In most of neural machine translation distillation or stealing scenarios, the highest-scoring hypothesis of the target model (teacher) is used to train a new model (student). If reference translations are also available, then better hypotheses (with respect to the references) can be oversampled and poor hypotheses either removed or undersampled. This paper explores the sampling method landscape (pruning, hypothesis oversampling and undersampling, deduplication and their combination) with English to Czech and English to German MT models using standard MT evaluation metrics. We show that careful oversampling and combination with the original data leads to better performance when compared to training only on the original or synthesized data or their direct combination.

Automatic Text Summarization (ATS) is the task of generating concise and fluent summaries from one or more documents. In this paper, we present IceSum, the first Icelandic corpus annotated with human-generated summaries. IceSum consists of 1,000 online news articles and their extractive summaries. We train and evaluate several neural network-based models on this dataset, comparing them against a selection of baseline methods. We find that an encoder-decoder model with a sequence-to-sequence based extractor obtains the best results, outperforming all baseline methods. Furthermore, we evaluate how the size of the training corpus affects the quality of the generated summaries. We release the corpus and the models with an open license.

Negation is a linguistic universal that poses difficulties for cognitive and computational processing. Despite many advances in text analytics, negation resolution remains an acute and continuously researched question in Natural Language Processing. Reliable negation parsing affects results in biomedical text mining, sentiment analysis, machine translation, and many other fields. The availability of multilingual pre-trained general representation models makes it possible to experiment with negation detection in languages that lack annotated data. In this work we test the performance of two state-of-the-art contextual representation models, Multilingual BERT and XLM-RoBERTa. We resolve negation scope by conducting zero-shot transfer between English, Spanish, French, and Russian. Our best result amounts to a token-level F1-score of 86.86% between Spanish and Russian. We correlate these results with a linguistic negation typology and lexical capacity of the models.

Neural networks are the state-of-the-art method of machine learning for many problems in NLP. Their success in machine translation and other NLP tasks is phenomenal, but their interpretability is challenging. We want to find out how neural networks represent meaning. In order to do this, we propose to examine the distribution of meaning in the vector space representation of words in neural networks trained for NLP tasks. Furthermore, we propose to consider various theories of meaning in the philosophy of language and to find a methodology that would enable us to connect these areas.

In this thesis proposal, we explore the application of event extraction to literary texts. Considering the lengths of literary documents modeling events in different granularities may be more adequate to extract meaningful information, as individual elements contribute little to the overall semantics. We adapt the concept of schemas as sequences of events all describing a single process, connected through shared participants extending it to for multiple schemas in a document. Segmentation of event sequences into schemas is approached by modeling event sequences, on such task as the narrative cloze task, the prediction of missing events in sequences. We propose building on sequences of event embeddings to form schema embeddings, thereby summarizing sections of documents using a single representation. This approach will allow for the comparisons of different sections of documents and entire literary works. Literature is a challenging domain based on its variety of genres, yet the representation of literary content has received relatively little attention.

Text simplification is a growing field with many potential useful applications. Training text simplification algorithms generally requires a lot of annotated data, however there are not many corpora suitable for this task. We propose a new unsupervised method for aligning text based on Doc2Vec embeddings and a new alignment algorithm, capable of aligning texts at different levels. Initial evaluation shows promising results for the new approach. We used the newly developed approach to create a new monolingual parallel corpus composed of the works of English early modern philosophers and their corresponding simplified versions.

We present a simple method for extending transformers to source-side trees. We define a number of masks that limit self-attention based on relationships among tree nodes, and we allow each attention head to learn which mask or masks to use. On translation from English to various low-resource languages, and translation in both directions between English and German, our method always improves over simple linearization of the source-side parse tree and almost always improves over a sequence-to-sequence baseline, by up to +2.1 BLEU.

One of the ways blind people understand their surroundings is by clicking images and relying on descriptions generated by image-captioning systems. Current work on captioning images for the visually impaired do not use the textual data present in the image when generating captions. This problem is critical as many visual scenes contain text, and 21% of the questions asked by blind people about the images they click pertain to the text present in them. In this work, we propose altering AoANet, a state-of-the-art image-captioning system, to leverage text detected in the image as an input feature. In addition, we use a pointer-generator network to copy detected text to the caption when tokens need to be reproduced accurately. Our model outperforms AoANet on the benchmark dataset VizWiz, giving a 35% and 16.2% performance improvement on CIDEr and SPICE scores, respectively.

Open-domain question answering aims at locating the answers to user-generated questions in massive collections of documents. Retriever-readers and knowledge graph approaches are two big families of solutions to this task. A retriever-reader first applies information retrieval techniques to locate a few passages that are likely to be relevant, and then feeds the retrieved text to a neural network reader to extract the answer. Alternatively, knowledge graphs can be constructed and queried to answer users’ questions. We propose an algorithm with a novel reader-retriever design that differs from both families. Our reader-retriever first uses an offline reader to read the corpus and generate collections of all answerable questions associated with their answers, and then uses an online retriever to respond to user queries by searching the pre-constructed question spaces for answers that are most likely to be asked in the given way. We further combine one retriever-reader and two reader-retrievers into a hybrid model called R6 for the best performance. Experiments with two large-scale public datasets show that R6 achieves state-of-the-art accuracy.

Meeting minutes record any subject matter discussed, decisions reached and actions taken at the meeting. The importance of automatic minuting cannot be overstated. In this paper, we present a sliding window approach to automatic generation of meeting minutes. It aims at addressing issues pertaining to the nature of spoken text, including the lengthy transcript and lack of document structure, which make it difficult to identify salient content to be included in meeting minutes. Our approach combines a sliding-window approach and a neural abstractive summarizer to navigate through the raw transcript to find salient content. The approach is evaluated on transcripts of natural meeting conversations, where we compare results obtained for human transcripts and two versions of automatic transcripts and discuss how and to what extent the summarizer succeeds at capturing salient content.

We propose semantic visualization as a linguistic visual analytic method. It can enable exploration and discovery over large datasets of complex networks by exploiting the semantics of the relations in them. This involves extracting information, applying parameter reduction operations, building hierarchical data representation and designing visualization. We also present the accompanying COVID-SemViz a searchable and interactive visualization system for knowledge exploration of COVID-19 data to demonstrate the application of our proposed method. In the user studies, users found that semantic visualization-powered COVID-SemViz is helpful in terms of finding relevant information and discovering unknown associations.

State-of-the-art transformer models have achieved robust performance on a variety of NLP tasks. Many of these approaches have employed domain agnostic pre-training tasks to train models that yield highly generalized sentence representations that can be fine-tuned for specific downstream tasks. We propose refining a pre-trained NLP model using the objective of detecting shuffled tokens. We use a sequential approach by starting with the pre-trained RoBERTa model and training it using our approach. Applying random shuffling strategy on the word-level, we found that our approach enables the RoBERTa model achieve better performance on 4 out of 7 GLUE tasks. Our results indicate that learning to detect shuffled tokens is a promising approach to learn more coherent sentence representations.

In this work, we explore generating morphologically enhanced word embeddings for Tamil, a highly agglutinative South Indian language with rich morphology that remains low-resource with regards to NLP tasks. We present here the first-ever word analogy dataset for Tamil, consisting of 4499 hand-curated word tetrads across 10 semantic and 13 morphological relation types. Using a rules-based segmenter to capture morphology as well as meta-embedding techniques, we train meta-embeddings that outperform existing baselines by 16% on our analogy task and appear to mitigate a previously observed trade-off between semantic and morphological accuracy.

Weakly-supervised text classification aims to induce text classifiers from only a few user-provided seed words. The vast majority of previous work assumes high-quality seed words are given. However, the expert-annotated seed words are sometimes non-trivial to come up with. Furthermore, in the weakly-supervised learning setting, we do not have any labeled document to measure the seed words’ efficacy, making the seed word selection process “a walk in the dark”. In this work, we remove the need for expert-curated seed words by first mining (noisy) candidate seed words associated with the category names. We then train interim models with individual candidate seed words. Lastly, we estimate the interim models’ error rate in an unsupervised manner. The seed words that yield the lowest estimated error rates are added to the final seed word set. A comprehensive evaluation of six binary classification tasks on four popular datasets demonstrates that the proposed method outperforms a baseline using only category name seed words and obtained comparable performance as a counterpart using expert-annotated seed words.

For a computer to naturally interact with a human, it needs to be human-like. In this paper, we propose a neural response generation model with multi-task learning of generation and classification, focusing on emotion. Our model based on BART (Lewis et al., 2020), a pre-trained transformer encoder-decoder model, is trained to generate responses and recognize emotions simultaneously. Furthermore, we weight the losses for the tasks to control the update of parameters. Automatic evaluations and crowdsourced manual evaluations show that the proposed model makes generated responses more emotionally aware.

Grammatical error correction (GEC) suffers from a lack of sufficient parallel data. Studies on GEC have proposed several methods to generate pseudo data, which comprise pairs of grammatical and artificially produced ungrammatical sentences. Currently, a mainstream approach to generate pseudo data is back-translation (BT). Most previous studies using BT have employed the same architecture for both the GEC and BT models. However, GEC models have different correction tendencies depending on the architecture of their models. Thus, in this study, we compare the correction tendencies of GEC models trained on pseudo data generated by three BT models with different architectures, namely, Transformer, CNN, and LSTM. The results confirm that the correction tendencies for each error type are different for every BT model. In addition, we investigate the correction tendencies when using a combination of pseudo data generated by different BT models. As a result, we find that the combination of different BT models improves or interpolates the performance of each error type compared with using a single BT model with different seeds.

The quality and quantity of parallel sentences are known as very important training data for constructing neural machine translation (NMT) systems. However, these resources are not available for many low-resource language pairs. Many existing methods need strong supervision are not suitable. Although several attempts at developing unsupervised models, they ignore the language-invariant between languages. In this paper, we propose an approach based on transfer learning to mine parallel sentences in the unsupervised setting. With the help of bilingual corpora of rich-resource language pairs, we can mine parallel sentences without bilingual supervision of low-resource language pairs. Experiments show that our approach improves the performance of mined parallel sentences compared with previous methods. In particular, we achieve excellent results at two real-world low-resource language pairs.

Recently, neural machine translation is widely used for its high translation accuracy, but it is also known to show poor performance at long sentence translation. Besides, this tendency appears prominently for low resource languages. We assume that these problems are caused by long sentences being few in the train data. Therefore, we propose a data augmentation method for handling long sentences. Our method is simple; we only use given parallel corpora as train data and generate long sentences by concatenating two sentences. Based on our experiments, we confirm improvements in long sentence translation by proposed data augmentation despite the simplicity. Moreover, the proposed method improves translation quality more when combined with back-translation.

Although research on emotion classification has significantly progressed in high-resource languages, it is still infancy for resource-constrained languages like Bengali. However, unavailability of necessary language processing tools and deficiency of benchmark corpora makes the emotion classification task in Bengali more challenging and complicated. This work proposes a transformer-based technique to classify the Bengali text into one of the six basic emotions: anger, fear, disgust, sadness, joy, and surprise. A Bengali emotion corpus consists of 6243 texts is developed for the classification task. Experimentation carried out using various machine learning (LR, RF, MNB, SVM), deep neural networks (CNN, BiLSTM, CNN+BiLSTM) and transformer (Bangla-BERT, m-BERT, XLM-R) based approaches. Experimental outcomes indicate that XLM-R outdoes all other techniques by achieving the highest weighted f_1-score of 69.73% on the test data.

This paper proposes a new abstractive document summarization model, hierarchical BART (Hie-BART), which captures hierarchical structures of a document (i.e., sentence-word structures) in the BART model. Although the existing BART model has achieved a state-of-the-art performance on document summarization tasks, the model does not have the interactions between sentence-level information and word-level information. In machine translation tasks, the performance of neural machine translation models has been improved by incorporating multi-granularity self-attention (MG-SA), which captures the relationships between words and phrases. Inspired by the previous work, the proposed Hie-BART model incorporates MG-SA into the encoder of the BART model for capturing sentence-word structures. Evaluations on the CNN/Daily Mail dataset show that the proposed Hie-BART model outperforms some strong baselines and improves the performance of a non-hierarchical BART model (+0.23 ROUGE-L).

In primary school, children’s books, as well as in modern language learning apps, multi-modal learning strategies like illustrations of terms and phrases are used to support reading comprehension. Also, several studies in educational psychology suggest that integrating cross-modal information will improve reading comprehension. We claim that state-of- he-art multi-modal transformers, which could be used in a language learner context to improve human reading, will perform poorly because of the short and relatively simple textual data those models are trained with. To prove our hypotheses, we collected a new multi-modal image-retrieval dataset based on data from Wikipedia. In an in-depth data analysis, we highlight the differences between our dataset and other popular datasets. Additionally, we evaluate several state-of-the-art multi-modal transformers on text-image retrieval on our dataset and analyze their meager results, which verify our claims.