United Arab Emirates

A young child&#39;s vocabulary size is correlated with their level of personal wellbeing and future academic success. Yet, interventions aimed at increasing early vocabulary would ideally be tailored to each individual child&#39;s needs and interests, and such personalization would be impossible without technological support. Here, we explore if and how natural language processing can be used to create individualized bedtime stories around target words to be learned by preschoolers. Generating stories from scratch is challenging and often results in stories of low quality. Thus, we propose an alternative approach: completing phrase-level gaps within prewritten stories. On this task, we explore the performance of GPT-3 with and without finetuning as well as with and without providing a word which is semantically related to the target word. Manual evaluation of the generated stories shows that GPT-3 and GPT-3-based models perform well on the task. Using GPT-3 without finetuning and including a context word into the prompt is the best performing approach.

EMNLP 2022

Generate Me a Bedtime Story: Leveraging Natural Language Processing for Early Vocabulary Enhancement

education generation gpt-3

A young child's vocabulary size is correlated with their level of personal wellbeing and future academic success. Yet, interventions aimed at increasing early vocabulary would ideally be tailored to each individual child's needs and interests, and such personalization would be impossible without technological support. Here, we explore if and how natural language processing can be used to create individualized bedtime stories around target words to be learned by preschoolers. Generating stories from scratch is challenging and often results in stories of low quality. Thus, we propose an alternative approach: completing phrase-level gaps within prewritten stories. On this task, we explore the performance of GPT-3 with and without finetuning as well as with and without providing a word which is semantically related to the target word. Manual evaluation of the generated stories shows that GPT-3 and GPT-3-based models perform well on the task. Using GPT-3 without finetuning and including a context word into the prompt is the best performing approach.

workshop paper

## 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.

In this paper, we introduce a conversational agent (chatbot) for Hindi-speaking youth called BELA—Bot for English Language Acquisition. Developed for young underprivileged students at an Indian non-profit, the agent supports both Hindi and Hinglish (code-switched Hindi and English, written primarily with English orthography) utterances. BELA has two interaction modes: a question-answering mode for classic English language learning tasks like word meanings, translations, reading passage comprehensions, etc., and an open-domain dialogue system mode to allow users to practice their language skills.

We present a high-level overview of the design of BELA, including the implementation details and the preliminary results of our early prototype. We also report the challenges in creating an English-language learning chatbot for a largely Hindi-speaking population.

BELA: Bot for English Language Acquisition

The various potential of children can be limited by language delay or language impairments. However, there are many instances where parents are unaware of the child's condition and do not obtain appropriate treatment as a result.  Additionally, experts collecting children's utterance to establish norms of language tests and evaluating children's language development level takes a significant amount of time and work. To address these issues, dependable automated screening tools are required. In this paper, we used pretrained LM to assist experts in quickly and objectively screening the language development level of children. Here, evaluating the language development level is to ensure that the child has the appropriate language abilities for his or her age, which is the same as the child's age. To do this, we analyzed the utterances of children according to age. Based on these findings, we use the standard deviations of the pretrained LM's probability as a score for children to screen their language development level. The experiment results showed very strong correlations between our proposed method and the Korean language test REVT (REVT-R, REVT-E), with Pearson correlation coefficient of 0.9888 and 0.9892, respectively.

Applicability of Pretrained Language Models: Automatic Screening for Children's Language Development Level

Traditional sentiment analysis is a sentence level or document-level task. However, a sentence or paragraph may contain multiple target terms with different sentiments, making sentiment prediction more challenging. Although pre-trained language models like BERT have been successful, incorporating dynamic semantic changes into aspect-based sentiment models remains difficult, especially for domain-specific sentiment analysis. To this end, in this paper, we propose a Term-Based Sentiment Analysis (TBSA), a novel method designed to learn Environmental, Social, and Governance (ESG) contexts based on a sustainability taxonomy for ESG aspect-oriented sentiment analysis. Notably, we introduce a technique enhancing the ESG term's attention, inspired by the success of attention-based neural networks in machine translation (Bahdanau et al., 2015) and Computer Vision (Bello et al., 2019). It enables the proposed model to focus on a small region of the sentences at each step and to reweigh the crucial terms for a better understanding of the ESG aspect-aware sentiment. Beyond the novelty in the model design, we propose a new dataset of 125,000+ ESG analyst annotated data points for sustainability term based sentiment classification, which derives from historical sustainability corpus data and expertise acquired by development finance institutions. Our extensive experiments combining the new method and the new dataset demonstrate the effectiveness of the Sustainability TBSA model with an accuracy of 91.30\% (90\% F1-score). Both internal and external business applications of our model show an evident potential for a significant positive impact toward furthering sustainable development goals (SDGs).

Transformers-Based Approach for a Sustainability Term-Based Sentiment Analysis (STBSA)

Hateful memes are a growing menace on social media. While the image and its corresponding text in a meme are related, they do not necessarily convey the same meaning when viewed individually. Hence, detecting hateful memes requires careful consideration of both visual and textual information. Multimodal pre-training can be beneficial for this task because it effectively captures the relationship between the image and the text by representing them in a similar feature space. Furthermore, it is essential to model the interactions between the image and text features through intermediate fusion. Most existing methods either employ multimodal pre-training or intermediate fusion, but not both. In this work, we propose the Hate-CLIPper architecture, which explicitly models the cross-modal interactions between the image and text representations obtained using Contrastive Language-Image Pre-training (CLIP) encoders via a feature interaction matrix (FIM). A simple classifier based on the FIM representation is able to achieve state-of-the-art performance on the Hateful Memes Challenge (HMC) dataset with an AUROC of 85.8, which even surpasses the human performance of 82.65. Experiments on other meme datasets such as Propaganda Memes and TamilMemes also demonstrate the generalizability of the proposed approach. Finally, we analyze the interpretability of the FIM representation and show that cross-modal interactions can indeed facilitate the learning of meaningful concepts. The code for this work is available at https://github.com/gokulkarthik/hateclipper

Hate-CLIPper: Multimodal Hateful Meme Classification based on Cross-modal Interaction of CLIP Features

In this paper, we describe a participatory, collaborative, and cost-effective process for creating translations in Oshiwambo, the most widely spoken African language in Namibia. We aim to (1) build a resource for language technology development, (2) bridge generational gaps in cultural and language knowledge, and at the same time (3) provide socio-economic opportunities through language preservation. The created data spans diverse topics of cultural importance, and comprises over 7.5k sentences written in the Oshindonga dialect and translated to English, the largest parallel corpus for Oshiwambo to-date. We show that it is very effective for machine translation, especially when combined with transfer learning.

Participatory Translations of Oshiwambo: Towards Culture Preservation with Language Technology

In the age of social media, where billions of internet users share information and opinions, the negative impact of pandemics is not limited to the physical world. It provokes a surge of incomplete, biased, and incorrect information, also known as an infodemic. This global infodemic jeopardizes measures to control the pandemic by creating panic, vaccine hesitancy, and fragmented social response. Platforms like Facebook allow advertisers to adapt their messaging to target different demographics and help alleviate or exacerbate the infodemic problem depending on their content. In this paper, we propose a minimally supervised multi-task learning framework for understanding messaging on Facebook related to the covid vaccine by identifying ad themes and moral foundations. Furthermore, we perform a more nuanced thematic analysis of messaging tactics of vaccine campaigns on social media so that policymakers can make better decisions on pandemic control.

Understanding COVID-19 Vaccine Campaign on Facebook using Minimal Supervision

Gender Bias in Meta-Embeddings

Challenges and Opportunities in Information Manipulation Detection: An Examination of Wartime Russian Media

Encoded text representations often capture sensitive attributes about individuals (e.g., race or gender), which raise privacy concerns and can make downstream models unfair to certain groups. In this work, we propose FEDERATE, an approach that combines ideas from differential privacy and adversarial training to learn private text representations which also induces fairer models. We empirically evaluate the trade-off between the privacy of the representations and the fairness and accuracy of the downstream model on four NLP datasets. Our results show that FEDERATE consistently improves upon previous methods, and thus suggest that privacy and fairness can positively reinforce each other.

Fair NLP Models with Differentially Private Text Encoders

An increasingly prevalent problem for intelligent technologies is text safety, as uncontrolled systems may generate recommendations to their users that lead to injury or life-threatening consequences. However, the degree of explicitness of a generated statement that can cause physical harm varies. In this paper, we distinguish types of text that can lead to physical harm and establish one particularly underexplored category: covertly unsafe text. Then, we further break down this category with respect to the system's information and discuss solutions to mitigate the generation of text in each of these subcategories. Ultimately, our work defines the problem of covertly unsafe language that causes physical harm and argues that this subtle yet dangerous issue needs to be prioritized by stakeholders and regulators. We highlight mitigation strategies to inspire future researchers to tackle this challenging problem and help improve safety within smart systems.

Generate Me a Bedtime Story: Leveraging Natural Language Processing for Early Vocabulary Enhancement

Next from EMNLP 2022

BELA: Bot for English Language Acquisition

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES