of starting off with existing pre-trained generic language models, and continue training them on Twitter corpora., Findings of EMNLP 2020. TweetEval benchmark available at https://github.com/cardiffnlp/tweeteval

Cross-lingual embeddings are vector space representations where word translations tend to be co-located. These representations enable learning transfer across languages, thus bridging the gap between data-rich languages such as English and others. In this paper, we present and evaluate a suite of cross-lingual embeddings for the English–Welsh language pair. To train the bilingual embeddings, a Welsh corpus of approximately 145 M words was combined with an English Wikipedia corpus. We used a bilingual dictionary to frame the problem of learning bilingual mappings as a supervised machine learning task, where a word vector space is first learned independently on a monolingual corpus, after which a linear alignment strategy is applied to map the monolingual embeddings to a common bilingual vector space. Two approaches were used to learn monolingual embeddings, including word2vec and fastText. Three cross-language alignment strategies were explored, including cosine similarity, inverted softmax and cross-domain similarity local scaling (CSLS). We evaluated different combinations of these approaches using two tasks, bilingual dictionary induction, and cross-lingual sentiment analysis. The best results were achieved using monolingual fastText embeddings and the CSLS metric. We also demonstrated that by including a few automatically translated training documents, the performance of a cross-lingual text classifier for Welsh can increase by approximately 20 percent points.

Lexical relation classification is the task of predicting whether a certain relation holds between a given pair of words. In this paper, we explore to which extent the current distributional landscape based on word embeddings provides a suitable basis for classification of collocations, i.e., pairs of words between which idiosyncratic lexical relations hold. First, we introduce a novel dataset with collocations categorized according to lexical functions. Second, we conduct experiments on a subset of this benchmark, comparing it in particular to the well known DiffVec dataset. In these experiments, in addition to simple word vector arithmetic operations, we also investigate the role of unsupervised relation vectors as a complementary input. While these relation vectors indeed help, we also show that lexical function classification poses a greater challenge than the syntactic and semantic relations that are typically used for benchmarks in the literature.

Word embeddings have become a standard resource in the toolset of any Natural Language Processing practitioner. While monolingual word embeddings encode information about words in the context of a particular language, cross-lingual embeddings define a multilingual space where word embeddings from two or more languages are integrated together. Current state-of-the-art approaches learn these embeddings by aligning two disjoint monolingual vector spaces through an orthogonal transformation which preserves the structure of the monolingual counterparts. In this work, we propose to apply an additional transformation after this initial alignment step, which aims to bring the vector representations of a given word and its translations closer to their average. Since this additional transformation is non-orthogonal, it also affects the structure of the monolingual spaces. We show that our approach both improves the integration of the monolingual spaces as well as the quality of the monolingual spaces themselves. Furthermore, because our transformation can be applied to an arbitrary number of languages, we are able to effectively obtain a truly multilingual space. The resulting (monolingual and multilingual) spaces show consistent gains over the current state-of-the-art in standard intrinsic tasks, namely dictionary induction and word similarity, as well as in extrinsic tasks such as cross-lingual hypernym discovery and cross-lingual natural language inference., Comment: 22 pages, 2 figures, 9 tables. Preprint submitted to Natural Language Engineering

Recently a number of unsupervised approaches have been proposed for learning vectors that capture the relationship between two words. Inspired by word embedding models, these approaches rely on co-occurrence statistics that are obtained from sentences in which the two target words appear. However, the number of such sentences is often quite small, and most of the words that occur in them are not relevant for characterizing the considered relationship. As a result, standard co-occurrence statistics typically lead to noisy relation vectors. To address this issue, we propose a latent variable model that aims to explicitly determine what words from the given sentences best characterize the relationship between the two target words. Relation vectors then correspond to the parameters of a simple unigram language model which is estimated from these words.

Comunicació presentada al 12th International Workshop on Semantic Evaluation (SemEval-2018), celebrat els dies 5 i 6 de juny de 2018 a Nova Orleans, EUA. This paper describes the results of the first shared task on Multilingual Emoji Prediction, organized as part of SemEval 2018. Given the text of a tweet, the task consists of predicting the most likely emoji to be used along such tweet. Two subtasks were proposed, one for English and one for Spanish, and participants were allowed to submit a system run to one or both subtasks. In total, 49 teams participated in the English subtask and 22 teams submitted a system run to the Spanish subtask. Evaluation was carried out emoji-wise, and the final ranking was based on macro F-Score. Data and further information about this task can be found at https://competitions. codalab.org/competitions/17344. Francesco B. and Horacio S. acknowledge support from the TUNER project (TIN2015-65308-C5-5-R, MINECO/FEDER, UE) and the Maria de Maeztu Units of Excellence Programme (MDM-2015-0502). The work of V. Patti and V. Basile was partially funded by the IHatePrejudice project (S1618 L2 BOSC 01).

Cross-lingual word embeddings are becoming increasingly important in multilingual NLP. Recently, it has been shown that these embeddings can be effectively learned by aligning two disjoint monolingual vector spaces through linear transformations, using no more than a small bilingual dictionary as supervision. In this work, we propose to apply an additional transformation after the initial alignment step, which moves cross-lingual synonyms towards a middle point between them. By applying this transformation our aim is to obtain a better cross-lingual integration of the vector spaces. In addition, and perhaps surprisingly, the monolingual spaces also improve by this transformation. This is in contrast to the original alignment, which is typically learned such that the structure of the monolingual spaces is preserved. Our experiments confirm that the resulting cross-lingual embeddings outperform state-of-the-art models in both monolingual and cross-lingual evaluation tasks., Comment: 11 pages, 4 tables, 1 figure. EMNLP 2018 camera-ready

Comunicació presentada a la Conference on Empirical Methods in Natural Language Processing, celebrada del 31 d'octubre al 4 de novembre de 2018 a Brussel·les, Bèlgica. Human language has evolved towards newer forms of communication such as social media, where emojis (i.e., ideograms bearing a visual meaning) play a key role. While there is an increasing body of work aimed at the computational modeling of emoji semantics, there is currently little understanding about what makes a computational model represent or predict a given emoji in a certain way. In this paper we propose a label-wise attention mechanism with which we attempt to better understand the nuances underlying emoji prediction. In addition to advantages in terms of interpretability, we show that our proposed architecture improves over standard baselines in emoji prediction, and does particularly well when predicting infrequent emojis. F. Barbieri and H. Saggion acknowledge support from the TUNER project (TIN2015-65308-C5-5- R, MINECO/FEDER, UE). Luis Espinosa-Anke, Jose Camacho-Collados and Steven Schockaert have been supported by ERC Starting Grant 637277.

Incorporating linguistic, world and common sense knowledge into AI/NLP systems is currently an important research area, with several open problems and challenges. At the same time, processing and storing this knowledge in lexical resources is not a straightforward task. This tutorial proposes to address these complementary goals from two methodological perspectives: the use of NLP methods to help the process of constructing and enriching lexical resources and the use of lexical resources for improving NLP applications. Two main types of audience can benefit from this tutorial: those working on language resources who are interested in becoming acquainted with automatic NLP techniques, with the end goal of speeding and/or easing up the process of resource curation; and on the other hand, researchers in NLP who would like to benefit from the knowledge of lexical resources to improve their systems and models. The slides of the tutorial are available at https://bitbucket.org/luisespinosa/lr-nlp/, Comment: NAACL 2018 Tutorial

Comunicació presentada a: Workshop on Semantic Deep Learning (SemDeep), celebrat amb ESWC 2017, del 28 de maig a l'1 de juny a Portoroz, Eslovènia. Music consumption habits as well as the Music market have changed dramatically due to the increasing popularity of digital audio and streaming services. Today, users are closer than ever to a vast number of songs, albums, artists and bands. However, the challenge remains in how to make sense of all the data available in the Music domain, and how current state of the art in Natural Language Processing and semantic technologies can contribute in Music Information Retrieval areas such as music recommendation, artist similarity or automatic playlist generation. In this paper, we present and evaluate a distributional sense-based embeddings model in the music domain, which can be easily used for these tasks, as well as a device for improving artist or album clus- tering. The model is trained on a disambiguated corpus linked to the MusicBrainz musical Knowledge Base, and following current knowledge-based approaches to sense-level embeddings, entity-related vectors are provided a la WordNet, concatenating the id of the entity and its mention. The model is evaluated both intrinsically and extrinsically in a supervised entity typing task, and released for the use and scrutiny of the community. We would like to thank the anonymous reviewers for their very helpful comments and suggestions for improving the quality of the manuscript. We also acknowledge support from the Spanish Minmistry of Economy and Competitiveness un- der the Maria de Maeztu Units of Excellence Programme (MDM-2015-0502) and under the TUNER project (TIN2015-65308-C5-5-R, MINECO/FEDER, UE).

Comunicació presentada al 4th SemWebEval Challenge at ESWC 2017, celebrat entre els dies 28 de maig i 1 de juny de 2017 a Portoroz, Eslovènia. The Open Knowledge Extraction Challenge invites researchers and practitioners from academia as well as industry to compete to the aim of pushing further the state of the art of knowledge extraction from text for the Semantic Web. The challenge has the ambition to provide a reference framework for research in this field by redefining a number of tasks typically from information and knowledge extraction by taking into account Semantic Web requirements and has the goal to test the performance of knowledge extraction systems. This year, the challenge goes in the third round and consists of three tasks which include named entity identification, typing and disambiguation by linking to a knowledge base depending on the task. The challenge makes use of small gold standard datasets that consist of manually curated documents and large silver standard datasets that consist of automatically generated synthetic documents. The performance measure of a participating system is twofold base on (1) Precision, Recall, F1-measure and on (2) Precision, Recall, F1-measure with respect to the runtime of the system. This work has been supported by the H2020 project HOBBIT (GA no. 688227) as well as the EuroStars projects DIESEL (project no. 01QE1512C) and QAMEL (project no. 01QE1549C). Also this work was partially funded by the Spanish Ministry of Economy and Competitiveness under the Maria de Maeztu Units of Excellence Programme (MDM-2015–0502).

Health information grows exponentially (doubling every 5 years), thus generating a sort of inflation of science, i.e. the generation of more knowledge than we can leverage. In an unprecedented data-driven shift, today doctors have no longer time to keep updated. This fact explains why only one in every five medical decisions is based strictly on evidence, which inevitably leads to variability. A good solution lies on clinical decision support systems, based on big data analysis. As the processing of large amounts of information gains relevance, automatic approaches become increasingly capable to see and correlate information further and better than the human mind can. In this context, healthcare professionals are increasingly counting on a new set of tools in order to deal with the growing information that becomes available to them on a daily basis. By allowing the grouping of collective knowledge and prioritizing “mindlines” against “guidelines”, these support systems are among the most promising applications of big data in health. In this demo paper we introduce Savana, an AI-enabled system based on Natural Language Processing (NLP) and Neural Networks, capable of, for instance, the automatic expansion of medical terminologies, thus enabling the re-use of information expressed in natural language in clinical reports. This automatized and precise digital extraction allows the generation of a real time information engine, which is currently being deployed in healthcare institutions, as well as clinical research and management.

Comunicació presentada a la 24th International World Wide Web Conference, celebrada els dies 18 a 22 de maig de 2015 a Florència, Itàlia. This paper presents a rule based approach to extracting relations from unstructured music text sources. The proposed approach identifies and disambiguates musical entities in text, such as songs, bands, persons, albums and music genres. Candidate relations are then obtained by traversing the dependency parsing tree of each sentence in the text with at least two identified entities. A set of syntactic rules based on part of speech tags are defined to filter out spurious and irrelevant relations. The extracted entities and relations are finally represented as a knowledge graph. We test our method on texts from songfacts.com, a website that provides tidbits with facts and stories about songs. The extracted relations are evaluated intrinsically by assessing their linguistic quality, as well as extrinsically by assessing the extent to which they map an existing music knowledge base. Our system produces a vast percentage of linguistically correct relations between entities, and is able to replicate a significant part of the knowledge base.

Paper presented at The 2015 Conference on Empirical Methods in Natural Language; 2015 Sept 17-21; Lisbon, Portugal. We present KB-UNIFY, a novel approach for integrating the output of different Open Information Extraction systems into a single unified and fully disambiguated knowledge repository. KB-UNIFY consists of three main steps: (1) disambiguation of relation argument pairs via a sensebased vector representation and a large unified sense inventory; (2) ranking of semantic relations according to their degree of specificity; (3) cross-resource relation alignment and merging based on the semantic similarity of domains and ranges. We tested KB-UNIFY on a set of four heterogeneous knowledge bases, obtaining high-quality results. We discuss and provide evaluations at each stage, and release output and evaluation data for the use and scrutiny of the community. The authors gratefully acknowledge the support of the ERC Starting Grant MultiJEDI No. 259234.

This paper presents a method for the generation of struc- tured data sources for music recommendation using information extracted from unstructured text sources. The proposed method identi es entities in text that are relevant to the music domain, and then extracts seman- tically meaningful relations between them. The extracted entities and re- lations are represented as a graph, from which the recommendations are computed. A major advantage of this approach is that the recommenda- tions can be conveyed to the user using natural language, thus providing an enhanced user experience. We test our method on texts from song- facts.com, a website that provides facts and stories about songs. The extracted relations are evaluated intrinsically by assessing their linguis- tic quality, as well as extrinsically by assessing the extent to which they map an existing music knowledge base. Finally, an experiment with real users is performed to assess the suitability of the extracted knowledge for music recommendation. Our method is able to extract relations between pair of musical entities with high precision, and the explanation of those relations to the user improves user satisfaction considerably.

Comunicació presentada a la International Conference on Applications of Natural Language to Data Bases/Information Systems celebrada del 18 al 20 de juny de 2014 a Montpellier, França. Definition Extraction (DE) is the task to automatically identify definitional knowledge in naturally-occurring text. This task has applications in ontology generation, glossary creation or question answering. Although the traditional approach to DE has been based on hand-crafted pattern-matching rules, recent methods incorporate learning algorithms in order to classify sentences as definitional or non-definitional. This paper presents a supervised approach to Definition Extraction in which only syntactic features derived from dependency relations are used. We model the problem as a classification task where each sentence has to be classified as being or not definitional. We compare our results with two well-known approaches: First, a supervised method based on Word-Class Lattices and second, an unsupervised approach based on mining recurrent patterns. Our competitive results suggest that syntactic information alone can contribute substantially to the development and improvement of DE systems. We also acknowledge partial support from project Dr. Inventor (FP7-ICT-2013.8.1 611383), programa Ramón y Cajal 2009 (RYC- 2009-04291), and project number TIN2012-38584-C06-03 from Ministerio de Economía y Competitividad, Secretaría de Estado de Investigación, Desarrollo e Innovación, Spain.

We present a corpus-based study of the Corporate Social Responsibility Reports issued by companies who operate in the Spanish Stock Market IBEX 35. Natural Language Processing methods are used for extracting informative collocations and measuring the textual complexity of these texts. Based on our results, we claim that (1) these reports are in general simpler and more redundant than general language, (2) there is a tendency towards the increasing use of the term sustainability to describe SR actions, and (3) integrated reports seem to be written for shareholders rather than the general public.

This paper proposes an approach to classify definitions as they appear in popularizing texts. Following the function theory of/nlexicography (Bergenholtz and Tarp, 2003), we propose a user-centered classification that breaks down definitions according to the way they are deployed in the text (Westerhout and Monachesi, 2007) and the information they encode (Sierra et al., 2006)./nWe hypothesize that different types of users can benefit from this classification, which covers a range of definitional styles, from the classic genus et differentia model to function-oriented definitions. The corpus used for this task consists of 50 transcripts from The Science Magazine Podcast (around 400k words), where 570 snippets containing definitional information have been manually annotated and classified.

L2 learners often produce “ungrammatical” word combinations such as, e.g., *give a suggestion or *make a walk. This is because of the “collocationality” of one of their items (the base) that limits the acceptance of collocates to express a specific meaning (‘perform’ above). We propose an algorithm that delivers, for a given base and the intended meaning of a collocate, the actual collocate lexeme(s) (make / take above). The algorithm exploits the/nlinear mapping between bases and collocates from examples and generates a collocation transformation matrix which is then applied to novel unseen cases. The evaluation shows a promising line of research in collocation discovery. The present work has been partially funded by the Spanish Ministry of Economy and Competitiveness/n(MINECO), through a predoctoral grant (BES-2012-057036) in the framework of the project HARenES (FFI2011-30219-C02-02), and by the European Commission under the grant number H2020–645012–RIA. We also acknowledge support from the Maria de Maeztu Excellence Program (MDM-2015-0502).

Automatically identifying definitional knowledge in text corpora (Definition Extraction or DE) is an important task with direct applications in, among others, Automatic Glossary Generation, Taxonomy Learning, Question Answering and Semantic Search. It is generally cast as a binary classification problem between definitional and non-definitional sentences. In this paper we present a set of neural architectures combining Convolutional and Recurrent Neural Networks, which are further enriched by incorporating linguistic information via syntactic dependencies. Our experimental results in the task of sentence classification, on two benchmarking DE datasets (one generic, one domain-specific), show that these models obtain consistent state of the art results. Furthermore, we demonstrate that models trained on clean Wikipedia-like definitions can successfully be applied to more noisy domain-specific corpora.

Comunicació presentada al 12th International Workshop on Semantic Evaluation (SemEval-2018), celebrat els dies 5 i 6 de juny de 2018 a Nova Orleans, EUA. This paper describes the SemEval 2018 Shared Task on Hypernym Discovery. We put forward this task as a complementary benchmark for modeling hypernymy, a problem which has traditionally been cast as a binary classification task, taking a pair of candidate words as input. Instead, our reformulated task is defined as follows: given an input term, retrieve (or discover) its suitable hypernyms from a target corpus. We proposed five different subtasks covering three languages (English, Spanish, and Italian), and two specific domains of knowledge in English (Medical and Music). Participants were allowed to compete in any or all of the subtasks. Overall, a total of 11 teams participated, with a total of 39 different systems submitted through all subtasks. Data, results and further information about the task can be found at https://competitions. codalab.org/competitions/17119. The authors gratefully acknowledge the economic support in the construction of the datasets from the Maria de Maeztu-UPF Grant provided to Horacio Saggion, Luis Espinosa-Anke and Sergio Oramas; Google Research through the Google Doctoral Fellowship in Natural Language Processing to Jose Camacho-Collados; and Bar-Ilan University through Vered Shwartz. This work is partially supported by the TUNER project (TIN2015- 65308-C5-5-R, MINECO/FEDER, UE), Spanish Ministry of Economy and Competitiveness.

We describe the system submitted to SemEval-2020 Task 6, Subtask 1. The aim of this subtask is to predict whether a given sentence contains a definition or not. Unsurprisingly, we found that strong results can be achieved by fine-tuning a pre-trained BERT language model. In this paper,we analyze the performance of this strategy. Among others, we show that results can be improved by using a two-step fine-tuning process, in which the BERT model is first fine-tuned on the full training set, and then further specialized towards a target domain.

Comunicació presentada al 10th International Workshop on Semantic Evaluation (SemEval 2016), celebrat els dies 16 i 17 de juny de 2016 a San Diego, EUA. This paper presents the participation of the TALN team in the Complex Word Identification Task of SemEval-2016 (Task 11). The purpose of the task was to determine if a word in a given sentence can be judged as complex or not by a certain target audience. To experiment with word complexity identification approaches, Task organizers provided a training set of 2,237 words judged as complex or not by 20 human evaluators, together with the sentence in which each word occurs. In our contribution we modelled each word to evaluate as a numeric vector populated with a set of lexical, semantic and contextual features that may help assess the complexity of a word. We trained a Random Forest classifier to automatically decide if each word is complex or not. We submitted two runs in which we respectively considered unweighted and weighted instances of complex words to train our classifier, where the weight of each instance is proportional to the number of evaluators that judged the word as complex. Our system scored as the third best performing one. This work is partly supported by the Spanish Ministry of Economy and Competitiveness under the Maria de Maeztu Units of Excellence Programme (MDM-2015-0502) and the ABLE-TO-INCLUDE Project (Competitivity and Innovation Programme of the European Commission, CIP-ICT-PSP-2013-7/621055).

The rate at which information about music is being created and shared on the web is growing exponentially. However, the challenge of making sense of all this data remains an open problem. In this paper, we present and evaluate an Information Extraction pipeline aimed at the construction of a Music Knowledge Base. Our approach starts off by collecting thousands of stories about songs from the songfacts.com website. Then, we combine a state-of-the-art Entity Linking tool and a linguistically motivated rule-based algorithm to extract semantic relations between entity pairs. Next, relations with similar semantics are grouped into clusters by exploiting syntactic dependencies. These relations are ranked thanks to a novel confidence measure based on statistical and linguistic evidence. Evaluation is carried out intrinsically, by assessing each component of the pipeline, as well as in an extrinsic task, in which we evaluate the contribution of natural language explanations in music recommendation. We demonstrate that our method is able to discover novel facts with high precision, which are missing in current generic as well as music-specific knowledge repositories. A system that constructs a Music Knowledge Base entirely from scratch.A method for clustering and scoring relations in a Relation Extraction pipeline.Reveals music facts absent from knowledge repositories (e.g. Wikipedia).Explains music recommendations in natural language.

Comunicació presentada a la 16th International Conference, (CICLing) celebrada del 14 al 20 d'abril de 2015 a El Caire, Egipte. Hypernym extraction is a crucial task for semantically motivated NLP tasks such as taxonomy and ontology learning, textual entailment or paraphrase identification. In this paper, we describe an approach to hypernym extraction from textual definitions, where machine-learning and post-classification refinement rules are combined. Our best-performing configuration shows competitive results compared to state-of-the-art systems in a well-known benchmarking dataset. The quality of our features is measured by combining them in different feature sets and by ranking them by their Information Gain score. Our experiments confirm that both syntactic and definitional information play a crucial role in the hypernym extraction task. This work is partially funded by the SKATER project, TIN2012-38584- C06-03, Ministerio de Economía y Competitividad, Secretaría de Estado de Investigación, Desarrollo e Innovación, España; and Dr. Inventor (FP7-ICT-2013.8.1 611383), programa Ramón y Cajal 2009 (RYC-2009-04291).

While the success of pre-trained language models has largely eliminated the need for high-quality static word vectors in many NLP applications, static word vectors continue to play an important role in tasks where word meaning needs to be modelled in the absence of linguistic context. In this paper, we explore how the contextualised embeddings predicted by BERT can be used to produce high-quality word vectors for such domains, in particular related to knowledge base completion, where our focus is on capturing the semantic properties of nouns. We find that a simple strategy of averaging the contextualised embeddings of masked word mentions leads to vectors that outperform the static word vectors learned by BERT, as well as those from standard word embedding models, in property induction tasks. We notice in particular that masking target words is critical to achieve this strong performance, as the resulting vectors focus less on idiosyncratic properties and more on general semantic properties. Inspired by this view, we propose a filtering strategy which is aimed at removing the most idiosyncratic mention vectors, allowing us to obtain further performance gains in property induction.

This paper describes the system submitted by the TALN-UPF team to SEMEVAL Task 17 (Taxonomy Extraction Evaluation). We present a method for automatically learning a taxonomy from a flat terminology, which benefits from a definition corpus obtained by querying the BabelNet semantic network. Then, we combine a machine-learning algorithm for term-hypernym extraction with linguistically-motivated heuristics for hypernym decomposition. Our approach performs well in terms of vertex coverage and newly added vertices, while it shows room for improvement in terms of graph topology, edge coverage and precision of novel edges. Paper presented at 9th International Workshop on Semantic Evaluation (SemEval 2015);2015 June 4-5; Denver, Colorado. This work is partially funded by the SKATER project, TIN2012-38584-C06-03, Ministerio de Economía y Competitividad, Secretaría de Estado de Investigación, Desarrollo e Innovación, España; and project Dr. Inventor (FP7-ICT-2013.8.1 611383).

Pre-trained neural language models (LMs) have achieved impressive results in various natural language processing tasks, across different languages. Surprisingly, this extends to the social media genre, despite the fact that social media often has very different characteristics from the language that LMs have seen during training. A particularly striking example is the performance of AraBERT, an LM for the Arabic language, which is successful in categorizing social media posts in Arabic dialects, despite only having been trained on Modern Standard Arabic. Our hypothesis in this paper is that the performance of LMs for social media can nonetheless be improved by incorporating static word vectors that have been specifically trained on social media. We show that a simple method for incorporating such word vectors is indeed successful in several Arabic and English benchmarks. Curiously, however, we also find that similar improvements are possible with word vectors that have been trained on traditional text sources (e.g. Wikipedia).

International audience; Scarcity of training data for task-oriented dialogue systemsis a well known problem that is usually tackled with costly and time-consuming manual data annotation. An alternative solution is to relyon automatic text generation which, although less accurate than humansupervision, has the advantage of being cheap and fast. Our contributionis twofold. First we show how to optimally train and control the generationof intent-specific sentences using a conditional variational autoencoder.Then we introduce a new protocol calledquery transferthat allows toleverage a large unlabelled dataset, possibly containing irrelevant queries,to extract relevant information. Comparison with two different baselinesshows that this method, in the appropriate regime, consistently improvesthe diversity of the generated queries without compromising their quality.We also demonstrate the effectiveness of our generation method as a dataaugmentation technique for language modelling tasks.