United Arab Emirates

This paper introduces the Forgotten Realms Wiki (FRW) data set and uses it for domain specific natural language generation. The data set was extracted from the Forgotten Realms Fandom wiki dedicated to the eponymous Dungeons and Dragons (D\&amp;D) game setting. The FRW data set is constituted of 11 sub-data sets in varying degrees of pre-processing including plain text, annotated text, linked graphs, info-box dictionary, and several embedding models. This is the first data set of this size for the D\&amp;D domain. We then present a pairwise similarity comparison benchmark and perform D\&amp;D domain specific natural language gener- ation using the corpus and evaluate the named entity classification with respect to the lore of Forgotten Realms.

EMNLP 2022

Synthesis and Evaluation of a Domain-specific Large Data Set for Dungeons \&amp; Dragons

wikipedia data set

domain specific text generation

dungeons and dragons

## Welcome to EMNLP 2022!
I am delighted to welcome you to EMNLP 2022! I believe this conference has been complicated beyond any precedent. Over the past year, it’s been thrilling to see the organization team approach each new puzzle with creativity and enthusiasm. We hope that those participating in Abu Dhabi as well as those joining remotely will leave the conference feeling newly inspired by the program and newly connected to our ever-growing community. Following EMNLP 2021 and major NLP conferences since, EMNLP 2022 is “hybrid,” serving both virtual and in-person participants.

Our key innovations for EMNLP 2022 include:

* EMNLP 2022 is “hybrid” in a second sense, as well: we allowed both direct and rolling review paper submissions, building on the pilot experiment of EMNLP 2021, which considered a small number of ARR submissions. 
* Familiar from NAACL but new to EMNLP, we’ve added an industry track.
* During the conference, “portals” will link virtual poster sessions to in-person conference participants during poster sessions each day.
* The first ACL-family conference in the United Arab Emirates.

 *Message from Noah A. Smith, University of Washington and Allen Institute for AI, Seattle, Washington, USA* 
***EMNLP 2022 General Chair***
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/9eec7d4a287ee18c278b08229290aa83.png)](https://drive.google.com/file/d/1OlPv6QBeo62VVTughj2jkiLeyHd1WnUt/view)
 
[![](https://assets.underline.io/uploads/markdown_image/1/image/a3db7a768409f05192210d98601edb25.png)](https://emnlp2022.rocket.chat/)

To access this site you need to register. Please register [here](https://2022.emnlp.org/registration/).

Register here

Welcome!
EMNLP 2022 will take place in Abu Dhabi from December 7th to December 11th, 2022. And it will be held in hybrid mode, both online and offline.

This paper introduces the Forgotten Realms Wiki (FRW) data set and uses it for domain specific natural language generation. The data set was extracted from the Forgotten Realms Fandom wiki dedicated to the eponymous Dungeons and Dragons (D\&D) game setting. The FRW data set is constituted of 11 sub-data sets in varying degrees of pre-processing including plain text, annotated text, linked graphs, info-box dictionary, and several embedding models. This is the first data set of this size for the D\&D domain. We then present a pairwise similarity comparison benchmark and perform D\&D domain specific natural language gener- ation using the corpus and evaluate the named entity classification with respect to the lore of Forgotten Realms.

Synthesis and Evaluation of a Domain-specific Large Data Set for Dungeons \& Dragons

workshop paper

Automatic personality detection has many applications in different fields. In this paper, since there is no substantial study on Persian speakers' personality detection, we present an efficient and high-performance personality detection system in Persian language. We collected text data from 3781 Twitter users, embedded them using fasttext and fed them to a neural network which comprises LSTM and attention layers. Finally, our model demonstrated significant improvement in at least two traits. Also, we used LIME to interpret how a derivative of our model predicts individual personality.

MBTI Personality Prediction Approach on Persian Twitter

Hate and offensive speech on social media is targeted to attack an individual or group of community based on protected characteristics such as gender, ethnicity, and religion. Hate and offensive speech on social media is a global problem that suffers the community especially, for an under-resourced language like Afaan Oromo language. One of the most widely spoken Cushitic language families is Afaan Oromo. Our objective is to develop and test a model used to detect and classify Afaan Oromo hate speech on social media. We developed numerous models that were used to detect and classify Afaan Oromo hate speech on social media by using different machine learning algorithms (classical, ensemble, and deep learning) with the combination of different feature extraction techniques such as BOW, TF-IDF, word2vec, and Keras Embedding layers. To perform the task, we required Afaan Oromo datasets, but the datasets were unavailable. By concentrating on four thematic areas of hate speech, such as gender, religion, race, and offensive speech, we were able to collect a total of 12,812 posts and comments from Facebook. BiLSTM with pre-trained word2vec feature extraction is an outperformed algorithm that achieves better accuracy of 0.84 and 0.88 for eight classes and two classes, respectively.

Afaan Oromo Hate Speech Detection and Classification on Social Media

Manglish (Malaysian Colloquial English) is a type of creole\textbackslash{}footnote\{According to Wikipedia, a creole language is a stable natural language developed from a mixture of different languages.\} widely used in Malaysia. Like other creoles, Manglish is under-resourced and ignored in NLP literature. Manglish is made up of English as source language but also draws parts of its grammar and vocabulary from its parent languages such as Malay, Chinese (including mandarin and dialects) and Tamil. Due to the severe lack of resources, this work leverages on related-resources to train language models on Manglish. We look into two main categories of related resources: (1) available corpora and (2) existing pre-trained language models. For corpora, Singlish (Singaporean Colloquial English) and various code-switched corpora are used; as for pre-trained language models, existing models such as Singlish and other code-switched LMs involving its `parent languages' are exploited. Given that Manglish is made up of resource-rich parent languages, i.e. English and Chinese, efficient processing of Manglish can allow us to tap into language technologies developed for these highly resourced languages.

Exploiting Available Resources for the Training of Manglish Language Models

Machine translation for Sranantongo, a low-resource Creole language spoken predominantly in Surinam, is virgin territory. In this study we create a general purpose machine translation system for Sranantongo to and from Dutch. In order to facilitate this research, we introduce the SRNcorpus, a collection of parallel Dutch-Sranantongo and monolingual Sranantongo data. We experiment with a wide range of proven machine translation methods. Our best models show the feasibility of machine translation for Sranantongo.

Towards a general purpose machine translation system for Sranantongo

Word sense disambiguation (WSD) plays an important role, in increasing the performance of NLP applications such as information retrieval, question answering, and machine translation. The manual disambiguation process by humans is tedious, prone to errors, expensive, and time-consuming. Low-resource languages such as Amharic do not have WordNet which makes the WSD task challenging. Previous works try to identify and develop Amharic word sense disambiguation using different approaches but still, there are no easy and automatic WSD tools.  Recent research uses handcrafted rules or directly fetching the meaning of an ambiguous word from the synset list in WordNet, but did not learn different representations from data automatically.. 
For this study, we collect the WSD dataset and employ contextual embeddings, namely AmRoBERTa to automatically disambiguate words to their proper sense. We employ CNN, Bi-LSTM, and BERT for the classification task and attain 90\%, 88\%, and 93\%  accuracy respectively. Based on our findings, BERT has better accuracy results than others.
When we use the masking technique of the pre-trained contextual embedding to find the correct sense, it attains 70\textbackslash{}\% accuracy. However, when we use the pre-trained contextual model and glosses to disambiguate the correct sense with embedding similarity, it achieves 71\textbackslash{}\% accuracy. The main contributions of our work are the release of new datasets and exploring the use of pre-trained contextual embeddings for the Amharic word sense disambiguation model.

Transfer Learning and Word Sense Disambiguation for Low-resource Language, the Case of Amharic

Gender subspace analysis, or using word vector representations to measure gender bias, is useful to study bias in machine learning models. However, direct application of gender subspace is challenging when the domain is shifted and when the size of a dataset is smaller. This study reports our qualitative and quantitative analyses of the learned gender subspace in a workplace communication dataset. We highlight how word sparsity and homonyms (i.e. words that have multiple semantic meaning) affect word embeddings, leading to degraded performance on subspace learning. We then propose word replacement and substitution to exclude non-gendered information from word embeddings in order to improve the robustness of gender subspace learned in domain-specific dataset for empirical gender bias analysis.

Boosting the Performance of Gender Subspace in Domain-Specific Gender Bias Analysis

Machine Translation (MT) is a sub-field of NLP that investigates how to use computer software to automatically translate text or speech from one language to another without human involvement. Nowadays, using deep neural networks for MT tasks has received a great deal of attention. This paper presents the first relatively large-scale Amharic-English parallel sentence dataset. Using these compiled data, we build bi-directional Amharic-English translation models by fine-tuning Facebook's multilingual pre-trained model achieving a BLEU score of 37.79 in Amharic-English translation and 32.74 in English-Amharic translation. Additionally, we explore the effects of Amharic homophone normalization on the MT task. The results show that normalization of Amharic homophone characters increases the performance of Amharic-English MT in both directions.

The Effect of Normalization for Bi-directional Amharic-English Neural Machine Translation

This paper presents an Amharic hate speech annotation using the Yandex Toloka crowdsourcing platform. The dataset is collected from 5 consecutive years 2018-2022, following some controversial events in Ethiopia that put the country in violence. We consider one-month data starting from the event date in each year. Accordingly, we annotate 5,400 tweets, nearly 1k tweets for each year. We explore the main challenges of crowdsourcing annotation for Amharic hate speech data collection using Toloka. We attain a Fleiss kappa score of 0.34 using three independent annotators that annotate the tweets where the gold label is determined using majority voting. Using the datasets, we build classification models using deep learning (LSTM and BiLSTM) and achieved a 0.49 F1-score for both models.  We will publicly release the dataset, source code, and models with a permissive license.

Challenges of Amharic Hate Speech Data Annotation Using Yandex Toloka Crowdsourcing Platform

In this work, we build a Question classification (QC) dataset from a social media platform, namely the Telegram public channel called @AskAnythingEthiopia. The platform allows asking questions that belong to various domains, like Politics, Music, Technology, Religion and so on. Questions are posted in Amharic, English, or Amharic in Latin script. Since the questions are posed in a mixed-code, we apply different strategies to pre-process the dataset. As part of the pre-processing tools, we build a Latin-to-Ethiopic-Script transliteration tool. We collect 8k Amharic and 24K Amharic but written in Latin script questions and develop deep learning-based question answering classifiers that attain an F-score of 57.79 in 20 different question categories. The datasets and pre-processing scripts are open-sourced to facilitate further research on the Amharic community-based question answering.

Question Answering Classification for Amharic Social Media Community Based Questions

In this ongoing work, we investigate a potential bias to associate terms related to ``female'' genitalia to reproduction, and terms related to ``male'' genitalia to sexual pleasure. While some of our suspicions are confirmed (stronger association of "female" genitalia to reproduction than to sexual pleasure), others are not (for example, our current data does not suggest a stronger association of reproduction to ``female'' genitalia than to ``male'' genitalia). We plan to continue this line of work by including more languages to get a better overview of the situation in languages other than English. Furthermore, we intend to critically assess cosine similarity as a proxy for association, as well as looking into (statistical) significance of the differences in cosine similarity that we found, and the effects of the skewed frequency distributions of the terms related to ``female'' and ``male'' genitals, and reproduction and sexual pleasure.

Downloads

Next from EMNLP 2022

MBTI Personality Prediction Approach on Persian Twitter

Similar lecture

Dungeons and Dragons as a Dialog Challenge for Artificial Intelligence