Measuring the similarity score between a pair of sentences in different languages is the essential requisite for multilingual sentence embedding methods. Predicting the similarity score consists of two sub-tasks, which are monolingual similarity eval
uation and multilingual sentence retrieval. However, conventional methods have mainly tackled only one of the sub-tasks and therefore showed biased performances. In this paper, we suggest a novel and strong method for multilingual sentence embedding, which shows performance improvement on both sub-tasks, consequently resulting in robust predictions of multilingual similarity scores. The suggested method consists of two parts: to learn semantic similarity of sentences in the pivot language and then to extend the learned semantic structure to different languages. To align semantic structures across different languages, we introduce a teacher-student network. The teacher network distills the knowledge of the pivot language to different languages of the student network. During the distillation, the parameters of the teacher network are updated with the slow-moving average. Together with the distillation and the parameter updating, the semantic structure of the student network can be directly aligned across different languages while preserving the ability to measure the semantic similarity. Thus, the multilingual training method drives performance improvement on multilingual similarity evaluation. The suggested model achieves the state-of-the-art performance on extended STS 2017 multilingual similarity evaluation as well as two sub-tasks, which are extended STS 2017 monolingual similarity evaluation and Tatoeba multilingual retrieval in 14 languages.
Byte-pair encoding (BPE) is a ubiquitous algorithm in the subword tokenization process of language models as it provides multiple benefits. However, this process is solely based on pre-training data statistics, making it hard for the tokenizer to han
dle infrequent spellings. On the other hand, though robust to misspellings, pure character-level models often lead to unreasonably long sequences and make it harder for the model to learn meaningful words. To alleviate these challenges, we propose a character-based subword module (char2subword) that learns the subword embedding table in pre-trained models like BERT. Our char2subword module builds representations from characters out of the subword vocabulary, and it can be used as a drop-in replacement of the subword embedding table. The module is robust to character-level alterations such as misspellings, word inflection, casing, and punctuation. We integrate it further with BERT through pre-training while keeping BERT transformer parameters fixed--and thus, providing a practical method. Finally, we show that incorporating our module to mBERT significantly improves the performance on the social media linguistic code-switching evaluation (LinCE) benchmark.
Current embedding-based large-scale retrieval models are trained with 0-1 hard label that indicates whether a query is relevant to a document, ignoring rich information of the relevance degree. This paper proposes to improve embedding-based retrieval
from the perspective of better characterizing the query-document relevance degree by introducing label enhancement (LE) for the first time. To generate label distribution in the retrieval scenario, we design a novel and effective supervised LE method that incorporates prior knowledge from dynamic term weighting methods into contextual embeddings. Our method significantly outperforms four competitive existing retrieval models and its counterparts equipped with two alternative LE techniques by training models with the generated label distribution as auxiliary supervision information. The superiority can be easily observed on English and Chinese large-scale retrieval tasks under both standard and cold-start settings.
Learning authors representations from their textual productions is now widely used to solve multiple downstream tasks, such as classification, link prediction or user recommendation. Author embedding methods are often built on top of either Doc2Vec (
Mikolov et al. 2014) or the Transformer architecture (Devlin et al. 2019). Evaluating the quality of these embeddings and what they capture is a difficult task. Most articles use either classification accuracy or authorship attribution, which does not clearly measure the quality of the representation space, if it really captures what it has been built for. In this paper, we propose a novel evaluation framework of author embedding methods based on the writing style. It allows to quantify if the embedding space effectively captures a set of stylistic features, chosen to be the best proxy of an author writing style. This approach gives less importance to the topics conveyed by the documents. It turns out that recent models are mostly driven by the inner semantic of authors' production. They are outperformed by simple baselines, based on state-of-the-art pretrained sentence embedding models, on several linguistic axes. These baselines can grasp complex linguistic phenomena and writing style more efficiently, paving the way for designing new style-driven author embedding models.
Position representation is crucial for building position-aware representations in Transformers. Existing position representations suffer from a lack of generalization to test data with unseen lengths or high computational cost. We investigate shifted
absolute position embedding (SHAPE) to address both issues. The basic idea of SHAPE is to achieve shift invariance, which is a key property of recent successful position representations, by randomly shifting absolute positions during training. We demonstrate that SHAPE is empirically comparable to its counterpart while being simpler and faster.
It is widely accepted that fine-tuning pre-trained language models usually brings about performance improvements in downstream tasks. However, there are limited studies on the reasons behind this effectiveness, particularly from the viewpoint of stru
ctural changes in the embedding space. Trying to fill this gap, in this paper, we analyze the extent to which the isotropy of the embedding space changes after fine-tuning. We demonstrate that, even though isotropy is a desirable geometrical property, fine-tuning does not necessarily result in isotropy enhancements. Moreover, local structures in pre-trained contextual word representations (CWRs), such as those encoding token types or frequency, undergo a massive change during fine-tuning. Our experiments show dramatic growth in the number of elongated directions in the embedding space, which, in contrast to pre-trained CWRs, carry the essential linguistic knowledge in the fine-tuned embedding space, making existing isotropy enhancement methods ineffective.
Narrative analysis is becoming increasingly important for a number of linguistic tasks including summarization, knowledge extraction, and question answering. We present a novel approach for narrative event representation using attention to re-context
ualize events across the whole story. Comparing to previous analysis we find an unexpected attachment of event semantics to predicate tokens within a popular transformer model. We test the utility of our approach on narrative completion prediction, achieving state of the art performance on Multiple Choice Narrative Cloze and scoring competitively on the Story Cloze Task.
In this work, we analyze the performance and properties of cross-lingual word embedding models created by mapping-based alignment methods. We use several measures of corpus and embedding similarity to predict BLI scores of cross-lingual embedding map
pings over three types of corpora, three embedding methods and 55 language pairs. Our experimental results corroborate that instead of mere size, the amount of common content in the training corpora is essential. This phenomenon manifests in that i) despite of the smaller corpus sizes, using only the comparable parts of Wikipedia for training the monolingual embedding spaces to be mapped is often more efficient than relying on all the contents of Wikipedia, ii) the smaller, in return less diversified Spanish Wikipedia works almost always much better as a training corpus for bilingual mappings than the ubiquitously used English Wikipedia.
Word embedding is essential for neural network models for various natural language processing tasks. Since the word embedding usually has a considerable size, in order to deploy a neural network model having it on edge devices, it should be effective
ly compressed. There was a study for proposing a block-wise low-rank approximation method for word embedding, called GroupReduce. Even if their structure is effective, the properties behind the concept of the block-wise word embedding compression were not sufficiently explored. Motivated by this, we improve GroupReduce in terms of word weighting and structuring. For word weighting, we propose a simple yet effective method inspired by the term frequency-inverse document frequency method and a novel differentiable method. Based on them, we construct a discriminative word embedding compression algorithm. In the experiments, we demonstrate that the proposed algorithm more effectively finds word weights than competitors in most cases. In addition, we show that the proposed algorithm can act like a framework through successful cooperation with quantization.
In cross-lingual language models, representations for many different languages live in the same space. Here, we investigate the linguistic and non-linguistic factors affecting sentence-level alignment in cross-lingual pretrained language models for 1
01 languages and 5,050 language pairs. Using BERT-based LaBSE and BiLSTM-based LASER as our models, and the Bible as our corpus, we compute a task-based measure of cross-lingual alignment in the form of bitext retrieval performance, as well as four intrinsic measures of vector space alignment and isomorphism. We then examine a range of linguistic, quasi-linguistic, and training-related features as potential predictors of these alignment metrics. The results of our analyses show that word order agreement and agreement in morphological complexity are two of the strongest linguistic predictors of cross-linguality. We also note in-family training data as a stronger predictor than language-specific training data across the board. We verify some of our linguistic findings by looking at the effect of morphological segmentation on English-Inuktitut alignment, in addition to examining the effect of word order agreement on isomorphism for 66 zero-shot language pairs from a different corpus. We make the data and code for our experiments publicly available.
