Spain

During the last two decades, we have progressively turned to the Internet and social media to
find news, entertain conversations and share opinion. Recently, OpenAI has developed a machine
learning system called GPT-2 for Generative Pre-trained Transformer-2, which can produce
deepfake texts. It can generate blocks of text based on brief writing prompts that look
like they were written by humans, facilitating the spread false or auto-generated text. In line
with this progress, and in order to counteract potential dangers, several methods have been proposed
for detecting text written by these language models. In this paper, we propose a transfer
learning based model that will be able to detect if an Arabic sentence is written by humans or
automatically generated by bots. Our dataset is based on tweets from a previous work, which
we have crawled and extended using the Twitter API. We used GPT2-Small-Arabic to generate
fake Arabic Sentences. For evaluation, we compared different recurrent neural network (RNN)
word embeddings based baseline models, namely: LSTM, BI-LSTM, GRU and BI-GRU, with a
transformer-based model. Our new transfer-learning model has obtained an accuracy up to 98%.
To the best of our knowledge, this work is the first study where ARABERT and GPT2 were
combined to detect and classify the Arabic auto-generated texts.

COLING 2020

BERT Transformer model for Detecting Arabic GPT2 Auto-Generated Tweets

auto-generated text detection

deep fake text

arabic language

gpt2

arabert

social media

stages 

COLING, the International Conference on Computational Linguistics, is one of the premier conferences for natural language processing and computational linguistics. Often grouped within the field of artificial intelligence, but actually pre-dating the development of artificial intelligence, advances in computational linguistics and natural language processing are now some of the major drivers behind the use of artificial intelligence for commercial and social applications – for example, on-line search, machine translation and with voice-assisted conversational devices.

First established in 1965, the biennial COLING conference is held in diverse parts of the globe and attracts participants from both top-ranked research centers and emerging countries. Today, the most important developments in our field are taking place not only in universities and academic research institutes, but also in industrial research departments and in technological startups. COLING conferences provide opportunities for all these communities to showcase their exciting developments.

COLING, the International Conference on Computational Linguistics, is one of the premier conferences for natural language processing and computational linguistics. Often grouped within the field of artificial intelligence, but actually pre-dating the development of artificial intelligence, advances in computational linguistics and natural language processing are now some of the major drivers behind the use of artificial intelligence for commercial and social applications – for example, on-line search, machine translation and with voice-assisted conversational devices.

poster

Globalization has caused the rise of the code-switching phenomenon among multilingual societies. In Arab countries, code-switching between Arabic and English has become frequent, especially through social media platforms. Consequently, research in Natural Language Processing (NLP) systems increased to tackle such a phenomenon. One of the significant challenges of developing code-switched NLP systems is the lack of data itself. We propose an open source trained bilingual contextual word embedding models of FLAIR, BERT, and ELECTRA.
We also propose a novel contextual word embedding model called KERMIT, which can efficiently map Arabic and English words inside one vector space in terms of data usage. We applied intrinsic and extrinsic evaluation methods to compare the performance of the models. Our results show that FLAIR and FastText achieve the highest results in the sentiment analysis task. However, KERMIT is the best-achieving model on the intrinsic evaluation and named entity recognition. Also, it outperforms the other transformer-based models on question answering task.

Contextual Embeddings for Arabic-English Code-Switched Data

Automatic categorization of short texts, such as news headlines and social media posts, has many applications ranging from content analysis to recommendation systems. In this paper, we use such text categorization i.e., labeling the social media posts to categories like `sports', `politics', `human-rights' among others, to showcase the efficacy of models across different sources and varieties of Arabic. In doing so, we show that diversifying the training data, whether by using diverse training data for the specific task (an increase of 21% macro F1) or using diverse data to pre-train a BERT model (26% macro F1), leads to overall improvements in classification effectiveness.
In our work, we also introduce two new Arabic text categorization datasets, where the first is composed of social media posts from a popular Arabic news channel that cover Twitter, Facebook, and YouTube, and the second is composed of tweets from popular Arabic accounts. The posts in the former are nearly exclusively authored in modern standard Arabic (MSA), while the tweets in the latter contain both MSA and dialectal Arabic.

Improving Arabic Text Categorization Using Transformer Training Diversification

Multi-sentence questions (MSQs) are sequences of questions connected by relations which, unlike sequences of standalone questions, need to be answered as a unit. Following Rhetorical Structure Theory (RST), we recognise that different ``question discourse relations'' between the subparts of MSQs reflect different speaker intents, and consequently elicit different answering strategies. Correctly identifying these relations is therefore a crucial step in automatically answering MSQs. We identify five different types of MSQs in English, and define five novel relations to describe them. We extract over 162,000 MSQs from Stack Exchange to enable future research. Finally, we implement a high-precision baseline classifier based on surface features.

Querent Intent in Multi-Sentence Questions

The development of linguistic corpora is fraught with various problems of annotation and representation. These constitute a very real challenge for the development and use of annotated corpora, but as yet not much literature exists on how to address the underlying problems. In this paper, we identify and discuss five sources of representation problems, which are independent though interrelated: ambiguity, variation, uncertainty, error and bias. We outline and characterize these sources, discussing how their improper treatment can have stark consequences for research outcomes. Finally, we discuss how an adequate treatment can inform corpus-related linguistic research, both computational and theoretical, improving the reliability of research results and NLP models, as well as informing the more general reproducibility issue.

Representation Problems in Linguistic Annotations: Ambiguity, Variation, Uncertainty, Error and Bias

This paper reports on the harvesting, analysis, and annotation of 20k documents from 4 different endangered language archives in 280 different low-resource languages. The documents are heterogeneous as to their provenance (holding archive, language, geographical area, creator) and internal structure (annotation types, metalanguages), but they have the ELAN-XML format in common. Typical annotations include sentence-level translations, morpheme-segmentation, morpheme-level translations, and parts-of-speech. The ELAN format gives a lot of freedom to document creators, and hence the data set is very heterogeneous. We use regularities in the ELAN format to arrive at a common internal representation of sentences, words, and morphemes, with translations into one or more additional languages. Building upon the paradigm of Linguistic Linked Open Data (LLOD, Chiarcos et al 2012), the document elements receive unique identifiers and are linked to other resources such as Glottolog for languages, Wikidata for semantic concepts, and the Leipzig Glossing Rules list for category abbreviations. We provide an RDF export in the LIGT format (Chiarcos & Ionov 2019), enabling uniform and interoperable access with some semantic enrichments to a formerly disparate resource type difficult to access. Two use cases (semantic search and colexification) are presented to show the viability of the approach.

Modelling and annotating interlinear glossed text from 280 different endangered languages as Linked Data with LIGT

It has become increasingly common for people to share cooking recipes on the Internet. Along with the increase in the number of shared recipes, there have been corresponding increases in recipe-related studies and datasets. However, there are still few datasets that provide linguistic annotations for the recipe-related studies even though such annotations should form the basis of the studies. This paper introduces a novel recipe-related dataset, named Cookpad Parsed Corpus, which contains linguistic annotations for Japanese recipes. We randomly extracted 500 recipes from the largest recipe-related dataset, the Cookpad Recipe Dataset, and annotated 4, 738 sentences in the recipes with morphemes, named entities, and dependency relations. This paper also reports benchmark results on our corpus for Japanese morphological analysis, named entity recognition, and dependency parsing. We show that there is still room for improvement in the analyses of recipes.

Cookpad Parsed Corpus: Linguistic Annotations of Japanese Recipes

Most annotation efforts assume that annotators will agree on labels, if the annotation categories are well-defined and documented in annotation guidelines. However, this is not always true. For instance, content-related questions such as ‘Is this sentence about topic X?’ are unlikely to elicit the same answer from all annotators. Additional specifications in the guidelines are helpful to some extent, but can soon get overspecified by rules that cannot be justified by a research question. In this study, we model the semantic category ‘illness’ and its use in a gradual way. For this purpose, we (i) ask many annotators (30 votes per item, 960 items) for their opinion in a crowdsourcing experiment, (ii) ask annotators to indicate their certainty with respect to their annotation, and (iii) compare this across two different text types. We show that results of multiple annotations and average annotator certainty correlate, but many ambiguities can only be captured if several people contribute. The annotated data allow us to filter for sentences with high or low agreement and analyze causes of disagreement, thus getting a better understanding of people’s perception of illness—as an example of a semantic category—as well as of the content of our annotated texts.

Modeling Ambiguity with Many Annotators and Self-Assessments of Annotator Certainty

Generating expert ground truth annotations of documents can be a very expensive process. However, such annotations are essential for training domain-specific keyphrase extraction models, especially when utilizing data-intensive deep learning models in unique domains such as real-estate. Therefore, it is critical to optimize the manual annotation process to maximize the quality of the annotations while minimizing the cost of manual labor. To address this need, we explore multiple annotation strategies including self-review and peer- review as well as various methods of resolving annotator disagreements. We evaluate these annotation strategies with respect to their cost and on the task of learning keyphrase extraction models applied with an experimental dataset in the real-estate domain. The results demonstrate that different annotation strategies should be considered depending on specific metrics such as precision and recall.

Understanding the Tradeoff between Cost and Quality of Expert Annotations for Keyphrase Extraction

This paper describes a novel annotation scheme specifically designed for a customer-service context where written interactions take place between a given user and the chatbot of an Italian telecommunication company. More specifically, the scheme aims to detect and highlight two aspects: the presence of errors in the conversation on both sides (i.e. customer and chatbot) and the “emotional load” of the conversation. This can be inferred from the presence of emotions of some kind (especially negative ones) in the customer messages, and from the possible empathic
responses provided by the agent. The dataset annotated according to this scheme is currently used to develop the prototype of a rule-based Natural Language Generation system aimed at improving the chatbot responses and the customer experience overall.

Annotating Errors and Emotions in Human-Chatbot Interactions in Italian

pyMMAX2 is an API for processing MMAX2 stand-off annotation data in Python. It provides a lightweight basis for the development of code which opens up the Java- and XML-based ecosystem of MMAX2 for more recent, Python-based NLP and data science methods. While pyMMAX2 is pure Python, and most functionality is implemented from scratch, the API re-uses the complex implementation of the essential business logic for MMAX2 annotation schemes by interfacing with the original MMAX2 Java libraries. pyMMAX2 is available for download at http://github.com/nlpAThits/pyMMAX2.

Downloads

Next from COLING 2020

Contextual Embeddings for Arabic-English Code-Switched Data

Similar lecture

Improving Arabic Text Categorization Using Transformer Training Diversification

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from COLING 2020

Contextual Embeddings for Arabic-English Code-Switched Data

Similar lecture

Improving Arabic Text Categorization Using Transformer Training Diversification

Downloads