Computer Science

Paraphrase generation plays an essential role in natural language process (NLP), and it has many downstream applications. However, training supervised paraphrase models requires many annotated paraphrase pairs, which are usually costly to obtain. On the other hand, the paraphrases generated by existing unsupervised approaches are usually syntactically similar to the source sentences and are limited in diversity. In this paper, we demonstrate that it is possible to generate syntactically various paraphrases without the need for annotated paraphrase pairs. We propose Syntactically controlled Paraphrase Generator (SynPG), an encoder-decoder based model that learns to disentangle the semantics and the syntax of a sentence from a collection of unannotated texts. The disentanglement enables SynPG to control the syntax of output paraphrases by manipulating the embedding in the syntactic space. Extensive experiments using automatic metrics and human evaluation show that SynPG performs better syntactic control than unsupervised baselines, while the quality of the generated paraphrases is competitive. We also demonstrate that the performance of SynPG is competitive or even better than supervised models when the unannotated data is large. Finally, we show that the syntactically controlled paraphrases generated by SynPG can be utilized for data augmentation to improve the robustness of NLP models.

Drawing causal conclusions from observational data requires making assumptions about the true data-generating process. Causal inference research typically considers low-dimensional data, such as categorical or numerical fields in structured medical records. High-dimensional and unstructured data such as natural language complicates the evaluation of causal inference methods; such evaluations rely on synthetic datasets with known causal effects. Models for natural language generation have been widely studied and perform well empirically. However, existing methods not immediately applicable to producing synthetic datasets for causal evaluations, as they do not allow for quantifying a causal effect on the text itself. In this work, we develop a framework for adapting existing generation models to produce synthetic text datasets with known causal effects. We use this framework to perform an empirical comparison of four recently-proposed methods for estimating causal effects from text data. We release our code and synthetic datasets.

Generative spoken language modeling involves learning jointly the acoustic and linguistic characteristics of a language from raw audio only (without text or labels). We introduce metrics to automatically evaluate the generated output in terms of acoustic and linguistic quality in two associated end-to-end tasks, respectively: speech resynthesis (repeating the speech input using the system's own voice), and speech generation (producing novel speech outputs conditional on a spoken prompt, or unconditionally), and validate these metrics with human judgment. We test baseline systems consisting of a discrete speech encoder (returning discrete, low bitrate, pseudo-text units), a generative language model (trained on pseudo-text units), and a speech decoder (generating a waveform from pseudo-text). By comparing three state-of-the-art unsupervised speech encoders (Contrastive Predictive Coding (CPC), wav2vec 2.0, HuBERT), and varying the number of discrete units (50, 100, 200), we investigate how the generative performance depends on the quality of the learned units as measured by unsupervised metrics (zero-shot probe tasks). We will open source our evaluation stack and baseline models.

We address the challenging problem of Natural Language Comprehension beyond plain-text documents by introducing the TILT neural network architecture which simultaneously learns layout information, visual features, and textual semantics. Contrary to previous approaches, we rely on a decoder capable of unifying a variety of problems involving natural language. The layout is represented as an attention bias and complemented with contextualized visual information, while the core of our model is a pretrained encoder-decoder Transformer. Our novel approach achieves state-of-the-art results in extracting information from documents and answering questions which demand layout understanding (DocVQA, CORD, WikiOps, SROIE). At the same time, we simplify the process by employing an end-to-end model.

To mitigate the negative effect of low quality training data on the performance of neural machine translation models, most existing strategies focus on filtering out harmful data before training starts. In this paper, we explore strategies that dynamically optimize data usage during the training process using the model's gradients on a small set of clean data. At each training step, our algorithm calculates the gradient alignment between the training data and the clean data to mask out data with negative alignment. Our method has a natural intuition: good training data should update the model parameters in a similar direction as the clean data. Experiments on three WMT language pairs show that our method brings significant improvement over strong baselines, and the improvements are generalizable across test data from different domains.

Graph structures are powerful tools for modeling the relationships between textual elements. Graph-of-Words (GoW) has been adopted in many Natural Language tasks to encode the association between terms. However, GoW provides few document-level relationships in cases when the connections between documents are also essential. For identifying sub-events on social media like Twitter, features from both word- and document-level can be useful as they supply different information of the event. We propose a hybrid Graph-of-Tweets (GoT) model which combines the word- and document-level structures for modeling Tweets. To compress large amount of raw data, we propose a graph merging method which utilizes FastText word embeddings to reduce the GoW. Furthermore, we present a novel method to construct GoT with the reduced GoW and a Mutual Information (MI) measure. Finally, we identify maximal cliques to extract popular sub-events. Our model showed promising results on condensing lexical-level information and capturing keywords of sub-events.

Story generation is a task that aims to automatically produce multiple sentences to make up a meaningful story. This task is challenging because it requires high-level understanding of semantic meaning of sentences and causality of story events. Naive sequence-to-sequence models generally fail to acquire such knowledge, as the logical correctness can hardly be guaranteed in a text generation model without the strategic planning. In this paper, we focus on planning a sequence of events assisted by event graphs, and use the events to guide the generator. Instead of using a sequence-to-sequence model to output a storyline as in some existing works, we propose to generate an event sequence by walking on an event graph. The event graphs are built automatically based on the corpus. To evaluate the proposed approach, we conduct human evaluation both on event planning and story generation. Based on large-scale human annotation results, our proposed approach is shown to produce more logically correct event sequences and stories.

Social media often acts as breeding grounds for different forms of offensive content. For low resource languages like Tamil, the situation is more complex due to the poor performance of multilingual or language-specific models and lack of proper benchmark datasets. Based on this shared task, Offensive Language Identification in Dravidian Languages at EACL 2021, we present an exhaustive exploration of different transformer models, We also provide a genetic algorithm technique for ensembling different models. Our ensembled models trained separately for each language secured the first position in Tamil, the second position in Kannada, and the first position in Malayalam sub-tasks. The models and codes are provided.

With the success of pre-trained language models in recent years, more and more researchers focus on opening the "black box" of these models. Following this interest, we carry out a qualitative and quantitative analysis of constituency grammar in attention heads of BERT and RoBERTa. We employ the syntactic distance method to extract implicit constituency grammar from the attention weights of each head. Our results show that there exist heads that can induce some grammar types much better than baselines, suggesting that some heads act as a proxy for constituency grammar. We also analyze how attention heads' constituency grammar inducing (CGI) ability changes after fine-tuning with two kinds of tasks, including sentence meaning similarity (SMS) tasks and natural language inference (NLI) tasks. Our results suggest that SMS tasks decrease the average CGI ability of upper layers, while NLI tasks increase it. Lastly, we investigate the connections between CGI ability and natural language understanding ability on QQP and MNLI tasks.

This paper introduces HeBERT and HebEMO. HeBERT is a Transformer-based model for modern Hebrew text, which relies on a BERT (Bidirectional Encoder Representations for Transformers) architecture. BERT has been shown to outperform alternative architectures in sentiment analysis, and is suggested to be particularly appropriate for MRLs. Analyzing multiple BERT specifications, we find that while model complexity correlates with high performance on language tasks that aim to understand terms in a sentence, a more-parsimonious model better captures the sentiment of entire sentence. Either way, out BERT-based language model outperforms all existing Hebrew alternatives on all common language tasks. HebEMO is a tool that uses HeBERT to detect polarity and extract emotions from Hebrew UGC. HebEMO is trained on a unique Covid-19-related UGC dataset that we collected and annotated for this study. Data collection and annotation followed an active learning procedure that aimed to maximize predictability. We show that HebEMO yields a high F1-score of 0.96 for polarity classification. Emotion detection reaches F1-scores of 0.78-0.97 for various target emotions, with the exception of surprise, which the model failed to capture (F1 = 0.41). These results are better than the best-reported performance, even among English-language models of emotion detection.