United Arab Emirates

Garden path sentences (i.e. “the horse raced past the barn fell”) are sentences that readers initially incorrectly parse, requiring partial or total re-analysis of the sentence structure. Given human difficulty in parsing garden paths, we aim to compare transformer language models’ performance on these sentences. We assess a BERT question-answering model’s performance on comprehension questions based on garden path and control sentences. We then further investigate the semantic roles assigned to arguments of verbs in garden path and control sentences by utilizing a probe task to directly assess which semantic role(s) BERT assigns. We find that BERT has relatively low performance in question answering based on garden path contexts, and the model incorrectly assigns semantic roles, aligning for the most part with human performance.

EMNLP 2022

The BERT Walked Down the Garden Path Assigned Semantic Roles

garden path sentences

probe task

model analysis

question answering

bert

## 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.

workshop paper

Automatic Speech Recognition (ASR) systems have become a very natural human-machine interaction in which it allows users to speak entries rather than punching numbers on a keypad. This study presents an ASR system for Amharic language that is one of the Ethiopian Languages. The current attempts on speech recognition systems require thousands of hours of transcribed speech dataset to reach adequate accuracy. However, the large majority of under-resourced spoken languages in general, Ethiopic languages in particular, have a very limited amount of labeled speech dataset. Thus, in this study, a self-supervised Transformer based Wave2Vec 2.0 approach has been conducted to build an ASR system. In order to pretrain the proposed model in an unsupervised approach, a total of more than 200 hours unlabeled speech have been collected. In addition, a total of 30 minutes of labeled speech dataset has been prepared. Then, a Wave2Vec model that is initialized with weighted parameters has been pre-trained on the unlabeled speech 
dataset. Then the pretrained model has been fine-tuned using a small amount of labeled speech dataset. A standard Word Error Rate (WER) evaluation metric has been used to evaluate the fine-tuned multilingual ASR model and it has 
shown a clear significant result.

Automatic Speech Recognition for Amharic Language using Self-Supervised

The main goal of this work is to develop sentence level automatic speech segmentation system for Amharic. Sentence segmentation is a process of identifying the end of a sentence. In this study, sentence segmentation system is 
implemented in to two approaches. In the first approach, we used an automatic tool for segmenting and labeling of Amharic speech data. Acoustic model is created using 
speech and their text scripts and compiling them into a statistical representation of sounds which makeup words. This is done through HMM modeling. The approach one automatic speech segmentation system is done by forced alignment. In this approach we used rule-based and AdaBoost to discriminate the true boundaries from false. In the second approach, we extracted prosodic features directly 
from speech waveform and also statistical method, AdaBoost, is used.

Prosody Based Automatic Speech segmentation for Amharic

The way sexual abuse is represented in the media can drive public opinion. We use NLP techniques to analyze representation of sexual abuse in New York Times (NYT) articles over three time periods, ending with the era of MeToo. We use computational methods to surface sentences in articles involving victims and perpetrators, which we annotate for victim and perpetrator demographic information. We use power and agency connotation frames to measure if and how ascribed power and agency of victims and perpetrators in news articles have changed over the years. We find that women are ascribed more agency in sentences describing abuse of women victims over time, and power ascribed to male perpetrators has decreased over time.

Before and beyond MeToo: Measuring changes in power and agency within sexual abuse news stories over time

In this work, we propose a flexible multimodal transformer architecture for classifying depression from social media posts containing text and images. We show that our model is able to surpass other methods by a large margin on a popular multimodal depression dataset from Twitter, obtaining state-of-the-art results of 95\% F1 score. Our model is flexible and can easily incorporate temporal information by manipulating the transformer positional encodings. Moreover, it can operate both on randomly sampled and unordered sets of posts, to be more robust to dataset noise, and on ordered collections of posts, accommodating the relative posting intervals without any major architectural modifications.

Detecting Depression on Twitter with a Time-Aware Multimodal Transformer

Out-of-Vocabulary (OOV) is a problem for Neural Machine Translation (NMT). OOV refers to words with a low occurrence in the training data or those absent from the training data. Word or phrase-based Data Augmentation (DA) techniques have been used to alleviate this. However, existing DA techniques have addressed only one of these OOV types and are limited to considering either syntactic or semantic constraints. We present a word/phrase replacement-based DA technique that considers both types of OOV, considering both syntactic and semantic properties of the words to guarantee fluency in the synthetic sentences. This technique was experimented with low resource language pair Sinhala-English.

Data Augmentation to Address Out-of-Vocabulary Problem in Low-Resource Sinhala-English Neural Machine Translation

Microaggressions (MAs) are subtle, often unintentional, form of prejudice. They often impact people of minority or marginalized backgrounds. Gender MAs are common in India but not well understood. To the best of my knowledge, there is no existing MAs dataset in Hindi. To tackle the problem of MAs, a dataset needs to be created with which a machine learning model can be trained. Enabling a model to learn MAs would allow for MAs detection and possibly correction in text. In this paper, I build (i) a small gender MAs dataset and crowdsource its labeling, and (ii) a novel pipeline to detect MAs in Hindi.

Tackling Gender Microaggressions in Hindi Text

Lexicon-based method for sentiment analysis provides explainable results compared to deep learning methods. However, the approach does not capture contextual information and provides sub-optimal performance. In this paper, we build domain-specific lexicons and shifter patterns that capture contextual information. We show that symbolic approaches can achieve competitive performance as deep neural networks for sentiment classification tasks with limited data. Our approach is useful in cases where an explanation is the most important criterion and not the accuracy. 



Domain-Specific Lexicon-Based Sentiment Analysis using Contextual Shifter Patterns

Large language models appear to learn facts from the large text corpora they are trained on. Such facts are encoded implicitly within their many parameters, making it difficult to verify or manipulate what knowledge has been learned. Language models have recently been extended to multilingual language models (MLLMs), enabling knowledge to be learned across hundreds of languages. Meanwhile, knowledge graphs contain facts in an explicit triple format, which require careful and costly curation and are only available in a few high-resource languages, restricting their research and application. To address these issues, we propose to enhance MLLMs with knowledge from multilingual knowledge graphs (MLKGs) so as to tackle language and knowledge graph tasks across many languages, including low-resource ones. Specifically, we introduce a lightweight adapter set to enhance MLLMs with cross-lingual entity alignment and facts from MLKGs for many languages. Experiments on common benchmarks show that such enhancement benefits both MLLMs and MLKGs, achieving: (1) comparable or improved performance for knowledge graph completion and entity alignment relative to baselines, especially for low-resource languages (for which knowledge graphs are unavailable); and (2) improved MLLM performance on language understanding tasks that require multilingual factual knowledge; all while maintaining performance on other general language tasks.

Adapters for Enhanced Modeling of Multilingual Knowledge and Text

Bilingual lexicon induction (BLI) with limited bilingual supervision is a crucial yet challenging task in multilingual NLP. Current state-of-the-art BLI methods rely on the induction of cross-lingual word embeddings (CLWEs) to capture cross-lingual word similarities; such CLWEs are obtained 1) via traditional static models (e.g., VecMap), or 2) by extracting type-level CLWEs from multilingual pretrained language models (mPLMs), or 3) through combining the former two options. In this work, we propose a novel semi-supervised post-hoc reranking method termed BLICEr (BLI with Cross-Encoder Reranking), applicable to any precalculated CLWE space, which improves their BLI capability. The key idea is to 'extract' cross-lingual lexical knowledge from mPLMs, and then combine it with the original CLWEs. This crucial step is done via 1) creating a word similarity dataset, comprising positive word pairs (i.e., true translations) and hard negative pairs induced from the original CLWE space, and then 2) fine-tuning an mPLM (e.g., mBERT or XLM-R) in a cross-encoder manner to predict the similarity scores. At inference, we 3) combine the similarity score from the original CLWE space with the score from the BLI-tuned cross-encoder. BLICEr establishes new state-of-the-art results on two standard BLI benchmarks spanning a wide spectrum of diverse languages: it substantially outperforms a series of strong baselines across the board. We also validate the robustness of BLICEr with different CLWEs. Our code is available at https://github.com/cambridgeltl/BLICEr.

Improving Bilingual Lexicon Induction with Cross-Encoder Reranking

In an effort to study the inductive biases of language models, numerous studies have attempted to use linguistically motivated tasks as a proxy of sorts, wherein performance on these tasks would imply an inductive bias towards a specific linguistic phenomenon. In this study, we attempt to analyse the inductive biases of language models with respect to natural language phenomena
in the context of building multilingual embedding spaces.
We sample corpora from 2 sources in 15 languages and train language models on pseudo-bilingual variants of each corpus, created by duplicating each corpus and shifting token indices for half the resulting corpus. We evaluate the cross-lingual capabilities of these LMs, and show that while correlations with language families tend to be weak, other corpus-level characteristics, such as type-token ratio, tend to be more strongly correlated. Finally, we show that multilingual spaces can be built, albeit less effectively, even when additional destructive perturbations are applied to the training corpora, implying that (effectively) bag-of-words models also have an inductive bias that is sufficient for inducing multilingual spaces.

Next from EMNLP 2022

Automatic Speech Recognition for Amharic Language using Self-Supervised

Similar lecture

The BERT Walked Down the Garden Path Assigned Semantic Roles