We propose Generation-Augmented Retrieval (GAR) for answering open-domain questions, which augments a query through text generation of heuristically discovered relevant contexts without external resources as supervision. We demonstrate that the generated contexts substantially enrich the semantics of the queries and GAR with sparse representations (BM25) achieves comparable or better performance than state-of-the-art dense retrieval methods such as DPR. We show that generating diverse contexts for a query is beneficial as fusing their results consistently yields better retrieval accuracy. Moreover, as sparse and dense representations are often complementary, GAR can be easily combined with DPR to achieve even better performance. GAR achieves state-of-the-art performance on Natural Questions and TriviaQA datasets under the extractive QA setup when equipped with an extractive reader, and consistently outperforms other retrieval methods when the same generative reader is used.
Current approaches to text generation largely rely on autoregressive models and maximum likelihood estimation. This paradigm leads to (i) diverse but low-quality samples due to mismatched learning objective and evaluation metric (likelihood vs. quality) and (ii) exposure bias due to mismatched history distributions (gold vs. model-generated). To alleviate these problems, we frame text generation as an offline reinforcement learning (RL) problem with expert demonstrations (i.e., the reference), where the goal is to maximize quality given model-generated histories. We propose GOLD (generation by off-policy learning from demonstrations): an easy-to-optimize algorithm that learns from the demonstrations by importance weighting. Intuitively, GOLD upweights confident tokens and downweights unconfident ones in the reference during training, avoiding optimization issues faced by prior RL approaches that rely on online data collection. According to both automatic and human evaluation, models trained by GOLD outperform those trained by MLE and policy gradient on summarization, question generation, and machine translation. Further, our models are less sensitive to decoding algorithms and alleviate exposure bias.
In recent years, the use of deep learning in language models gained much attention. Some research projects claim that they can generate text that can be interpreted as human-writing, enabling new possibilities in many application areas. Among the different areas related to language processing, one of the most notable in applying this type of modeling is programming languages. For years, the Machine Learning community has been researching this software engineering area, pursuing goals like applying different approaches to auto-complete, generate, fix, or evaluate code programmed by humans. Considering the increasing popularity of the Deep-Learning-enabled language models approach, we detected a lack of empirical papers that compare different deep learning architectures to create and use language models based on programming code. This paper compares different neural network architectures like AWD-LSTMs, AWD-QRNNs, and Transformer while using transfer learning and different tokenizations to see how they behave in building language models using a Python dataset for code generation and filling mask tasks. Considering the results, we discuss each approach's different strengths and weaknesses and what gaps we find to evaluate the language models or apply them in a real programming context.
Despite the success of existing referenced metrics (e.g., BLEU and MoverScore), they correlate poorly with human judgments for open-ended text generation including story or dialog generation because of the notorious one-to-many issue: there are many plausible outputs for the same input, which may differ substantially in literal or semantics from the limited number of given references. To alleviate this issue, we propose UNION, a learnable unreferenced metric for evaluating open-ended story generation, which measures the quality of a generated story without any reference. Built on top of BERT, UNION is trained to distinguish human-written stories from negative samples and recover the perturbation in negative stories. We propose an approach of constructing negative samples by mimicking the errors commonly observed in existing NLG models, including repeated plots, conflicting logic, and long-range incoherence. Experiments on two story datasets demonstrate that UNION is a reliable measure for evaluating the quality of generated stories, which correlates better with human judgments and is more generalizable than existing state-of-the-art metrics.
The performance of autoregressive models on natural language generation tasks has dramatically improved due to the adoption of deep, self-attentive architectures. However, these gains have come at the cost of hindering inference speed, making state-of-the-art models cumbersome to deploy in real-world, time-sensitive settings. We develop a compression technique for autoregressive models that is driven by an imitation learning perspective on knowledge distillation. The algorithm is designed to address the exposure bias problem. On prototypical language generation tasks such as translation and summarization, our method consistently outperforms other distillation algorithms, such as sequence-level knowledge distillation. Student models trained with our method attain 1.4 to 4.8 BLEU/ROUGE points higher than those trained from scratch, while increasing inference speed by up to 14 times in comparison to the teacher model.
This work studies the widely adopted ancestral sampling algorithms for auto-regressive language models, which is not widely studied in the literature. We use the quality-diversity (Q-D) trade-off to investigate three popular sampling algorithms (top-k, nucleus and tempered sampling). We focus on the task of open-ended language generation. We first show that the existing sampling algorithms have similar performance. After carefully inspecting the transformations defined by different sampling algorithms, we identify three key properties that are shared among them: entropy reduction, order preservation, and slope preservation. To validate the importance of the identified properties, we design two sets of new sampling algorithms: one set in which each algorithm satisfies all three properties, and one set in which each algorithm violates at least one of the properties. We compare their performance with existing sampling algorithms, and find that violating the identified properties could lead to drastic performance degradation, as measured by the Q-D trade-off. On the other hand, we find that the set of sampling algorithms that satisfies these properties performs on par with the existing sampling algorithms. Our data and code are available at https://github.com/moinnadeem/characterizing-sampling-algorithms
Natural language generation provides designers with methods for automatically generating text, e.g. for creating summaries, chatbots and game content. In practise, text generators are often either learned and hard to interpret, or created by hand using techniques such as grammars and templates. In this paper, we introduce a novel grammar induction algorithm for learning interpretable grammars for generative purposes, called Gitta. We also introduce the novel notion of template trees to discover latent templates in corpora to derive these generative grammars. By using existing human-created grammars, we found that the algorithm can reasonably approximate these grammars using only a few examples. These results indicate that Gitta could be used to automatically learn interpretable and easily modifiable grammars, and thus provide a stepping stone for human-machine co-creation of generative models.
Brain decoding, understood as the process of mapping brain activities to the stimuli that generated them, has been an active research area in the last years. In the case of language stimuli, recent studies have shown that it is possible to decode fMRI scans into an embedding of the word a subject is reading. However, such word embeddings are designed for natural language processing tasks rather than for brain decoding. Therefore, they limit our ability to recover the precise stimulus. In this work, we propose to directly classify an fMRI scan, mapping it to the corresponding word within a fixed vocabulary. Unlike existing work, we evaluate on scans from previously unseen subjects. We argue that this is a more realistic setup and we present a model that can decode fMRI data from unseen subjects. Our model achieves 5.22% Top-1 and 13.59% Top-5 accuracy in this challenging task, significantly outperforming all the considered competitive baselines. Furthermore, we use the decoded words to guide language generation with the GPT-2 model. This way, we advance the quest for a system that translates brain activities into coherent text.
Plug-and-play language models (PPLMs) enable topic-conditioned natural language generation by pairing large pre-trained generators with attribute models used to steer the predicted token distribution towards the selected topic. Despite their computational efficiency, PPLMs require large amounts of labeled texts to effectively balance generation fluency and proper conditioning, making them unsuitable for low-resource settings. We present ETC-NLG, an approach leveraging topic modeling annotations to enable fully-unsupervised End-to-end Topic-Conditioned Natural Language Generation over emergent topics in unlabeled document collections. We first test the effectiveness of our approach in a low-resource setting for Italian, evaluating the conditioning for both topic models and gold annotations. We then perform a comparative evaluation of ETC-NLG for Italian and English using a parallel corpus. Finally, we propose an automatic approach to estimate the effectiveness of conditioning on the generated utterances.
Systems that generate natural language text from abstract meaning representations such as AMR are typically evaluated using automatic surface matching metrics that compare the generated texts to reference texts from which the input meaning representations were constructed. We show that besides well-known issues from which such metrics suffer, an additional problem arises when applying these metrics for AMR-to-text evaluation, since an abstract meaning representation allows for numerous surface realizations. In this work we aim to alleviate these issues by proposing $\mathcal{M}\mathcal{F}_\beta$, a decomposable metric that builds on two pillars. The first is the principle of meaning preservation $\mathcal{M}$: it measures to what extent a given AMR can be reconstructed from the generated sentence using SOTA AMR parsers and applying (fine-grained) AMR evaluation metrics to measure the distance between the original and the reconstructed AMR. The second pillar builds on a principle of (grammatical) form $\mathcal{F}$ that measures the linguistic quality of the generated text, which we implement using SOTA language models. In two extensive pilot studies we show that fulfillment of both principles offers benefits for AMR-to-text evaluation, including explainability of scores. Since $\mathcal{M}\mathcal{F}_\beta$ does not necessarily rely on gold AMRs, it may extend to other text generation tasks.
We present the Language Interpretability Tool (LIT), an open-source platform for visualization and understanding of NLP models. We focus on core questions about model behavior: Why did my model make this prediction? When does it perform poorly? What happens under a controlled change in the input? LIT integrates local explanations, aggregate analysis, and counterfactual generation into a streamlined, browser-based interface to enable rapid exploration and error analysis. We include case studies for a diverse set of workflows, including exploring counterfactuals for sentiment analysis, measuring gender bias in coreference systems, and exploring local behavior in text generation. LIT supports a wide range of models--including classification, seq2seq, and structured prediction--and is highly extensible through a declarative, framework-agnostic API. LIT is under active development, with code and full documentation available at https://github.com/pair-code/lit.
Dialogue systems in open domain have achieved great success due to the easily obtained single-turn corpus and the development of deep learning, but the multi-turn scenario is still a challenge because of the frequent coreference and information omission. In this paper, we investigate the incomplete utterance restoration which has brought general improvement over multi-turn dialogue systems in recent studies. Meanwhile, jointly inspired by the autoregression for text generation and the sequence labeling for text editing, we propose a novel semi autoregressive generator (SARG) with the high efficiency and flexibility. Moreover, experiments on two benchmarks show that our proposed model significantly outperforms the state-of-the-art models in terms of quality and inference speed.
The recent advances in language modeling significantly improved the generative capabilities of deep neural models: in 2019 OpenAI released GPT-2, a pre-trained language model that can autonomously generate coherent, non-trivial and human-like text samples. Since then, ever more powerful text generative models have been developed. Adversaries can exploit these tremendous generative capabilities to enhance social bots that will have the ability to write plausible deepfake messages, hoping to contaminate public debate. To prevent this, it is crucial to develop deepfake social media messages detection systems. However, to the best of our knowledge no one has ever addressed the detection of machine-generated texts on social networks like Twitter or Facebook. With the aim of helping the research in this detection field, we collected the first dataset of \real deepfake tweets, TweepFake. It is real in the sense that each deepfake tweet was actually posted on Twitter. We collected tweets from a total of 23 bots, imitating 17 human accounts. The bots are based on various generation techniques, i.e., Markov Chains, RNN, RNN+Markov, LSTM, GPT-2. We also randomly selected tweets from the humans imitated by the bots to have an overall balanced dataset of 25,572 tweets (half human and half bots generated). The dataset is publicly available on Kaggle. Lastly, we evaluated 13 deepfake text detection methods (based on various state-of-the-art approaches) to both demonstrate the challenges that Tweepfake poses and create a solid baseline of detection techniques. We hope that TweepFake can offer the opportunity to tackle the deepfake detection on social media messages as well.
Song lyrics convey a meaningful story in a creative manner with complex rhythmic patterns. Researchers have been successful in generating and analyisng lyrics for poetry and songs in English and Chinese. But there are no works which explore the Hindi language datasets. Given the popularity of Hindi songs across the world and the ambiguous nature of romanized Hindi script, we propose Bollyrics, an automatic lyric generator for romanized Hindi songs. We propose simple techniques to capture rhyming patterns before and during the model training process in Hindi language. The dataset and codes are available publicly at https://github.com/lingo-iitgn/Bollyrics.
Graph-to-text generation aims to generate fluent texts from graph-based data. In this paper, we investigate two recently proposed pretrained language models (PLMs) and analyze the impact of different task-adaptive pretraining strategies for PLMs in graph-to-text generation. We present a study across three graph domains: meaning representations, Wikipedia knowledge graphs (KGs) and scientific KGs. We show that the PLMs BART and T5 achieve new state-of-the-art results and that task-adaptive pretraining strategies improve their performance even further. In particular, we report new state-of-the-art BLEU scores of 49.72 on LDC2017T10, 59.70 on WebNLG, and 25.66 on AGENDA datasets - a relative improvement of 31.8%, 4.5%, and 42.4%, respectively. In an extensive analysis, we identify possible reasons for the PLMs' success on graph-to-text tasks. We find evidence that their knowledge about true facts helps them perform well even when the input graph representation is reduced to a simple bag of node and edge labels.
Automatic generation of high-quality commit messages for code commits can substantially facilitate software developers' works and coordination. However, the semantic gap between source code and natural language poses a major challenge for the task. Several studies have been proposed to alleviate the challenge but none explicitly involves code contextual information during commit message generation. Specifically, existing research adopts static embedding for code tokens, which maps a token to the same vector regardless of its context. In this paper, we propose a novel Contextualized code representation learning strategy for commit message Generation (CoreGen). CoreGen first learns contextualized code representations which exploit the contextual information behind code commit sequences. The learned representations of code commits built upon Transformer are then fine-tuned for downstream commit message generation. Experiments on the benchmark dataset demonstrate the superior effectiveness of our model over the baseline models with at least 28.18% improvement in terms of BLEU-4 score. Furthermore, we also highlight the future opportunities in training contextualized code representations on larger code corpus as a solution to low-resource tasks and adapting the contextualized code representation framework to other code-to-text generation tasks.
It has been a common approach to pre-train a language model on a large corpus and fine-tune it on task-specific data. In practice, we observe that fine-tuning a pre-trained model on a small dataset may lead to over- and/or under-estimation problem. In this paper, we propose MC-Tailor, a novel method to alleviate the above issue in text generation tasks by truncating and transferring the probability mass from over-estimated regions to under-estimated ones. Experiments on a variety of text generation datasets show that MC-Tailor consistently and significantly outperforms the fine-tuning approach. Our code is available at this url.
This papers revisits the receptive theory in context of computational creativity. It presents a case study of a Paranoid Transformer - a fully autonomous text generation engine with raw output that could be read as the narrative of a mad digital persona without any additional human post-filtering. We describe technical details of the generative system, provide examples of output and discuss the impact of receptive theory, chance discovery and simulation of fringe mental state on the understanding of computational creativity.
Conversational recommender systems (CRS) aim to recommend high-quality items to users through interactive conversations. Although several efforts have been made for CRS, two major issues still remain to be solved. First, the conversation data itself lacks of sufficient contextual information for accurately understanding users' preference. Second, there is a semantic gap between natural language expression and item-level user preference. To address these issues, we incorporate both word-oriented and entity-oriented knowledge graphs (KG) to enhance the data representations in CRSs, and adopt Mutual Information Maximization to align the word-level and entity-level semantic spaces. Based on the aligned semantic representations, we further develop a KG-enhanced recommender component for making accurate recommendations, and a KG-enhanced dialog component that can generate informative keywords or entities in the response text. Extensive experiments have demonstrated the effectiveness of our approach in yielding better performance on both recommendation and conversation tasks.
We present DART, an open domain structured DAta Record to Text generation dataset with over 82k instances (DARTs). Data-to-Text annotations can be a costly process, especially when dealing with tables which are the major source of structured data and contain nontrivial structures. To this end, we propose a procedure of extracting semantic triples from tables that encodes their structures by exploiting the semantic dependencies among table headers and the table title. Our dataset construction framework effectively merged heterogeneous sources from open domain semantic parsing and dialogue-act-based meaning representation tasks by utilizing techniques such as: tree ontology annotation, question-answer pair to declarative sentence conversion, and predicate unification, all with minimum post-editing. We present systematic evaluation on DART as well as new state-of-the-art results on WebNLG 2017 to show that DART (1) poses new challenges to existing data-to-text datasets and (2) facilitates out-of-domain generalization. Our data and code can be found at https://github.com/Yale-LILY/dart.