Word embeddings (WE) have been shown to capture biases from the text they are trained on, which has led to the development of several bias measurement metrics and bias mitigation algorithms (i.e., methods that transform the embedding space to reduce bias). This study identifies three confounding factors that hinder the comparison of bias mitigation algorithms with bias measurement metrics: (1) reliance on different word sets when applying bias mitigation algorithms, (2) leakage between training words employed by mitigation methods and evaluation words used by metrics, and (3) inconsistencies in normalization transformations between mitigation algorithms. We propose a very simple comparison methodology that carefully controls for word sets and vector normalization to address these factors. We conduct a component isolation experiment to assess how each component of our methodology impacts bias measurement. After comparing the bias mitigation algorithms using our comparison methodology, we observe increased consistency between different debiasing algorithms when evaluated using our approach.

Large language models (LLM) are now a very common and successful path to approach language and retrieval tasks. While these LLM achieve surprisingly good results it is a challenge to use them on more constrained resources. Techniques to compress these LLM into smaller and faster models have emerged for English or Multilingual settings, but it is still a challenge for other languages. In fact, Spanish is the second language with most native speakers but lacks of these kind of resources. In this work, we evaluate all the models publicly available for Spanish on a set of 6 tasks and then, by leveraging on Knowledge Distillation, we present Speedy Gonzales, a collection of inference-efficient task-specific language models based on the ALBERT architecture. All of our models (fine-tuned and distilled) are publicly available on: https://huggingface.co/dccuchile.

Named Entity Recognition (NER) is an important task in Natural Language Processing that aims to identify text spans belonging to predefined categories. Traditional NER systems ignore nested entities, which are entities contained in other entity mentions. Although several methods have been proposed to address this case, most of them rely on complex task-specific structures and ignore potentially useful baselines for the task. We argue that this creates an overly optimistic impression of their performance. This paper revisits the Multiple LSTM-CRF (MLC) model, a simple, overlooked, yet powerful approach based on training independent sequence labeling models for each entity type. Extensive experiments with three nested NER corpora show that, regardless of the simplicity of this model, its performance is better or at least as well as more sophisticated methods. Furthermore, we show that the MLC architecture achieves state-of-the-art results in the Chilean Waiting List corpus by including pre-trained language models. In addition, we implemented an open-source library that computes task-specific metrics for nested NER. The results suggest that metrics used in previous work do not measure well the ability of a model to detect nested entities, while our metrics provide new evidence on how existing approaches handle the task.

We present the first shared task on semantic change discovery and detection in Spanish. We create the first dataset of Spanish words manually annotated by semantic change using the DURel framewok (Schlechtweg et al., 2018). The task is divided in two phases: 1) graded change discovery, and 2) binary change detection. In addition to introducing a new language for this task, the main novelty with respect to the previous tasks consists in predicting and evaluating changes for all vocabulary words in the corpus. Six teams participated in phase 1 and seven teams in phase 2 of the shared task, and the best system obtained a Spearman rank correlation of 0.735 for phase 1 and an F1 score of 0.735 for phase 2. We describe the systems developed by the competing teams, highlighting the techniques that were particularly useful.

In recent years there have been considerable advances in pre-trained language models, where non-English language versions have also been made available. Due to their increasing use, many lightweight versions of these models (with reduced parameters) have also been released to speed up training and inference times. However, versions of these lighter models (e.g., ALBERT, DistilBERT) for languages other than English are still scarce. In this paper we present ALBETO and DistilBETO, which are versions of ALBERT and DistilBERT pre-trained exclusively on Spanish corpora. We train several versions of ALBETO ranging from 5M to 223M parameters and one of DistilBETO with 67M parameters. We evaluate our models in the GLUES benchmark that includes various natural language understanding tasks in Spanish. The results show that our lightweight models achieve competitive results to those of BETO (Spanish-BERT) despite having fewer parameters. More specifically, our larger ALBETO model outperforms all other models on the MLDoc, PAWS-X, XNLI, MLQA, SQAC and XQuAD datasets. However, BETO remains unbeaten for POS and NER. As a further contribution, all models are publicly available to the community for future research.

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.

The quality of the annotated data directly influences in the success of supervised NLP models. However, creating annotated datasets is often time-consuming and expensive. Although the annotation tool takes an important role, we know little about how it influences annotation quality. We compare the quality of annotations for the task of chat-untangling made by non-experts annotators using two different tools. The first is SLATE, an existing command-line based tool, and the second is Parlay, a new tool we developed that integrates mouse interaction and visual links. Our experimental results indicate that, while both tools perform similarly in terms of annotation quality, Parlay offers a significantly better user experience.

We present a new task in educational NLP, recommend the best interventions to help special needs education professionals to work with students with different disabilities. We use the professionals’ observations of the students together with the students diagnosis and other chosen interventions to predict the best interventions for Chilean special needs students.

To avoid the “meaning conflation deficiency” of word embeddings, a number of models have aimed to embed individual word senses. These methods at one time performed well on tasks such as word sense induction (WSI), but they have since been overtaken by task-specific techniques which exploit contextualized embeddings. However, sense embeddings and contextualization need not be mutually exclusive. We introduce PolyLM, a method which formulates the task of learning sense embeddings as a language modeling problem, allowing contextualization techniques to be applied. PolyLM is based on two underlying assumptions about word senses: firstly, that the probability of a word occurring in a given context is equal to the sum of the probabilities of its individual senses occurring; and secondly, that for a given occurrence of a word, one of its senses tends to be much more plausible in the context than the others. We evaluate PolyLM on WSI, showing that it performs considerably better than previous sense embedding techniques, and matches the current state-of-the-art specialized WSI method despite having six times fewer parameters. Code and pre-trained models are available at https://github.com/AlanAnsell/PolyLM.

We present a system for the task of unsupervised lexical change detection: given a target word and two corpora spanning different periods of time, automatically detects whether the word has lost or gained senses from one corpus to another. Our system employs the temporal referencing method to obtain compatible representations of target words in different periods of time. This is done by concatenating corpora of different periods and performing a temporal referencing of target words i.e., treating occurrences of target words in different periods as two independent tokens. Afterwards, we train word embeddings on the joint corpus and compare the referenced vectors of each target word using cosine similarity. Our submission was ranked 7th among 34 teams for subtask 1, obtaining an average accuracy of 0.637, only 0.050 points behind the first ranked system.

Māori loanwords are widely used in New Zealand English for various social functions by New Zealanders within and outside of the Māori community. Motivated by the lack of linguistic resources for studying how Māori loanwords are used in social media, we present a new corpus of New Zealand English tweets. We collected tweets containing selected Māori words that are likely to be known by New Zealanders who do not speak Māori. Since over 30% of these words turned out to be irrelevant, we manually annotated a sample of our tweets into relevant and irrelevant categories. This data was used to train machine learning models to automatically filter out irrelevant tweets.

We present the SemEval-2018 Task 1: Affect in Tweets, which includes an array of subtasks on inferring the affectual state of a person from their tweet. For each task, we created labeled data from English, Arabic, and Spanish tweets. The individual tasks are: 1. emotion intensity regression, 2. emotion intensity ordinal classification, 3. valence (sentiment) regression, 4. valence ordinal classification, and 5. emotion classification. Seventy-five teams (about 200 team members) participated in the shared task. We summarize the methods, resources, and tools used by the participating teams, with a focus on the techniques and resources that are particularly useful. We also analyze systems for consistent bias towards a particular race or gender. The data is made freely available to further improve our understanding of how people convey emotions through language.

This paper examines the task of detecting intensity of emotion from text. We create the first datasets of tweets annotated for anger, fear, joy, and sadness intensities. We use a technique called best–worst scaling (BWS) that improves annotation consistency and obtains reliable fine-grained scores. We show that emotion-word hashtags often impact emotion intensity, usually conveying a more intense emotion. Finally, we create a benchmark regression system and conduct experiments to determine: which features are useful for detecting emotion intensity; and, the extent to which two emotions are similar in terms of how they manifest in language.

We present the first shared task on detecting the intensity of emotion felt by the speaker of a tweet. We create the first datasets of tweets annotated for anger, fear, joy, and sadness intensities using a technique called best–worst scaling (BWS). We show that the annotations lead to reliable fine-grained intensity scores (rankings of tweets by intensity). The data was partitioned into training, development, and test sets for the competition. Twenty-two teams participated in the shared task, with the best system obtaining a Pearson correlation of 0.747 with the gold intensity scores. We summarize the machine learning setups, resources, and tools used by the participating teams, with a focus on the techniques and resources that are particularly useful for the task. The emotion intensity dataset and the shared task are helping improve our understanding of how we convey more or less intense emotions through language.