Константин Вячеславович Воронцов

This paper introduces an approach to measuring the cognitive complexity of texts on various language levels. While standard readability indices are based on the linear combination of primary statistics, our general approach allows us to estimate complexity on morphological, lexical, syntactic, and discursive levels. Each model is defined by the tokens for the specific language level and the complexity function of a single token. We then use the reference collection of moderately complex texts and the quantile-based approach to spot the abnormally rare tokens. The proposed supervised ensemble, based on the ElasticNet model, incorporates models from all language levels. Having collected a labeled dataset through crowdsourcing, consisting of pairs of articles from the Russian Wikipedia, we consider several models and ensembles and compare them to common baselines. Suggested models are flexible due to the freedom in choosing the reference collection. The described experiments confirm the competitiveness of the proposed approach, as the ensembles demonstrate the best target metric value.

Opinion mining is a popular task, that is applied, for example, to determine news polarisation and identify product review classes. Our task is unsupervised clusterization of opinionated texts, in particular news on political events. Many papers that tackle this issue use generative models based on lexical features. Our goal is to determine the entities defying an opinion amongst lexical, syntactic and semantic features as well as their compositions. More specifically, we test the hypothesis that an opinion is determined by the composition of the mentioned facts (SPO triples), the semantic roles of the words and the sentiment lexicon used in it. In this paper we formalise this task and prove that using a composition of the above features provides the best quality when clusterising opinionated texts. To test this hypothesis we have gathered and labelled two corpuses of news on political events and proposed a set of unsupervised algorithms for extracting the features.

, : . , , , . ,

This article proposes a new approach for building topic models on unbalanced collections in topic modelling, based on the existing methods and our experiments with such methods. Real-world data collections contain topics in various proportions, and often documents of the relatively small theme become distributed all over the larger topics instead of being grouped into one topic. To address this issue, we design a new regularizer for and matrices in probabilistic Latent Semantic Analysis (pLSA) model. We make sure this regularizer increases the quality of topic models, trained on unbalanced collections. Besides, we conceptually support this regularizer by our experiments.

Real-time monitoring of scientific papers and technological news requires fast processing of complicated search demands motivated by thematically relevant information acquisition. For this case, the authors develop an exploratory search engine based on probabilistic hierarchical topic modeling. Topic model gives a low dimensional sparse interpretable vector representation (topical embedding) of a text, which is used for ranking documents by their similarity to the query. They explore several ways of comparing topical vectors including searching with thematically homogeneous text segments. Topical hierarchies are built using the regularized EM-algorithm from BigARTM project. The topic-based search achieves better precision and recall than other approaches (TF-IDF, fastText, LSTM, BERT) and even human assessors who spend up to an hour to complete the same search task. They also discover that blending hierarchical topic vectors with neural pretrained embeddings is a promising way of enriching both models that helps to get precision and recall higher than 90

Topic modeling is an area of text analysis actively developing over the past 20 years. A probabilistic topic model (PTM) finds a set of hidden topics from a collection of text documents. It defines each topic as a probability distribution over words and describes each document as a probability mixture of topic distributions. Learning algorithms for topic models are usually based on Bayesian inference or log-likelihood maximization. In both cases, EM-like algorithms are used. In this paper, we propose to replace the logarithm in the log-likelihood by an arbitrary smooth loss function. We prove that such a modification preserves both the structure of the algorithm and compatibility with any regularizers in terms of additive regularization of topic models (ARTM). Moreover, in the case of a linear loss, the Estep becomes much faster due to the omission of a normalization. We study combinations of the fast and usual E-steps and compare them to regularization using different number of topics in both offline and online versions of EM-algorithm. For an empirical comparison of the algorithms, we estimate perplexity, coherence, and learning time. We use an efficient parallel implementation of the EM-algorithm from the BigARTM open-source library. We show that in most cases the two-stage strategy wins, which uses fast E-steps at the beginning of iterations, then proceeds with usual E-steps.

We explore a recently proposed question answering system. We developed a high speed modification based on dividing the question answering system into three consecutive stages. The first step is to find the candidate documents that most likely contain the answer to the question. The second step is to rank sentences by the probability of having a correct answer to the question. The third step is to find the exact phrase that answers the question. At the third step we used a recently proposed recurrent bidirectional neural network predicting the beginning and the end of a response. In this paper we showed that the proposed question answering system allows to speed up its work without significant losses in the quality. For each step we also explored the feature space construction techniques allowing to improve the final quality.

With the increasing popularity of cashless payment methods for everyday, seasonal and special expenses popular banks accumulate huge amount of data about customer operations. In the article, we report a successful application of topic modelling to extract behaviour patterns from the data. The resulting models are built with BigARTM framework: flexible and efficient tool for topic modelling. The framework allows us to experiment with various models including PLSA, LDA and beyond. Results demonstrate ability of the approach to aggregate information about behaviour patterns of different customer groups. The results analysis allows to see the topics of such people clusters varying from travellers to mortgage holders. Moreover, low-dementional embeddings of the customers, which was given with topic model, were studied. We display that the client vector representations store demographic information as well as source data. We also test for a best way of preparing data for the model with metric above in mind.

In the exploratory search paradigm of information retrieval, the user has a complicated search demand that can not be formulated in a short query. The user collects thematically relevant information iteratively in a “query-browse-refine” process being motivated by learning, understanding, and knowledge acquisition purposes. We consider an elementary step of this scenario in which the search intent can be expressed by a long text query. For this case, we develop an exploratory search engine based on probabilistic topic modeling. Topic model gives a low-dimensional sparse interpretable vector representation (topical embedding) of a text. The search engine uses these embeddings for ranking documents by their similarity to the query. We show that performing only one query, the topic-based search engine achieves better precision and recall that human assessors do spending up to one hour in a conventional browse-refine loop. We use additive regularization for topic modeling (ARTM) to make the model simultaneously sparse, decorrelated, n-gram, multimodal and hierarchical. We show experimentally that each of these features of the model is important to achieve precision and recall higher than 90

This paper introduces a new approach to estimating the text document complexity.Common readability indices are based on average length of sentences and words.In contrast to these methods, we propose to count the number of rare words occurring abnormally often in the document.We use the reference corpus of texts and the quantile approach in order to determine what words are rare, and what frequencies are abnormal.We construct a general text complexity model, which can be adjusted for the specific task, and introduce two special models.The experimental design is based on a set of thematically similar pairs of Wikipedia articles, labeled using crowdsourcing.The experiments demonstrate the competitiveness of the proposed approach.