Is the output softmax layer, which is adopted by most language models (LMs), always the best way to compute the next word probability? Given so many attention layers in a modern transformer-based LM, are the pointer networks redundant nowadays? In this study, we discover that the answers to both questions are no. This is because the softmax bottleneck sometimes prevents the LMs from predicting the desired distribution and the pointer networks can be used to break the bottleneck efficiently. Based on the finding, we propose several softmax alternatives by simplifying the pointer networks and accelerating the word-by-word rerankers. In GPT-2, our proposals are significantly better and more efficient than mixture of softmax, a state-of-the-art softmax alternative. In summarization experiments, without very significantly decreasing its training/testing speed, our best method based on T5-Small improves factCC score by 2 points in CNN/DM and XSUM dataset, and improves MAUVE scores by 30% in BookSum paragraph-level dataset.

Ensembling BERT models often significantly improves accuracy, but at the cost of significantly more computation and memory footprint. In this work, we propose Multi-CLS BERT, a novel ensembling method for CLS-based prediction tasks that is almost as efficient as a single BERT model. Multi-CLS BERT uses multiple CLS tokens with a parameterization and objective that encourages their diversity. Thus instead of fine-tuning each BERT model in an ensemble (and running them all at test time), we need only fine-tune our single Multi-CLS BERT model (and run the one model at test time, ensembling just the multiple final CLS embeddings). To test its effectiveness, we build Multi-CLS BERT on top of a state-of-the-art pretraining method for BERT (Aroca-Ouellette and Rudzicz, 2020). In experiments on GLUE and SuperGLUE we show that our Multi-CLS BERT reliably improves both overall accuracy and confidence estimation. When only 100 training samples are available in GLUE, the Multi-CLS BERT_Base model can even outperform the corresponding BERT_Large model. We analyze the behavior of our Multi-CLS BERT, showing that it has many of the same characteristics and behavior as a typical BERT 5-way ensemble, but with nearly 4-times less computation and memory.

Given a citation in the body of a research paper, cited text identification aims to find the sentences in the cited paper that are most relevant to the citing sentence. The task is fundamentally one of sentence matching, where affinity is often assessed by a cosine similarity between sentence embeddings. However, (a) sentences may not be well-represented by a single embedding because they contain multiple distinct semantic aspects, and (b) good matches may not require a strong match in all aspects. To overcome these limitations, we propose a simple and efficient unsupervised method for cited text identification that adapts an asymmetric similarity measure to allow partial matches of multiple aspects in both sentences. On the CL-SciSumm dataset we find that our method outperforms a baseline symmetric approach, and, surprisingly, also outperforms all supervised and unsupervised systems submitted to past editions of CL-SciSumm Shared Task 1a.

Neural language models (LMs) such as GPT-2 estimate the probability distribution over the next word by a softmax over the vocabulary. The softmax layer produces the distribution based on the dot products of a single hidden state and the embeddings of words in the vocabulary. However, we discover that this single hidden state cannot produce all probability distributions regardless of the LM size or training data size because the single hidden state embedding cannot be close to the embeddings of all the possible next words simultaneously when there are other interfering word embeddings between them. In this work, we demonstrate the importance of this limitation both theoretically and practically. Our work not only deepens our understanding of softmax bottleneck and mixture of softmax (MoS) but also inspires us to propose multi-facet softmax (MFS) to address the limitations of MoS. Extensive empirical analyses confirm our findings and show that against MoS, the proposed MFS achieves two-fold improvements in the perplexity of GPT-2 and BERT.

Universal schema (USchema) assumes that two sentence patterns that share the same entity pairs are similar to each other. This assumption is widely adopted for solving various types of relation extraction (RE) tasks. Nevertheless, each sentence pattern could contain multiple facets, and not every facet is similar to all the facets of another sentence pattern co-occurring with the same entity pair. To address the violation of the USchema assumption, we propose multi-facet universal schema that uses a neural model to represent each sentence pattern as multiple facet embeddings and encourage one of these facet embeddings to be close to that of another sentence pattern if they co-occur with the same entity pair. In our experiments, we demonstrate that multi-facet embeddings significantly outperform their single-facet embedding counterpart, compositional universal schema (CUSchema) (Verga et al., 2016), in distantly supervised relation extraction tasks. Moreover, we can also use multiple embeddings to detect the entailment relation between two sentence patterns when no manual label is available.

Large Transformer-based language models can aid human authors by suggesting plausible continuations of text written so far. However, current interactive writing assistants do not allow authors to guide text generation in desired topical directions. To address this limitation, we design a framework that displays multiple candidate upcoming topics, of which a user can select a subset to guide the generation. Our framework consists of two components: (1) a method that produces a set of candidate topics by predicting the centers of word clusters in the possible continuations, and (2) a text generation model whose output adheres to the chosen topics. The training of both components is self-supervised, using only unlabeled text. Our experiments demonstrate that our topic options are better than those of standard clustering approaches, and our framework often generates fluent sentences related to the chosen topics, as judged by automated metrics and crowdsourced workers.

For many business applications, we often seek to analyze sentiments associated with any arbitrary aspects of commercial products, despite having a very limited amount of labels or even without any labels at all. However, existing aspect target sentiment classification (ATSC) models are not trainable if annotated datasets are not available. Even with labeled data, they fall short of reaching satisfactory performance. To address this, we propose simple approaches that better solve ATSC with natural language prompts, enabling the task under zero-shot cases and enhancing supervised settings, especially for few-shot cases. Under the few-shot setting for SemEval 2014 Task 4 laptop domain, our method of reformulating ATSC as an NLI task outperforms supervised SOTA approaches by up to 24.13 accuracy points and 33.14 macro F1 points. Moreover, we demonstrate that our prompts could handle implicitly stated aspects as well: our models reach about 77% accuracy on detecting sentiments for aspect categories (e.g., food), which do not necessarily appear within the text, even though we trained the models only with explicitly mentioned aspect terms (e.g., fajitas) from just 16 reviews - while the accuracy of the no-prompt baseline is only around 65%.

Materials science literature contains millions of materials synthesis procedures described in unstructured natural language text. Large-scale analysis of these synthesis procedures would facilitate deeper scientific understanding of materials synthesis and enable automated synthesis planning. Such analysis requires extracting structured representations of synthesis procedures from the raw text as a first step. To facilitate the training and evaluation of synthesis extraction models, we introduce a dataset of 230 synthesis procedures annotated by domain experts with labeled graphs that express the semantics of the synthesis sentences. The nodes in this graph are synthesis operations and their typed arguments, and labeled edges specify relations between the nodes. We describe this new resource in detail and highlight some specific challenges to annotating scientific text with shallow semantic structure. We make the corpus available to the community to promote further research and development of scientific information extraction systems.

Modeling hypernymy, such as poodle is-a dog, is an important generalization aid to many NLP tasks, such as entailment, relation extraction, and question answering. Supervised learning from labeled hypernym sources, such as WordNet, limits the coverage of these models, which can be addressed by learning hypernyms from unlabeled text. Existing unsupervised methods either do not scale to large vocabularies or yield unacceptably poor accuracy. This paper introduces distributional inclusion vector embedding (DIVE), a simple-to-implement unsupervised method of hypernym discovery via per-word non-negative vector embeddings which preserve the inclusion property of word contexts. In experimental evaluations more comprehensive than any previous literature of which we are aware—evaluating on 11 datasets using multiple existing as well as newly proposed scoring functions—we find that our method provides up to double the precision of previous unsupervised methods, and the highest average performance, using a much more compact word representation, and yielding many new state-of-the-art results.

Word sense induction (WSI), which addresses polysemy by unsupervised discovery of multiple word senses, resolves ambiguities for downstream NLP tasks and also makes word representations more interpretable. This paper proposes an accurate and efficient graph-based method for WSI that builds a global non-negative vector embedding basis (which are interpretable like topics) and clusters the basis indexes in the ego network of each polysemous word. By adopting distributional inclusion vector embeddings as our basis formation model, we avoid the expensive step of nearest neighbor search that plagues other graph-based methods without sacrificing the quality of sense clusters. Experiments on three datasets show that our proposed method produces similar or better sense clusters and embeddings compared with previous state-of-the-art methods while being significantly more efficient.