Despite their strong performance on many tasks, pre-trained language models have been shown to struggle on out-of-distribution compositional generalization. Meanwhile, recent work has shown considerable improvements on many NLP tasks from model scaling. Can scaling up model size also improve compositional generalization in semantic parsing? We evaluate encoder-decoder models up to 11B parameters and decoder-only models up to 540B parameters, and compare model scaling curves for three different methods for applying a pre-trained language model to a new task: fine-tuning all parameters, prompt tuning, and in-context learning. We observe that fine-tuning generally has flat or negative scaling curves on out-of-distribution compositional generalization in semantic parsing evaluations. In-context learning has positive scaling curves, but is generally outperformed by much smaller fine-tuned models. Prompt-tuning can outperform fine-tuning, suggesting further potential improvements from scaling as it exhibits a more positive scaling curve. Additionally, we identify several error trends that vary with model scale. For example, larger models are generally better at modeling the syntax of the output space, but are also more prone to certain types of overfitting. Overall, our study highlights limitations of current techniques for effectively leveraging model scale for compositional generalization, while our analysis also suggests promising directions for future work.

Many neural end-to-end systems today do not rely on syntactic parse trees, as much of the information that parse trees provide is encoded in the parameters of pretrained models. Lessons learned from parsing technologies and from taking a multilingual perspective, however, are still relevant even for end-to-end models. This talk will describe work that relies on compositionality in semantic parsing and in reading comprehension requiring numerical reasoning. We’ll then describe a new dataset that requires advances in multilingual modeling, and some approaches designed to better model morphology than off-the-shelf subword models that make some progress on these challenges.

Pretrained neural models such as BERT, when fine-tuned to perform natural language inference (NLI), often show high accuracy on standard datasets, but display a surprising lack of sensitivity to word order on controlled challenge sets. We hypothesize that this issue is not primarily caused by the pretrained model’s limitations, but rather by the paucity of crowdsourced NLI examples that might convey the importance of syntactic structure at the fine-tuning stage. We explore several methods to augment standard training sets with syntactically informative examples, generated by applying syntactic transformations to sentences from the MNLI corpus. The best-performing augmentation method, subject/object inversion, improved BERT’s accuracy on controlled examples that diagnose sensitivity to word order from 0.28 to 0.73, without affecting performance on the MNLI test set. This improvement generalized beyond the particular construction used for data augmentation, suggesting that augmentation causes BERT to recruit abstract syntactic representations.

Natural language inference (NLI) data has proven useful in benchmarking and, especially, as pretraining data for tasks requiring language understanding. However, the crowdsourcing protocol that was used to collect this data has known issues and was not explicitly optimized for either of these purposes, so it is likely far from ideal. We propose four alternative protocols, each aimed at improving either the ease with which annotators can produce sound training examples or the quality and diversity of those examples. Using these alternatives and a fifth baseline protocol, we collect and compare five new 8.5k-example training sets. In evaluations focused on transfer learning applications, our results are solidly negative, with models trained on our baseline dataset yielding good transfer performance to downstream tasks, but none of our four new methods (nor the recent ANLI) showing any improvements over that baseline. In a small silver lining, we observe that all four new protocols, especially those where annotators edit *pre-filled* text boxes, reduce previously observed issues with annotation artifacts.

We introduce a novel method of generating synthetic question answering corpora by combining models of question generation and answer extraction, and by filtering the results to ensure roundtrip consistency. By pretraining on the resulting corpora we obtain significant improvements on SQuAD2 and NQ, establishing a new state-of-the-art on the latter. Our synthetic data generation models, for both question generation and answer extraction, can be fully reproduced by finetuning a publicly available BERT model on the extractive subsets of SQuAD2 and NQ. We also describe a more powerful variant that does full sequence-to-sequence pretraining for question generation, obtaining exact match and F1 at less than 0.1% and 0.4% from human performance on SQuAD2.

Reading comprehension models have been successfully applied to extractive text answers, but it is unclear how best to generalize these models to abstractive numerical answers. We enable a BERT-based reading comprehension model to perform lightweight numerical reasoning. We augment the model with a predefined set of executable ‘programs’ which encompass simple arithmetic as well as extraction. Rather than having to learn to manipulate numbers directly, the model can pick a program and execute it. On the recent Discrete Reasoning Over Passages (DROP) dataset, designed to challenge reading comprehension models, we show a 33% absolute improvement by adding shallow programs. The model can learn to predict new operations when appropriate in a math word problem setting (Roy and Roth, 2015) with very few training examples.

The rise of neural networks, and particularly recurrent neural networks, has produced significant advances in part-of-speech tagging accuracy. One characteristic common among these models is the presence of rich initial word encodings. These encodings typically are composed of a recurrent character-based representation with dynamically and pre-trained word embeddings. However, these encodings do not consider a context wider than a single word and it is only through subsequent recurrent layers that word or sub-word information interacts. In this paper, we investigate models that use recurrent neural networks with sentence-level context for initial character and word-based representations. In particular we show that optimal results are obtained by integrating these context sensitive representations through synchronized training with a meta-model that learns to combine their states.

We report results from the SR’18 Shared Task, a new multilingual surface realisation task organised as part of the ACL’18 Workshop on Multilingual Surface Realisation. As in its English-only predecessor task SR’11, the shared task comprised two tracks with different levels of complexity: (a) a shallow track where the inputs were full UD structures with word order information removed and tokens lemmatised; and (b) a deep track where additionally, functional words and morphological information were removed. The shallow track was offered in ten, and the deep track in three languages. Systems were evaluated (a) automatically, using a range of intrinsic metrics, and (b) by human judges in terms of readability and meaning similarity. This report presents the evaluation results, along with descriptions of the SR’18 tracks, data and evaluation methods. For full descriptions of the participating systems, please see the separate system reports elsewhere in this volume.

English part-of-speech taggers regularly make egregious errors related to noun-verb ambiguity, despite having achieved 97%+ accuracy on the WSJ Penn Treebank since 2002. These mistakes have been difficult to quantify and make taggers less useful to downstream tasks such as translation and text-to-speech synthesis. This paper creates a new dataset of over 30,000 naturally-occurring non-trivial examples of noun-verb ambiguity. Taggers within 1% of each other when measured on the WSJ have accuracies ranging from 57% to 75% accuracy on this challenge set. Enhancing the strongest existing tagger with contextual word embeddings and targeted training data improves its accuracy to 89%, a 14% absolute (52% relative) improvement. Downstream, using just this enhanced tagger yields a 28% reduction in error over the prior best learned model for homograph disambiguation for textto-speech synthesis.

The Conference on Computational Natural Language Learning (CoNLL) features a shared task, in which participants train and test their learning systems on the same data sets. In 2017, the task was devoted to learning dependency parsers for a large number of languages, in a real-world setting without any gold-standard annotation on input. All test sets followed a unified annotation scheme, namely that of Universal Dependencies. In this paper, we define the task and evaluation methodology, describe how the data sets were prepared, report and analyze the main results, and provide a brief categorization of the different approaches of the participating systems.

We show that small and shallow feed-forward neural networks can achieve near state-of-the-art results on a range of unstructured and structured language processing tasks while being considerably cheaper in memory and computational requirements than deep recurrent models. Motivated by resource-constrained environments like mobile phones, we showcase simple techniques for obtaining such small neural network models, and investigate different tradeoffs when deciding how to allocate a small memory budget.

Parsers that parametrize over wider scopes are generally more accurate than edge-factored models. For graph-based non-projective parsers, wider factorizations have so far implied large increases in the computational complexity of the parsing problem. This paper introduces a “crossing-sensitive” generalization of a third-order factorization that trades off complexity in the model structure (i.e., scoring with features over multiple edges) with complexity in the output structure (i.e., producing crossing edges). Under this model, the optimal 1-Endpoint-Crossing tree can be found in O(n4) time, matching the asymptotic run-time of both the third-order projective parser and the edge-factored 1-Endpoint-Crossing parser. The crossing-sensitive third-order parser is significantly more accurate than the third-order projective parser under many experimental settings and significantly less accurate on none.

Dependency parsing algorithms capable of producing the types of crossing dependencies seen in natural language sentences have traditionally been orders of magnitude slower than algorithms for projective trees. For 95.8–99.8% of dependency parses in various natural language treebanks, whenever an edge is crossed, the edges that cross it all have a common vertex. The optimal dependency tree that satisfies this 1-Endpoint-Crossing property can be found with an O(n4) parsing algorithm that recursively combines forests over intervals with one exterior point. 1-Endpoint-Crossing trees also have natural connections to linguistics and another class of graphs that has been studied in NLP.

While the web provides a fantastic linguistic resource, collecting and processing data at web-scale is beyond the reach of most academic laboratories. Previous research has relied on search engines to collect online information, but this is hopelessly inefficient for building large-scale linguistic resources, such as lists of named-entity types or clusters of distributionally similar words. An alternative to processing web-scale text directly is to use the information provided in an N-gram corpus. An N-gram corpus is an efficient compression of large amounts of text. An N-gram corpus states how often each sequence of words (up to length N) occurs. We propose tools for working with enhanced web-scale N-gram corpora that include richer levels of source annotation, such as part-of-speech tags. We describe a new set of search tools that make use of these tags, and collectively lower the barrier for lexical learning and ambiguity resolution at web-scale. They will allow novel sources of information to be applied to long-standing natural language challenges.