Can pretrained language models (PLMs) generate derivationally complex words? We present the first study investigating this question, taking BERT as the example PLM. We examine BERT's derivational capabilities in different settings, ranging from using the unmodified pretrained model to full finetuning. Our best model, DagoBERT (Derivationally and generatively optimized BERT), clearly outperforms the previous state of the art in derivation generation (DG). Furthermore, our experiments show that the input segmentation crucially impacts BERT's derivational knowledge, suggesting that the performance of PLMs could be further improved if a morphologically informed vocabulary of units were used.
It is well known that the standard likelihood training and approximate decoding objectives in neural text generation models lead to less human-like responses for open-ended tasks such as language modeling and story generation. In this paper we have analyzed limitations of these models for abstractive document summarization and found that these models are highly prone to hallucinate content that is unfaithful to the input document. We conducted a large scale human evaluation of several neural abstractive summarization systems to better understand the types of hallucinations they produce. Our human annotators found substantial amounts of hallucinated content in all model generated summaries. However, our analysis does show that pretrained models are better summarizers not only in terms of raw metrics, i.e., ROUGE, but also in generating faithful and factual summaries as evaluated by humans. Furthermore, we show that textual entailment measures better correlate with faithfulness than standard metrics, potentially leading the way to automatic evaluation metrics as well as training and decoding criteria.
The purpose of unconditional text generation is to train a model with real sentences, then generate novel sentences of the same quality and diversity as the training data. However, when different metrics are used for comparing the methods of unconditional text generation, contradictory conclusions are drawn. The difficulty is that both the diversity and quality of the sample should be considered simultaneously when the models are evaluated. To solve this problem, a novel metric of distributional discrepancy (DD) is designed to evaluate generators based on the discrepancy between the generated and real training sentences. However, it cannot compute the DD directly because the distribution of real sentences is unavailable. Thus, we propose a method for estimating the DD by training a neural-network-based text classifier. For comparison, three existing metrics, bi-lingual evaluation understudy (BLEU) versus self-BLEU, language model score versus reverse language model score, and Fr\'{e}chet embedding distance, along with the proposed DD, are used to evaluate two popular generative models of long short-term memory and generative pretrained transformer 2 on both syntactic and real data. Experimental results show that DD is significantly better than the three existing metrics for ranking these generative models.
Evaluation of cross-lingual encoders is usually performed either via zero-shot cross-lingual transfer in supervised downstream tasks or via unsupervised cross-lingual textual similarity. In this paper, we concern ourselves with reference-free machine translation (MT) evaluation where we directly compare source texts to (sometimes low-quality) system translations, which represents a natural adversarial setup for multilingual encoders. Reference-free evaluation holds the promise of web-scale comparison of MT systems. We systematically investigate a range of metrics based on state-of-the-art cross-lingual semantic representations obtained with pretrained M-BERT and LASER. We find that they perform poorly as semantic encoders for reference-free MT evaluation and identify their two key limitations, namely, (a) a semantic mismatch between representations of mutual translations and, more prominently, (b) the inability to punish "translationese", i.e., low-quality literal translations. We propose two partial remedies: (1) post-hoc re-alignment of the vector spaces and (2) coupling of semantic-similarity based metrics with target-side language modeling. In segment-level MT evaluation, our best metric surpasses reference-based BLEU by 5.7 correlation points.
This paper investigates how to effectively incorporate a pre-trained masked language model (MLM), such as BERT, into an encoder-decoder (EncDec) model for grammatical error correction (GEC). The answer to this question is not as straightforward as one might expect because the previous common methods for incorporating a MLM into an EncDec model have potential drawbacks when applied to GEC. For example, the distribution of the inputs to a GEC model can be considerably different (erroneous, clumsy, etc.) from that of the corpora used for pre-training MLMs; however, this issue is not addressed in the previous methods. Our experiments show that our proposed method, where we first fine-tune a MLM with a given GEC corpus and then use the output of the fine-tuned MLM as additional features in the GEC model, maximizes the benefit of the MLM. The best-performing model achieves state-of-the-art performances on the BEA-2019 and CoNLL-2014 benchmarks. Our code is publicly available at: https://github.com/kanekomasahiro/bert-gec.
Sequential fine-tuning and multi-task learning are methods aiming to incorporate knowledge from multiple tasks; however, they suffer from catastrophic forgetting and difficulties in dataset balancing. To address these shortcomings, we propose AdapterFusion, a new two stage learning algorithm that leverages knowledge from multiple tasks. First, in the knowledge extraction stage we learn task specific parameters called adapters, that encapsulate the task-specific information. We then combine the adapters in a separate knowledge composition step. We show that by separating the two stages, i.e., knowledge extraction and knowledge composition, the classifier can effectively exploit the representations learned from multiple tasks in a non-destructive manner. We empirically evaluate AdapterFusion on 16 diverse NLU tasks, and find that it effectively combines various types of knowledge at different layers of the model. We show that our approach outperforms traditional strategies such as full fine-tuning as well as multi-task learning. Our code and adapters are available at AdapterHub.ml.
We present HERO, a novel framework for large-scale video+language omni-representation learning. HERO encodes multimodal inputs in a hierarchical structure, where local context of a video frame is captured by a Cross-modal Transformer via multimodal fusion, and global video context is captured by a Temporal Transformer. In addition to standard Masked Language Modeling (MLM) and Masked Frame Modeling (MFM) objectives, we design two new pre-training tasks: (i) Video-Subtitle Matching (VSM), where the model predicts both global and local temporal alignment; and (ii) Frame Order Modeling (FOM), where the model predicts the right order of shuffled video frames. HERO is jointly trained on HowTo100M and large-scale TV datasets to gain deep understanding of complex social dynamics with multi-character interactions. Comprehensive experiments demonstrate that HERO achieves new state of the art on multiple benchmarks over Text-based Video/Video-moment Retrieval, Video Question Answering (QA), Video-and-language Inference and Video Captioning tasks across different domains. We also introduce two new challenging benchmarks How2QA and How2R for Video QA and Retrieval, collected from diverse video content over multimodalities.
Language model fine-tuning is essential for modern natural language processing, but is computationally expensive and time-consuming. Further, the effectiveness of fine-tuning is limited by the inclusion of training examples that negatively affect performance. Here we present a general fine-tuning method that we call information gain filtration for improving the overall training efficiency and final performance of language model fine-tuning. We define the information gain of an example as the improvement on a test metric after training on that example. A secondary learner is then trained to approximate this quantity. During fine-tuning, this learner selects informative examples and skips uninformative ones. We show that our method has consistent improvement across datasets, fine-tuning tasks, and language model architectures. For example, we achieve a median perplexity of 54.0 on a books dataset compared to 57.3 for standard fine-tuning. We present statistical evidence that offers insight into the improvements of our method over standard fine-tuning. The generality of our method leads us to propose a new paradigm for language model fine-tuning -- we encourage researchers to release pretrained secondary learners on common corpora to promote efficient and effective fine-tuning, thereby improving the performance and reducing the overall energy footprint of language model fine-tuning.
A standard approach to evaluating language models analyzes how models assign probabilities to valid versus invalid syntactic constructions (i.e. is a grammatical sentence more probable than an ungrammatical sentence). Our work uses ambiguous relative clause attachment to extend such evaluations to cases of multiple simultaneous valid interpretations, where stark grammaticality differences are absent. We compare model performance in English and Spanish to show that non-linguistic biases in RNN LMs advantageously overlap with syntactic structure in English but not Spanish. Thus, English models may appear to acquire human-like syntactic preferences, while models trained on Spanish fail to acquire comparable human-like preferences. We conclude by relating these results to broader concerns about the relationship between comprehension (i.e. typical language model use cases) and production (which generates the training data for language models), suggesting that necessary linguistic biases are not present in the training signal at all.
We rely on arguments in our daily lives to deliver our opinions and base them on evidence, making them more convincing in turn. However, finding and formulating arguments can be challenging. In this work, we train a language model for argument generation that can be controlled on a fine-grained level to generate sentence-level arguments for a given topic, stance, and aspect. We define argument aspect detection as a necessary method to allow this fine-granular control and crowdsource a dataset with 5,032 arguments annotated with aspects. Our evaluation shows that our generation model is able to generate high-quality, aspect-specific arguments. Moreover, these arguments can be used to improve the performance of stance detection models via data augmentation and to generate counter-arguments. We publish all datasets and code to fine-tune the language model.
Multilingual pre-trained Transformers, such as mBERT (Devlin et al., 2019) and XLM-RoBERTa (Conneau et al., 2020a), have been shown to enable the effective cross-lingual zero-shot transfer. However, their performance on Arabic information extraction (IE) tasks is not very well studied. In this paper, we pre-train a customized bilingual BERT, dubbed GigaBERT, that is designed specifically for Arabic NLP and English-to-Arabic zero-shot transfer learning. We study GigaBERT's effectiveness on zero-short transfer across four IE tasks: named entity recognition, part-of-speech tagging, argument role labeling, and relation extraction. Our best model significantly outperforms mBERT, XLM-RoBERTa, and AraBERT (Antoun et al., 2020) in both the supervised and zero-shot transfer settings. We have made our pre-trained models publicly available at https://github.com/lanwuwei/GigaBERT.
The scarcity of large parallel corpora is an important obstacle for neural machine translation. A common solution is to exploit the knowledge of language models (LM) trained on abundant monolingual data. In this work, we propose a novel approach to incorporate a LM as prior in a neural translation model (TM). Specifically, we add a regularization term, which pushes the output distributions of the TM to be probable under the LM prior, while avoiding wrong predictions when the TM "disagrees" with the LM. This objective relates to knowledge distillation, where the LM can be viewed as teaching the TM about the target language. The proposed approach does not compromise decoding speed, because the LM is used only at training time, unlike previous work that requires it during inference. We present an analysis of the effects that different methods have on the distributions of the TM. Results on two low-resource machine translation datasets show clear improvements even with limited monolingual data.
Suspense is a crucial ingredient of narrative fiction, engaging readers and making stories compelling. While there is a vast theoretical literature on suspense, it is computationally not well understood. We compare two ways for modelling suspense: surprise, a backward-looking measure of how unexpected the current state is given the story so far; and uncertainty reduction, a forward-looking measure of how unexpected the continuation of the story is. Both can be computed either directly over story representations or over their probability distributions. We propose a hierarchical language model that encodes stories and computes surprise and uncertainty reduction. Evaluating against short stories annotated with human suspense judgements, we find that uncertainty reduction over representations is the best predictor, resulting in near-human accuracy. We also show that uncertainty reduction can be used to predict suspenseful events in movie synopses.
Opinion summarization is the automatic creation of text reflecting subjective information expressed in multiple documents, such as user reviews of a product. The task is practically important and has attracted a lot of attention. However, due to the high cost of summary production, datasets large enough for training supervised models are lacking. Instead, the task has been traditionally approached with extractive methods that learn to select text fragments in an unsupervised or weakly-supervised way. Recently, it has been shown that abstractive summaries, potentially more fluent and better at reflecting conflicting information, can also be produced in an unsupervised fashion. However, these models, not being exposed to actual summaries, fail to capture their essential properties. In this work, we show that even a handful of summaries is sufficient to bootstrap generation of the summary text with all expected properties, such as writing style, informativeness, fluency, and sentiment preservation. We start by training a conditional Transformer language model to generate a new product review given other available reviews of the product. The model is also conditioned on review properties that are directly related to summaries; the properties are derived from reviews with no manual effort. In the second stage, we fine-tune a plug-in module that learns to predict property values on a handful of summaries. This lets us switch the generator to the summarization mode. We show on Amazon and Yelp datasets that our approach substantially outperforms previous extractive and abstractive methods in automatic and human evaluation.
By introducing a small set of additional parameters, a probe learns to solve specific linguistic tasks (e.g., dependency parsing) in a supervised manner using feature representations (e.g., contextualized embeddings). The effectiveness of such probing tasks is taken as evidence that the pre-trained model encodes linguistic knowledge. However, this approach of evaluating a language model is undermined by the uncertainty of the amount of knowledge that is learned by the probe itself. Complementary to those works, we propose a parameter-free probing technique for analyzing pre-trained language models (e.g., BERT). Our method does not require direct supervision from the probing tasks, nor do we introduce additional parameters to the probing process. Our experiments on BERT show that syntactic trees recovered from BERT using our method are significantly better than linguistically-uninformed baselines. We further feed the empirically induced dependency structures into a downstream sentiment classification task and find its improvement compatible with or even superior to a human-designed dependency schema.
Argumentation is a type of discourse where speakers try to persuade their audience about the reasonableness of a claim by presenting supportive arguments. Most work in argument mining has focused on modeling arguments in monologues. We propose a computational model for argument mining in online persuasive discussion forums that brings together the micro-level (argument as product) and macro-level (argument as process) models of argumentation. Fundamentally, this approach relies on identifying relations between components of arguments in a discussion thread. Our approach for relation prediction uses contextual information in terms of fine-tuning a pre-trained language model and leveraging discourse relations based on Rhetorical Structure Theory. We additionally propose a candidate selection method to automatically predict what parts of one's argument will be targeted by other participants in the discussion. Our models obtain significant improvements compared to recent state-of-the-art approaches using pointer networks and a pre-trained language model.
We propose transfer learning as a method for analyzing the encoding of grammatical structure in neural language models. We train LSTMs on non-linguistic data and evaluate their performance on natural language to assess which kinds of data induce generalizable structural features that LSTMs can use for natural language. We find that training on non-linguistic data with latent structure (MIDI music or Java code) improves test performance on natural language, despite no overlap in surface form or vocabulary. To pinpoint the kinds of abstract structure that models may be encoding to lead to this improvement, we run similar experiments with two artificial parentheses languages: one which has a hierarchical recursive structure, and a control which has paired tokens but no recursion. Surprisingly, training a model on either of these artificial languages leads to the same substantial gains when testing on natural language. Further experiments on transfer between natural languages controlling for vocabulary overlap show that zero-shot performance on a test language is highly correlated with typological syntactic similarity to the training language, suggesting that representations induced by pre-training correspond to the cross-linguistic syntactic properties. Our results provide insights into the ways that neural models represent abstract syntactic structure, and also about the kind of structural inductive biases which allow for natural language acquisition.
Transformers are powerful for sequence modeling. Nearly all state-of-the-art language models and pre-trained language models are based on the Transformer architecture. However, it distinguishes sequential tokens only with the token position index. We hypothesize that better contextual representations can be generated from the Transformer with richer positional information. To verify this, we propose a segment-aware Transformer (Segatron), by replacing the original token position encoding with a combined position encoding of paragraph, sentence, and token. We first introduce the segment-aware mechanism to Transformer-XL, which is a popular Transformer-based language model with memory extension and relative position encoding. We find that our method can further improve the Transformer-XL base model and large model, achieving 17.1 perplexity on the WikiText-103 dataset. We further investigate the pre-training masked language modeling task with Segatron. Experimental results show that BERT pre-trained with Segatron (SegaBERT) can outperform BERT with vanilla Transformer on various NLP tasks, and outperforms RoBERTa on zero-shot sentence representation learning.
How does language model pretraining help transfer learning? We consider a simple ablation technique for determining the impact of each pretrained layer on transfer task performance. This method, partial reinitialization, involves replacing different layers of a pretrained model with random weights, then finetuning the entire model on the transfer task and observing the change in performance. This technique reveals that in BERT, layers with high probing performance on downstream GLUE tasks are neither necessary nor sufficient for high accuracy on those tasks. Furthermore, the benefit of using pretrained parameters for a layer varies dramatically with finetuning dataset size: parameters that provide tremendous performance improvement when data is plentiful may provide negligible benefits in data-scarce settings. These results reveal the complexity of the transfer learning process, highlighting the limitations of methods that operate on frozen models or single data samples.
An increasing number of people in the world today speak a mixed-language as a result of being multilingual. However, building a speech recognition system for code-switching remains difficult due to the availability of limited resources and the expense and significant effort required to collect mixed-language data. We therefore propose a new learning method, meta-transfer learning, to transfer learn on a code-switched speech recognition system in a low-resource setting by judiciously extracting information from high-resource monolingual datasets. Our model learns to recognize individual languages, and transfer them so as to better recognize mixed-language speech by conditioning the optimization on the code-switching data. Based on experimental results, our model outperforms existing baselines on speech recognition and language modeling tasks, and is faster to converge.