Transformer models cannot easily scale to long sequences due to their O(Nˆ2) time and space complexity. This has led to Transformer variants seeking to lower computational complexity, such as Longformer and Performer. While such models have theoretically greater efficiency, their effectiveness on real NLP tasks has not been well studied. We benchmark 7 variants of Transformer models on 5 difficult NLP tasks and 7 datasets. We design experiments to isolate the effect of pretraining and hyperparameter settings, to focus on their capacity for long-range attention. Moreover, we present various methods to investigate attention behaviors to illuminate model details beyond metric scores. We find that the modified attention in long-range transformers has advantages on content selection and query-guided decoding, but they come with previously unrecognized drawbacks such as insufficient attention to distant tokens and accumulated approximation error.

The Transformer architecture has led to significant gains in machine translation. However, most studies focus on only sentence-level translation without considering the context dependency within documents, leading to the inadequacy of document-level coherence. Some recent research tried to mitigate this issue by introducing an additional context encoder or translating with multiple sentences or even the entire document. Such methods may lose the information on the target side or have an increasing computational complexity as documents get longer. To address such problems, we introduce a recurrent memory unit to the vanilla Transformer, which supports the information exchange between the sentence and previous context. The memory unit is recurrently updated by acquiring information from sentences, and passing the aggregated knowledge back to subsequent sentence states. We follow a two-stage training strategy, in which the model is first trained at the sentence level and then finetuned for document-level translation. We conduct experiments on three popular datasets for document-level machine translation and our model has an average improvement of 0.91 s-BLEU over the sentence-level baseline. We also achieve state-of-the-art results on TED and News, outperforming the previous work by 0.36 s-BLEU and 1.49 d-BLEU on average.

Recent work has shown advantages of incorporating knowledge graphs (KGs) into BERT for various NLP tasks. One common way is to feed entity embeddings as an additional input during pre-training. There are two limitations to such a method. First, to train the entity embeddings to include rich information of factual knowledge, it typically requires access to the entire KG. This is challenging for KGs with daily changes (e.g., Wikidata). Second, it requires a large scale pre-training corpus with entity annotations and high computational cost during pre-training. In this work, we efficiently construct entity embeddings only from the type knowledge, that does not require access to the entire KG. Although the entity embeddings contain only local information, they perform very well when combined with context. Furthermore, we show that our entity embeddings, constructed from BERT’s input embeddings, can be directly incorporated into the fine-tuning phase without requiring any specialized pre-training. In addition, these entity embeddings can also be constructed on the fly without requiring a large memory footprint to store them. Finally, we propose task-specific models that incorporate our entity embeddings for entity linking, entity typing, and relation classification. Experiments show that our models have comparable or superior performance to existing models while being more resource efficient.

Multilingual pretrained models have shown strong cross-lingual transfer ability. Some works used code-switching sentences, which consist of tokens from multiple languages, to enhance the cross-lingual representation further, and have shown success in many zero-shot cross-lingual tasks. However, code-switched tokens are likely to cause grammatical incoherence in newly substituted sentences, and negatively affect the performance on token-sensitive tasks, such as Part-of-Speech (POS) tagging and Named-Entity-Recognition (NER). This paper mitigates the limitation of the code-switching method by not only making the token replacement but considering the similarity between the context and the switched tokens so that the newly substituted sentences are grammatically consistent during both training and inference. We conduct experiments on cross-lingual POS and NER over 30+ languages, and demonstrate the effectiveness of our method by outperforming the mBERT by 0.95 and original code-switching method by 1.67 on F1 scores.

Building pretrained language models is considered expensive and data-intensive, but must we increase dataset size to achieve better performance? We propose an alternative to larger training sets by automatically identifying smaller yet domain-representative subsets. We extend Cynical Data Selection, a statistical sentence scoring method that conditions on a representative target domain corpus. As an example, we treat the OntoNotes corpus as a target domain and pretrain a RoBERTa-like encoder from a cynically selected subset of the Pile. On both perplexity and across several downstream tasks in the target domain, it consistently outperforms random selection with 20x less data, 3x fewer training iterations, and 2x less estimated cloud compute cost, validating the recipe of automatic document selection for LM pretraining.

Character-aware neural language models can capture the relationship between words by exploiting character-level information and are particularly effective for languages with rich morphology. However, these models are usually biased towards information from surface forms. To alleviate this problem, we propose a simple and effective method to improve a character-aware neural language model by forcing a character encoder to produce word-based embeddings under Skip-gram architecture in a warm-up step without extra training data. We empirically show that the resulting character-aware neural language model achieves obvious improvements of perplexity scores on typologically diverse languages, that contain many low-frequency or unseen words.

This work presents multi-modal deep SVDD (mSVDD) for one-class text classification. By extending the uni-modal SVDD to a multiple modal one, we build mSVDD with multiple hyperspheres, that enable us to build a much better description for target one-class data. Additionally, the end-to-end architecture of mSVDD can jointly handle neural feature learning and one-class text learning. We also introduce a mechanism for incorporating negative supervision in the absence of real negative data, which can be beneficial to the mSVDD model. We conduct experiments on Reuters and 20 Newsgroup datasets, and the experimental results demonstrate that mSVDD outperforms uni-modal SVDD and mSVDD can get further improvements when negative supervision is incorporated.

We propose a simple and effective method for incorporating word clusters into the Continuous Bag-of-Words (CBOW) model. Specifically, we propose to replace infrequent input and output words in CBOW model with their clusters. The resulting cluster-incorporated CBOW model produces embeddings of frequent words and a small amount of cluster embeddings, which will be fine-tuned in downstream tasks. We empirically show our replacing method works well on several downstream tasks. Through our analysis, we show that our method might be also useful for other similar models which produce word embeddings.

We propose a simple and effective method to inject word-level information into character-aware neural language models. Unlike previous approaches which usually inject word-level information at the input of a long short-term memory (LSTM) network, we inject it into the softmax function. The resultant model can be seen as a combination of character-aware language model and simple word-level language model. Our injection method can also be used together with previous methods. Through the experiments on 14 typologically diverse languages, we empirically show that our injection method, when used together with the previous methods, works better than the previous methods, including a gating mechanism, averaging, and concatenation of word vectors. We also provide a comprehensive comparison of these injection methods.

Scholars in inter-disciplinary fields like the Digital Humanities are increasingly interested in semantic annotation of specialized corpora. Yet, under-resourced languages, imperfect or noisily structured data, and user-specific classification tasks make it difficult to meet their needs using off-the-shelf models. Manual annotation of large corpora from scratch, meanwhile, can be prohibitively expensive. Thus, we propose an active learning solution for named entity recognition, attempting to maximize a custom model’s improvement per additional unit of manual annotation. Our system robustly handles any domain or user-defined label set and requires no external resources, enabling quality named entity recognition for Humanities corpora where such resources are not available. Evaluating on typologically disparate languages and datasets, we reduce required annotation by 20-60% and greatly outperform a competitive active learning baseline.

Sequence labeling of biomedical entities, e.g., side effects or phenotypes, was a long-term task in BioNLP and MedNLP communities. Thanks to effects made among these communities, adverse reaction NER has developed dramatically in recent years. As an illuminative application, to achieve knowledge discovery via the combination of the text mining result and bioinformatics idea shed lights on the pharmacological mechanism research.

This paper explores the automatic learning of distributed representations of the target’s context for semantic frame labeling with target-based neural model. We constrain the whole sentence as the model’s input without feature extraction from the sentence. This is different from many previous works in which local feature extraction of the targets is widely used. This constraint makes the task harder, especially with long sentences, but also makes our model easily applicable to a range of resources and other similar tasks. We evaluate our model on several resources and get the state-of-the-art result on subtask 2 of SemEval 2015 task 15. Finally, we extend the task to word-sense disambiguation task and we also achieve a strong result in comparison to state-of-the-art work.