EMNLP 2022 论文列表

We propose a simple refactoring of multi-choice question answering (MCQA) tasks as a series of binary classifications. The MCQA task is generally performed by scoring each (question, answer) pair normalized over all the pairs, and then selecting the answer from the pair that yield the highest score. For n answer choices, this is equivalent to an n-class classification setup where only one class (true answer) is correct. We instead show that classifying (question, true answer) as positive instances and (question, false answer) as negative instances is significantly more effective across various models and datasets. We show the efficacy of our proposed approach in different tasks – abductive reasoning, commonsense question answering, science question answering, and sentence completion. Our DeBERTa binary classification model reaches the top or close to the top performance on public leaderboards for these tasks. The source code of the proposed approach is available at https://github.com/declare-lab/TEAM.

Many language generation models are now available for a wide range of generation tasks, including machine translation and summarization. Combining such diverse models may lead to further progress, but ensembling generation models is challenging during inference: conventional ensembling methods (e.g., shallow fusion) require that the models share vocabulary/tokenization schemes. We introduce Twist decoding, a simple and general text generation algorithm that benefits from diverse models at inference time. Our method does not assume the vocabulary, tokenization or even generation order is shared. Our extensive evaluations on machine translation and scientific paper summarization demonstrate that Twist decoding substantially outperforms each model decoded in isolation over various scenarios, including cases where domain-specific and general-purpose models are both available. Twist decoding also consistently outperforms the popular reranking heuristic where output candidates from one model are rescored by another. We hope that our work will encourage researchers and practitioners to examine generation models collectively, not just independently, and to seek out models with complementary strengths to the currently available models.

Non-Fungible Tokens (NFTs) are a relatively unexplored class of assets. Designing strategies to forecast NFT trends is an intricate task due to its extremely volatile nature. The market is largely driven by public sentiment and “hype”, which in turn has a high correlation with conversations taking place on social media platforms like Twitter. Prior work done for modelling stock market data does not take into account the extent of impact certain highly influential tweets and their authors can have on the market. Building on these limitations and the nature of the NFT market, we propose a novel reach-aware temporal learning approach to make predictions for forecasting future trends in the NFT market. We perform experiments on a new dataset consisting of over 1.3 million tweets and 180 thousand NFT transactions spanning over 15 NFT collections curated by us. Our model (TA-NFT) outperforms other state-of-the-art methods by an average of 36%. Through extensive quantitative and ablative analysis, we demonstrate the ability of our approach as a practical method for predicting NFT trends.

Pre-trained language models have achieved remarkable successes in natural language processing tasks, coming at the cost of increasing model size. To address this issue, knowledge distillation (KD) has been widely applied to compress language models. However, typical KD approaches for language models have overlooked the difficulty of training examples, suffering from incorrect teacher prediction transfer and sub-efficient training. In this paper, we propose a novel KD framework, Tutor-KD, which improves the distillation effectiveness by controlling the difficulty of training examples during pre-training. We introduce a tutor network that generates samples that are easy for the teacher but difficult for the student, with training on a carefully designed policy gradient method. Experimental results show that Tutor-KD significantly and consistently outperforms the state-of-the-art KD methods with variously sized student models on the GLUE benchmark, demonstrating that the tutor can effectively generate training examples for the student.

Simultaneous machine translation (SiMT) starts its translation before reading the whole source sentence and employs either fixed or adaptive policy to generate the target sentence. Compared to the fixed policy, the adaptive policy achieves better latency-quality tradeoffs by adopting a flexible translation policy. If the policy can evaluate rationality before taking action, the probability of incorrect actions will also decrease. However, previous methods lack evaluation of actions before taking them. In this paper, we propose a method of performing the adaptive policy via integrating post-evaluation into the fixed policy. Specifically, whenever a candidate token is generated, our model will evaluate the rationality of the next action by measuring the change in the source content. Our model will then take different actions based on the evaluation results. Experiments on three translation tasks show that our method can exceed strong baselines under all latency.

Long samples of text from neural language models can be of poor quality. Truncation sampling algorithms–like top-p or top-k—address this by setting some words’ probabilities to zero at each step. This work investigates why these methods are important, and how to improve them. We propose thinking of a neural language model as a mixture of a true distribution and a smoothing distribution that avoids infinite perplexity. In this light, truncation algorithms aim to perform desmoothing, estimating a subset of the support of the true distribution. Finding a good subset is crucial: we show that top-p unnecessarily truncates high-probability words, for example causing it to truncate all words but Trump for a document that starts with Donald. We introduce eta-sampling, which truncates words below an entropy-dependent probability threshold. Compared to previous algorithms, our eta-sampling generates more plausible long documents according to humans, is better at breaking out of repetition, and behaves more reasonably on a battery of test distributions.

Clinical trials are essential for drug development but are extremely expensive and time-consuming to conduct. It is beneficial to study similar historical trials when designing a clinical trial. However, lengthy trial documents and lack of labeled data make trial similarity search difficult. We propose a zero-shotclinical trial retrieval method, called Trial2Vec, which learns through self-supervision without the need for annotating similar clinical trials. Specifically, the meta-structure of trial documents (e.g., title, eligibility criteria, target disease) along with clinical knowledge (e.g., UMLS knowledge base) are leveraged to automatically generate contrastive samples. Besides, encodes trial documents considering meta-structure thus producing compact embeddings aggregating multi-aspect information from the whole document. We show that our method yields medically interpretable embeddings by visualization and it gets 15% average improvement over the best baselines on precision/recall for trial retrieval, which is evaluated on our labeled 1600 trial pairs. In addition, we prove the pretrained embeddings benefit the downstream trial outcome prediction task over 240k trials. Software is available at https://github.com/RyanWangZf/Trial2Vec.

We present MolT5 - a self-supervised learning framework for pretraining models on a vast amount of unlabeled natural language text and molecule strings. MolT5 allows for new, useful, and challenging analogs of traditional vision-language tasks, such as molecule captioning and text-based de novo molecule generation (altogether: translation between molecules and language), which we explore for the first time. Since MolT5 pretrains models on single-modal data, it helps overcome the chemistry domain shortcoming of data scarcity. Furthermore, we consider several metrics, including a new cross-modal embedding-based metric, to evaluate the tasks of molecule captioning and text-based molecule generation. Our results show that MolT5-based models are able to generate outputs, both molecules and captions, which in many cases are high quality.

The Annals of Joseon Dynasty (AJD) contain the daily records of the Kings of Joseon, the 500-year kingdom preceding the modern nation of Korea.The Annals were originally written in an archaic Korean writing system, ‘Hanja’, and were translated into Korean from 1968 to 1993.The resulting translation was however too literal and contained many archaic Korean words; thus, a new expert translation effort began in 2012. Since then, the records of only one king have been completed in a decade.In parallel, expert translators are working on English translation, also at a slow pace and produced only one king’s records in English so far.Thus, we propose H2KE, a neural machine translation model, that translates historical documents in Hanja to more easily understandable Korean and to English.Built on top of multilingual neural machine translation, H2KE learns to translate a historical document written in Hanja, from both a full dataset of outdated Korean translation and a small dataset of more recently translated contemporary Korean and English.We compare our method against two baselines:a recent model that simultaneously learns to restore and translate Hanja historical documentand a Transformer based model trained only on newly translated corpora.The experiments reveal that our method significantly outperforms the baselines in terms of BLEU scores for both contemporary Korean and English translations.We further conduct extensive human evaluation which shows that our translation is preferred over the original expert translations by both experts and non-expert Korean speakers.

Pretrained, large, generative language models (LMs) have had great success in a wide range of sequence tagging and structured prediction tasks. Casting a sequence tagging task as a Seq2Seq one requires deciding the formats of the input and output sequences. However, we lack a principled understanding of the trade-offs associated with these formats (such as the effect on model accuracy, sequence length, multilingual generalization, hallucination). In this paper, we rigorously study different formats one could use for casting input text sentences and their output labels into the input and target (i.e., output) of a Seq2Seq model. Along the way, we introduce a new format, which we show to to be both simpler and more effective. Additionally the new format demonstrates significant gains in the multilingual settings – both zero-shot transfer learning and joint training. Lastly, we find that the new format is more robust and almost completely devoid of hallucination – an issue we find common in existing formats. With well over a 1000 experiments studying 14 different formats, over 7 diverse public benchmarks – including 3 multilingual datasets spanning 7 languages – we believe our findings provide a strong empirical basis in understanding how we should tackle sequence tagging tasks.

We investigate the knowledge graph entity typing task which aims at inferring plausible entity types. In this paper, we propose a novel Transformer-based Entity Typing (TET) approach, effectively encoding the content of neighbours of an entity by means of a transformer mechanism. More precisely, TET is composed of three different mechanisms: a local transformer allowing to infer missing entity types by independently encoding the information provided by each of its neighbours; a global transformer aggregating the information of all neighbours of an entity into a single long sequence to reason about more complex entity types; and a context transformer integrating neighbours content in a differentiated way through information exchange between neighbour pairs, while preserving the graph structure. Furthermore, TET uses information about class membership of types to semantically strengthen the representation of an entity. Experiments on two real-world datasets demonstrate the superior performance of TET compared to the state-of-the-art.

Causal transformer language models (LMs), such as GPT-3, typically require some form of positional encoding, such as positional embeddings. However, we show that LMs without any explicit positional encoding are still competitive with standard models and that this phenomenon is robust across different datasets, model sizes, and sequence lengths.Probing experiments reveal that such models acquire an implicit notion of absolute positions throughout the network, effectively compensating for the missing information.We conjecture that causal attention enables the model to infer the number of predecessors that each token can attend to, thereby approximating its absolute position.Our findings indicate that causal LMs might derive positional awareness not only from the explicit positioning mechanism but also from the effects of the causal mask.

Transformer-based language models (LMs) are at the core of modern NLP, but their internal prediction construction process is opaque and largely not understood. In this work, we make a substantial step towards unveiling this underlying prediction process, by reverse-engineering the operation of the feed-forward network (FFN) layers, one of the building blocks of transformer models. We view the token representation as a changing distribution over the vocabulary, and the output from each FFN layer as an additive update to that distribution. Then, we analyze the FFN updates in the vocabulary space, showing that each update can be decomposed to sub-updates corresponding to single FFN parameter vectors, each promoting concepts that are often human-interpretable. We then leverage these findings for controlling LM predictions, where we reduce the toxicity of GPT2 by almost 50%, and for improving computation efficiency with a simple early exit rule, saving 20% of computation on average.

Recent research shows synthetic data as a source of supervision helps pretrained language models (PLM) transfer learning to new target tasks/domains. However, this idea is less explored for spatial language. We provide two new data resources on multiple spatial language processing tasks. The first dataset is synthesized for transfer learning on spatial question answering (SQA) and spatial role labeling (SpRL). Compared to previous SQA datasets, we include a larger variety of spatial relation types and spatial expressions. Our data generation process is easily extendable with new spatial expression lexicons. The second one is a real-world SQA dataset with human-generated questions built on an existing corpus with SPRL annotations. This dataset can be used to evaluate spatial language processing models in realistic situations. We show pretraining with automatically generated data significantly improves the SOTA results on several SQA and SPRL benchmarks, particularly when the training data in the target domain is small.

In this work, we investigate transfer learning from semantic role labeling (SRL) to event argument extraction (EAE), considering their similar argument structures. We view the extraction task as a role querying problem, unifying various methods into a single framework. There are key discrepancies on role labels and distant arguments between semantic role and event argument annotations. To mitigate these discrepancies, we specify natural language-like queries to tackle the label mismatch problem and devise argument augmentation to recover distant arguments. We show that SRL annotations can serve as a valuable resource for EAE, and a template-based slot querying strategy is especially effective for facilitating the transfer. In extensive evaluations on two English EAE benchmarks, our proposed model obtains impressive zero-shot results by leveraging SRL annotations, reaching nearly 80% of the fullysupervised scores. It further provides benefits in low-resource cases, where few EAE annotations are available. Moreover, we show that our approach generalizes to cross-domain and multilingual scenarios.

Sanskrit Word Segmentation (SWS) is essential in making digitized texts available and in deploying downstream tasks. It is, however, non-trivial because of the sandhi phenomenon that modifies the characters at the word boundaries, and needs special treatment. Existing lexicon driven approaches for SWS make use of Sanskrit Heritage Reader, a lexicon-driven shallow parser, to generate the complete candidate solution space, over which various methods are applied to produce the most valid solution. However, these approaches fail while encountering out-of-vocabulary tokens. On the other hand, purely engineering methods for SWS have made use of recent advances in deep learning, but cannot make use of the latent word information on availability. To mitigate the shortcomings of both families of approaches, we propose Transformer based Linguistically Informed Sanskrit Tokenizer (TransLIST) consisting of (1) a module that encodes the character input along with latent-word information, which takes into account the sandhi phenomenon specific to SWS and is apt to work with partial or no candidate solutions, (2) a novel soft-masked attention to prioritize potential candidate words and (3) a novel path ranking algorithm to rectify the corrupted predictions. Experiments on the benchmark datasets for SWS show that TransLIST outperforms the current state-of-the-art system by an average 7.2 points absolute gain in terms of perfect match (PM) metric.

Zero-Resource Cross-Lingual Named Entity Recognition aims at training an NER model of the target language using only labeled source language data and unlabeled target language data. Existing methods are mainly divided into three categories: model transfer based, data transfer based and knowledge transfer based. Each method has its own disadvantages, and combining more than one of them often leads to better performance. However, the performance of data transfer based methods is often limited by inevitable noise in the translation process. To handle the problem, we propose a framework named TransAdv to mitigate lexical and syntactic errors of word-by-word translated data, better utilizing the data by multi-level adversarial learning and multi-model knowledge distillation. Extensive experiments are conducted over 6 target languages with English as the source language, and the results show that TransAdv achieves competitive performance to the state-of-the-art models.

Knowledge graph embedding methods are important for the knowledge graph completion (or link prediction) task.One state-of-the-art method, PairRE, leverages two separate vectors to model complex relations (i.e., 1-to-N, N-to-1, and N-to-N) in knowledge graphs. However, such a method strictly restricts entities on the hyper-ellipsoid surfaces which limits the optimization of entity distribution, leading to suboptimal performance of knowledge graph completion. To address this issue, we propose a novel score function TranSHER, which leverages relation-specific translations between head and tail entities to relax the constraint of hyper-ellipsoid restrictions. By introducing an intuitive and simple relation-specific translation, TranSHER can provide more direct guidance on optimization and capture more semantic characteristics of entities with complex relations. Experimental results show that TranSHER achieves state-of-the-art performance on link prediction and generalizes well to datasets in different domains and scales. Our codes are public available athttps://github.com/yizhilll/TranSHER.

Knowledge graph embedding (KGE) aims to learn continuous vector representations of relations and entities in knowledge graph (KG). Recently, transition-based KGE methods have become popular and achieved promising performance. However, scoring patterns like TransE are not suitable for complex scenarios where the same entity pair has different relations. Although some models attempt to employ entity-relation interaction or projection to improve entity representation for one-to-many/many-to-one/many-to-many complex relations, they still continue the traditional scoring pattern, where only a single relation vector in the relation part is used to translate the head entity to the tail entity or their variants. And recent research shows that entity representation only needs to consider entities and their interactions to achieve better performance. Thus, in this paper, we propose a novel transition-based method, TranS, for KGE. The single relation vector of the relation part in the traditional scoring pattern is replaced by the synthetic relation representation with entity-relation interactions to solve these issues. And the entity part still retains its independence through entity-entity interactions. Experiments on a large KG dataset, ogbl-wikikg2, show that our model achieves state-of-the-art results.