NAACL 2021 Papers

Grounding natural language instructions on the web to perform previously unseen tasks enables accessibility and automation. We introduce a task and dataset to train AI agents from open-domain, step-by-step instructions originally written for people. We build RUSS (Rapid Universal Support Service) to tackle this problem. RUSS consists of two models: First, a BERT-LSTM with pointers parses instructions to WebLang, a domain-specific language we design for grounding natural language on the web. Then, a grounding model retrieves the unique IDs of any webpage elements requested in the WebLang. RUSS may interact with the user through a dialogue (e.g. ask for an address) or execute a web operation (e.g. click a button) inside the web runtime. To augment training, we synthesize natural language instructions mapped to WebLang. Our dataset consists of 80 different customer service problems from help websites, with a total of 741 step-by-step instructions and their corresponding actions. RUSS achieves 76.7% end-to-end accuracy predicting agent actions from single instructions. It outperforms state-of-the-art models that directly map instructions to actions without WebLang. Our user study shows that RUSS is preferred by actual users over web navigation.

We introduce a grey-box adversarial attack and defence framework for sentiment classification. We address the issues of differentiability, label preservation and input reconstruction for adversarial attack and defence in one unified framework. Our results show that once trained, the attacking model is capable of generating high-quality adversarial examples substantially faster (one order of magnitude less in time) than state-of-the-art attacking methods. These examples also preserve the original sentiment according to human evaluation. Additionally, our framework produces an improved classifier that is robust in defending against multiple adversarial attacking methods. Code is available at: https://github.com/ibm-aur-nlp/adv-def-text-dist.

Nowadays, with many e-commerce platforms conducting global business, e-commerce search systems are required to handle product retrieval under multilingual scenarios. Moreover, comparing with maintaining per-country specific e-commerce search systems, having an universal system across countries can further reduce the operational and computational costs, and facilitate business expansion to new countries. In this paper, we introduce an universal end-to-end multilingual retrieval system, and discuss our learnings and technical details when training and deploying the system to serve billion-scale product retrieval for e-commerce search. In particular, we propose a multilingual graph attention based retrieval network by leveraging recent advances in transformer-based multilingual language models and graph neural network architectures to capture the interactions between search queries and items in e-commerce search. Offline experiments on five countries data show that our algorithm outperforms the state-of-the-art baselines by 35% recall and 25% mAP on average. Moreover, the proposed model shows significant increase of conversion/revenue in online A/B experiments and has been deployed in production for multiple countries.

Recent work on aspect-level sentiment classification has demonstrated the efficacy of incorporating syntactic structures such as dependency trees with graph neural networks (GNN), but these approaches are usually vulnerable to parsing errors. To better leverage syntactic information in the face of unavoidable errors, we propose a simple yet effective graph ensemble technique, GraphMerge, to make use of the predictions from different parsers. Instead of assigning one set of model parameters to each dependency tree, we first combine the dependency relations from different parses before applying GNNs over the resulting graph. This allows GNN models to be robust to parse errors at no additional computational cost, and helps avoid overparameterization and overfitting from GNN layer stacking by introducing more connectivity into the ensemble graph. Our experiments on the SemEval 2014 Task 4 and ACL 14 Twitter datasets show that our GraphMerge model not only outperforms models with single dependency tree, but also beats other ensemble models without adding model parameters.

We study the problem of Event Causality Identification (ECI) to detect causal relation between event mention pairs in text. Although deep learning models have recently shown state-of-the-art performance for ECI, they are limited to the intra-sentence setting where event mention pairs are presented in the same sentences. This work addresses this issue by developing a novel deep learning model for document-level ECI (DECI) to accept inter-sentence event mention pairs. As such, we propose a graph-based model that constructs interaction graphs to capture relevant connections between important objects for DECI in input documents. Such interaction graphs are then consumed by graph convolutional networks to learn document context-augmented representations for causality prediction between events. Various information sources are introduced to enrich the interaction graphs for DECI, featuring discourse, syntax, and semantic information. Our extensive experiments show that the proposed model achieves state-of-the-art performance on two benchmark datasets.

Named entity linking (NEL) or mapping “strings” to “things” in a knowledge base is a fundamental preprocessing step in systems that require knowledge of entities such as information extraction and question answering. In this work, we lay out and investigate two challenges faced by individuals or organizations building NEL systems. Can they directly use an off-the-shelf system? If not, how easily can such a system be repurposed for their use case? First, we conduct a study of off-the-shelf commercial and academic NEL systems. We find that most systems struggle to link rare entities, with commercial solutions lagging their academic counterparts by 10%+. Second, for a use case where the NEL model is used in a sports question-answering (QA) system, we investigate how to close the loop in our analysis by repurposing the best off-the-shelf model (Bootleg) to correct sport-related errors. We show how tailoring a simple technique for patching models using weak labeling can provide a 25% absolute improvement in accuracy of sport-related errors.

Typical fact verification models use retrieved written evidence to verify claims. Evidence sources, however, often change over time as more information is gathered and revised. In order to adapt, models must be sensitive to subtle differences in supporting evidence. We present VitaminC, a benchmark infused with challenging cases that require fact verification models to discern and adjust to slight factual changes. We collect over 100,000 Wikipedia revisions that modify an underlying fact, and leverage these revisions, together with additional synthetically constructed ones, to create a total of over 400,000 claim-evidence pairs. Unlike previous resources, the examples in VitaminC are contrastive, i.e., they contain evidence pairs that are nearly identical in language and content, with the exception that one supports a given claim while the other does not. We show that training using this design increases robustness—improving accuracy by 10% on adversarial fact verification and 6% on adversarial natural language inference (NLI). Moreover, the structure of VitaminC leads us to define additional tasks for fact-checking resources: tagging relevant words in the evidence for verifying the claim, identifying factual revisions, and providing automatic edits via factually consistent text generation.

There are common semantics shared across text and images. Given a sentence in a source language, whether depicting the visual scene helps translation into a target language? Existing multimodal neural machine translation methods (MNMT) require triplets of bilingual sentence - image for training and tuples of source sentence - image for inference. In this paper, we propose ImagiT, a novel machine translation method via visual imagination. ImagiT first learns to generate visual representation from the source sentence, and then utilizes both source sentence and the “imagined representation” to produce a target translation. Unlike previous methods, it only needs the source sentence at the inference time. Experiments demonstrate that ImagiT benefits from visual imagination and significantly outperforms the text-only neural machine translation baselines. Further analysis reveals that the imagination process in ImagiT helps fill in missing information when performing the degradation strategy.

Evaluating the quality of responses generated by open-domain conversation systems is a challenging task. This is partly because there can be multiple appropriate responses to a given dialogue history. Reference-based metrics that rely on comparisons to a set of known correct responses often fail to account for this variety, and consequently correlate poorly with human judgment. To address this problem, researchers have investigated the possibility of assessing response quality without using a set of known correct responses. RUBER demonstrated that an automatic response evaluation model could be made using unsupervised learning for the next-utterance prediction (NUP) task. For the unsupervised learning of such model, we propose a method of manipulating a golden response to create a new negative response that is designed to be inappropriate within the context while maintaining high similarity with the original golden response. We find, from our experiments on English datasets, that using the negative samples generated by our method alongside random negative samples can increase the model’s correlation with human evaluations. The process of generating such negative samples is automated and does not rely on human annotation.

Given the diversity of the candidates and complexity of job requirements, and since interviewing is an inherently subjective process, it is an important task to ensure consistent, uniform, efficient and objective interviews that result in high quality recruitment. We propose an interview assistant system to automatically, and in an objective manner, select an optimal set of technical questions (from question banks) personalized for a candidate. This set can help a human interviewer to plan for an upcoming interview of that candidate. We formalize the problem of selecting a set of questions as an integer linear programming problem and use standard solvers to get a solution. We use knowledge graph as background knowledge in this formulation, and derive our objective functions and constraints from it. We use candidate’s resume to personalize the selection of questions. We propose an intrinsic evaluation to compare a set of suggested questions with actually asked questions. We also use expert interviewers to comparatively evaluate our approach with a set of reasonable baselines.

Understanding and executing natural language instructions in a grounded domain is one of the hallmarks of artificial intelligence. In this paper, we focus on instruction understanding in the blocks world domain and investigate the language understanding abilities of two top-performing systems for the task. We aim to understand if the test performance of these models indicates an understanding of the spatial domain and of the natural language instructions relative to it, or whether they merely over-fit spurious signals in the dataset. We formulate a set of expectations one might have from an instruction following model and concretely characterize the different dimensions of robustness such a model should possess. Despite decent test performance, we find that state-of-the-art models fall short of these expectations and are extremely brittle. We then propose a learning strategy that involves data augmentation and show through extensive experiments that the proposed learning strategy yields models that are competitive on the original test set while satisfying our expectations much better.

The input vocabulary and the representations learned are crucial to the performance of neural NLP models. Using the full vocabulary results in less explainable and more memory intensive models, with the embedding layer often constituting the majority of model parameters. It is thus common to use a smaller vocabulary to lower memory requirements and construct more interpertable models. We propose a vocabulary selection method that views words as members of a team trying to maximize the model’s performance. We apply power indices from cooperative game theory, including the Shapley value and Banzhaf index, that measure the relative importance of individual team members in accomplishing a joint task. We approximately compute these indices to identify the most influential words. Our empirical evaluation examines multiple NLP tasks, including sentence and document classification, question answering and textual entailment. We compare to baselines that select words based on frequency, TF-IDF and regression coefficients under L1 regularization, and show that this game-theoretic vocabulary selection outperforms all baseline on a range of different tasks and datasets.

Neural abstractive summarization models are flexible and can produce coherent summaries, but they are sometimes unfaithful and can be difficult to control. While previous studies attempt to provide different types of guidance to control the output and increase faithfulness, it is not clear how these strategies compare and contrast to each other. In this paper, we propose a general and extensible guided summarization framework (GSum) that can effectively take different kinds of external guidance as input, and we perform experiments across several different varieties. Experiments demonstrate that this model is effective, achieving state-of-the-art performance according to ROUGE on 4 popular summarization datasets when using highlighted sentences as guidance. In addition, we show that our guided model can generate more faithful summaries and demonstrate how different types of guidance generate qualitatively different summaries, lending a degree of controllability to the learned models.

Human innovation in language, such as inventing new words, is a challenge for pretrained language models. We assess the ability of one large model, GPT-3, to process new words and decide on their meaning. We create a set of nonce words and prompt GPT-3 to generate their dictionary definitions. We find GPT-3 produces plausible definitions that align with human judgments. Moreover, GPT-3’s definitions are sometimes preferred to those invented by humans, signaling its intriguing ability not just to adapt, but to add to the evolving vocabulary of the English language.

Named Entity Recognition (NER) remains difficult in real-world settings; current challenges include short texts (low context), emerging entities, and complex entities (e.g. movie names). Gazetteer features can help, but results have been mixed due to challenges with adding extra features, and a lack of realistic evaluation data. It has been shown that including gazetteer features can cause models to overuse or underuse them, leading to poor generalization. We propose GEMNET, a novel approach for gazetteer knowledge integration, including (1) a flexible Contextual Gazetteer Representation (CGR) encoder that can be fused with any word-level model; and (2) a Mixture-of- Experts gating network that overcomes the feature overuse issue by learning to conditionally combine the context and gazetteer features, instead of assigning them fixed weights. To comprehensively evaluate our approaches, we create 3 large NER datasets (24M tokens) reflecting current challenges. In an uncased setting, our methods show large gains (up to +49% F1) in recognizing difficult entities compared to existing baselines. On standard benchmarks, we achieve a new uncased SOTA on CoNLL03 and WNUT17.

With advances in neural language models, the focus of linguistic steganography has shifted from edit-based approaches to generation-based ones. While the latter’s payload capacity is impressive, generating genuine-looking texts remains challenging. In this paper, we revisit edit-based linguistic steganography, with the idea that a masked language model offers an off-the-shelf solution. The proposed method eliminates painstaking rule construction and has a high payload capacity for an edit-based model. It is also shown to be more secure against automatic detection than a generation-based method while offering better control of the security/payload capacity trade-off.

The lack of publicly available evaluation data for low-resource languages limits progress in Spoken Language Understanding (SLU). As key tasks like intent classification and slot filling require abundant training data, it is desirable to reuse existing data in high-resource languages to develop models for low-resource scenarios. We introduce xSID, a new benchmark for cross-lingual (x) Slot and Intent Detection in 13 languages from 6 language families, including a very low-resource dialect. To tackle the challenge, we propose a joint learning approach, with English SLU training data and non-English auxiliary tasks from raw text, syntax and translation for transfer. We study two setups which differ by type and language coverage of the pre-trained embeddings. Our results show that jointly learning the main tasks with masked language modeling is effective for slots, while machine translation transfer works best for intent classification.

Understanding how news media frame political issues is important due to its impact on public attitudes, yet hard to automate. Computational approaches have largely focused on classifying the frame of a full news article while framing signals are often subtle and local. Furthermore, automatic news analysis is a sensitive domain, and existing classifiers lack transparency in their predictions. This paper addresses both issues with a novel semi-supervised model, which jointly learns to embed local information about the events and related actors in a news article through an auto-encoding framework, and to leverage this signal for document-level frame classification. Our experiments show that: our model outperforms previous models of frame prediction; we can further improve performance with unlabeled training data leveraging the semi-supervised nature of our model; and the learnt event and actor embeddings intuitively corroborate the document-level predictions, providing a nuanced and interpretable article frame representation.

We release FoolMeTwice (FM2 for short), a large dataset of challenging entailment pairs collected through a fun multi-player game. Gamification encourages adversarial examples, drastically lowering the number of examples that can be solved using “shortcuts” compared to other popular entailment datasets. Players are presented with two tasks. The first task asks the player to write a plausible claim based on the evidence from a Wikipedia page. The second one shows two plausible claims written by other players, one of which is false, and the goal is to identify it before the time runs out. Players “pay” to see clues retrieved from the evidence pool: the more evidence the player needs, the harder the claim. Game-play between motivated players leads to diverse strategies for crafting claims, such as temporal inference and diverting to unrelated evidence, and results in higher quality data for the entailment and evidence retrieval tasks. We open source the dataset and the game code.