It is often difficult to reliably evaluate models which generate text. Among them, text style transfer is a particularly difficult to evaluate, because its success depends on a number of parameters. We conduct an evaluation of a large number of models on a detoxification task. We explore the relations between the manual and automatic metrics and find that there is only weak correlation between them, which is dependent on the type of model which generated text. Automatic metrics tend to be less reliable for better-performing models. However, our findings suggest that, ChrF and BertScore metrics can be used as a proxy for human evaluation of text detoxification to some extent.

We present a novel pipeline for the collection of parallel data for the detoxification task. We collect non-toxic paraphrases for over 10,000 English toxic sentences. We also show that this pipeline can be used to distill a large existing corpus of paraphrases to get toxic-neutral sentence pairs. We release two parallel corpora which can be used for the training of detoxification models. To the best of our knowledge, these are the first parallel datasets for this task. We describe our pipeline in detail to make it fast to set up for a new language or domain, thus contributing to faster and easier development of new parallel resources. We train several detoxification models on the collected data and compare them with several baselines and state-of-the-art unsupervised approaches. We conduct both automatic and manual evaluations. All models trained on parallel data outperform the state-of-the-art unsupervised models by a large margin. This suggests that our novel datasets can boost the performance of detoxification systems.

Text style transfer and paraphrasing of texts are actively growing areas of NLP, dozens of methods for solving these tasks have been recently introduced. In both tasks, the system is supposed to generate a text which should be semantically similar to the input text. Therefore, these tasks are dependent on methods of measuring textual semantic similarity. However, it is still unclear which measures are the best to automatically evaluate content preservation between original and generated text. According to our observations, many researchers still use BLEU-like measures, while there exist more advanced measures including neural-based that significantly outperform classic approaches. The current problem is the lack of a thorough evaluation of the available measures. We close this gap by conducting a large-scale computational study by comparing 57 measures based on different principles on 19 annotated datasets. We show that measures based on cross-encoder models outperform alternative approaches in almost all cases. We also introduce the Mutual Implication Score (MIS), a measure that uses the idea of paraphrasing as a bidirectional entailment and outperforms all other measures on the paraphrase detection task and performs on par with the best measures in the text style transfer task.

Paraphrase identification task can be easily challenged by changing word order, e.g. as in “Can a good person become bad?”. While for English this problem was tackled by the PAWS dataset (Zhang et al., 2019), datasets for Russian paraphrase detection lack non-paraphrase examples with high lexical overlap. We present RuPAWS, the first adversarial dataset for Russian paraphrase identification. Our dataset consists of examples from PAWS translated to the Russian language and manually annotated by native speakers. We compare it to the largest available dataset for Russian ParaPhraser and show that the best available paraphrase identifiers for the Russian language fail on the RuPAWS dataset. At the same time, the state-of-the-art paraphrasing model RuBERT trained on both RuPAWS and ParaPhraser obtains high performance on the RuPAWS dataset while maintaining its accuracy on the ParaPhraser benchmark. We also show that RuPAWS can measure the sensitivity of models to word order and syntax structure since simple baselines fail even when given RuPAWS training samples.

We present two novel unsupervised methods for eliminating toxicity in text. Our first method combines two recent ideas: (1) guidance of the generation process with small style-conditional language models and (2) use of paraphrasing models to perform style transfer. We use a well-performing paraphraser guided by style-trained language models to keep the text content and remove toxicity. Our second method uses BERT to replace toxic words with their non-offensive synonyms. We make the method more flexible by enabling BERT to replace mask tokens with a variable number of words. Finally, we present the first large-scale comparative study of style transfer models on the task of toxicity removal. We compare our models with a number of methods for style transfer. The models are evaluated in a reference-free way using a combination of unsupervised style transfer metrics. Both methods we suggest yield new SOTA results.

This work describes the participation of the Skoltech NLP group team (Sk) in the Toxic Spans Detection task at SemEval-2021. The goal of the task is to identify the most toxic fragments of a given sentence, which is a binary sequence tagging problem. We show that fine-tuning a RoBERTa model for this problem is a strong baseline. This baseline can be further improved by pre-training the RoBERTa model on a large dataset labeled for toxicity at the sentence level. While our solution scored among the top 20% participating models, it is only 2 points below the best result. This suggests the viability of our approach.

We present a system for answering comparative questions (Is X better than Y with respect to Z?) in natural language. Answering such questions is important for assisting humans in making informed decisions. The key component of our system is a natural language interface for comparative QA that can be used in personal assistants, chatbots, and similar NLP devices. Comparative QA is a challenging NLP task, since it requires collecting support evidence from many different sources, and direct comparisons of rare objects may be not available even on the entire Web. We take the first step towards a solution for such a task offering a testbed for comparative QA in natural language by probing several methods, making the three best ones available as an online demo.

The vast majority of the existing approaches for taxonomy enrichment apply word embeddings as they have proven to accumulate contexts (in a broad sense) extracted from texts which are sufficient for attaching orphan words to the taxonomy. On the other hand, apart from being large lexical and semantic resources, taxonomies are graph structures. Combining word embeddings with graph structure of taxonomy could be of use for predicting taxonomic relations. In this paper we compare several approaches for attaching new words to the existing taxonomy which are based on the graph representations with the one that relies on fastText embeddings. We test all methods on Russian and English datasets, but they could be also applied to other wordnets and languages.

Not all topics are equally “flammable” in terms of toxicity: a calm discussion of turtles or fishing less often fuels inappropriate toxic dialogues than a discussion of politics or sexual minorities. We define a set of sensitive topics that can yield inappropriate and toxic messages and describe the methodology of collecting and labelling a dataset for appropriateness. While toxicity in user-generated data is well-studied, we aim at defining a more fine-grained notion of inappropriateness. The core of inappropriateness is that it can harm the reputation of a speaker. This is different from toxicity in two respects: (i) inappropriateness is topic-related, and (ii) inappropriate message is not toxic but still unacceptable. We collect and release two datasets for Russian: a topic-labelled dataset and an appropriateness-labelled dataset. We also release pre-trained classification models trained on this data.

Disambiguation of word senses in context is easy for humans, but is a major challenge for automatic approaches. Sophisticated supervised and knowledge-based models were developed to solve this task. However, (i) the inherent Zipfian distribution of supervised training instances for a given word and/or (ii) the quality of linguistic knowledge representations motivate the development of completely unsupervised and knowledge-free approaches to word sense disambiguation (WSD). They are particularly useful for under-resourced languages which do not have any resources for building either supervised and/or knowledge-based models. In this paper, we present a method that takes as input a standard pre-trained word embedding model and induces a fully-fledged word sense inventory, which can be used for disambiguation in context. We use this method to induce a collection of sense inventories for 158 languages on the basis of the original pre-trained fastText word embeddings by Grave et al., (2018), enabling WSD in these languages. Models and system are available online.

Ontologies, taxonomies, and thesauri have always been in high demand in a large number of NLP tasks. However, most studies are focused on the creation of lexical resources rather than the maintenance of the existing ones and keeping them up-to-date. In this paper, we address the problem of taxonomy enrichment. Namely, we explore the possibilities of taxonomy extension in a resource-poor setting and present several methods which are applicable to a large number of languages. We also create novel English and Russian datasets for training and evaluating taxonomy enrichment systems and describe a technique of creating such datasets for other languages.

We explore different model architectures for the WMT 19 shared task on word-level quality estimation of automatic translation. We start with a model similar to Shef-bRNN, which we modify by using conditional random fields for sequence labelling. Additionally, we use a different approach for labelling gaps and source words. We further develop this model by including features from different sources such as BERT, baseline features for the task and transformer encoders. We evaluate the performance of our models on the English-German dataset for the corresponding shared task.

Adoption of messaging communication and voice assistants has grown rapidly in the last years. This creates a demand for tools that speed up prototyping of feature-rich dialogue systems. An open-source library DeepPavlov is tailored for development of conversational agents. The library prioritises efficiency, modularity, and extensibility with the goal to make it easier to develop dialogue systems from scratch and with limited data available. It supports modular as well as end-to-end approaches to implementation of conversational agents. Conversational agent consists of skills and every skill can be decomposed into components. Components are usually models which solve typical NLP tasks such as intent classification, named entity recognition or pre-trained word vectors. Sequence-to-sequence chit-chat skill, question answering skill or task-oriented skill can be assembled from components provided in the library.

We suggest a new language-independent architecture of robust word vectors (RoVe). It is designed to alleviate the issue of typos, which are common in almost any user-generated content, and hinder automatic text processing. Our model is morphologically motivated, which allows it to deal with unseen word forms in morphologically rich languages. We present the results on a number of Natural Language Processing (NLP) tasks and languages for the variety of related architectures and show that proposed architecture is typo-proof.

We report the results of the WMT18 shared task on Quality Estimation, i.e. the task of predicting the quality of the output of machine translation systems at various granularity levels: word, phrase, sentence and document. This year we include four language pairs, three text domains, and translations produced by both statistical and neural machine translation systems. Participating teams from ten institutions submitted a variety of systems to different task variants and language pairs.

This paper presents our work towards a novel approach for Quality Estimation (QE) of machine translation based on sequences of adjacent words, the so-called phrases. This new level of QE aims to provide a natural balance between QE at word and sentence-level, which are either too fine grained or too coarse levels for some applications. However, phrase-level QE implies an intrinsic challenge: how to segment a machine translation into sequence of words (contiguous or not) that represent an error. We discuss three possible segmentation strategies to automatically extract erroneous phrases. We evaluate these strategies against annotations at phrase-level produced by humans, using a new dataset collected for this purpose.

We present Marmot~― a new toolkit for quality estimation (QE) of machine translation output. Marmot contains utilities targeted at quality estimation at the word and phrase level. However, due to its flexibility and modularity, it can also be extended to work at the sentence level. In addition, it can be used as a framework for extracting features and learning models for many common natural language processing tasks. The tool has a set of state-of-the-art features for QE, and new features can easily be added. The tool is open-source and can be downloaded from https://github.com/qe-team/marmot/

This paper presents a new active learning technique for machine translation based on quality estimation of automatically translated sentences. It uses an error-driven strategy, i.e., it assumes that the more errors an automatically translated sentence contains, the more informative it is for the translation system. Our approach is based on a quality estimation technique which involves a wider range of features of the source text, automatic translation, and machine translation system compared to previous work. In addition, we enhance the machine translation system training data with post-edited machine translations of the sentences selected, instead of simulating this using previously created reference translations. We found that re-training systems with additional post-edited data yields higher quality translations regardless of the selection strategy used. We relate this to the fact that post-editions tend to be closer to source sentences as compared to references, making the rule extraction process more reliable.