NAACL 2019 论文列表

Current research on spoken language translation (SLT) has to confront with the scarcity of sizeable and publicly available training corpora. This problem hinders the adoption of neural end-to-end approaches, which represent the state of the art in the two parent tasks of SLT: automatic speech recognition and machine translation. To fill this gap, we created MuST-C, a multilingual speech translation corpus whose size and quality will facilitate the training of end-to-end systems for SLT from English into 8 languages. For each target language, MuST-C comprises at least 385 hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual transcriptions and translations. Together with a description of the corpus creation methodology (scalable to add new data and cover new languages), we provide an empirical verification of its quality and SLT results computed with a state-of-the-art approach on each language direction.

Language identification for code-switching (CS), the phenomenon of alternating between two or more languages in conversations, has traditionally been approached under the assumption of a single language per token. However, if at least one language is morphologically rich, a large number of words can be composed of morphemes from more than one language (intra-word CS). In this paper, we extend the language identification task to the subword-level, such that it includes splitting mixed words while tagging each part with a language ID. We further propose a model for this task, which is based on a segmental recurrent neural network. In experiments on a new Spanish–Wixarika dataset and on an adapted German–Turkish dataset, our proposed model performs slightly better than or roughly on par with our best baseline, respectively. Considering only mixed words, however, it strongly outperforms all baselines.

Recent work has shown that visual context improves cross-lingual sense disambiguation for nouns. We extend this line of work to the more challenging task of cross-lingual verb sense disambiguation, introducing the MultiSense dataset of 9,504 images annotated with English, German, and Spanish verbs. Each image in MultiSense is annotated with an English verb and its translation in German or Spanish. We show that cross-lingual verb sense disambiguation models benefit from visual context, compared to unimodal baselines. We also show that the verb sense predicted by our best disambiguation model can improve the results of a text-only machine translation system when used for a multimodal translation task.

Machine learning has shown promise for automatic detection of Alzheimer’s disease (AD) through speech; however, efforts are hampered by a scarcity of data, especially in languages other than English. We propose a method to learn a correspondence between independently engineered lexicosyntactic features in two languages, using a large parallel corpus of out-of-domain movie dialogue data. We apply it to dementia detection in Mandarin Chinese, and demonstrate that our method outperforms both unilingual and machine translation-based baselines. This appears to be the first study that transfers feature domains in detecting cognitive decline.

The task of retrieving clips within videos based on a given natural language query requires cross-modal reasoning over multiple frames. Prior approaches such as sliding window classifiers are inefficient, while text-clip similarity driven ranking-based approaches such as segment proposal networks are far more complicated. In order to select the most relevant video clip corresponding to the given text description, we propose a novel extractive approach that predicts the start and end frames by leveraging cross-modal interactions between the text and video - this removes the need to retrieve and re-rank multiple proposal segments. Using recurrent networks we encode the two modalities into a joint representation which is then used in different variants of start-end frame predictor networks. Through extensive experimentation and ablative analysis, we demonstrate that our simple and elegant approach significantly outperforms state of the art on two datasets and has comparable performance on a third.

We demonstrate the surprising strength of unimodal baselines in multimodal domains, and make concrete recommendations for best practices in future research. Where existing work often compares against random or majority class baselines, we argue that unimodal approaches better capture and reflect dataset biases and therefore provide an important comparison when assessing the performance of multimodal techniques. We present unimodal ablations on three recent datasets in visual navigation and QA, seeing an up to 29% absolute gain in performance over published baselines.

We present a novel method for mapping unrestricted text to knowledge graph entities by framing the task as a sequence-to-sequence problem. Specifically, given the encoded state of an input text, our decoder directly predicts paths in the knowledge graph, starting from the root and ending at the the target node following hypernym-hyponym relationships. In this way, and in contrast to other text-to-entity mapping systems, our model outputs hierarchically structured predictions that are fully interpretable in the context of the underlying ontology, in an end-to-end manner. We present a proof-of-concept experiment with encouraging results, comparable to those of state-of-the-art systems.

Over the last few years, there has been growing interest in learning models for physically grounded language understanding tasks, such as the popular blocks world domain. These works typically view this problem as a single-step process, in which a human operator gives an instruction and an automated agent is evaluated on its ability to execute it. In this paper we take the first step towards increasing the bandwidth of this interaction, and suggest a protocol for including advice, high-level observations about the task, which can help constrain the agent’s prediction. We evaluate our approach on the blocks world task, and show that even simple advice can help lead to significant performance improvements. To help reduce the effort involved in supplying the advice, we also explore model self-generated advice which can still improve results.

We propose a new visual grounding task called Visual Query Detection (VQD). In VQD, the task is to localize a variable number of objects in an image where the objects are specified in natural language. VQD is related to visual referring expression comprehension, where the task is to localize only one object. We propose the first algorithms for VQD, and we evaluate them on both visual referring expression datasets and our new VQDv1 dataset.

Training models to map natural language instructions to programs, given target world supervision only, requires searching for good programs at training time. Search is commonly done using beam search in the space of partial programs or program trees, but as the length of the instructions grows finding a good program becomes difficult. In this work, we propose a search algorithm that uses the target world state, known at training time, to train a critic network that predicts the expected reward of every search state. We then score search states on the beam by interpolating their expected reward with the likelihood of programs represented by the search state. Moreover, we search not in the space of programs but in a more compressed state of program executions, augmented with recent entities and actions. On the SCONE dataset, we show that our algorithm dramatically improves performance on all three domains compared to standard beam search and other baselines.

Current predominant neural machine translation (NMT) models often have a deep structure with large amounts of parameters, making these models hard to train and easily suffering from over-fitting. A common practice is to utilize a validation set to evaluate the training process and select the best checkpoint. Average and ensemble techniques on checkpoints can lead to further performance improvement. However, as these methods do not affect the training process, the system performance is restricted to the checkpoints generated in original training procedure. In contrast, we propose an online knowledge distillation method. Our method on-the-fly generates a teacher model from checkpoints, guiding the training process to obtain better performance. Experiments on several datasets and language pairs show steady improvement over a strong self-attention-based baseline system. We also provide analysis on data-limited setting against over-fitting. Furthermore, our method leads to an improvement in a machine reading experiment as well.

Neural Networks trained with gradient descent are known to be susceptible to catastrophic forgetting caused by parameter shift during the training process. In the context of Neural Machine Translation (NMT) this results in poor performance on heterogeneous datasets and on sub-tasks like rare phrase translation. On the other hand, non-parametric approaches are immune to forgetting, perfectly complementing the generalization ability of NMT. However, attempts to combine non-parametric or retrieval based approaches with NMT have only been successful on narrow domains, possibly due to over-reliance on sentence level retrieval. We propose a novel n-gram level retrieval approach that relies on local phrase level similarities, allowing us to retrieve neighbors that are useful for translation even when overall sentence similarity is low. We complement this with an expressive neural network, allowing our model to extract information from the noisy retrieved context. We evaluate our Semi-parametric NMT approach on a heterogeneous dataset composed of WMT, IWSLT, JRC-Acquis and OpenSubtitles, and demonstrate gains on all 4 evaluation sets. The Semi-parametric nature of our approach also opens the door for non-parametric domain adaptation, demonstrating strong inference-time adaptation performance on new domains without the need for any parameter updates.

Modern Machine Translation (MT) systems perform remarkably well on clean, in-domain text. However most of the human generated text, particularly in the realm of social media, is full of typos, slang, dialect, idiolect and other noise which can have a disastrous impact on the accuracy of MT. In this paper we propose methods to enhance the robustness of MT systems by emulating naturally occurring noise in otherwise clean data. Synthesizing noise in this manner we are ultimately able to make a vanilla MT system more resilient to naturally occurring noise, partially mitigating loss in accuracy resulting therefrom.

We introduce a curriculum learning approach to adapt generic neural machine translation models to a specific domain. Samples are grouped by their similarities to the domain of interest and each group is fed to the training algorithm with a particular schedule. This approach is simple to implement on top of any neural framework or architecture, and consistently outperforms both unadapted and adapted baselines in experiments with two distinct domains and two language pairs.

Recent research has discovered that a shared bilingual word embedding space can be induced by projecting monolingual word embedding spaces from two languages using a self-learning paradigm without any bilingual supervision. However, it has also been shown that for distant language pairs such fully unsupervised self-learning methods are unstable and often get stuck in poor local optima due to reduced isomorphism between starting monolingual spaces. In this work, we propose a new robust framework for learning unsupervised multilingual word embeddings that mitigates the instability issues. We learn a shared multilingual embedding space for a variable number of languages by incrementally adding new languages one by one to the current multilingual space. Through the gradual language addition the method can leverage the interdependencies between the new language and all other languages in the current multilingual space. We find that it is beneficial to project more distant languages later in the iterative process. Our fully unsupervised multilingual embedding spaces yield results that are on par with the state-of-the-art methods in the bilingual lexicon induction (BLI) task, and simultaneously obtain state-of-the-art scores on two downstream tasks: multilingual document classification and multilingual dependency parsing, outperforming even supervised baselines. This finding also accentuates the need to establish evaluation protocols for cross-lingual word embeddings beyond the omnipresent intrinsic BLI task in future work.

Beam search optimization (Wiseman and Rush, 2016) resolves many issues in neural machine translation. However, this method lacks principled stopping criteria and does not learn how to stop during training, and the model naturally prefers longer hypotheses during the testing time in practice since they use the raw score instead of the probability-based score. We propose a novel ranking method which enables an optimal beam search stop- ping criteria. We further introduce a structured prediction loss function which penalizes suboptimal finished candidates produced by beam search during training. Experiments of neural machine translation on both synthetic data and real languages (German→English and Chinese→English) demonstrate our pro- posed methods lead to better length and BLEU score.

We propose a new automatic evaluation metric for machine translation. Our proposed metric is obtained by adjusting the Earth Mover’s Distance (EMD) to the evaluation task. The EMD measure is used to obtain the distance between two probability distributions consisting of some signatures having a feature and a weight. We use word embeddings, sentence-level tf-idf, and cosine similarity between two word embeddings, respectively, as the features, weight, and the distance between two features. Results show that our proposed metric can evaluate machine translation based on word meaning. Moreover, for distance, cosine similarity and word position information are used to address word-order differences. We designate this metric as Word Embedding-Based automatic MT evaluation using Word Position Information (WE_WPI). A meta-evaluation using WMT16 metrics shared task set indicates that our WE_WPI achieves the highest correlation with human judgment among several representative metrics.

Stance detection in twitter aims at mining user stances expressed in a tweet towards a single or multiple target entities. To tackle this problem, most of the prior studies have been explored the traditional deep learning models, e.g., LSTM and GRU. However, in compared to these traditional approaches, recently proposed densely connected Bi-LSTM and nested LSTMs architectures effectively address the vanishing-gradient and overfitting problems as well as dealing with long-term dependencies. In this paper, we propose a neural ensemble model that adopts the strengths of these two LSTM variants to learn better long-term dependencies, where each module coupled with an attention mechanism that amplifies the contribution of important elements in the final representation. We also employ a multi-kernel convolution on top of them to extract the higher-level tweet representations. Results of extensive experiments on single and multi-target stance detection datasets show that our proposed method achieves substantial improvement over the current state-of-the-art deep learning based methods.

In relation extraction with distant supervision, noisy labels make it difficult to train quality models. Previous neural models addressed this problem using an attention mechanism that attends to sentences that are likely to express the relations. We improve such models by combining the distant supervision data with an additional directly-supervised data, which we use as supervision for the attention weights. We find that joint training on both types of supervision leads to a better model because it improves the model’s ability to identify noisy sentences. In addition, we find that sigmoidal attention weights with max pooling achieves better performance over the commonly used weighted average attention in this setup. Our proposed method achieves a new state-of-the-art result on the widely used FB-NYT dataset.