Existing models for cross-domain named entity recognition (NER) rely on numerous unlabeled corpus or labeled NER training data in target domains. However, collecting data for low-resource target domains is not only expensive but also time-consuming. Hence, we propose a cross-domain NER model that does not use any external resources. We first introduce a Multi-Task Learning (MTL) by adding a new objective function to detect whether tokens are named entities or not. We then introduce a framework called Mixture of Entity Experts (MoEE) to improve the robustness for zero-resource domain adaptation. Finally, experimental results show that our model outperforms strong unsupervised cross-domain sequence labeling models, and the performance of our model is close to that of the state-of-the-art model which leverages extensive resources.
We train neural machine translation (NMT) models from English to six target languages, using NMT encoder representations to predict ancestor constituent labels of source language words. We find that NMT encoders learn similar source syntax regardless of NMT target language, relying on explicit morphosyntactic cues to extract syntactic features from source sentences. Furthermore, the NMT encoders outperform RNNs trained directly on several of the constituent label prediction tasks, suggesting that NMT encoder representations can be used effectively for natural language tasks involving syntax. However, both the NMT encoders and the directly-trained RNNs learn substantially different syntactic information from a probabilistic context-free grammar (PCFG) parser. Despite lower overall accuracy scores, the PCFG often performs well on sentences for which the RNN-based models perform poorly, suggesting that RNN architectures are constrained in the types of syntax they can learn.
Probabilistic word embeddings have shown effectiveness in capturing notions of generality and entailment, but there is very little work on doing the analogous type of investigation for sentences. In this paper we define probabilistic models that produce distributions for sentences. Our best-performing model treats each word as a linear transformation operator applied to a multivariate Gaussian distribution. We train our models on paraphrases and demonstrate that they naturally capture sentence specificity. While our proposed model achieves the best performance overall, we also show that specificity is represented by simpler architectures via the norm of the sentence vectors. Qualitative analysis shows that our probabilistic model captures sentential entailment and provides ways to analyze the specificity and preciseness of individual words.
In this paper, we provide an alternate perspective on word representations, by reinterpreting the dimensions of the vector space of a word embedding as a collection of features. In this reinterpretation, every component of the word vector is normalized against all the word vectors in the vocabulary. This idea now allows us to view each vector as an n-tuple (akin to a fuzzy set), where n is the dimensionality of the word representation and each element represents the probability of the word possessing a feature. Indeed, this representation enables the use fuzzy set theoretic operations, such as union, intersection and difference. Unlike previous attempts, we show that this representation of words provides a notion of similarity which is inherently asymmetric and hence closer to human similarity judgements. We compare the performance of this representation with various benchmarks, and explore some of the unique properties including function word detection, detection of polysemous words, and some insight into the interpretability provided by set theoretic operations.
Recent works have discussed the extent to which emergent languages can exhibit properties of natural languages particularly learning compositionality. In this paper, we investigate the learning biases that affect the efficacy and compositionality in multi-agent communication in addition to the communicative bandwidth. Our foremost contribution is to explore how the capacity of a neural network impacts its ability to learn a compositional language. We additionally introduce a set of evaluation metrics with which we analyze the learned languages. Our hypothesis is that there should be a specific range of model capacity and channel bandwidth that induces compositional structure in the resulting language and consequently encourages systematic generalization. While we empirically see evidence for the bottom of this range, we curiously do not find evidence for the top part of the range and believe that this is an open question for the community.
We propose a geometric framework for learning meta-embeddings of words from different embedding sources. Our framework transforms the embeddings into a common latent space, where, for example, simple averaging or concatenation of different embeddings (of a given word) is more amenable. The proposed latent space arises from two particular geometric transformations - source embedding specific orthogonal rotations and a common Mahalanobis metric scaling. Empirical results on several word similarity and word analogy benchmarks illustrate the efficacy of the proposed framework.
Work on projection-based induction of cross-lingual word embedding spaces (CLWEs) predominantly focuses on the improvement of the projection (i.e., mapping) mechanisms. In this work, in contrast, we show that a simple method for post-processing monolingual embedding spaces facilitates learning of the cross-lingual alignment and, in turn, substantially improves bilingual lexicon induction (BLI). The post-processing method we examine is grounded in the generalisation of first- and second-order monolingual similarities to the nth-order similarity. By post-processing monolingual spaces before the cross-lingual alignment, the method can be coupled with any projection-based method for inducing CLWE spaces. We demonstrate the effectiveness of this simple monolingual post-processing across a set of 15 typologically diverse languages (i.e., 15*14 BLI setups), and in combination with two different projection methods.
We apply small perturbations to word embeddings and minimize the resultant adversarial risk to regularize the model. We exploit a novel combination of two different approaches to estimate these perturbations: 1) using the true label and 2) using the model prediction. Without relying on any human-crafted features, knowledge bases, or additional datasets other than the target datasets, our model boosts the fine-tuning performance of RoBERTa, achieving competitive results on multiple reading comprehension datasets that require commonsense inference.
Skip-Gram is a simple, but effective, model to learn a word embedding mapping by estimating a conditional probability distribution for each word of the dictionary. In the context of Information Geometry, these distributions form a Riemannian statistical manifold, where word embeddings are interpreted as vectors in the tangent bundle of the manifold. In this paper we show how the choice of the geometry on the manifold allows impacts on the performances both on intrinsic and extrinsic tasks, in function of a deformation parameter alpha.
Fine-tuned variants of BERT are able to achieve state-of-the-art accuracy on many natural language processing tasks, although at significant computational costs. In this paper, we verify BERT’s effectiveness for document classification and investigate the extent to which BERT-level effectiveness can be obtained by different baselines, combined with knowledge distillation—a popular model compression method. The results show that BERT-level effectiveness can be achieved by a single-layer LSTM with at least 40× fewer FLOPS and only ∼3% parameters. More importantly, this study analyzes the limits of knowledge distillation as we distill BERT’s knowledge all the way down to linear models—a relevant baseline for the task. We report substantial improvement in effectiveness for even the simplest models, as they capture the knowledge learnt by BERT.
End-to-end models trained on natural language inference (NLI) datasets show low generalization on out-of-distribution evaluation sets. The models tend to learn shallow heuristics due to dataset biases. The performance decreases dramatically on diagnostic sets measuring compositionality or robustness against simple heuristics. Existing solutions for this problem employ dataset augmentation which has the drawbacks of being applicable to only a limited set of adversaries and at worst hurting the model performance on other adversaries not included in the augmentation set. Instead, our proposed solution is to improve sentence understanding (hence out-of-distribution generalization) with joint learning of explicit semantics. We show that a BERT based model trained jointly on English semantic role labeling (SRL) and NLI achieves significantly higher performance on external evaluation sets measuring generalization performance.
The task of automatic misogyny identification and categorization has not received as much attention as other natural language tasks have, even though it is crucial for identifying hate speech in social Internet interactions. In this work, we address this sentence classification task from a representation learning perspective, using both a bidirectional LSTM and BERT optimized with the following metric learning loss functions: contrastive loss, triplet loss, center loss, congenerous cosine loss and additive angular margin loss. We set new state-of-the-art for the task with our fine-tuned BERT, whose sentence embeddings can be compared with a simple cosine distance, and we release all our code as open source for easy reproducibility. Moreover, we find that almost every loss function performs equally well in this setting, matching the regular cross entropy loss.
Recent work showed that embeddings from related languages can improve the performance of sequence tagging, even for monolingual models. In this analysis paper, we investigate whether the best auxiliary language can be predicted based on language distances and show that the most related language is not always the best auxiliary language. Further, we show that attention-based meta-embeddings can effectively combine pre-trained embeddings from different languages for sequence tagging and set new state-of-the-art results for part-of-speech tagging in five languages.
Although temporal tagging is still dominated by rule-based systems, there have been recent attempts at neural temporal taggers. However, all of them focus on monolingual settings. In this paper, we explore multilingual methods for the extraction of temporal expressions from text and investigate adversarial training for aligning embedding spaces to one common space. With this, we create a single multilingual model that can also be transferred to unseen languages and set the new state of the art in those cross-lingual transfer experiments.
In this paper we present a comparison between the linguistic knowledge encoded in the internal representations of a contextual Language Model (BERT) and a contextual-independent one (Word2vec). We use a wide set of probing tasks, each of which corresponds to a distinct sentence-level feature extracted from different levels of linguistic annotation. We show that, although BERT is capable of understanding the full context of each word in an input sequence, the implicit knowledge encoded in its aggregated sentence representations is still comparable to that of a contextual-independent model. We also find that BERT is able to encode sentence-level properties even within single-word embeddings, obtaining comparable or even superior results than those obtained with sentence representations.
Multilingual BERT (mBERT) trained on 104 languages has shown surprisingly good cross-lingual performance on several NLP tasks, even without explicit cross-lingual signals. However, these evaluations have focused on cross-lingual transfer with high-resource languages, covering only a third of the languages covered by mBERT. We explore how mBERT performs on a much wider set of languages, focusing on the quality of representation for low-resource languages, measured by within-language performance. We consider three tasks: Named Entity Recognition (99 languages), Part-of-speech Tagging and Dependency Parsing (54 languages each). mBERT does better than or comparable to baselines on high resource languages but does much worse for low resource languages. Furthermore, monolingual BERT models for these languages do even worse. Paired with similar languages, the performance gap between monolingual BERT and mBERT can be narrowed. We find that better models for low resource languages require more efficient pretraining techniques or more data.
The attention mechanism has quickly become ubiquitous in NLP. In addition to improving performance of models, attention has been widely used as a glimpse into the inner workings of NLP models. The latter aspect has in the recent years become a common topic of discussion, most notably in recent work of Jain and Wallace; Wiegreffe and Pinter. With the shortcomings of using attention weights as a tool of transparency revealed, the attention mechanism has been stuck in a limbo without concrete proof when and whether it can be used as an explanation. In this paper, we provide an explanation as to why attention has seen rightful critique when used with recurrent networks in sequence classification tasks. We propose a remedy to these issues in the form of a word level objective and our findings give credibility for attention to provide faithful interpretations of recurrent models.
Pre-trained universal feature extractors, such as BERT for natural language processing and VGG for computer vision, have become effective methods for improving deep learning models without requiring more labeled data. While effective, feature extractors like BERT may be prohibitively large for some deployment scenarios. We explore weight pruning for BERT and ask: how does compression during pre-training affect transfer learning? We find that pruning affects transfer learning in three broad regimes. Low levels of pruning (30-40%) do not affect pre-training loss or transfer to downstream tasks at all. Medium levels of pruning increase the pre-training loss and prevent useful pre-training information from being transferred to downstream tasks. High levels of pruning additionally prevent models from fitting downstream datasets, leading to further degradation. Finally, we observe that fine-tuning BERT on a specific task does not improve its prunability. We conclude that BERT can be pruned once during pre-training rather than separately for each task without affecting performance.
Word embeddings have become a staple of several natural language processing tasks, yet much remains to be understood about their properties. In this work, we analyze word embeddings in terms of their principal components and arrive at a number of novel and counterintuitive observations. In particular, we characterize the utility of variance explained by the principal components as a proxy for downstream performance. Furthermore, through syntactic probing of the principal embedding space, we show that the syntactic information captured by a principal component does not correlate with the amount of variance it explains. Consequently, we investigate the limitations of variance based embedding post-processing algorithms and demonstrate that such post-processing is counter-productive in sentence classification and machine translation tasks. Finally, we offer a few precautionary guidelines on applying variance based embedding post-processing and explain why non-isotropic geometry might be integral to word embedding performance.
Many natural language processing (NLP) tasks involve reasoning with textual spans, including question answering, entity recognition, and coreference resolution. While extensive research has focused on functional architectures for representing words and sentences, there is less work on representing arbitrary spans of text within sentences. In this paper, we conduct a comprehensive empirical evaluation of six span representation methods using eight pretrained language representation models across six tasks, including two tasks that we introduce. We find that, although some simple span representations are fairly reliable across tasks, in general the optimal span representation varies by task, and can also vary within different facets of individual tasks. We also find that the choice of span representation has a bigger impact with a fixed pretrained encoder than with a fine-tuned encoder.
While large-scale pretraining has achieved great success in many NLP tasks, it has not been fully studied whether external linguistic knowledge can improve data-driven models. In this work, we introduce sememe knowledge into Transformer and propose three sememe-enhanced Transformer models. Sememes, by linguistic definition, are the minimum semantic units of language, which can well represent implicit semantic meanings behind words. Our experiments demonstrate that introducing sememe knowledge into Transformer can consistently improve language modeling and downstream tasks. The adversarial test further demonstrates that sememe knowledge can substantially improve model robustness.
We evaluate the compositionality of general-purpose sentence encoders by proposing two different metrics to quantify compositional understanding capability of sentence encoders. We introduce a novel metric, Polarity Sensitivity Scoring (PSS), which utilizes sentiment perturbations as a proxy for measuring compositionality. We then compare results from PSS with those obtained via our proposed extension of a metric called Tree Reconstruction Error (TRE) (CITATION) where compositionality is evaluated by measuring how well a true representation producing model can be approximated by a model that explicitly combines representations of its primitives.
In CCG and other highly lexicalized grammars, supertagging a sentence’s words with their lexical categories is a critical step for efficient parsing. Because of the high degree of lexicalization in these grammars, the lexical categories can be very complex. Existing approaches to supervised CCG supertagging treat the categories as atomic units, even when the categories are not simple; when they encounter words with categories unseen during training, their guesses are accordingly unsophisticated. In this paper, we make use of the primitives and operators that constitute the lexical categories of categorial grammars. Instead of opaque labels, we treat lexical categories themselves as linear sequences. We present an LSTM-based model that replaces standard word-level classification with prediction of a sequence of primitives, similarly to LSTM decoders. Our model obtains state-of-the-art word accuracy for single-task English CCG supertagging, increases parser coverage and F1, and is able to produce novel categories. Analysis shows a synergistic effect between this decomposed view and incorporation of prediction history.
We evaluate named entity representations of BERT-based NLP models by investigating their robustness to replacements from the same typed class in the input. We highlight that on several tasks while such perturbations are natural, state of the art trained models are surprisingly brittle. The brittleness continues even with the recent entity-aware BERT models. We also try to discern the cause of this non-robustness, considering factors such as tokenization and frequency of occurrence. Then we provide a simple method that ensembles predictions from multiple replacements while jointly modeling the uncertainty of type annotations and label predictions. Experiments on three NLP tasks shows that our method enhances robustness and increases accuracy on both natural and adversarial datasets.