Marcelo Mendoza

Vladimir Araujo, Andrés Carvallo, Souvik Kundu et al.

Due to the success of pre-trained language models, versions of languages other than English have been released in recent years. This fact implies the need for resources to evaluate these models. In the case of Spanish, there are few ways to systematically assess the models' quality. In this paper, we narrow the gap by building two evaluation benchmarks. Inspired by previous work (Conneau and Kiela, 2018; Chen et al., 2019), we introduce Spanish SentEval and Spanish DiscoEval, aiming to assess the capabilities of stand-alone and discourse-aware sentence representations, respectively. Our benchmarks include considerable pre-existing and newly constructed datasets that address different tasks from various domains. In addition, we evaluate and analyze the most recent pre-trained Spanish language models to exhibit their capabilities and limitations. As an example, we discover that for the case of discourse evaluation tasks, mBERT, a language model trained on multiple languages, usually provides a richer latent representation than models trained only with documents in Spanish. We hope our contribution will motivate a fairer, more comparable, and less cumbersome way to evaluate future Spanish language models.

Marcelo Mendoza, Naim Bro

From administrative registers of last names in Santiago, Chile, we create a surname affinity network that encodes socioeconomic data. This network is a multi-relational graph with nodes representing surnames and edges representing the prevalence of interactions between surnames by socioeconomic decile. We model the prediction of links as a knowledge base completion problem, and find that sharing neighbors is highly predictive of the formation of new links. Importantly, We distinguish between grounded neighbors and neighbors in the embedding space, and find that the latter is more predictive of tie formation. The paper discusses the implications of this finding in explaining the high levels of elite endogamy in Santiago.

Andrés Carvallo, Tamara Quiroga, Carlos Aspillaga et al.

While civilized users employ social media to stay informed and discuss daily occurrences, haters perceive these platforms as fertile ground for attacking groups and individuals. The prevailing approach to counter this phenomenon involves detecting such attacks by identifying toxic language. Effective platform measures aim to report haters and block their network access. In this context, employing hate speech detection methods aids in identifying these attacks amidst vast volumes of text, which are impossible for humans to analyze manually. In our study, we expand upon the usual hate speech detection methods, typically based on text classifiers, to develop a Named Entity Recognition (NER) System for Identity Groups. To achieve this, we created a dataset that allows extending a conventional NER to recognize identity groups. Consequently, our tool not only detects whether a sentence contains an attack but also tags the sentence tokens corresponding to the mentioned group. Results indicate that the model performs competitively in identifying groups with an average f1-score of 0.75, outperforming in identifying ethnicity attack spans with an f1-score of 0.80 compared to other identity groups. Moreover, the tool shows an outstanding generalization capability to minority classes concerning sexual orientation and gender, achieving an f1-score of 0.77 and 0.72, respectively. We tested the utility of our tool in a case study on social media, annotating and comparing comments from Facebook related to news mentioning identity groups. The case study reveals differences in the types of attacks recorded, effectively detecting named entities related to the categories of the analyzed news articles. Entities are accurately tagged within their categories, with a negligible error rate for inter-category tagging.

Vladimir Araujo, Andrés Villa, Marcelo Mendoza et al.

Current language models are usually trained using a self-supervised scheme, where the main focus is learning representations at the word or sentence level. However, there has been limited progress in generating useful discourse-level representations. In this work, we propose to use ideas from predictive coding theory to augment BERT-style language models with a mechanism that allows them to learn suitable discourse-level representations. As a result, our proposed approach is able to predict future sentences using explicit top-down connections that operate at the intermediate layers of the network. By experimenting with benchmarks designed to evaluate discourse-related knowledge using pre-trained sentence representations, we demonstrate that our approach improves performance in 6 out of 11 tasks by excelling in discourse relationship detection.

Ignacio Tampe Palma, Marcelo Mendoza, Evangelos Milios

Summarization has usually relied on gold standard summaries to train extractive or abstractive models. Social media brings a hurdle to summarization techniques since it requires addressing a multi-document multi-author approach. We address this challenging task by introducing a novel method that generates abstractive summaries of online news discussions. Our method extends a BERT-based architecture, including an attention encoding that fed comments' likes during the training stage. To train our model, we define a task which consists of reconstructing high impact comments based on popularity (likes). Accordingly, our model learns to summarize online discussions based on their most relevant comments. Our novel approach provides a summary that represents the most relevant aspects of a news item that users comment on, incorporating the social context as a source of information to summarize texts in online social networks. Our model is evaluated using ROUGE scores between the generated summary and each comment on the thread. Our model, including the social attention encoding, significantly outperforms both extractive and abstractive summarization methods based on such evaluation.

Carlos Aspillaga, Marcelo Mendoza, Alvaro Soto

The field of natural language understanding has experienced exponential progress in the last few years, with impressive results in several tasks. This success has motivated researchers to study the underlying knowledge encoded by these models. Despite this, attempts to understand their semantic capabilities have not been successful, often leading to non-conclusive, or contradictory conclusions among different works. Via a probing classifier, we extract the underlying knowledge graph of nine of the most influential language models of the last years, including word embeddings, text generators, and context encoders. This probe is based on concept relatedness, grounded on WordNet. Our results reveal that all the models encode this knowledge, but suffer from several inaccuracies. Furthermore, we show that the different architectures and training strategies lead to different model biases. We conduct a systematic evaluation to discover specific factors that explain why some concepts are challenging. We hope our insights will motivate the development of models that capture concepts more precisely.

Alfredo Silva, Marcelo Mendoza

Word embeddings are vital descriptors of words in unigram representations of documents for many tasks in natural language processing and information retrieval. The representation of queries has been one of the most critical challenges in this area because it consists of a few terms and has little descriptive capacity. Strategies such as average word embeddings can enrich the queries' descriptive capacity since they favor the identification of related terms from the continuous vector representations that characterize these approaches. We propose a data-driven strategy to combine word embeddings. We use Idf combinations of embeddings to represent queries, showing that these representations outperform the average word embeddings recently proposed in the literature. Experimental results on benchmark data show that our proposal performs well, suggesting that data-driven combinations of word embeddings are a promising line of research in ad-hoc information retrieval.

Ignacio Espinoza, Marcelo Mendoza, Pablo Ortega et al.

Opinions in forums and social networks are released by millions of people due to the increasing number of users that use Web 2.0 platforms to opine about brands and organizations. For enterprises or government agencies it is almost impossible to track what people say producing a gap between user needs/expectations and organizations actions. To bridge this gap we create Viscovery, a platform for opinion summarization and trend tracking that is able to analyze a stream of opinions recovered from forums. To do this we use dynamic topic models, allowing to uncover the hidden structure of topics behind opinions, characterizing vocabulary dynamics. We extend dynamic topic models for incremental learning, a key aspect needed in Viscovery for model updating in near-real time. In addition, we include in Viscovery sentiment analysis, allowing to separate positive/negative words for a specific topic at different levels of granularity. Viscovery allows to visualize representative opinions and terms in each topic. At a coarse level of granularity, the dynamic of the topics can be analyzed using a 2D topic embedding, suggesting longitudinal topic merging or segmentation. In this paper we report our experience developing this platform, sharing lessons learned and opportunities that arise from the use of sentiment analysis and topic modeling in real world applications.

Carlos Castillo, Gianmarco De Francisci Morales, Marcelo Mendoza et al.

We perform an automatic analysis of television news programs, based on the closed captions that accompany them. Specifically, we collect all the news broadcasted in over 140 television channels in the US during a period of six months. We start by segmenting, processing, and annotating the closed captions automatically. Next, we focus on the analysis of their linguistic style and on mentions of people using NLP methods. We present a series of key insights about news providers, people in the news, and we discuss the biases that can be uncovered by automatic means. These insights are contrasted by looking at the data from multiple points of view, including qualitative assessment.