AACL 2020 Papers

Our analysis of large summarization datasets indicates that redundancy is a very serious problem when summarizing long documents. Yet, redundancy reduction has not been thoroughly investigated in neural summarization. In this work, we systematically explore and compare different ways to deal with redundancy when summarizing long documents. Specifically, we organize existing methods into categories based on when and how the redundancy is considered. Then, in the context of these categories, we propose three additional methods balancing non-redundancy and importance in a general and flexible way. In a series of experiments, we show that our proposed methods achieve the state-of-the-art with respect to ROUGE scores on two scientific paper datasets, Pubmed and arXiv, while reducing redundancy significantly.

Many top-performing image captioning models rely solely on object features computed with an object detection model to generate image descriptions. However, recent studies propose to directly use scene graphs to introduce information about object relations into captioning, hoping to better describe interactions between objects. In this work, we thoroughly investigate the use of scene graphs in image captioning. We empirically study whether using additional scene graph encoders can lead to better image descriptions and propose a conditional graph attention network (C-GAT), where the image captioning decoder state is used to condition the graph updates. Finally, we determine to what extent noise in the predicted scene graphs influence caption quality. Overall, we find no significant difference between models that use scene graph features and models that only use object detection features across different captioning metrics, which suggests that existing scene graph generation models are still too noisy to be useful in image captioning. Moreover, although the quality of predicted scene graphs is very low in general, when using high quality scene graphs we obtain gains of up to 3.3 CIDEr compared to a strong Bottom-Up Top-Down baseline.

Systematic Generalization refers to a learning algorithm’s ability to extrapolate learned behavior to unseen situations that are distinct but semantically similar to its training data. As shown in recent work, state-of-the-art deep learning models fail dramatically even on tasks for which they are designed when the test set is systematically different from the training data. We hypothesize that explicitly modeling the relations between objects in their contexts while learning their representations will help achieve systematic generalization. Therefore, we propose a novel method that learns objects’ contextualized embeddings with dynamic message passing conditioned on the input natural language and end-to-end trainable with other downstream deep learning modules. To our knowledge, this model is the first one that significantly outperforms the provided baseline and reaches state-of-the-art performance on grounded SCAN (gSCAN), a grounded natural language navigation dataset designed to require systematic generalization in its test splits.

Learning specific hands-on skills such as cooking, car maintenance, and home repairs increasingly happens via instructional videos. The user experience with such videos is known to be improved by meta-information such as time-stamped annotations for the main steps involved. Generating such annotations automatically is challenging, and we describe here two relevant contributions. First, we construct and release a new dense video captioning dataset, Video Timeline Tags (ViTT), featuring a variety of instructional videos together with time-stamped annotations. Second, we explore several multimodal sequence-to-sequence pretraining strategies that leverage large unsupervised datasets of videos and caption-like texts. We pretrain and subsequently finetune dense video captioning models using both YouCook2 and ViTT. We show that such models generalize well and are robust over a wide variety of instructional videos.

Fine-tuning (FT) pre-trained sentence embedding models on small datasets has been shown to have limitations. In this paper we show that concatenating the embeddings from the pre-trained model with those from a simple sentence embedding model trained only on the target data, can improve over the performance of FT for few-sample tasks. To this end, a linear classifier is trained on the combined embeddings, either by freezing the embedding model weights or training the classifier and embedding models end-to-end. We perform evaluation on seven small datasets from NLP tasks and show that our approach with end-to-end training outperforms FT with negligible computational overhead. Further, we also show that sophisticated combination techniques like CCA and KCCA do not work as well in practice as concatenation. We provide theoretical analysis to explain this empirical observation.

Recent years have seen important advances in the quality of state-of-the-art models, but this has come at the expense of models becoming less interpretable. This survey presents an overview of the current state of Explainable AI (XAI), considered within the domain of Natural Language Processing (NLP). We discuss the main categorization of explanations, as well as the various ways explanations can be arrived at and visualized. We detail the operations and explainability techniques currently available for generating explanations for NLP model predictions, to serve as a resource for model developers in the community. Finally, we point out the current gaps and encourage directions for future work in this important research area.

Pairwise data automatically constructed from weakly supervised signals has been widely used for training deep learning models. Pairwise datasets such as parallel texts can have uneven quality levels overall, but usually contain data subsets that are more useful as learning examples. We present two methods to refine data that are aimed to obtain that kind of subsets in a self-supervised way. Our methods are based on iteratively training dual-encoder models to compute similarity scores. We evaluate our methods on de-noising parallel texts and training neural machine translation models. We find that: (i) The self-supervised refinement achieves most machine translation gains in the first iteration, but following iterations further improve its intrinsic evaluation. (ii) Machine translations can improve the de-noising performance when combined with selection steps. (iii) Our methods are able to reach the performance of a supervised method. Being entirely self-supervised, our methods are well-suited to handle pairwise data without the need of prior knowledge or human annotations.

Word embedding methods have become the de-facto way to represent words, having been successfully applied to a wide array of natural language processing tasks. In this paper, we explore the hypothesis that embedding methods can also be effectively used to represent spatial locations. Using a new dataset consisting of the location trajectories of 729 students over a seven month period and text data related to those locations, we implement several strategies to create location embeddings, which we then use to create embeddings of the sequences of locations a student has visited. To identify the surface level properties captured in the representations, we propose a number of probing tasks such as the presence of a specific location in a sequence or the type of activities that take place at a location. We then leverage the representations we generated and employ them in more complex downstream tasks ranging from predicting a student’s area of study to a student’s depression level, showing the effectiveness of these location embeddings.

Transformer-based pre-trained language models (PLMs) have dramatically improved the state of the art in NLP across many tasks. This has led to substantial interest in analyzing the syntactic knowledge PLMs learn. Previous approaches to this question have been limited, mostly using test suites or probes. Here, we propose a novel fully unsupervised parsing approach that extracts constituency trees from PLM attention heads. We rank transformer attention heads based on their inherent properties, and create an ensemble of high-ranking heads to produce the final tree. Our method is adaptable to low-resource languages, as it does not rely on development sets, which can be expensive to annotate. Our experiments show that the proposed method often outperform existing approaches if there is no development set present. Our unsupervised parser can also be used as a tool to analyze the grammars PLMs learn implicitly. For this, we use the parse trees induced by our method to train a neural PCFG and compare it to a grammar derived from a human-annotated treebank.

Recent work in unsupervised parsing has tried to incorporate visual information into learning, but results suggest that these models need linguistic bias to compete against models that only rely on text. This work proposes grammar induction models which use visual information from images for labeled parsing, and achieve state-of-the-art results on grounded grammar induction on several languages. Results indicate that visual information is especially helpful in languages where high frequency words are more broadly distributed. Comparison between models with and without visual information shows that the grounded models are able to use visual information for proposing noun phrases, gathering useful information from images for unknown words, and achieving better performance at prepositional phrase attachment prediction.

Attention-based encoder-decoder models have achieved great success in neural machine translation tasks. However, the lengths of the target sequences are not explicitly predicted in these models. This work proposes length prediction as an auxiliary task and set up a sub-network to obtain the length information from the encoder. Experimental results show that the length prediction sub-network brings improvements over the strong baseline system and that the predicted length can be used as an alternative to length normalization during decoding.

Approaching named entities transliteration as a Neural Machine Translation (NMT) problem is common practice. While many have applied various NMT techniques to enhance machine transliteration models, few focus on the linguistic features particular to the relevant languages. In this paper, we investigate the effect of incorporating phonetic features for English-to-Chinese transliteration under the multi-task learning (MTL) setting—where we define a phonetic auxiliary task aimed to improve the generalization performance of the main transliteration task. In addition to our system, we also release a new English-to-Chinese dataset and propose a novel evaluation metric which considers multiple possible transliterations given a source name. Our results show that the multi-task model achieves similar performance as the previous state of the art with a model of a much smaller size.

Predicting the quality of machine translation has traditionally been addressed with language-specific models, under the assumption that the quality label distribution or linguistic features exhibit traits that are not shared across languages. An obvious disadvantage of this approach is the need for labelled data for each given language pair. We challenge this assumption by exploring different approaches to multilingual Quality Estimation (QE), including using scores from translation models. We show that these outperform single-language models, particularly in less balanced quality label distributions and low-resource settings. In the extreme case of zero-shot QE, we show that it is possible to accurately predict quality for any given new language from models trained on other languages. Our findings indicate that state-of-the-art neural QE models based on powerful pre-trained representations generalise well across languages, making them more applicable in real-world settings.

This paper presents a new dataset, B-SHARP, that can be used to develop NLP models for the detection of Mild Cognitive Impairment (MCI) known as an early sign of Alzheimer’s disease. Our dataset contains 1-2 min speech segments from 326 human subjects for 3 topics, (1) daily activity, (2) room environment, and (3) picture description, and their transcripts so that a total of 650 speech segments are collected. Given the B-SHARP dataset, several hierarchical text classification models are developed that jointly learn combinatory features across all 3 topics. The best performance of 74.1% is achieved by an ensemble model that adapts 3 types of transformer encoders. To the best of our knowledge, this is the first work that builds deep learning-based text classification models on multiple contents for the detection of MCI.

In this paper, we analyse the challenges of Chinese content scoring in comparison to English. As a review of prior work for Chinese content scoring shows a lack of open-access data in the field, we present two short-answer data sets for Chinese. The Chinese Educational Short Answers data set (CESA) contains 1800 student answers for five science-related questions. As a second data set, we collected ASAP-ZH with 942 answers by re-using three existing prompts from the ASAP data set. We adapt a state-of-the-art content scoring system for Chinese and evaluate it in several settings on these data sets. Results show that features on lower segmentation levels such as character n-grams tend to have better performance than features on token level.

This work studies the widely adopted ancestral sampling algorithms for auto-regressive language models. We use the quality-diversity (Q-D) trade-off to investigate three popular sampling methods (top-k, nucleus and tempered sampling). We focus on the task of open-ended language generation, and first show that the existing sampling algorithms have similar performance. By carefully inspecting the transformations defined by different sampling algorithms, we identify three key properties that are shared among them: entropy reduction, order preservation, and slope preservation. To validate the importance of the identified properties, we design two sets of new sampling methods: one set in which each algorithm satisfies all three properties, and one set in which each algorithm violates at least one of the properties. We compare their performance with existing algorithms, and find that violating the identified properties could lead to drastic performance degradation, as measured by the Q-D trade-off. On the other hand, we find that the set of sampling algorithms that satisfy these properties performs on par with the existing sampling algorithms.

Modern task-oriented dialog systems need to reliably understand users’ intents. Intent detection is even more challenging when moving to new domains or new languages, since there is little annotated data. To address this challenge, we present a suite of pretrained intent detection models which can predict a broad range of intended goals from many actions because they are trained on wikiHow, a comprehensive instructional website. Our models achieve state-of-the-art results on the Snips dataset, the Schema-Guided Dialogue dataset, and all 3 languages of the Facebook multilingual dialog datasets. Our models also demonstrate strong zero- and few-shot performance, reaching over 75% accuracy using only 100 training examples in all datasets.

Abstractive community detection is an important spoken language understanding task, whose goal is to group utterances in a conversation according to whether they can be jointly summarized by a common abstractive sentence. This paper provides a novel approach to this task. We first introduce a neural contextual utterance encoder featuring three types of self-attention mechanisms. We then train it using the siamese and triplet energy-based meta-architectures. Experiments on the AMI corpus show that our system outperforms multiple energy-based and non-energy based baselines from the state-of-the-art. Code and data are publicly available.

Unsupervised style transfer in text has previously been explored through the sentiment transfer task. The task entails inverting the overall sentiment polarity in a given input sentence, while preserving its content. From the Aspect-Based Sentiment Analysis (ABSA) task, we know that multiple sentiment polarities can often be present together in a sentence with multiple aspects. In this paper, the task of aspect-level sentiment controllable style transfer is introduced, where each of the aspect-level sentiments can individually be controlled at the output. To achieve this goal, a BERT-based encoder-decoder architecture with saliency weighted polarity injection is proposed, with unsupervised training strategies, such as ABSA masked-language-modelling. Through both automatic and manual evaluation, we show that the system is successful in controlling aspect-level sentiments.