EMNLP 2021 Papers

Recent progress in generative language models has enabled machines to generate astonishingly realistic texts. While there are many legitimate applications of such models, there is also a rising need to distinguish machine-generated texts from human-written ones (e.g., fake news detection). However, to our best knowledge, there is currently no benchmark environment with datasets and tasks to systematically study the so-called ”Turing Test” problem for neural text generation methods. In this work, we present the TURINGBENCH benchmark environment, which is comprised of (1) a dataset with 200K human- or machine-generated samples across 20 labels Human, GPT-1, GPT-2_small, GPT-2_medium, GPT-2_large,GPT-2_xl, GPT-2_PyTorch, GPT-3, GROVER_base, GROVER_large, GROVER_mega, CTRL, XLM, XLNET_base, XLNET_large, FAIR_wmt19, FAIR_wmt20, TRANSFORMER_XL, PPLM_distil, PPLM_gpt2, (2) two benchmark tasks–i.e., Turing Test (TT) and Authorship Attribution (AA), and (3) a website with leaderboards. Our preliminary experimental results using TURINGBENCH show that GPT-3 and FAIR_wmt20 are the current winners, among all language models tested, in generating the most human-like indistinguishable texts with the lowest F1 score by five state-of-the-art TT detection models. The TURINGBENCH is available at: https://turingbench.ist.psu.edu/

Learning sentence embeddings often requires a large amount of labeled data. However, for most tasks and domains, labeled data is seldom available and creating it is expensive. In this work, we present a new state-of-the-art unsupervised method based on pre-trained Transformers and Sequential Denoising Auto-Encoder (TSDAE) which outperforms previous approaches by up to 6.4 points. It can achieve up to 93.1% of the performance of in-domain supervised approaches. Further, we show that TSDAE is a strong domain adaptation and pre-training method for sentence embeddings, significantly outperforming other approaches like Masked Language Model. A crucial shortcoming of previous studies is the narrow evaluation: Most work mainly evaluates on the single task of Semantic Textual Similarity (STS), which does not require any domain knowledge. It is unclear if these proposed methods generalize to other domains and tasks. We fill this gap and evaluate TSDAE and other recent approaches on four different datasets from heterogeneous domains.

Human conversations naturally evolve around different topics and fluently move between them. In research on dialog systems, the ability to actively and smoothly transition to new topics is often ignored. In this paper we introduce TIAGE, a new topic-shift aware dialog benchmark constructed utilizing human annotations on topic shifts. Based on TIAGE, we introduce three tasks to investigate different scenarios of topic-shift modeling in dialog settings: topic-shift detection, topic-shift triggered response generation and topic-aware dialog generation. Experiments on these tasks show that the topic-shift signals in TIAGE are useful for topic-shift response generation. On the other hand, dialog systems still struggle to decide when to change topic. This indicates further research is needed in topic-shift aware dialog modeling.

As an essential form of knowledge representation, taxonomies are widely used in various downstream natural language processing tasks. However, with the continuously rising of new concepts, many existing taxonomies are unable to maintain coverage by manual expansion. In this paper, we propose TEMP, a self-supervised taxonomy expansion method, which predicts the position of new concepts by ranking the generated taxonomy-paths. For the first time, TEMP employs pre-trained contextual encoders in taxonomy construction and hypernym detection problems. Experiments prove that pre-trained contextual embeddings are able to capture hypernym-hyponym relations. To learn more detailed differences between taxonomy-paths, we train the model with dynamic margin loss by a novel dynamic margin function. Extensive evaluations exhibit that TEMP outperforms prior state-of-the-art taxonomy expansion approaches by 14.3% in accuracy and 15.8% in mean reciprocal rank on three public benchmarks.

To alleviate label scarcity in Named Entity Recognition (NER) task, distantly supervised NER methods are widely applied to automatically label data and identify entities. Although the human effort is reduced, the generated incomplete and noisy annotations pose new challenges for learning effective neural models. In this paper, we propose a novel dictionary extension method which extracts new entities through the type expanded model. Moreover, we design a multi-granularity boundary-aware network which detects entity boundaries from both local and global perspectives. We conduct experiments on different types of datasets, the results show that our model outperforms previous state-of-the-art distantly supervised systems and even surpasses the supervised models.

Joint extraction of entities and relations from unstructured texts to form factual triples is a fundamental task of constructing a Knowledge Base (KB). A common method is to decode triples by predicting entity pairs to obtain the corresponding relation. However, it is still challenging to handle this task efficiently, especially for the overlapping triple problem. To address such a problem, this paper proposes a novel efficient entities and relations extraction model called TDEER, which stands for Translating Decoding Schema for Joint Extraction of Entities and Relations. Unlike the common approaches, the proposed translating decoding schema regards the relation as a translating operation from subject to objects, i.e., TDEER decodes triples as subject + relation → objects. TDEER can naturally handle the overlapping triple problem, because the translating decoding schema can recognize all possible triples, including overlapping and non-overlapping triples. To enhance model robustness, we introduce negative samples to alleviate error accumulation at different stages. Extensive experiments on public datasets demonstrate that TDEER produces competitive results compared with the state-of-the-art (SOTA) baselines. Furthermore, the computation complexity analysis indicates that TDEER is more efficient than powerful baselines. Especially, the proposed TDEER is 2 times faster than the recent SOTA models. The code is available at https://github.com/4AI/TDEER.

Although distributed learning has increasingly gained attention in terms of effectively utilizing local devices for data privacy enhancement, recent studies show that publicly shared gradients in the training process can reveal the private training data (gradient leakage) to a third-party. We have, however, no systematic understanding of the gradient leakage mechanism on the Transformer based language models. In this paper, as the first attempt, we formulate the gradient attack problem on the Transformer-based language models and propose a gradient attack algorithm, TAG, to reconstruct the local training data. Experimental results on Transformer, TinyBERT4, TinyBERT6 BERT_BASE, and BERT_LARGE using GLUE benchmark show that compared with DLG, TAG works well on more weight distributions in reconstructing training data and achieves 1.5x recover rate and 2.5x ROUGE-2 over prior methods without the need of ground truth label. TAG can obtain up to 90% data by attacking gradients in CoLA dataset. In addition, TAG is stronger than previous approaches on larger models, smaller dictionary size, and smaller input length. We hope the proposed TAG will shed some light on the privacy leakage problem in Transformer-based NLP models.

The paradigm of pre-training followed by finetuning has become a standard procedure for NLP tasks, with a known problem of domain shift between the pre-training and downstream corpus. Previous works have tried to mitigate this problem with additional pre-training, either on the downstream corpus itself when it is large enough, or on a manually curated unlabeled corpus of a similar domain. In this paper, we address the problem for the case when the downstream corpus is too small for additional pre-training. We propose TADPOLE, a task adapted pre-training framework based on data selection techniques adapted from Domain Adaptation. We formulate the data selection as an anomaly detection problem that unlike existing methods works well when the downstream corpus is limited in size. It results in a scalable and efficient unsupervised technique that eliminates the need for any manual data curation. We evaluate our framework on eight tasks across four different domains: Biomedical, Computer Science, News, and Movie reviews, and compare its performance against competitive baseline techniques from the area of Domain Adaptation. Our framework outperforms all the baseline methods. On small datasets with less than 5K training examples, we get a gain of 1.82% in performance with additional pre-training for only 5% steps compared to the originally pre-trained models. It also compliments some of the other techniques such as data augmentation known for boosting performance when downstream corpus is small; highest performance is achieved when data augmentation is combined with task adapted pre-training.

Transformers are the dominant architecture in NLP, but their training and fine-tuning is still very challenging. In this paper, we present the design and implementation of a visual analytic framework for assisting researchers in such process, by providing them with valuable insights about the model’s intrinsic properties and behaviours. Our framework offers an intuitive overview that allows the user to explore different facets of the model (e.g., hidden states, attention) through interactive visualization, and allows a suite of built-in algorithms that compute the importance of model components and different parts of the input sequence. Case studies and feedback from a user focus group indicate that the framework is useful, and suggest several improvements. Our framework is available at: https://github.com/raymondzmc/T3-Vis.

We analyze the grounded SCAN (gSCAN) benchmark, which was recently proposed to study systematic generalization for grounded language understanding. First, we study which aspects of the original benchmark can be solved by commonly used methods in multi-modal research. We find that a general-purpose Transformer-based model with cross-modal attention achieves strong performance on a majority of the gSCAN splits, surprisingly outperforming more specialized approaches from prior work. Furthermore, our analysis suggests that many of the remaining errors reveal the same fundamental challenge in systematic generalization of linguistic constructs regardless of visual context. Second, inspired by this finding, we propose challenging new tasks for gSCAN by generating data to incorporate relations between objects in the visual environment. Finally, we find that current models are surprisingly data inefficient given the narrow scope of commands in gSCAN, suggesting another challenge for future work.

Basic-level categories (BLC) are an important psycholinguistic concept introduced by Rosch et al. (1976); they are defined as the most inclusive categories for which a concrete mental image of the category as a whole can be formed, and also as those categories which are acquired early in life. Rosch’s original algorithm for detecting BLC (called cue-validity) is based on the availability of semantic features such as “has tail” for “cat”, and has remained untested at large. An at-scale algorithm for the automatic determination of BLC exists, but it operates without Rosch-style semantic features, and is thus unable to verify Rosch’s hypothesis. We present the first method for the detection of BLC at scale that makes use of Rosch-style semantic features. For both English and Mandarin, we test three methods of generating such features for any synset within Wordnet (WN): extraction of textual features from Wikipedia pages, Distributional Memory (DM) and BART. The best of our methods outperforms the current SoA in BLC detection, with an accuracy of English BLC detection of 75.0%, and of Mandarin BLC detection 80.7% on a test set. When applied to all of WordNet, our model predicts that 1,118 synsets in English Wordnet (1.4%) are BLC, far fewer than existing methods, and with a precision improvement of over 200% over these. As well as confirming the usefulness of Rosch’s cue validity algorithm, we also developed and evaluated our own new indicator for BLC, which models the fact that BLC features tend to be BLC themselves.

Coupled with the availability of large scale datasets, deep learning architectures have enabled rapid progress on the Question Answering task. However, most of those datasets are in English, and the performances of state-of-the-art multilingual models are significantly lower when evaluated on non-English data. Due to high data collection costs, it is not realistic to obtain annotated data for each language one desires to support. We propose a method to improve the Cross-lingual Question Answering performance without requiring additional annotated data, leveraging Question Generation models to produce synthetic samples in a cross-lingual fashion. We show that the proposed method allows to significantly outperform the baselines trained on English data only. We report a new state-of-the-art on four datasets: MLQA, XQuAD, SQuAD-it and PIAF (fr).

Previous works on syntactically controlled paraphrase generation heavily rely on large-scale parallel paraphrase data that is not easily available for many languages and domains. In this paper, we take this research direction to the extreme and investigate whether it is possible to learn syntactically controlled paraphrase generation with nonparallel data. We propose a syntactically-informed unsupervised paraphrasing model based on conditional variational auto-encoder (VAE) which can generate texts in a specified syntactic structure. Particularly, we design a two-stage learning method to effectively train the model using non-parallel data. The conditional VAE is trained to reconstruct the input sentence according to the given input and its syntactic structure. Furthermore, to improve the syntactic controllability and semantic consistency of the pre-trained conditional VAE, we fine-tune it using syntax controlling and cycle reconstruction learning objectives, and employ Gumbel-Softmax to combine these new learning objectives. Experiment results demonstrate that the proposed model trained only on non-parallel data is capable of generating diverse paraphrases with specified syntactic structure. Additionally, we validate the effectiveness of our method for generating syntactically adversarial examples on the sentiment analysis task.

Generative conversation systems tend to produce meaningless and generic responses, which significantly reduce the user experience. In order to generate informative and diverse responses, recent studies proposed to fuse knowledge to improve informativeness and adopt latent variables to enhance the diversity. However, utilizing latent variables will lead to the inaccuracy of knowledge in the responses, and the dissemination of wrong knowledge will mislead the communicators. To address this problem, we propose a Syntactically Diverse Adversarial Network (SDAN) for knowledge-grounded conversation model. SDAN contains an adversarial hierarchical semantic network to keep the semantic coherence, a knowledge-aware network to attend more related knowledge for improving the informativeness and a syntactic latent variable network to generate syntactically diverse responses. Additionally, in order to increase the controllability of syntax, we adopt adversarial learning to decouple semantic and syntactic representations. Experimental results show that our model can not only generate syntactically diverse and knowledge-accurate responses but also significantly achieve the balance between improving the syntactic diversity and maintaining the knowledge accuracy.

Joint entity and relation extraction is challenging due to the complex interaction of interaction between named entity recognition and relation extraction. Although most existing works tend to jointly train these two tasks through a shared network, they fail to fully utilize the interdependence between entity types and relation types. In this paper, we design a novel synchronous dual network (SDN) with cross-type attention via separately and interactively considering the entity types and relation types. On the one hand, SDN adopts two isomorphic bi-directional type-attention LSTM to encode the entity type enhanced representations and the relation type enhanced representations, respectively. On the other hand, SDN explicitly models the interdependence between entity types and relation types via cross-type attention mechanism. In addition, we also propose a new multi-task learning strategy via modeling the interaction of two types of information. Experiments on NYT and WebNLG datasets verify the effectiveness of the proposed model, achieving state-of-the-art performance.

Code-switching (CS), a ubiquitous phenomenon due to the ease of communication it offers in multilingual communities still remains an understudied problem in language processing. The primary reasons behind this are: (1) minimal efforts in leveraging large pretrained multilingual models, and (2) the lack of annotated data. The distinguishing case of low performance of multilingual models in CS is the intra-sentence mixing of languages leading to switch points. We first benchmark two sequence labeling tasks – POS and NER on 4 different language pairs with a suite of pretrained models to identify the problems and select the best performing char-BERT model among them (addressing (1)). We then propose a self training method to repurpose the existing pretrained models using a switch-point bias by leveraging unannotated data (addressing (2)). We finally demonstrate that our approach performs well on both tasks by reducing the gap between the switch point performance while retaining the overall performance on two distinct language pairs in both the tasks. We plan to release our models and the code for all our experiments.

Unfair stereotypical biases (e.g., gender, racial, or religious biases) encoded in modern pretrained language models (PLMs) have negative ethical implications for widespread adoption of state-of-the-art language technology. To remedy for this, a wide range of debiasing techniques have recently been introduced to remove such stereotypical biases from PLMs. Existing debiasing methods, however, directly modify all of the PLMs parameters, which – besides being computationally expensive – comes with the inherent risk of (catastrophic) forgetting of useful language knowledge acquired in pretraining. In this work, we propose a more sustainable modular debiasing approach based on dedicated debiasing adapters, dubbed ADELE. Concretely, we (1) inject adapter modules into the original PLM layers and (2) update only the adapters (i.e., we keep the original PLM parameters frozen) via language modeling training on a counterfactually augmented corpus. We showcase ADELE, in gender debiasing of BERT: our extensive evaluation, encompassing three intrinsic and two extrinsic bias measures, renders ADELE, very effective in bias mitigation. We further show that – due to its modular nature – ADELE, coupled with task adapters, retains fairness even after large-scale downstream training. Finally, by means of multilingual BERT, we successfully transfer ADELE, to six target languages.

Large language models have shown promising results in zero-shot settings. For example, they can perform multiple choice tasks simply by conditioning on a question and selecting the answer with the highest probability. However, ranking by string probability can be problematic due to surface form competition—wherein different surface forms compete for probability mass, even if they represent the same underlying concept in a given context, e.g. “computer” and “PC.” Since probability mass is finite, this lowers the probability of the correct answer, due to competition from other strings that are valid answers (but not one of the multiple choice options). We introduce Domain Conditional Pointwise Mutual Information, an alternative scoring function that directly compensates for surface form competition by simply reweighing each option according to its a priori likelihood within the context of a specific task. It achieves consistent gains in zero-shot performance over both calibrated and uncalibrated scoring functions on all GPT-2 and GPT-3 models on a variety of multiple choice datasets.

While contrastive learning is proven to be an effective training strategy in computer vision, Natural Language Processing (NLP) is only recently adopting it as a self-supervised alternative to Masked Language Modeling (MLM) for improving sequence representations. This paper introduces SupCL-Seq, which extends the supervised contrastive learning from computer vision to the optimization of sequence representations in NLP. By altering the dropout mask probability in standard Transformer architectures (e.g. BERT-base), for every representation (anchor), we generate augmented altered views. A supervised contrastive loss is then utilized to maximize the system’s capability of pulling together similar samples (e.g., anchors and their altered views) and pushing apart the samples belonging to the other classes. Despite its simplicity, SupCL-Seq leads to large gains in many sequence classification tasks on the GLUE benchmark compared to a standard BERT-base, including 6% absolute improvement on CoLA, 5.4% on MRPC, 4.7% on RTE and 2.6% on STS-B. We also show consistent gains over self-supervised contrastively learned representations, especially in non-semantic tasks. Finally we show that these gains are not solely due to augmentation, but rather to a downstream optimized sequence representation.