We introduce SPUD (Semantically Perturbed Universal Dependencies), a framework for creating nonce treebanks for the multilingual Universal Dependencies (UD) corpora. SPUD data satisfies syntactic argument structure, provides syntactic annotations, and ensures grammaticality via language-specific rules. We create nonce data in Arabic, English, French, German, and Russian, and demonstrate two use cases of SPUD treebanks. First, we investigate the effect of nonce data on word co-occurrence statistics, as measured by perplexity scores of autoregressive (ALM) and masked language models (MLM). We find that ALM scores are significantly more affected by nonce data than MLM scores. Second, we show how nonce data affects the performance of syntactic dependency probes. We replicate the findings of Müller-Eberstein et al. (2022) on nonce test data and show that the performance declines on both MLMs and ALMs wrt. original test data. However, a majority of the performance is kept, suggesting that the probe indeed learns syntax independently from semantics.

Neural models and deep learning techniques have predominantly been used in many tasks of natural language processing (NLP), including automatic readability assessment (ARA). They apply deep transfer learning and enjoy high accuracy. However, most of the models still cannot leverage long dependence such as inter-sentential topic-level or document-level information because of their structure and computational cost. Moreover, neural models usually have low interpretability. In this paper, we propose a generalization of passage-level, corpus-level, document-level and topic-level features. In our experiments, we show the effectiveness of “Statistical Lexical Spread (SLS)” features when combined with IDF (inverse document frequency) and TF-IDF (term frequency–inverse document frequency), which adds a topological perspective (inter-document) to readability to complement the typological approaches (intra-document) used in traditional readability formulas. Interestingly, simply adding these features in BERT models outperformed state-of-the-art systems trained on a large number of hand-crafted features derived from heavy linguistic processing. In analysis, we show that SLS is also easy-to-interpret because SLS computes lexical features, which appear explicitly in texts, compared to parameters in neural models.

This paper addresses the task of semantic frame induction based on pre-trained language models (LMs). The current state of the art is to directly use contextualized embeddings from models such as BERT and to cluster them in a two step clustering process (first lemma-internal, then over all verb tokens in the data set). We propose not to use the LM’s embeddings as such but rather to refine them via some transformer-based denoising autoencoder. The resulting embeddings allow to obtain competitive results while clustering them in a single pass. This shows clearly that the autoendocer allows to already concentrate on the information that is relevant for distinguishing event types.

In this paper, we investigate to which extent contextual neural language models (LMs) implicitly learn syntactic structure. More concretely, we focus on constituent structure as represented in the Penn Treebank (PTB). Using standard probing techniques based on diagnostic classifiers, we assess the accuracy of representing constituents of different categories within the neuron activations of a LM such as RoBERTa. In order to make sure that our probe focuses on syntactic knowledge and not on implicit semantic generalizations, we also experiment on a PTB version that is obtained by randomly replacing constituents with each other while keeping syntactic structure, i.e., a semantically ill-formed but syntactically well-formed version of the PTB. We find that 4 pretrained transfomer LMs obtain high performance on our probing tasks even on manipulated data, suggesting that semantic and syntactic knowledge in their representations can be separated and that constituency information is in fact learned by the LM. Moreover, we show that a complete constituency tree can be linearly separated from LM representations.

Proper dialect identification is important for a variety of Arabic NLP applications. In this paper, we present a method for rapidly constructing a tweet dataset containing a wide range of country-level Arabic dialects —covering 18 different countries in the Middle East and North Africa region. Our method relies on applying multiple filters to identify users who belong to different countries based on their account descriptions and to eliminate tweets that either write mainly in Modern Standard Arabic or mostly use vulgar language. The resultant dataset contains 540k tweets from 2,525 users who are evenly distributed across 18 Arab countries. Using intrinsic evaluation, we show that the labels of a set of randomly selected tweets are 91.5% accurate. For extrinsic evaluation, we are able to build effective country level dialect identification on tweets with a macro-averaged F1-score of 60.6% across 18 classes.

Detecting offensive language on Twitter has many applications ranging from detecting/predicting bullying to measuring polarization. In this paper, we focus on building a large Arabic offensive tweet dataset. We introduce a method for building a dataset that is not biased by topic, dialect, or target. We produce the largest Arabic dataset to date with special tags for vulgarity and hate speech. We thoroughly analyze the dataset to determine which topics, dialects, and gender are most associated with offensive tweets and how Arabic speakers useoffensive language. Lastly, we conduct many experiments to produce strong results (F1 =83.2) on the dataset using SOTA techniques.

User stance detection entails ascertaining the position of a user towards a target, such as an entity, topic, or claim. Recent work that employs unsupervised classification has shown that performing stance detection on vocal Twitter users, who have many tweets on a target, can be highly accurate (+98%). However, such methods perform poorly or fail completely for less vocal users, who may have authored only a few tweets about a target. In this paper, we tackle stance detection for such users using two approaches. In the first approach, we improve user-level stance detection by representing tweets using contextualized embeddings, which capture latent meanings of words in context. We show that this approach outperforms two strong baselines and achieves 89.6% accuracy and 91.3% macro F-measure on eight controversial topics. In the second approach, we expand the tweets of a given user using their Twitter timeline tweets, which may not be topically relevant, and then we perform unsupervised classification of the user, which entails clustering a user with other users in the training set. This approach achieves 95.6% accuracy and 93.1% macro F-measure.

This paper addresses the question to which extent neural contextual language models such as BERT implicitly represent complex semantic properties. More concretely, the paper shows that the neuron activations obtained from processing an English sentence provide discriminative features for predicting the (non-)causativity of the event denoted by the verb in a simple linear classifier. A layer-wise analysis reveals that the relevant properties are mostly learned in the higher layers. Moreover, further experiments show that appr. 10% of the neuron activations are enough to already predict causativity with a relatively high accuracy.

In this paper, we describe our efforts at OSACT Shared Task on Offensive Language Detection. The shared task consists of two subtasks: offensive language detection (Subtask A) and hate speech detection (Subtask B). For offensive language detection, a system combination of Support Vector Machines (SVMs) and Deep Neural Networks (DNNs) achieved the best results on development set, which ranked 1st in the official results for Subtask A with F1-score of 90.51% on the test set. For hate speech detection, DNNs were less effective and a system combination of multiple SVMs with different parameters achieved the best results on development set, which ranked 4th in official results for Subtask B with F1-macro score of 80.63% on the test set.

This paper describes the systems submitted by the Arabic Language Technology group (ALT) at SemEval-2020 Task 12: Multilingual Offensive Language Identification in Social Media. We focus on sub-task A (Offensive Language Identification) for two languages: Arabic and English. Our efforts for both languages achieved more than 90% macro-averaged F1-score on the official test set. For Arabic, the best results were obtained by a system combination of Support Vector Machine, Deep Neural Network, and fine-tuned Bidirectional Encoder Representations from Transformers (BERT). For English, the best results were obtained by fine-tuning BERT.

Arabic text is typically written without short vowels (or diacritics). However, their presence is required for properly verbalizing Arabic and is hence essential for applications such as text to speech. There are two types of diacritics, namely core-word diacritics and case-endings. Most previous works on automatic Arabic diacritic recovery rely on a large number of manually engineered features, particularly for case-endings. In this work, we present a unified character level sequence-to-sequence deep learning model that recovers both types of diacritics without the use of explicit feature engineering. Specifically, we employ a standard neural machine translation setup on overlapping windows of words (broken down into characters), and then we use voting to select the most likely diacritized form of a word. The proposed model outperforms all previous state-of-the-art systems. Our best settings achieve a word error rate (WER) of 4.49% compared to the state-of-the-art of 12.25% on a standard dataset.

Short vowels, aka diacritics, are more often omitted when writing different varieties of Arabic including Modern Standard Arabic (MSA), Classical Arabic (CA), and Dialectal Arabic (DA). However, diacritics are required to properly pronounce words, which makes diacritic restoration (a.k.a. diacritization) essential for language learning and text-to-speech applications. In this paper, we present a system for diacritizing MSA, CA, and two varieties of DA, namely Moroccan and Tunisian. The system uses a character level sequence-to-sequence deep learning model that requires no feature engineering and beats all previous SOTA systems for all the Arabic varieties that we test on.

When speakers code-switch between their native language and a second language or language variant, they follow a syntactic pattern where words and phrases from the embedded language are inserted into the matrix language. This paper explores the possibility of utilizing this pattern in improving code-switching identification between Modern Standard Arabic (MSA) and Egyptian Arabic (EA). We try to answer the question of how strong is the POS signal in word-level code-switching identification. We build a deep learning model enriched with linguistic features (including POS tags) that outperforms the state-of-the-art results by 1.9% on the development set and 1.0% on the test set. We also show that in intra-sentential code-switching, the selection of lexical items is constrained by POS categories, where function words tend to come more often from the dialectal language while the majority of content words come from the standard language.

This paper describes the QC-GO team submission to the MADAR Shared Task Subtask 1 (travel domain dialect identification) and Subtask 2 (Twitter user location identification). In our participation in both subtasks, we explored a number of approaches and system combinations to obtain the best performance for both tasks. These include deep neural nets and heuristics. Since individual approaches suffer from various shortcomings, the combination of different approaches was able to fill some of these gaps. Our system achieves F1-Scores of 66.1% and 67.0% on the development sets for Subtasks 1 and 2 respectively.

We present ongoing work on data-driven parsing of German and French with Lexicalized Tree Adjoining Grammars. We use a supertagging approach combined with deep learning. We show the challenges of extracting LTAG supertags from the French Treebank, introduce the use of left- and right-sister-adjunction, present a neural architecture for the supertagger, and report experiments of n-best supertagging for French and German.

This paper describes our system submission to the CALCS 2018 shared task on named entity recognition on code-switched data for the language variant pair of Modern Standard Arabic and Egyptian dialectal Arabic. We build a a Deep Neural Network that combines word and character-based representations in convolutional and recurrent networks with a CRF layer. The model is augmented with stacked layers of enriched information such pre-trained embeddings, Brown clusters and named entity gazetteers. Our system is ranked second among those participating in the shared task achieving an FB1 average of 70.09%.

In this paper, we describe Mumpitz, the system we submitted to the PARSEME Shared task on automatic identification of verbal multiword expressions (VMWEs). Mumpitz consists of a Bidirectional Recurrent Neural Network (BRNN) with Long Short-Term Memory (LSTM) units and a heuristic that leverages the dependency information provided in the PARSEME corpus data to differentiate VMWEs in a sentence. We submitted results for seven languages in the closed track of the task and for one language in the open track. For the open track we used the same system, but with pretrained instead of randomly initialized word embeddings to improve the system performance.

This paper describes our system submission to the SemEval 2018 Task 10 on Capturing Discriminative Attributes. Given two concepts and an attribute, the task is to determine whether the attribute is semantically related to one concept and not the other. In this work we assume that discriminative attributes can be detected by discovering the association (or lack of association) between a pair of words. The hypothesis we test in this contribution is whether the semantic difference between two pairs of concepts can be treated in terms of measuring the distance between words in a vector space, or can simply be obtained as a by-product of word co-occurrence counts.

Arabic dialects do not just share a common koiné, but there are shared pan-dialectal linguistic phenomena that allow computational models for dialects to learn from each other. In this paper we build a unified segmentation model where the training data for different dialects are combined and a single model is trained. The model yields higher accuracies than dialect-specific models, eliminating the need for dialect identification before segmentation. We also measure the degree of relatedness between four major Arabic dialects by testing how a segmentation model trained on one dialect performs on the other dialects. We found that linguistic relatedness is contingent with geographical proximity. In our experiments we use SVM-based ranking and bi-LSTM-CRF sequence labeling.

The automated processing of Arabic Dialects is challenging due to the lack of spelling standards and to the scarcity of annotated data and resources in general. Segmentation of words into its constituent parts is an important processing building block. In this paper, we show how a segmenter can be trained using only 350 annotated tweets using neural networks without any normalization or use of lexical features or lexical resources. We deal with segmentation as a sequence labeling problem at the character level. We show experimentally that our model can rival state-of-the-art methods that rely on additional resources.

This paper describes our system submission to the CogALex-2016 Shared Task on Corpus-Based Identification of Semantic Relations. Our system won first place for Task-1 and second place for Task-2. The evaluation results of our system on the test set is 88.1% (79.0% for TRUE only) f-measure for Task-1 on detecting semantic similarity, and 76.0% (42.3% when excluding RANDOM) for Task-2 on identifying finer-grained semantic relations. In our experiments, we try word analogy, linear regression, and multi-task Convolutional Neural Networks (CNNs) with word embeddings from publicly available word vectors. We found that linear regression performs better in the binary classification (Task-1), while CNNs have better performance in the multi-class semantic classification (Task-2). We assume that word analogy is more suited for deterministic answers rather than handling the ambiguity of one-to-many and many-to-many relationships. We also show that classifier performance could benefit from balancing the distribution of labels in the training data.

In this paper, we describe our effort in the development and annotation of a large scale corpus containing code-switched data. Until recently, very limited effort has been devoted to develop computational approaches or even basic linguistic resources to support research into the processing of Moroccan Darija.

Dans cet article, nous présentons une modélisation de la morphologie dérivationnelle de l’arabe utilisant le cadre métagrammatical offert par XMG. Nous démontrons que l’utilisation de racines et patrons abstraits comme morphèmes atomiques sous-spécifiés offre une manière élégante de traiter l’interaction entre morphologie et sémantique.

Arabic is a language known for its rich and complex morphology. Although many research projects have focused on the problem of Arabic morphological analysis using different techniques and approaches, very few have addressed the issue of generation of fully inflected words for the purpose of text authoring. Available open-source spell checking resources for Arabic are too small and inadequate. Ayaspell, for example, the official resource used with OpenOffice applications, contains only 300,000 fully inflected words. We try to bridge this critical gap by creating an adequate, open-source and large-coverage word list for Arabic containing 9,000,000 fully inflected surface words. Furthermore, from a large list of valid forms and invalid forms we create a character-based tri-gram language model to approximate knowledge about permissible character clusters in Arabic, creating a novel method for detecting spelling errors. Testing of this language model gives a precision of 98.2% at a recall of 100%. We take our research a step further by creating a context-independent spelling correction tool using a finite-state automaton that measures the edit distance between input words and candidate corrections, the Noisy Channel Model, and knowledge-based rules. Our system performs significantly better than Hunspell in choosing the best solution, but it is still below the MS Spell Checker.