Papers

On-device directional hearing requires audio source separation from a given direction while achieving stringent human-imperceptible latency requirements. While neural nets can achieve significantly better performance than traditional beamformers, all existing models fall short of supporting low-latency causal inference on computationally-constrained wearables. We present DeepBeam, a hybrid model that combines traditional beamformers with a custom lightweight neural net. The former reduces the computational burden of the latter and also improves its generalizability, while the latter is designed to further reduce the memory and computational overhead to enable real-time and low-latency operations. Our evaluation shows comparable performance to state-of-the-art causal inference models on synthetic data while achieving a 5x reduction of model size, 4x reduction of computation per second, 5x reduction in processing time and generalizing better to real hardware data. Further, our real-time hybrid model runs in 8 ms on mobile CPUs designed for low-power wearable devices and achieves an end-to-end latency of 17.5 ms.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

In this paper, we propose a novel method to extend sequence-to-sequence models to accurately process sequences much longer than the ones used during training while being sample- and resource-efficient, supported by thorough experimentation. To investigate the effectiveness of our method, we apply it to the task of correcting documents already processed with Optical Character Recognition (OCR) systems using sequence-to-sequence models based on characters. We test our method on nine languages of the ICDAR 2019 competition on post-OCR text correction and achieve a new state-of-the-art performance in five of them. The strategy with the best performance involves splitting the input document in character n-grams and combining their individual corrections into the final output using a voting scheme that is equivalent to an ensemble of a large number of sequence models. We further investigate how to weigh the contributions from each one of the members of this ensemble. Our code for post-OCR correction is shared at https://github.com/jarobyte91/post_ocr_correction.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Text content created by humans or language models is often stolen or misused by adversaries. Tracing text provenance can help claim the ownership of text content or identify the malicious users who distribute misleading content like machine-generated fake news. There have been some attempts to achieve this, mainly based on watermarking techniques. Specifically, traditional text watermarking methods embed watermarks by slightly altering text format like line spacing and font, which, however, are fragile to cross-media transmissions like OCR. Considering this, natural language watermarking methods represent watermarks by replacing words in original sentences with synonyms from handcrafted lexical resources (e.g., WordNet), but they do not consider the substitution’s impact on the overall sentence's meaning. Recently, a transformer-based network was proposed to embed watermarks by modifying the unobtrusive words (e.g., function words), which also impair the sentence's logical and semantic coherence. Besides, one well-trained network fails on other different types of text content. To address the limitations mentioned above, we propose a natural language watermarking scheme based on context-aware lexical substitution (LS). Specifically, we employ BERT to suggest LS candidates by inferring the semantic relatedness between the candidates and the original sentence. Based on this, a selection strategy in terms of synchronicity and substitutability is further designed to test whether a word is exactly suitable for carrying the watermark signal. Extensive experiments demonstrate that, under both objective and subjective metrics, our watermarking scheme can well preserve the semantic integrity of original sentences and has a better transferability than existing methods. Besides, the proposed LS approach outperforms the state-of-the-art approach on the Stanford Word Substitution Benchmark.

poster session 5 @ red 5, poster session 12 @ red 5, oral session 12 @ red 5, poster session 5, poster session 12, oral session 12

Cross-lingual pre-training has achieved great successes using monolingual and bilingual plain text corpora. However, most pre-trained models neglect multilingual knowledge, which is language agnostic but comprises abundant cross-lingual structure alignment. In this paper, we propose XLM-K, a cross-lingual language model incorporating multilingual knowledge in pre-training. XLM-K augments existing multilingual pre-training with two knowledge tasks, namely Masked Entity Prediction Task and Object Entailment Task. We evaluate XLM-K on MLQA, NER and XNLI. Experimental results clearly demonstrate significant improvements over existing multilingual language models. The results on MLQA and NER exhibit the superiority of XLM-K in knowledge related tasks. The success in XNLI shows a better cross-lingual transferability obtained in XLM-K. What is more, we provide a detailed probing analysis to confirm the desired knowledge captured in our pre-training regimen.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Neuroimaging studies have identified multiple brain regions that are associated with semantic and syntactic processing when comprehending language. However, existing methods cannot explore the neural correlates of fine-grained word syntactic features, such as part-of-speech and dependency relations. This paper proposes an alternative framework to study how different word syntactic features are represented in the brain. To separate each syntactic feature, we propose a feature elimination method, called Mean Vector Null space Projection (MVNP). This method can remove a specific feature from word representations, resulting in one-feature-removed representations. Then we respectively associate one-feature-removed and the original word vectors with brain imaging data to explore how the brain represents the removed feature. This paper for the first time studies the cortical representations of multiple fine-grained syntactic features simultaneously and suggests some possible contributions of several brain regions to the complex division of syntactic processing. These findings indicate that the brain foundations of syntactic information processing might be broader than those suggested by classical studies.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Financial forecasting has been an important and active area of machine learning research because of the challenges it presents and the potential rewards that even minor improvements in prediction accuracy or forecasting may entail. Traditionally, financial forecasting has heavily relied on quantitative indicators and metrics derived from structured financial statements. Earnings conference call data, including text and audio, is an important source of unstructured data that has been used for various prediction tasks using deep earning and related approaches. However, current deep learning-based methods are limited in the way that they deal with numeric data; numbers are typically treated as plain-text tokens without taking advantage of their underlying numeric structure. This paper describes a numeric-oriented hierarchical transformer model (NumHTML) to predict stock returns, and financial risk using multi-modal aligned earnings calls data by taking advantage of the different categories of numbers (monetary, temporal, percentages etc.) and their magnitude. We present the results of a comprehensive evaluation of NumHTML against several state-of-the-art baselines using a real-world publicly available dataset. The results indicate that NumHTML significantly outperforms the current state-of-the-art across a variety of evaluation metrics and that it has the potential to offer significant financial gains in a practical trading context.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Syntax has been shown to benefit Coreference Resolution from incorporating long-range dependencies and structured information captured by syntax trees, either in traditional statistical machine learning based systems or recently proposed neural models. However, most leading systems use only dependency trees. We argue that constituent trees also encode important information, such as explicit span-boundary signals captured by nested multi-word phrases, extra linguistic labels and hierarchical structures useful for detecting anaphora. In this work, we propose a simple yet effective graph-based method to incorporate constituent syntactic structures. Moreover, we also explore to utilise higher-order neighbourhood information to encode rich structures in constituent trees. A novel message propagation mechanism is therefore proposed to enable information flow among elements in syntax trees. Experiments on the English and Chinese portions of OntoNotes 5.0 benchmark show that our proposed model either beats a strong baseline or achieves new state-of-the-art performance.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Citing and describing related literature are crucial to scientific writing. Many existing approaches show encouraging performance in citation recommendation, but are unable to accomplish the more challenging and onerous task of citation text generation. In this paper, we propose a novel disentangled representation based model DisenCite to automatically generate the citation text through integrating paper text and citation graph. A key novelty of our method compared with existing approaches is to generate context-specific citation text, empowering the generation of different types of citations for the same paper. In particular, we first build and make available a graph enhanced contextual citation dataset (GCite) with 25K edges in different types characterized by citation contained sections over 4.8K research papers. Based on this dataset, we encode each paper according to both textual contexts and structure information in the heterogeneous citation graph. The resulted paper representations are then disentangled by the mutual information regularization between this paper and its neighbors in graph. Extensive experiments demonstrate the superior performance of our method comparing to state-of-the-art approaches. We further conduct ablation and case studies to reassure that the improvement of our method comes from generating the context-specific citation through incorporating the citation graph.

poster session 5 @ red 5, poster session 12 @ red 5, oral session 5 @ red 5, poster session 5, poster session 12, oral session 5

Neural language models show vulnerability to adversarial examples which are semantically similar to their original counterparts with a few words replaced by their synonyms. A common way to improve model robustness is adversarial training which follows two steps—collecting adversarial examples by attacking a target model, and fine-tuning the model on the augmented dataset with these adversarial examples. The objective of traditional adversarial training is to make a model produce the same correct predictions on an original/adversarial example pair. However, the consistency between model decision-makings on two similar texts is ignored. We argue that a robust model should behave consistently on original/adversarial example pairs, that is making the same predictions (what) based on the same reasons (how) which can be reflected by consistent interpretations. In this work, we propose a novel feature-level adversarial training method named FLAT. FLAT aims at improving model robustness in terms of both predictions and interpretations. FLAT incorporates variational word masks in neural networks to learn global word importance and play as a bottleneck teaching the model to make predictions based on important words. FLAT explicitly shoots at the vulnerability problem caused by the mismatch between model understandings on the replaced words and their synonyms in original/adversarial example pairs by regularizing the corresponding global word importance scores. Experiments show the effectiveness of FLAT in improving the robustness with respect to both predictions and interpretations of four neural network models (LSTM, CNN, BERT, and DeBERTa) to two adversarial attacks on four text classification tasks. The models trained via FLAT also show better robustness than baseline models on unforeseen adversarial examples across different attacks.

poster session 5 @ red 5, poster session 12 @ red 5, oral session 5 @ red 5, poster session 5, poster session 12, oral session 5

Nowadays, due to the breakthrough in natural language generation (NLG), including machine translation, document summarization, image captioning, etc NLG models have been encapsulated in cloud APIs to serve over half a billion people worldwide and process over one hundred billion word generations per day. Thus, NLG APIs have already become essential profitable services in many commercial companies. Due to the substantial financial and intellectual investments, service providers adopt a pay-as-you-use policy to promote sustainable market growth. However, recent works have shown that cloud platforms suffer from financial losses imposed by model extraction attacks, which aim to imitate the functionality and utility of the victim services, thus violating the intellectual property (IP) of cloud APIs. This work targets at protecting IP of NLG APIs by identifying the attackers who have utilized watermarked responses from the victim NLG APIs. However, most existing watermarking techniques are not directly amenable for IP protection of NLG APIs. To bridge this gap, we first present a novel watermarking method for text generation APIs by conducting lexical modification to the original outputs. Compared with the competitive baselines, our watermark approach achieves better identifiable performance in terms of p-value, with fewer semantic losses. In addition, our watermarks are more understandable and intuitive to humans than the baselines. Finally, the empirical studies show our approach is also applicable to queries from different domains, and is effective on the attacker trained on a mixture of the corpus which includes less than 10% watermarked samples.

poster session 5 @ red 5, poster session 12 @ red 5, oral session 12 @ red 5, poster session 5, poster session 12, oral session 12

Lack of external knowledge makes empathetic dialogue systems difficult to perceive implicit emotions and learn emotional interactions from limited dialogue history. To address the above problems, we propose to leverage external knowledge, including commonsense knowledge and emotional lexical knowledge, to explicitly understand and express emotions in empathetic dialogue generation. We first enrich the dialogue history by jointly interacting with external knowledge and construct an emotional context graph. Then we learn emotional context representations from the knowledge-enriched emotional context graph and distill emotional signals, which are the prerequisites to predicate emotions expressed in responses. Finally, to generate the empathetic response, we propose an emotional cross-attention mechanism to learn the emotional dependencies from the emotional context graph. Extensive experiments conducted on a benchmark dataset verify the effectiveness of the proposed method. In addition, we find the performance of our method can be further improved by integrating with a pre-trained model that works orthogonally.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Metaphor is not only a linguistic phenomenon but also reflects the concept projection between source and target domains in human cognition. Previous sequence tagging-based metaphor identification methods could not model the concept projection, resulting in a limitation that the outputs of these models are unexplainable in the predictions of the metaphoricity labels. In this work, we propose the first explainable metaphor identification model, inspired by Conceptual Metaphor Theory. The model is based on statistic learning, a lexical resource, and a novel reward mechanism. Our model can identify the metaphoricity on the word-pair level, and explain the predicted metaphoricity labels via learned concept mappings. The use of the reward mechanism allows the model to learn the optimal concept mappings without knowing their true labels. Our method is also applicable for the concepts that are out of training domains by using the lexical resource. The automatically generated concept mappings demonstrate the implicit human thoughts in metaphoric expressions. Our experiments show the effectiveness of the proposed model in metaphor identification, and concept mapping tasks, respectively.

poster session 5 @ red 5, poster session 12 @ red 5, oral session 5 @ red 5, poster session 5, poster session 12, oral session 5

Implicit discourse relation recognition (IDRR) is a challenging but crucial task in discourse analysis. Most existing methods train multiple models to predict multi-level labels independently, while ignoring the dependence between hierarchically structured labels. In this paper, we consider multi-level IDRR as a conditional label sequence generation task and propose a Label Dependence-aware Sequence Generation Model (LDSGM) for it. Specifically, we first design a label attentive encoder to learn the global representation of an input instance and its level-specific contexts, where the label dependence is integrated to obtain better label embeddings. Then, we employ a label sequence decoder to output the predicted labels in a top-down manner, where the predicted higher-level labels are directly used to guide the label prediction at the current level. We further develop a mutual learning enhanced training method to exploit the label dependence in a bottom-up direction, which is captured by an auxiliary decoder introduced during training. Experimental results on the PDTB dataset show that our model achieves the state-of-the-art performance on multi-level IDRR. We release our code at https://github.com/nlpersECJTU/LDSGM.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

We study the problem of online learning with human feedback in the human-in-the-loop machine translation, in which the human translators revise the machine-generated translations and then the corrected translations are used to improve the neural machine translation (NMT) system. However, previous methods require online model updating or additional translation memory networks to achieve high-quality performance, making them inflexible and inefficient in practice. In this paper, we propose a novel non-parametric online learning method without changing the model structure. This approach introduces two k-nearest-neighbor (knn) modules: one module memorizes the human feedback, which is the correct sentences provided by human translators, while the other balances the usage of the history human feedback and original NMT models adaptively. Experiments conducted on EMEA and JRC-Acquis benchmarks demonstrate that our proposed method obtains substantial improvements on translation accuracy and achieves better adaptation performance with less repeating human correction operations.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Graph convolutional networks have been successfully applied to the task of event detection. However, existing works rely heavily on a fixed syntactic parse tree structure from an external parser. In addition, the information content extracted for aggregation is determined simply by the (syntactic) edge direction or type but irrespective of what semantics the vertices have, which is somewhat rigid. With this work, we propose a novel graph convolutional method that combines an adaptive graph generation technique and a multi-channel selection strategy. The adaptive graph generation technique enables the gradients to pass through the graph sampling layer by using the ST-Gumbel-Softmax trick. The multi-channel selection strategy allows two adjacent vertices to automatically determine which information channels to get through for information extraction and aggregation. The proposed method achieves the state-of-the-art performance on ACE2005 dataset.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

The task of Grammatical Error Correction (GEC) has received remarkable attention with wide applications in Natural Language Processing (NLP) in recent years. While one of the key principles of GEC is to keep the correct parts unchanged and avoid over-correction, previous sequence-to-sequence (seq2seq) models generate results from scratch, which are not guaranteed to follow the original sentence structure and may suffer from the over-correction problem. In the meantime, the recently proposed sequence tagging models can overcome the over-correction problem by only generating edit operations, but are conditioned on human designed language-specific tagging labels. In this paper, we combine the pros and alleviate the cons of both models by proposing a novel Sequence-to-Action (S2A) module. The S2A module jointly takes the source and target sentences as input, and is able to automatically generate a token-level action sequence before predicting each token, where each action is generated from three choices named \texttt{SKIP}, \texttt{COPY} and \texttt{GEN}erate. Then the actions are fused with the basic seq2seq framework to provide final predictions. We conduct experiments on the benchmark datasets of both English and Chinese GEC tasks. Our model consistently outperforms the seq2seq baselines, while being able to significantly alleviate the over-correction problem as well as holding better generality and diversity in the generation results compared to the sequence tagging models.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Slides are broadly used to present the research works and there are several studies on the problem of automatic slides generation. However, the lack of dataset hinders further research. In this paper, we construct a benchmark dataset for the problem of slides generation from scholarly papers. The dataset is fine-grained and consists of aligned pairs of single slide and specific region of a paper. Then we deploy several baseline models and conduct preliminary experiments. The results show that this task is challenging and awaits more exploration. The dataset and code will be released.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Recent work has shown that current text classification models are vulnerable to small adversarial perturbation to inputs, and adversarial training that re-trains the models with the support of adversarial examples is the most popular way to alleviate the impact of the perturbation. However, current adversarial training methods have two principal problems: worse model generalization and ineffective defending against other text attacks. In this paper, we propose a \textbf{K}eyword-bias-aware \textbf{A}dversarial \textbf{T}ext \textbf{G}eneration model (KATG) that implicitly generates adversarial sentences using a generator-discriminator structure. Instead of using a benign sentence to generate an adversarial sentence, the KATG model utilizes extra multiple benign sentences (namely prior sentences) to guide adversarial sentence generation. Furthermore, to cover more perturbation used in existing attacks, a keyword-bias-aware sampling is proposed to select sentences containing biased words as prior sentences. Besides, to effectively utilize prior sentences, a generative flow mechanism is proposed to construct latent semantic space and learn a latent representation for the prior sentences. Experiments demonstrate that adversarial sentences generated by our KATG model can strengthen the victim model's robustness and generalization.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Common image-text joint understanding techniques presume that image and associated text can universally be characterized by a single implicit model. However, co-occurring images and text can be related in qualitatively different ways, and explicitly modeling this could improve the performance of current joint understanding models. In this paper, we train a \cmcm for text-to-image retrieval task. Our analysis shows that models trained with image--text coherence relations can retrieve images originally paired with target text more often than coherence-agnostic models. We also show via human evaluation that the images retrieved by the proposed coherence-aware model are preferred over a coherence-agnostic baseline by a huge margin. Our findings provide insights into the ways that different modalities communicate and the role of coherence relations in capturing commonsense inferences in text and imagery.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12

Recent methods in speech and language technology pre-train very large models which are fine-tuned for specific tasks. However, the benefits of such large models are often limited to a few resource rich languages of the world. In this work, we make multiple contributions towards building ASR systems for low resource languages from the Indian subcontinent. First, we curate 16,000 hours of raw speech data for 40 Indian languages from a wide variety of domains including education, news, technology, and finance. Second, using this raw speech data we pre-train several variants of wav2vec style models for 40 Indian languages. Third, we analyze the pretrained models to find key features: code-book vectors of similar sounding phonemes are shared across languages, representations across layers are discriminative of the language family, and attention heads often pay attention within small local windows. Fourth, we fine-tune this model for downstream ASR for 9 languages and obtain state-of-the-art results on 3 public datasets, including on very low-resource languages such as Sinhala and Nepali. Our work establishes that multilingual pretraining is an effective strategy for building ASR systems for the linguistically diverse speakers of the Indian subcontinent.

poster session 5 @ red 5, poster session 12 @ red 5, oral session 12 @ red 5, poster session 5, poster session 12, oral session 12

State-of-the-art deep NLP models have achieved impressive improvements on many tasks. However, they are found to be vulnerable to some perturbations. Before they are widely adopted, the fundamental issues of robustness need to be addressed. In this paper, we design a robustness enhancement method to defend against word substitution perturbation, whose basic idea is to fight perturbation with perturbation. We find that: although many well-trained deep models are not robust in the setting of the presence of adversarial samples, they satisfy weak robustness. That means they can handle most non-crafted perturbations well. Taking advantage of the weak robustness property of deep models, we utilize non-crafted perturbations to resist the adversarial perturbations crafted by attackers. Our method contains two main stages. The first stage is using randomized perturbation to conform the input to the data distribution. The second stage is using randomized perturbation to eliminate the instability of prediction results and enhance the robustness guarantee. Experimental results show that our method can significantly improve the ability of deep models to resist the state-of-the-art adversarial attacks while maintaining the prediction performance on the original clean data.

poster session 5 @ red 5, poster session 12 @ red 5, poster session 5, poster session 12