| { | |
| "paper_id": "2021", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T01:10:52.590649Z" | |
| }, | |
| "title": "Russian Paraphrasers: Paraphrase with Transformers", | |
| "authors": [ | |
| { | |
| "first": "Alena", | |
| "middle": [], | |
| "last": "Fenogenova", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "SberDevices", | |
| "location": { | |
| "settlement": "Sberbank", | |
| "country": "Russia" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper focuses on generation methods for paraphrasing in the Russian language. There are several transformer-based models (Russian and multilingual) trained on a collected corpus of paraphrases. We compare different models, contrast the quality of paraphrases using different ranking methods and apply paraphrasing methods in the context of augmentation procedure for different tasks. The contributions of the work are the combined paraphrasing dataset, fine-tuned generated models for Russian paraphrasing task and additionally the open source tool for simple usage of the paraphrasers.", | |
| "pdf_parse": { | |
| "paper_id": "2021", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper focuses on generation methods for paraphrasing in the Russian language. There are several transformer-based models (Russian and multilingual) trained on a collected corpus of paraphrases. We compare different models, contrast the quality of paraphrases using different ranking methods and apply paraphrasing methods in the context of augmentation procedure for different tasks. The contributions of the work are the combined paraphrasing dataset, fine-tuned generated models for Russian paraphrasing task and additionally the open source tool for simple usage of the paraphrasers.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "One of the prominent features of any natural language is its diversity. Variability and ambiguity of natural languages lead to infinity of sequence combinations and one can always form a new sentence that has never been said before. However, there are approaches to automatic construction of texts with roughly the same meaning: paraphrases. Paraphrasing is expressing the meaning of an input sequence in alternative ways while maintaining grammatical and syntactical correctness. Paraphrases are of a great use in diverse applications on downstream NLP tasks and are presented in two main task forms: 1) Paraphrase identification -detecting if a pair of text inputs has the same meaning; classification task. 2) Paraphrase generation -producing paraphrases allows for the creation of more varied and fluent text; generation task.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The identification of paraphrases is very useful in many tasks, such as multi-document summarization (identifying paraphrases allows to condense information repeated across documents), question answering (checking the sequences of the tests, keyword matching to find answers), semantic parsing and search (to find the same queries or documents) and many others (Lewis et al., 2020) .", | |
| "cite_spans": [ | |
| { | |
| "start": 361, | |
| "end": 381, | |
| "text": "(Lewis et al., 2020)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this work we will discuss paraphrase generation applicability. Paraphrase generation is used in different NLP applications (for example, in chatbots to diversify responses (Lippe et al., 2020) ) and sub-tasks. Paraphrasers can be used to augment datasets with new data. For question answering systems, paraphrasing questions can not just increase the number of data examples for training ML-models (Xu et al., 2020) , but are also used to match them with key words in the knowledge base. Paraphrasers can help generate adversarial examples to evaluate model robustness -increasing the stability of ML-models: training models on a wide variety of examples in different styles, with different sentiment, but the same meaning or intent of the user. The demand for targeting paraphrasers for generating specific writing styles is also trending now (Xu et al., 2012; Bolshakov and Gelbukh, 2004) . This type of paraphrasing performs different types of style transfer, such as changing style from rude to polite, or from professional to simple language.", | |
| "cite_spans": [ | |
| { | |
| "start": 175, | |
| "end": 195, | |
| "text": "(Lippe et al., 2020)", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 401, | |
| "end": 418, | |
| "text": "(Xu et al., 2020)", | |
| "ref_id": "BIBREF28" | |
| }, | |
| { | |
| "start": 847, | |
| "end": 864, | |
| "text": "(Xu et al., 2012;", | |
| "ref_id": "BIBREF29" | |
| }, | |
| { | |
| "start": 865, | |
| "end": 893, | |
| "text": "Bolshakov and Gelbukh, 2004)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "There are some general approaches for paraphrase generation. Rule-based approaches (Meteer and Shaked, 1988) and data-driven methods (Madnani and Dorr, 2010) are the oldest ones. Currently, the most common approach is to consider the task as supervised learning using sequence-to-sequence models (Gupta et al., 2018) . The unsupervised approaches (Niu et al., 2020) are also very common. Other methods proposed include use of Deep Reinforcement Learning (Qian et al., 2019; Siddique et al., 2020) . Fine-tuning with large language models such as GPT2 is also a valuable approach that can be considered supervised (Witteveen and Andrews, 2019) or unsupervised (Hegde and Patil, 2020) .", | |
| "cite_spans": [ | |
| { | |
| "start": 83, | |
| "end": 108, | |
| "text": "(Meteer and Shaked, 1988)", | |
| "ref_id": "BIBREF15" | |
| }, | |
| { | |
| "start": 133, | |
| "end": 157, | |
| "text": "(Madnani and Dorr, 2010)", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 296, | |
| "end": 316, | |
| "text": "(Gupta et al., 2018)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 347, | |
| "end": 365, | |
| "text": "(Niu et al., 2020)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 454, | |
| "end": 473, | |
| "text": "(Qian et al., 2019;", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 474, | |
| "end": 496, | |
| "text": "Siddique et al., 2020)", | |
| "ref_id": "BIBREF25" | |
| }, | |
| { | |
| "start": 613, | |
| "end": 642, | |
| "text": "(Witteveen and Andrews, 2019)", | |
| "ref_id": "BIBREF27" | |
| }, | |
| { | |
| "start": 659, | |
| "end": 682, | |
| "text": "(Hegde and Patil, 2020)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "The majority of the resources and methods for paraphrasing are proposed for the English language. For the Russian language there were several attempts of paraphrase corpora creation (Pronoza et al., 2015; Gudkov et al., 2020) . In 2016 the collection of the Russian paraphrase corpus and the Paraphrase Detection Shared Task (Pivovarova et al., 2017) were organized, which attracted attention to the topic and led to a number of further works on the identification of paraphrases (Kuratov and Arkhipov, 2019) and sentence similarity experiments (Kravchenko, 2017; Boyarsky and Kanevsky, 2017) .", | |
| "cite_spans": [ | |
| { | |
| "start": 182, | |
| "end": 204, | |
| "text": "(Pronoza et al., 2015;", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 205, | |
| "end": 225, | |
| "text": "Gudkov et al., 2020)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 325, | |
| "end": 350, | |
| "text": "(Pivovarova et al., 2017)", | |
| "ref_id": "BIBREF17" | |
| }, | |
| { | |
| "start": 480, | |
| "end": 508, | |
| "text": "(Kuratov and Arkhipov, 2019)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 545, | |
| "end": 563, | |
| "text": "(Kravchenko, 2017;", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 564, | |
| "end": 592, | |
| "text": "Boyarsky and Kanevsky, 2017)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this paper, we compare different language models for paraphrase generation in Russian, namely rugpt2-large, rugpt3-large, and multilingual models -mT5. We prove that all these models can generate good Russian paraphrases and test different ranking methods on generated examples. We provide the combined paraphrasing dataset, finetuned generated models for Russian paraphrasing task, augmentation experiments on data for common NLP tasks, and additionally present the open source tool for user-friendly usage of the Russian paraphrasers.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "This paper is structured as follows: in section 2 we present the methodology -the dataset we use 2.1, models we fine-tune 2.2, and range strategies for paraphrasers output 2.3; section 3 is devoted to evaluation and analysis of the paraphraser performance and results -the models scores 3.1, the augmentation procedure with paraphrasers 3.2, and 3.3 the discussion about the results in the context of paraphrase application; and section 4 concludes the paper.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Language models achieve impressive results when trained in a semi-supervised manner to predict the next word or words in a sequence. They can be fine-tuned and used for a variety of downstream NLP tasks (text classification, sentiment analysis, NER etc.). Good examples of such large language models that can be used for text generation are GPT-2, GPT-3, and mT5. In this section, we present our experiments with these models for Russian trained on the prepared dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Methodology", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Historically there are several approaches that have been used to construct paraphrasing datasets. The event-based approach was chosen for the creation of the Russian paraphrase corpus (Pivovarova et al., 2017) . For experiments in this paper the dataset we use consists of two main parts: 1) news data from ParaPhraserPlus 1 for train and validation set and Shared task golden test for test set 2) conversational data from subtitles 2 (that were generated in an argument-distribution approach) and dialogues of users with chatbots (further in the text called speech). The distribution of the parts and data sizes are presented in Table 1 . The test set was checked manually and further in the evaluation we assume that golden set contains high quality paraphrases to compare with.", | |
| "cite_spans": [ | |
| { | |
| "start": 184, | |
| "end": 209, | |
| "text": "(Pivovarova et al., 2017)", | |
| "ref_id": "BIBREF17" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 630, | |
| "end": 637, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Dataset", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "Thus, the dataset presents two domains: informal style (speech subset, also presented in question form) and formal (news headlines). The speech subset of the data was checked for grammatical errors and typos with Yandex.Speller 3 . It was also filtered by metrics ROUGE-L (Lin, 2004) with threshold between 0.95 and 0.5. The example of the data is presented in Figure 1 .", | |
| "cite_spans": [ | |
| { | |
| "start": 272, | |
| "end": 283, | |
| "text": "(Lin, 2004)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 361, | |
| "end": 369, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Dataset", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The news subset of the corpus was converted into the format of sentence pairs: sentence i == sentenceparaphrase i . Additionally, we automatically checked the cases when the information in the paraphrase sentence was excessive, for instance, sentence 1 Jose Mourinho on the verge of being fired at Manchester United. and sentence 2 Mourinho could be fired if Manchester United lose to Burnley on Saturday. The second sentence contains more information about the game, it is timing and the opposing team; in data it is permissible to have extra information in the reference sentence, but not in the paraphrase. Our experiments show that the generative models (fine-tuned on such structured data) generated more diverse sentences with absolutely out of control new information and names that could not be defined as paraphrases. It was the reason for the filtration of the pairs, where paraphrase sentence has length much longer than reference sentence or contains significantly more NER, date, and address information (the tool natasha 4 was used to detect entities). We set thresholds empirically and not strictly in order to exclude extremely inappropriate cases and kept the sentences where the entities or their number are the same, such as, Poroshenko asked the head of Turkey not to recognize the presidential elections in the Crimea and Poroshenko urged Erdogan not to recognize the presidential elections in Crimea.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Dataset", | |
| "sec_num": "2.1" | |
| }, | |
| { | |
| "text": "The idea of paraphrase generation is to build a model that reads a sequence of words and then generates a different sequence of words with the same meaning. Paraphrase generation task can be defined as generating a target sentence T for a reference sentence P where the newly generated target sentence T is semantically similar to reference sentence P .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "We chose three pre-trained language models that are available for Russian:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "1. ruGPT2-large 5 is a model by SberDevices team trained as a Russian analogue of Open-AI GPT-2 model (Radford et al., 2019) . GPT-2 is an auto-regressive model, has up-to 1.5 Billion parameters, was trained on 40GB of Internet text to predict the next word. ruGPT2 was trained on 1024 context length with transformers on 170GB data on 64 GPUs 3 weeks.", | |
| "cite_spans": [ | |
| { | |
| "start": 102, | |
| "end": 124, | |
| "text": "(Radford et al., 2019)", | |
| "ref_id": "BIBREF20" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "2. ruGPT3-large is almost analogous to famous GPT-3 (Brown et al., 2020) . ruGPT3 was 4 https://github.com/natasha/natasha 5 https://github.com/sberbank-ai/ru-gpts trained on Internet text on 1024 context length with transformers on 80 billion tokens around 3 epochs, and then was fine-tuned on 2048 context.", | |
| "cite_spans": [ | |
| { | |
| "start": 46, | |
| "end": 72, | |
| "text": "GPT-3 (Brown et al., 2020)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "3. mT5 (Xue et al., 2020) -Multilingual T5 (mT5) by Google is a massively multilingual pre-trained text-to-text transformer model, trained on the mC4 corpus, covering 101 languages including Russian. We trained three mT5 models on the same data: mT5-small, mT5-base and mT5-large.", | |
| "cite_spans": [ | |
| { | |
| "start": 7, | |
| "end": 25, | |
| "text": "(Xue et al., 2020)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "We used Huggingface Transformers Library 6 to fine-tune the models on a sentence reconstruction task to generate paraphrases. Input data for GPTbased models were in the format:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "< s > P i === T i < /s > .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "Input data for mT5 models contained the sequence \"rephrase: \" and looked like the following:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "perephrasiruj : P i < /s > and target format: < s > T i < /s > .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "All the models were trained on a single GPU Tesla V100-SXM3 32 Gb for 3-5 epochs takes 28 minutes per epoch; validation set's perplexity was used to do early stopping.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "Once the model was fine-tuned, it was able to produce paraphrases in a way it was trained. If one fed in any reference phrase with the same sequence token \"===\" or \"rephrase:\", the model generated paraphrases on demand.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Models", | |
| "sec_num": "2.2" | |
| }, | |
| { | |
| "text": "After the model was trained, we sampled from the model from test sentences as conditional input. It allowed to generate different multiple candidate sentences for the single reference sentence. We have tested different parameters (we use the interface of Hugging face, so the parameters are basic for generation: top p , top k sampling parameters, temperature, etc.), but finally used temperature = 1.0, top k = 10, top p = 0.9, maxlength = length(P ) + 10, repetitionpenalty = 1.5 for GPT-based models and temperature = 1.0, top k = 50, top p = 0.95, maxlength = 150, repetitionpenalty = 1.5for mT5 models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "Still the quality of generating multiple outputs varies, and we can select from n examples (where n = 10) the best quality paraphrases based on a number of criteria or one of the range strategies to filter output down to a set of satisfactory results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "We suggest 5 types of the candidate range: 1) cosine sentence similarity between reference sentence and generated one; 2) pairwise cosine sentence similarity between n generated sentences and the reference; 3) syntax based approach; 4) BLEU best; 5) ROUGE-L best.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "The first two strategies are based on sentence similarity scores received with SentenceTransformers 7 . It is a Python framework for state-of-the-art sentence and text embeddings, created based on the initial paper Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (Reimers and Gurevych, 2019) . One can use this framework to compute sentence / text embeddings for more than 100 languages. These embeddings can then be compared with cosine-similarity e.g. to find sentences with a similar meaning. We used paraphrase-xlmr-multilingual-v1 Thakur et al., 2020) model for paraphrase identification task (paraphrase mining). It is a multilingual version (including Russian) of distilrobertabase-paraphrase-v1 (multilingual knowledge distilled version of multilingual Universal Sentence Encoder), trained on parallel data for 50+ languages. In the first strategy we ranged pairwise n candidates comparing cosine sentence similarities between a reference sentence and the generated one and chose the best ones (or set a distance threshold from which we are confident it is a good paraphrase).", | |
| "cite_spans": [ | |
| { | |
| "start": 278, | |
| "end": 306, | |
| "text": "(Reimers and Gurevych, 2019)", | |
| "ref_id": "BIBREF21" | |
| }, | |
| { | |
| "start": 551, | |
| "end": 571, | |
| "text": "Thakur et al., 2020)", | |
| "ref_id": "BIBREF26" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "In the second strategy we used paraphrase mining 8 , the task of finding paraphrases in a corpus of sentences. The framework allows to find paraphrases in a list of sentences. Thus, we input all generated sentences and the reference one and model outputs the paraphrases.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "The syntax based approach is based on the idea that arguments in the reference sentences and in the target sentence will be the same. Thus we can count the number of syntactic subjects and repeated tokens in both sentences and range the most coincidental ones. For syntax parsing we used Deeppavlov Bert Syntargus model 9 . The final two range strategies were using ROUGE-L scores and BLEU pairwise scores to choose the best from n candidates. We eliminated candidates with scores more than 0.9 and less than 0.3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "The most stable range strategy in our experiments was the first one -cosine sentence similarity between a reference sentence and the generated candidate.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Candidate range", | |
| "sec_num": "2.3" | |
| }, | |
| { | |
| "text": "We propose a two-step evaluation procedure: 1) universal metrics between gold testset examples and generated models outputs and 2) application of paraphrasers on downstream tasks where we augment data. Additionally, we will discuss the quality of the paraphrases evaluated by humans on subset of examples.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluation", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "To measure the quality of paraphrase generation we used average ROUGE-L, BLEU-n metrics and average cosine distance between reference and generated (calculated with model for paraphrase identification task) sentences. BLEU-n calculates n gram overlap (unigrams, bigrams and trigrams), ROUGE-L measures the longest matching sequence.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Results", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "Thus, we first ranged candidates as described in section 2.3, counted average scores between them for each example in the testset and got average scores for all testset examples. The results of the models are presented in Table 3 . It is worth mentioning that in the process of ranging candidates we eliminated examples that were very similar by Levenshtein distance with the reference sentence (the cases when paraphraser changes case or adds punctuation symbols are not what we want).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 222, | |
| "end": 229, | |
| "text": "Table 3", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Results", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "We can observe that the golden set results have the highest scores, still the average results of the filtered models are high. The best results by ROUGE and BLEU scores are demonstrated by the mT5small model, however it is interesting that the mt5base and large models scores are lower, while the average candidates cosine similarity in these models is higher. It is due to the fact that if we explore generated sequences we find out that the mT5 model generates sequences that do not have great variability. For example, it is likely to generate sentences that differ only in punctuation symbols or prepositions from the reference sentence. In other words, the metrics of average cosine similarity is more reliable when paraphrases are expected to be more diverse. Thus, in order to choose the best model one need to pay attention to the metrics which are more appropriate for one's task.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Results", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "The range step of the candidates is of a great importance. We took the results of mT5-small and gpt3 models. In Table 2 we present the scores depending on the different range strategies. One can see that the results vary a lot. Without filtration GPT-based model performed much worse by all the metrics. The mT5 model after filtration had even higher scores by BLEU and ROUGE metrics, but they decreased with average cosine similarity. Therefore, depending on the model and the result one expects, the range strategy should be different.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 112, | |
| "end": 119, | |
| "text": "Table 2", | |
| "ref_id": "TABREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Results", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "In addition to general metrics, we tested if augmenting the training data with the use of paraphrasers could help to improve the performance of the model on the down-stream tasks. For this purpose we applied fine-tuned models to paraphrase examples in the training samples and, thus, augmenting the training data.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "To demonstrate how paraphrases perform with default parameters on a down-stream task, we chose the following datasets:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "1. RuSentiment 10 (Rogers et al., 2018) -dataset for sentiment analysis of social media posts in Russian.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "2. TERRa (Shavrina et al., 2020) -Textual Entailment Recognition for Russian, a part of Russian SuperGLUE benchmark 11 . This task requires to recognize, given two text fragments, whether the meaning of one text is entailed (can be inferred) from the other text. To augment data we paraphrased the premise in each sample, kept the hypothesis and the labels, but shuffled the extended training set.", | |
| "cite_spans": [ | |
| { | |
| "start": 9, | |
| "end": 32, | |
| "text": "(Shavrina et al., 2020)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "3. DaNetQA (Glushkova et al., 2020) -Russian yes/no Question Answering Dataset, a part of Russian SuperGLUE benchmark (Shavrina et al., 2020) . In this dataset we paraphrased only questions and kept the paragraphs in the original format.", | |
| "cite_spans": [ | |
| { | |
| "start": 11, | |
| "end": 35, | |
| "text": "(Glushkova et al., 2020)", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 118, | |
| "end": 141, | |
| "text": "(Shavrina et al., 2020)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "We took mT5-base model with range strategy of pairwise cosine similarity and default parameters. For each task we created a baseline solution as an example of paraphrasers applicability on simple setups for common tasks. For DaNetQA we made simple sequence classification with DeepPavlov/rubert-base-cased embeddings trained 10 epochs. For ruSentiment we used a Logistic regression classifier as a baseline. The tf-idf baseline provided by the organizers was used for TERRa.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "We can see in Table 4 that the results are slightly different for all the tasks. On the TERRa task there was an increase in the performance. However, in ruSentiment we observed decrease of performance on the test set, as well as in DaNetQA, where the quality was almost the same. During the evaluation procedure the performance on training set for all the tasks was increasing. It is worth to mention that we use paraphrasers from the library with the default, same parameters for all three tasks, and even with them the results do not decrease significantly.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 14, | |
| "end": 21, | |
| "text": "Table 4", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The results of the experiment are quite controversial. On the one hand, we did not observe a significant decrease in the performance on the set, on the other hand, we suppose to see increase of performance with larger sizes of the dataset. One of possible explanations for this is that there was no new information in the added training set, the labels and the meaning were the same, which caused better performance during the training stage and possible overfitting. Examples of sentiment data are very short, with specific lexicon, emojis etc., which also could influence the results. DaNetQA dataset assumes YES/NO questions format, while the paraphraser could change the form of the question heavily and decrease the performance. Additionally, we believe that for every downstream task it is essential to choose a model and parameters more appropriate for the data on each step: generation, ranking, and evaluation. However, these hypotheses need further augmentation testing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data augmentation", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The generated sentences are of different quality; all of the fine-tuned models are able to produce appropriate paraphrases, as well as some of them contain extra information, some typos, agreement errors or different meaning. In Figure 2 one can see the best candidates examples for three of the models. Table 5 represents the distribution of three classes: 1) good, 2) bad and 3) paraphrasers with extra information or some grammatical errors. We took 50 examples from testset, generated the paraphrases with each model, took best candidates and manually checked the number of examples for each class. The GPT2 is more stable; GPT3 is tend to produce more diverse paraphrases and add extra information that changes sense or makes it controversial; mT5 model makes more mistakes or instead changes the reference sentence not much.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 229, | |
| "end": 237, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 304, | |
| "end": 311, | |
| "text": "Table 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Discussion", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "The number of inappropriate candidates is significant and the procedure of candidates range and model parameters setup is crucial and should be specific for every task where we want to use paraphrasers generated on large language models. Each model has its own generating traits. For instance, GPT-based models are likely to generate more off-top sentences and the diversity of their answers is high. We also noticed the tendency of GPT-based models to change the quality of generation examples depending on the max length. mT5 models prefer to change sentences in small pieces: change argument in the sentence on its synonym or add/cut more punctuation and symbols. Therefore, mT5-base results are rated higher with BLEU and ROUGE scores, but the examples do not differ much from reference sentences. The suitability of multilingual models to Russian has no doubts, the results are comparable. Additionally mT5-models are much faster in generation than GPT-based.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Discussion", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "We believe the paraphrasers will be useful in many applications, thus we provide the dataset and fine-tuned models in hugging-face format in open source. The library with paraphrasers and some of range strategies for them is also available 12 . We hope everyone can find the perfect Russian paraphraser for oneself.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Discussion", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "Good Extra info/Typos Bad GPT2 70% 17% 13% GPT3 56% 34% 10% mT5-base 63% 21% 16% Table 5 : Human evaluation of paraphrasers performance. All results were scored manually by people. The distribution is presented in percentage.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 81, | |
| "end": 88, | |
| "text": "Table 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Model", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Paraphrase generation with large language models achieves impressive results. Our experiments show that both multilingual and Russian-oriented models are able to quickly learn the task of paraphrasing through fine-tuning training on a prepared Russian set of paraphrase examples. This paper contributions are the corpus of paraphrases, 5 fine-tuned models for the Russian language, comparison of them, range strategies for finding the best candidates, and the open source library in Python for convenient use of the pre-trained paraphrasers.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In future work, we would like to further explore the effectiveness of generated paraphrasers for different augmentation experiments and evaluate the models robustness in terms of reconstruction and generated paraphrases.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "http://paraphraser.ru/download/ 2 https://github.com/rysshe/paraphrase/tree/master/data 3 https://yandex.ru/dev/speller/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://huggingface.co/", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://github.com/UKPLab/sentence-transformers", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://www.sbert.net/examples/applications/paraphrasemining/README.html 9 http://docs.deeppavlov.ai/en/master/features/models/syntaxparser.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://text-machine.cs.uml.edu/projects/rusentiment/ 11 https://russiansuperglue.com/tasks", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "https://github.com/RussianNLP/russianparaphrasers", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Synonymous paraphrasing using wordnet and internet", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Igor", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alexander", | |
| "middle": [], | |
| "last": "Bolshakov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Gelbukh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "International Conference on Application of Natural Language to Information Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "312--323", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Igor A Bolshakov and Alexander Gelbukh. 2004. Syn- onymous paraphrasing using wordnet and internet. In International Conference on Application of Nat- ural Language to Information Systems, pages 312- 323. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Effect of semantic parsing depth on the identification of paraphrases in russian texts", | |
| "authors": [ | |
| { | |
| "first": "Kirill", | |
| "middle": [], | |
| "last": "Boyarsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Eugeni", | |
| "middle": [], | |
| "last": "Kanevsky", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Conference on Artificial Intelligence and Natural Language", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "226--241", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kirill Boyarsky and Eugeni Kanevsky. 2017. Effect of semantic parsing depth on the identification of para- phrases in russian texts. In Conference on Artificial Intelligence and Natural Language, pages 226-241. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Language models are few-shot learners", | |
| "authors": [ | |
| { | |
| "first": "Benjamin", | |
| "middle": [], | |
| "last": "Tom B Brown", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nick", | |
| "middle": [], | |
| "last": "Mann", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Melanie", | |
| "middle": [], | |
| "last": "Ryder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jared", | |
| "middle": [], | |
| "last": "Subbiah", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Prafulla", | |
| "middle": [], | |
| "last": "Kaplan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Arvind", | |
| "middle": [], | |
| "last": "Dhariwal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Pranav", | |
| "middle": [], | |
| "last": "Neelakantan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Girish", | |
| "middle": [], | |
| "last": "Shyam", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Amanda", | |
| "middle": [], | |
| "last": "Sastry", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Askell", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2005.14165" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Danetqa: a yes/no question answering dataset for the russian language", | |
| "authors": [ | |
| { | |
| "first": "Taisia", | |
| "middle": [], | |
| "last": "Glushkova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alexey", | |
| "middle": [], | |
| "last": "Machnev", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alena", | |
| "middle": [], | |
| "last": "Fenogenova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tatiana", | |
| "middle": [], | |
| "last": "Shavrina", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ekaterina", | |
| "middle": [], | |
| "last": "Artemova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dmitry I", | |
| "middle": [], | |
| "last": "Ignatov", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2010.02605" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Taisia Glushkova, Alexey Machnev, Alena Fenogen- ova, Tatiana Shavrina, Ekaterina Artemova, and Dmitry I Ignatov. 2020. Danetqa: a yes/no question answering dataset for the russian language. arXiv preprint arXiv:2010.02605.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Automatically ranked russian paraphrase corpus for text generation", | |
| "authors": [ | |
| { | |
| "first": "Vadim", | |
| "middle": [], | |
| "last": "Gudkov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Olga", | |
| "middle": [], | |
| "last": "Mitrofanova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Elizaveta", | |
| "middle": [], | |
| "last": "Filippskikh", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2006.09719" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Vadim Gudkov, Olga Mitrofanova, and Elizaveta Fil- ippskikh. 2020. Automatically ranked russian para- phrase corpus for text generation. arXiv preprint arXiv:2006.09719.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "A deep generative framework for paraphrase generation", | |
| "authors": [ | |
| { | |
| "first": "Ankush", | |
| "middle": [], | |
| "last": "Gupta", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Arvind", | |
| "middle": [], | |
| "last": "Agarwal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Prawaan", | |
| "middle": [], | |
| "last": "Singh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Piyush", | |
| "middle": [], | |
| "last": "Rai", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the AAAI Conference on Artificial Intelligence", | |
| "volume": "32", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ankush Gupta, Arvind Agarwal, Prawaan Singh, and Piyush Rai. 2018. A deep generative framework for paraphrase generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Unsupervised paraphrase generation using pre-trained language models", | |
| "authors": [ | |
| { | |
| "first": "Chaitra", | |
| "middle": [], | |
| "last": "Hegde", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shrikumar", | |
| "middle": [], | |
| "last": "Patil", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2006.05477" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chaitra Hegde and Shrikumar Patil. 2020. Unsuper- vised paraphrase generation using pre-trained lan- guage models. arXiv preprint arXiv:2006.05477.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Paraphrase detection using machine translation and textual similarity algorithms", | |
| "authors": [ | |
| { | |
| "first": "Dmitry", | |
| "middle": [], | |
| "last": "Kravchenko", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Conference on artificial intelligence and natural language", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "277--292", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dmitry Kravchenko. 2017. Paraphrase detection us- ing machine translation and textual similarity algo- rithms. In Conference on artificial intelligence and natural language, pages 277-292. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Adaptation of deep bidirectional multilingual transformers for russian language", | |
| "authors": [ | |
| { | |
| "first": "Yuri", | |
| "middle": [], | |
| "last": "Kuratov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mikhail", | |
| "middle": [], | |
| "last": "Arkhipov", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1905.07213" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yuri Kuratov and Mikhail Arkhipov. 2019. Adaptation of deep bidirectional multilingual transformers for russian language. arXiv preprint arXiv:1905.07213.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Pre-training via paraphrasing", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2006.15020" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Pre-training via paraphrasing. arXiv preprint arXiv:2006.15020.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Rouge: A package for automatic evaluation of summaries", | |
| "authors": [ | |
| { | |
| "first": "Chin-Yew", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Text summarization branches out", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "74--81", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Diversifying task-oriented dialogue response generation with prototype guided paraphrasing", | |
| "authors": [ | |
| { | |
| "first": "Phillip", | |
| "middle": [], | |
| "last": "Lippe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Pengjie", | |
| "middle": [], | |
| "last": "Ren", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hinda", | |
| "middle": [], | |
| "last": "Haned", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bart", | |
| "middle": [], | |
| "last": "Voorn", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Maarten", | |
| "middle": [], | |
| "last": "De Rijke", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2008.03391" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Phillip Lippe, Pengjie Ren, Hinda Haned, Bart Voorn, and Maarten de Rijke. 2020. Diversify- ing task-oriented dialogue response generation with prototype guided paraphrasing. arXiv preprint arXiv:2008.03391.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Generating phrasal and sentential paraphrases: A survey of data-driven methods", | |
| "authors": [ | |
| { | |
| "first": "Nitin", | |
| "middle": [], | |
| "last": "Madnani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bonnie", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Dorr", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Computational Linguistics", | |
| "volume": "36", | |
| "issue": "3", | |
| "pages": "341--387", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nitin Madnani and Bonnie J Dorr. 2010. Generat- ing phrasal and sentential paraphrases: A survey of data-driven methods. Computational Linguistics, 36(3):341-387.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Figure 2: Generated examples for sentence Bolshoi Theater violinist dies after falling into orchestra pit. + is a good generated paraphrase, -is not appropriate paraphrase, extra info means the generally good paraphrase but it contains more information than the reference ones", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Figure 2: Generated examples for sentence Bolshoi Theater violinist dies after falling into orchestra pit. + is a good generated paraphrase, -is not appropriate paraphrase, extra info means the generally good paraphrase but it contains more information than the reference ones.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Strategies for effective paraphrasing", | |
| "authors": [ | |
| { | |
| "first": "Marie", | |
| "middle": [], | |
| "last": "Meteer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Varda", | |
| "middle": [], | |
| "last": "Shaked", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1988, | |
| "venue": "International Conference on Computational Linguistics", | |
| "volume": "2", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Marie Meteer and Varda Shaked. 1988. Strategies for effective paraphrasing. In Coling Budapest 1988 Volume 2: International Conference on Computa- tional Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Nitish Shirish Keskar, and Caiming Xiong. 2020. Unsupervised paraphrase generation via dynamic blocking", | |
| "authors": [ | |
| { | |
| "first": "Tong", | |
| "middle": [], | |
| "last": "Niu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Semih", | |
| "middle": [], | |
| "last": "Yavuz", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yingbo", | |
| "middle": [], | |
| "last": "Zhou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Huan", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2010.12885" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tong Niu, Semih Yavuz, Yingbo Zhou, Huan Wang, Nitish Shirish Keskar, and Caiming Xiong. 2020. Unsupervised paraphrase generation via dynamic blocking. arXiv preprint arXiv:2010.12885.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Paraphraser: Russian paraphrase corpus and shared task", | |
| "authors": [ | |
| { | |
| "first": "Lidia", | |
| "middle": [], | |
| "last": "Pivovarova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ekaterina", | |
| "middle": [], | |
| "last": "Pronoza", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Elena", | |
| "middle": [], | |
| "last": "Yagunova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anton", | |
| "middle": [], | |
| "last": "Pronoza", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Conference on Artificial Intelligence and Natural Language", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "211--225", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Lidia Pivovarova, Ekaterina Pronoza, Elena Yagunova, and Anton Pronoza. 2017. Paraphraser: Russian paraphrase corpus and shared task. In Conference on Artificial Intelligence and Natural Language, pages 211-225. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Construction of a russian paraphrase corpus: unsupervised paraphrase extraction", | |
| "authors": [ | |
| { | |
| "first": "Ekaterina", | |
| "middle": [], | |
| "last": "Pronoza", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Elena", | |
| "middle": [], | |
| "last": "Yagunova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anton", | |
| "middle": [], | |
| "last": "Pronoza", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Russian Summer School in Information Retrieval", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "146--157", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ekaterina Pronoza, Elena Yagunova, and Anton Pronoza. 2015. Construction of a russian paraphrase corpus: unsupervised paraphrase extraction. In Rus- sian Summer School in Information Retrieval, pages 146-157. Springer.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Exploring diverse expressions for paraphrase generation", | |
| "authors": [ | |
| { | |
| "first": "Lihua", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Lin", | |
| "middle": [], | |
| "last": "Qiu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Weinan", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Xin", | |
| "middle": [], | |
| "last": "Jiang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yong", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3164--3173", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Lihua Qian, Lin Qiu, Weinan Zhang, Xin Jiang, and Yong Yu. 2019. Exploring diverse expressions for paraphrase generation. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3164-3173.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Language models are unsupervised multitask learners", | |
| "authors": [ | |
| { | |
| "first": "Alec", | |
| "middle": [], | |
| "last": "Radford", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Rewon", | |
| "middle": [], | |
| "last": "Child", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "David", | |
| "middle": [], | |
| "last": "Luan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dario", | |
| "middle": [], | |
| "last": "Amodei", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ilya", | |
| "middle": [], | |
| "last": "Sutskever", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "OpenAI blog", | |
| "volume": "1", | |
| "issue": "8", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.", | |
| "links": null | |
| }, | |
| "BIBREF21": { | |
| "ref_id": "b21", | |
| "title": "Sentencebert: Sentence embeddings using siamese bertnetworks", | |
| "authors": [ | |
| { | |
| "first": "Nils", | |
| "middle": [], | |
| "last": "Reimers", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Iryna", | |
| "middle": [], | |
| "last": "Gurevych", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nils Reimers and Iryna Gurevych. 2019. Sentence- bert: Sentence embeddings using siamese bert- networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF22": { | |
| "ref_id": "b22", | |
| "title": "Making monolingual sentence embeddings multilingual using knowledge distillation", | |
| "authors": [ | |
| { | |
| "first": "Nils", | |
| "middle": [], | |
| "last": "Reimers", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Iryna", | |
| "middle": [], | |
| "last": "Gurevych", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nils Reimers and Iryna Gurevych. 2020. Making monolingual sentence embeddings multilingual us- ing knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natu- ral Language Processing. Association for Computa- tional Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF23": { | |
| "ref_id": "b23", | |
| "title": "Rusentiment: An enriched sentiment analysis dataset for social media in russian", | |
| "authors": [ | |
| { | |
| "first": "Anna", | |
| "middle": [], | |
| "last": "Rogers", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alexey", | |
| "middle": [], | |
| "last": "Romanov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anna", | |
| "middle": [], | |
| "last": "Rumshisky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Svitlana", | |
| "middle": [], | |
| "last": "Volkova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mikhail", | |
| "middle": [], | |
| "last": "Gronas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alex", | |
| "middle": [], | |
| "last": "Gribov", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 27th international conference on computational linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "755--763", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Anna Rogers, Alexey Romanov, Anna Rumshisky, Svitlana Volkova, Mikhail Gronas, and Alex Gribov. 2018. Rusentiment: An enriched sentiment analy- sis dataset for social media in russian. In Proceed- ings of the 27th international conference on compu- tational linguistics, pages 755-763.", | |
| "links": null | |
| }, | |
| "BIBREF24": { | |
| "ref_id": "b24", | |
| "title": "Andrey Chertok, and Andrey Evlampiev. 2020. Russiansuperglue: A russian language understanding evaluation benchmark", | |
| "authors": [ | |
| { | |
| "first": "Tatiana", | |
| "middle": [], | |
| "last": "Shavrina", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alena", | |
| "middle": [], | |
| "last": "Fenogenova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anton", | |
| "middle": [], | |
| "last": "Emelyanov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Denis", | |
| "middle": [], | |
| "last": "Shevelev", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ekaterina", | |
| "middle": [], | |
| "last": "Artemova", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Valentin", | |
| "middle": [], | |
| "last": "Malykh", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Vladislav", | |
| "middle": [], | |
| "last": "Mikhailov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Maria", | |
| "middle": [], | |
| "last": "Tikhonova", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2010.15925" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Tatiana Shavrina, Alena Fenogenova, Anton Emelyanov, Denis Shevelev, Ekaterina Arte- mova, Valentin Malykh, Vladislav Mikhailov, Maria Tikhonova, Andrey Chertok, and Andrey Evlampiev. 2020. Russiansuperglue: A russian language un- derstanding evaluation benchmark. arXiv preprint arXiv:2010.15925.", | |
| "links": null | |
| }, | |
| "BIBREF25": { | |
| "ref_id": "b25", | |
| "title": "Unsupervised paraphrasing via deep reinforcement learning", | |
| "authors": [ | |
| { | |
| "first": "Samet", | |
| "middle": [], | |
| "last": "Ab Siddique", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Vagelis", | |
| "middle": [], | |
| "last": "Oymak", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Hristidis", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1800--1809", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "AB Siddique, Samet Oymak, and Vagelis Hristidis. 2020. Unsupervised paraphrasing via deep rein- forcement learning. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1800-1809.", | |
| "links": null | |
| }, | |
| "BIBREF26": { | |
| "ref_id": "b26", | |
| "title": "Augmented sbert: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks", | |
| "authors": [ | |
| { | |
| "first": "Nandan", | |
| "middle": [], | |
| "last": "Thakur", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nils", | |
| "middle": [], | |
| "last": "Reimers", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Johannes", | |
| "middle": [], | |
| "last": "Daxenberger", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Iryna", | |
| "middle": [], | |
| "last": "Gurevych", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2010.08240" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. 2020. Augmented sbert: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. arXiv preprint arXiv:2010.08240.", | |
| "links": null | |
| }, | |
| "BIBREF27": { | |
| "ref_id": "b27", | |
| "title": "Paraphrasing with large language models", | |
| "authors": [ | |
| { | |
| "first": "Sam", | |
| "middle": [], | |
| "last": "Witteveen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Martin", | |
| "middle": [], | |
| "last": "Andrews", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:1911.09661" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sam Witteveen and Martin Andrews. 2019. Paraphras- ing with large language models. arXiv preprint arXiv:1911.09661.", | |
| "links": null | |
| }, | |
| "BIBREF28": { | |
| "ref_id": "b28", | |
| "title": "Autoqa: From databases to qa semantic parsers with only synthetic training data", | |
| "authors": [ | |
| { | |
| "first": "Silei", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Sina", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Giovanni", | |
| "middle": [], | |
| "last": "Semnani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Monica", | |
| "middle": [ | |
| "S" | |
| ], | |
| "last": "Campagna", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Lam", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2010.04806" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Silei Xu, Sina J Semnani, Giovanni Campagna, and Monica S Lam. 2020. Autoqa: From databases to qa semantic parsers with only synthetic training data. arXiv preprint arXiv:2010.04806.", | |
| "links": null | |
| }, | |
| "BIBREF29": { | |
| "ref_id": "b29", | |
| "title": "Paraphrasing for style", | |
| "authors": [ | |
| { | |
| "first": "Wei", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alan", | |
| "middle": [], | |
| "last": "Ritter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bill", | |
| "middle": [], | |
| "last": "Dolan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ralph", | |
| "middle": [], | |
| "last": "Grishman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Colin", | |
| "middle": [], | |
| "last": "Cherry", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of COLING 2012", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "2899--2914", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Wei Xu, Alan Ritter, Bill Dolan, Ralph Grishman, and Colin Cherry. 2012. Paraphrasing for style. In Pro- ceedings of COLING 2012, pages 2899-2914.", | |
| "links": null | |
| }, | |
| "BIBREF30": { | |
| "ref_id": "b30", | |
| "title": "Aditya Barua, and Colin Raffel. 2020. mt5: A massively multilingual pre-trained text-to-text transformer", | |
| "authors": [ | |
| { | |
| "first": "Linting", | |
| "middle": [], | |
| "last": "Xue", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Noah", | |
| "middle": [], | |
| "last": "Constant", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Adam", | |
| "middle": [], | |
| "last": "Roberts", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mihir", | |
| "middle": [], | |
| "last": "Kale", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Rami", | |
| "middle": [], | |
| "last": "Al-Rfou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Aditya", | |
| "middle": [], | |
| "last": "Siddhant", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "arXiv": [ | |
| "arXiv:2010.11934" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Linting Xue, Noah Constant, Adam Roberts, Mi- hir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mt5: A mas- sively multilingual pre-trained text-to-text trans- former. arXiv preprint arXiv:2010.11934.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "Dialogues data example" | |
| }, | |
| "TABREF1": { | |
| "num": null, | |
| "content": "<table/>", | |
| "type_str": "table", | |
| "text": "The dataset statistics and distribution.", | |
| "html": null | |
| }, | |
| "TABREF3": { | |
| "num": null, | |
| "content": "<table><tr><td colspan=\"6\">Paraphraser Cosine similarity BLEU-1 BLEU-2 BLEU-3 ROUGE-L</td></tr><tr><td>golden set</td><td>0.848</td><td>0.57</td><td>0.43</td><td>0.28</td><td>0.58</td></tr><tr><td>mT5-small</td><td>0.781</td><td>0.49</td><td>0.35</td><td>0.21</td><td>0.49</td></tr><tr><td>mT5-base</td><td>0.798</td><td>0.35</td><td>0.23</td><td>0.12</td><td>0.37</td></tr><tr><td>mT5-large</td><td>0.802</td><td>0.40</td><td>0.25</td><td>0.12</td><td>0.41</td></tr><tr><td>rugpt2</td><td>0.717</td><td>0.43</td><td>0.29</td><td>0.17</td><td>0.44</td></tr><tr><td>rugpt3</td><td>0.754</td><td>0.41</td><td>0.27</td><td>0.15</td><td>0.42</td></tr></table>", | |
| "type_str": "table", | |
| "text": "Scores of the mT5-small model and rugpt3 with different range strategies.", | |
| "html": null | |
| }, | |
| "TABREF4": { | |
| "num": null, | |
| "content": "<table><tr><td>Dataset</td><td>Orig Aug</td><td>+examples</td></tr><tr><td colspan=\"2\">DaNetQA 0.621 0.62</td><td>1750</td></tr><tr><td>ruSent</td><td colspan=\"2\">0.674 0.666 5550</td></tr><tr><td>TERRa</td><td colspan=\"2\">0.471 0.475 5600</td></tr></table>", | |
| "type_str": "table", | |
| "text": "Scores of the models average", | |
| "html": null | |
| }, | |
| "TABREF5": { | |
| "num": null, | |
| "content": "<table><tr><td>: Augmentation results on test sets of the mT5-</td></tr><tr><td>base model and number of generated examples that</td></tr><tr><td>were added.</td></tr></table>", | |
| "type_str": "table", | |
| "text": "", | |
| "html": null | |
| } | |
| } | |
| } | |
| } |