In this paper, we describe our end-to-end multilingual speech translation system submitted to the IWSLT 2021 evaluation campaign on the Multilingual Speech Translation shared task. Our system is built by leveraging transfer learning across modalities
, tasks and languages. First, we leverage general-purpose multilingual modules pretrained with large amounts of unlabelled and labelled data. We further enable knowledge transfer from the text task to the speech task by training two tasks jointly. Finally, our multilingual model is finetuned on speech translation task-specific data to achieve the best translation results. Experimental results show our system outperforms the reported systems, including both end-to-end and cascaded based approaches, by a large margin. In some translation directions, our speech translation results evaluated on the public Multilingual TEDx test set are even comparable with the ones from a strong text-to-text translation system, which uses the oracle speech transcripts as input.
Multi-head attention has each of the attention heads collect salient information from different parts of an input sequence, making it a powerful mechanism for sequence modeling. Multilingual and multi-domain learning are common scenarios for sequence
modeling, where the key challenge is to maximize positive transfer and mitigate negative transfer across languages and domains. In this paper, we find that non-selective attention sharing is sub-optimal for achieving good generalization across all languages and domains. We further propose attention sharing strategies to facilitate parameter sharing and specialization in multilingual and multi-domain sequence modeling. Our approach automatically learns shared and specialized attention heads for different languages and domains to mitigate their interference. Evaluated in various tasks including speech recognition, text-to-text and speech-to-text translation, the proposed attention sharing strategies consistently bring gains to sequence models built upon multi-head attention. For speech-to-text translation, our approach yields an average of $+2.0$ BLEU over $13$ language directions in multilingual setting and $+2.0$ BLEU over $3$ domains in multi-domain setting.
Multilingual machine translation has attracted much attention recently due to its support of knowledge transfer among languages and the low cost of training and deployment compared with numerous bilingual models. A known challenge of multilingual mod
els is the negative language interference. In order to enhance the translation quality, deeper and wider architectures are applied to multilingual modeling for larger model capacity, which suffers from the increased inference cost at the same time. It has been pointed out in recent studies that parameters shared among languages are the cause of interference while they may also enable positive transfer. Based on these insights, we propose an adaptive and sparse architecture for multilingual modeling, and train the model to learn shared and language-specific parameters to improve the positive transfer and mitigate the interference. The sparse architecture only activates a subnetwork which preserves inference efficiency, and the adaptive design selects different subnetworks based on the input languages. Evaluated on multilingual translation across multiple public datasets, our model outperforms strong baselines in terms of translation quality without increasing the inference cost.
We study a new application for text generation -- idiomatic sentence generation -- which aims to transfer literal phrases in sentences into their idiomatic counterparts. Inspired by psycholinguistic theories of idiom use in ones native language, we p
ropose a novel approach for this task, which retrieves the appropriate idiom for a given literal sentence, extracts the span of the sentence to be replaced by the idiom, and generates the idiomatic sentence by using a neural model to combine the retrieved idiom and the remainder of the sentence. Experiments on a novel dataset created for this task show that our model is able to effectively transfer literal sentences into idiomatic ones. Furthermore, automatic and human evaluations show that for this task, the proposed model outperforms a series of competitive baseline models for text generation.
The meaning of a word is closely linked to sociocultural factors that can change over time and location, resulting in corresponding meaning changes. Taking a global view of words and their meanings in a widely used language, such as English, may requ
ire us to capture more refined semantics for use in time-specific or location-aware situations, such as the study of cultural trends or language use. However, popular vector representations for words do not adequately include temporal or spatial information. In this work, we present a model for learning word representation conditioned on time and location. In addition to capturing meaning changes over time and location, we require that the resulting word embeddings retain salient semantic and geometric properties. We train our model on time- and location-stamped corpora, and show using both quantitative and qualitative evaluations that it can capture semantics across time and locations. We note that our model compares favorably with the state-of-the-art for time-specific embedding, and serves as a new benchmark for location-specific embeddings.
In this paper, we study machine reading comprehension (MRC) on long texts, where a model takes as inputs a lengthy document and a question and then extracts a text span from the document as an answer. State-of-the-art models tend to use a pretrained
transformer model (e.g., BERT) to encode the joint contextual information of document and question. However, these transformer-based models can only take a fixed-length (e.g., 512) text as its input. To deal with even longer text inputs, previous approaches usually chunk them into equally-spaced segments and predict answers based on each segment independently without considering the information from other segments. As a result, they may form segments that fail to cover the correct answer span or retain insufficient contexts around it, which significantly degrades the performance. Moreover, they are less capable of answering questions that need cross-segment information. We propose to let a model learn to chunk in a more flexible way via reinforcement learning: a model can decide the next segment that it wants to process in either direction. We also employ recurrent mechanisms to enable information to flow across segments. Experiments on three MRC datasets -- CoQA, QuAC, and TriviaQA -- demonstrate the effectiveness of our proposed recurrent chunking mechanisms: we can obtain segments that are more likely to contain complete answers and at the same time provide sufficient contexts around the ground truth answers for better predictions.
