Dominican Republic

Masked language modeling (MLM), a self-supervised pretraining objective, is widely used in natural language processing for learning text representations. MLM trains a model to predict a random sample of input tokens that have been replaced by a [MASK] placeholder in a multi-class setting over the entire vocabulary. When pretraining, it is common to use alongside MLM other auxiliary objectives on the token or sequence level to improve downstream performance (e.g. next sentence prediction). However, no previous work so far has attempted in examining whether other simpler linguistically intuitive or not objectives can be used standalone as main pretraining objectives. In this paper, we explore five simple pretraining objectives based on token-level classification tasks as replacements of MLM. Empirical results on GLUE and SQuAD show that our proposed methods achieve comparable or better performance to MLM using a BERT-BASE architecture. We further validate our methods using smaller models, showing that pretraining a model with 41% of the BERT-BASE&#39;s parameters, BERT-MEDIUM results in only a 1% drop in GLUE scores with our best objective.

EMNLP 2021

Frustratingly Simple Pretraining Alternatives to Masked Language Modeling

EMNLP 2021 is planned to be a hybrid event in Punta Cana, Dominican Republic, with both on-site and fully virtual participation possible. The experience for on-site participants would closely approximate a normal pre-COVID *ACL conference, with 5-6 thematically organized parallel sessions and live Q/A and interactive discussion immediately after the talks. Presentations by virtual participants will be equitably interleaved with those of on-site participants, projected on the auditorium screens as if on-site, and also followed immediately by live Q/A and interactive discussion at a time during reasonable waking hours for the virtual presenter. For all participants, on-site and virtual, who are unable to attend a session due to either time-zone issues or because they are participating in another session live, talk recordings and slides will be available online at a minimum after the live presentation (and in many cases before as well), and questions may be submitted in advance on session-specific discussion boards and answered live in session with the usual visual aids if desired.

<iframe style="width:700px;height:400px" src="https://online.fliphtml5.com/ebtyf/ceby/"  seamless="seamless" scrolling="no" frameborder="0" allowtransparency="true" allowfullscreen="true" ></iframe>

Please Note: The EMNLP registration system is not currently connected to the underline site as we are still in the process of building out EMNLP 2021. You will receive access instructions from underline the week of November 1st. 

Access is given only to EMNLP upon registration, if you have not registered please do so [here](https://2021.emnlp.org/registration).

Registered attendees will receive access the week of November 1st.

EMNLP 2021 is planned to be a hybrid event in Punta Cana, Dominican Republic, with both on-site and fully virtual participation possible.

Masked language modeling (MLM), a self-supervised pretraining objective, is widely used in natural language processing for learning text representations. MLM trains a model to predict a random sample of input tokens that have been replaced by a [MASK] placeholder in a multi-class setting over the entire vocabulary. When pretraining, it is common to use alongside MLM other auxiliary objectives on the token or sequence level to improve downstream performance (e.g. next sentence prediction). However, no previous work so far has attempted in examining whether other simpler linguistically intuitive or not objectives can be used standalone as main pretraining objectives. In this paper, we explore five simple pretraining objectives based on token-level classification tasks as replacements of MLM. Empirical results on GLUE and SQuAD show that our proposed methods achieve comparable or better performance to MLM using a BERT-BASE architecture. We further validate our methods using smaller models, showing that pretraining a model with 41% of the BERT-BASE's parameters, BERT-MEDIUM results in only a 1% drop in GLUE scores with our best objective.

poster

Text-Based Games (TBGs) have emerged as important testbeds for reinforcement learning (RL) in the natural language domain. Previous methods using LSTM-based action policies are uninterpretable and often overfit the training games showing poor performance to unseen test games. We present SymboLic Action policy for Textual Environments (SLATE), that learns interpretable action policy rules from symbolic abstractions of textual observations for improved generalization. We outline a method for end-to-end differentiable symbolic rule learning and show that such symbolic policies outperform previous state-of-the-art methods in text-based RL for the coin collector environment from 5-10x fewer training games. Additionally, our method provides human-understandable policy rules that can be readily verified for their logical consistency and can be easily debugged.

Neuro-Symbolic Approaches for Text-Based Policy Learning

Inspired by mutual information (MI) based feature selection in SVMs and logistic regression, in this paper, we propose MI-based layer-wise pruning: for each layer of a multi-layer neural network, neurons with higher values of MI with respect to preserved neurons in the upper layer are preserved. Starting from the top softmax layer, layer-wise pruning proceeds in a top-down fashion until reaching the bottom word embedding layer. The proposed pruning strategy offers merits over weight-based pruning techniques: (1) it avoids irregular memory access since representations and matrices can be squeezed into their smaller but dense counterparts, leading to greater speedup; (2) in a manner of top-down pruning, the proposed method operates from a more global perspective based on training signals in the top layer, and prunes each layer by propagating the effect of global signals through layers, leading to better performances at the same sparsity level. Extensive experiments show that at the same sparsity level, the proposed strategy offers both greater speedup and higher performances than weight-based pruning methods (e.g., magnitude pruning, movement pruning).

Layer-wise Model Pruning based on Mutual Information

Short text classification is a fundamental task in natural language processing. It is hard due to the lack of context information and labeled data in practice. In this paper, we propose a new method called SHINE, which is based on graph neural network (GNN), for short text classification. First, we model the short text dataset as a hierarchical heterogeneous graph consisting of word-level component graphs which introduce more semantic and syntactic information. Then, we dynamically learn a short document graph that facilitates effective label propagation among similar short texts. Thus, comparing with existing GNN-based methods, SHINE can better exploit interactions between nodes of the same types and capture similarities between short texts. Extensive experiments on various benchmark short text datasets show that SHINE consistently outperforms state-of-the-art methods, especially with fewer labels.

Hierarchical Heterogeneous Graph Representation Learning for Short Text Classification

We introduce EfficientCL, a memory-efficient continual pretraining method that applies contrastive learning with novel data augmentation and curriculum learning. For data augmentation, we stack two types of operation sequentially: cutoff and PCA jittering. While pretraining steps proceed, we apply curriculum learning by incrementing the augmentation degree for each difficulty step. After data augmentation is finished, contrastive learning is applied on projected embeddings of original and augmented examples. When finetuned on GLUE benchmark, our model outperforms baseline models, especially for sentence-level tasks. Additionally, this improvement is capable with only 70% of computational memory compared to the baseline model.

Efficient Contrastive Learning via Novel Data Augmentation and Curriculum Learning

In this paper, we propose an annotated sentiment analysis dataset made of informally written Bangla texts. This dataset comprises public comments on news and videos collected from social media covering 13 different domains, including politics, education, and agriculture. These comments are labeled with one of the polarity labels, namely positive, negative, and neutral. One significant characteristic of the dataset is that each of the comments is noisy in terms of the mix of dialects and grammatical incorrectness. Our experiments to develop a benchmark classification system show that hand-crafted lexical features provide superior performance than neural network and pretrained language models.

SentNoB: A Dataset for Analysing Sentiment on Noisy Bangla Texts

While the production of information in the European early modern period is a well-researched topic, the question how people were engaging with the information explosion that occurred in early modern Europe, is still underexposed. This paper presents the annotations and experiments aimed at exploring whether we can automatically extract media-related information (source, perception, and receiver) from a corpus of early modern Dutch chronicles in order to get insight in the mediascape of early modern middle-class people from a historic perspective. In a number of classification experiments with Conditional Random Fields, three categories of features are tested: (i) raw and binary word embedding features, (ii) lexicon features, and (iii) character features. Overall, the classifier that uses raw embeddings performs slightly better. However, given that the best F-scores are around 0.60, we conclude that the machine learning approach needs to be combined with a close reading approach for the results to be useful to answer history research questions.

The Early Modern Dutch Mediascape. Detecting Media Mentions in Chronicles Using Word Embeddings and CRF

Although olfactory references play a crucial role in our cultural memory, only few works in NLP have tried to capture them from a computational perspective. Currently, the main challenge is not much the development of technological components for olfactory information extraction, given recent advances in semantic processing and natural language understanding, but rather the lack of a theoretical framework to capture this information from a linguistic point of view, as a preliminary step towards the development of automated systems.
Therefore, in this work we present the annotation guidelines, developed with the help of history scholars and domain experts, aimed at capturing all the relevant elements involved in olfactory situations or events described in texts. These guidelines have been inspired by FrameNet annotation, but underwent some adaptations, which are detailed in this paper. Furthermore, we present a case study concerning the annotation of olfactory situations in English historical travel writings describing trips to Italy. An analysis of the most frequent role fillers show that olfactory descriptions pertain to some typical domains such as religion, food, nature, ancient past, poor sanitation, all supporting the creation of a stereotypical imagery related to Italy. On the other hand, positive feelings triggered by smells are prevalent, and contribute to framing travels to Italy as an exciting experience involving all senses.

FrameNet-like Annotation of Olfactory Information in Texts

Pretrained language models like BERT have advanced the state of the art for many NLP tasks. For resource-rich languages, one has the choice between a number of language-specific models, while multilingual models are also worth considering. These models are well known for their crosslingual performance, but have also shown competitive in-language performance on some tasks. We consider monolingual and multilingual models from the perspective of historical texts, and in particular for texts enriched with editorial notes: how do language models deal with the historical and editorial content in these texts? 
We present a new Named Entity Recognition dataset for Dutch based on 17th and 18th century United East India Company (VOC) reports extended with modern editorial notes. 
Our experiments with multilingual and Dutch pretrained language models confirm the crosslingual abilities of multilingual models while showing that all language models can leverage mixed-variant data. In particular, language models successfully incorporate notes for the prediction of entities in historical texts. We also find that multilingual models outperform monolingual models on our data, but that this superiority is linked to the task at hand: multilingual models lose their advantage when confronted with more semantical tasks.

Batavia asked for advice. Pretrained language models for Named Entity Recognition in historical texts.

We report on an inter-annotator agreement experiment involving instances of text reuse focusing on the well-known case of biblical intertextuality in medieval literature. We target the application use case of literary scholars whose aim is to document instances of biblical references in the "apparatus fontium" of a prospective digital edition. We develop a Bayesian implementation of Cohen's kappa for multiple annotators that allows us to assess the influence of various contextual effects on the inter-annotator agreement, producing both more robust estimates of the agreement indices as well as insights into the annotation process that leads to the estimated indices. As a result, we are able to produce a novel and nuanced estimation of inter-annotator agreement in the context of intertextuality, exploring the challenges that arise from manually annotating a dataset of biblical references in the writings of Bernard of Clairvaux. Among others, our method is able to unveil the fact that the obtained agreement depends heavily on the biblical source book of the proposed reference, as well as the underlying algorithm used to retrieve the candidate match. %In particular, a semantically motivated algorithm was shown to retrieve candidate pairs that under circumstances are less controversial than those retrieved by an alternative text alignment algorithm, with a slight bias towards literal reuse styles.
Finally, a discussion of the hurdles encountered by annotators supplements the results of the statistical analysis, contributing a qualitative insight into the difficulties involved in the identification of literary text reuse.

Quantifying Contextual Aspects of Inter-annotator Agreement in Intertextuality Research

In this work, we design an end-to-end model for poetry generation based on conditioned recurrent neural network (RNN) language models whose goal is to learn stylistic features (poem length, sentiment, alliteration, and rhyming) from examples alone. We show this model successfully learns the ‘meaning’ of length and sentiment, as we can control it to generate longer or shorter as well as more positive or more negative poems. However, the model does not grasp sound phenomena like alliter-
ation and rhyming, but instead exploits low-level statistical cues. Possible reasons include the size of the training data, the relatively low frequency and difficulty of these sublexical phenomena as well as model biases. We show that more recent GPT-2 models also have problems learning sublexical phenomena such as rhyming from examples alone.

Frustratingly Simple Pretraining Alternatives to Masked Language Modeling

Downloads

Next from EMNLP 2021

Neuro-Symbolic Approaches for Text-Based Policy Learning

Similar lecture

Fine-Tuned Transformers Show Clusters of Similar Representations Across Layers