Factual consistency detection has gotten raised attention in the task of abstractive summarization. Many existing works rely on synthetic training data, which may not accurately reflect or match the inconsistencies produced by summarization models. In this paper, we first systematically analyze the shortcomings of the current methods in synthesizing inconsistent summaries. Current synthesis methods may fail to produce inconsistencies of coreference errors and discourse errors, per our quantitative and qualitative study. Then, employing the parameter-efficient finetuning (PEFT) technique, we discover that a competitive factual consistency detector can be achieved using thousands of real model-generated summaries with human annotations. Our study demonstrates the importance of real machine-generated texts with human annotation in NLG evaluation as our model outperforms the SOTA on the CoGenSumm, FactCC, Frank, and SummEval datasets.

Automated summary quality assessment falls into two categories: reference-based and reference-free. Reference-based metrics, historically deemed more accurate due to the additional information provided by human-written references, are limited by their reliance on human input. In this paper, we hypothesize that the comparison methodologies used by some reference-based metrics to evaluate a system summary against its corresponding reference can be effectively adapted to assess it against its source document, thereby transforming these metrics into reference-free ones. Experimental results support this hypothesis. After being repurposed reference-freely, the zero-shot BERTScore using the pretrained DeBERTa-large-MNLI model of <0.5B parameters consistently outperforms its original reference-based version across various aspects on the SummEval and Newsroom datasets. It also excels in comparison to most existing reference-free metrics and closely competes with zero-shot summary evaluators based on GPT-3.5.

Image and text retrieval is one of the foundational tasks in the vision and language domain with multiple real-world applications. State-of-the-art contrastive approaches, e.g. CLIP, ALIGN, represent images and texts as dense embeddings and calculate the similarity in the dense embedding space as the matching score. On the other hand, sparse semantic features like bag-of-words models are more interpretable, but believed to suffer from inferior accuracy than dense representations. In this work, we show that it is possible to build a sparse semantic representation that is as powerful as, or even better than, dense presentations. We extend the CLIP model and build a sparse text and image representation (STAIR), where the image and text are mapped to a sparse token space. Each token in the space is a (sub-)word in the vocabulary, which is not only interpretable but also easy to integrate with existing information retrieval systems. STAIR model significantly outperforms a CLIP model with +4.9% and +4.3% absolute Recall@1 improvement on COCO-5k text→image and image→text retrieval respectively. It also achieved better performance on both of ImageNet zero-shot and linear probing compared to CLIP.

Canonical automatic summary evaluation metrics, such as ROUGE, focus on lexical similarity which cannot well capture semantics nor linguistic quality and require a reference summary which is costly to obtain. Recently, there have been a growing number of efforts to alleviate either or both of the two drawbacks. In this paper, we present a proof-of-concept study to a weakly supervised summary evaluation approach without the presence of reference summaries. Massive data in existing summarization datasets are transformed for training by pairing documents with corrupted reference summaries. In cross-domain tests, our strategy outperforms baselines with promising improvements, and show a great advantage in gauging linguistic qualities over all metrics.

While BERT is an effective method for learning monolingual sentence embeddings for semantic similarity and embedding based transfer learning BERT based cross-lingual sentence embeddings have yet to be explored. We systematically investigate methods for learning multilingual sentence embeddings by combining the best methods for learning monolingual and cross-lingual representations including: masked language modeling (MLM), translation language modeling (TLM), dual encoder translation ranking, and additive margin softmax. We show that introducing a pre-trained multilingual language model dramatically reduces the amount of parallel training data required to achieve good performance by 80%. Composing the best of these methods produces a model that achieves 83.7% bi-text retrieval accuracy over 112 languages on Tatoeba, well above the 65.5% achieved by LASER, while still performing competitively on monolingual transfer learning benchmarks. Parallel data mined from CommonCrawl using our best model is shown to train competitive NMT models for en-zh and en-de. We publicly release our best multilingual sentence embedding model for 109+ languages at https://tfhub.dev/google/LaBSE.

It has been shown that dual encoders trained on one domain often fail to generalize to other domains for retrieval tasks. One widespread belief is that the bottleneck layer of a dual encoder, where the final score is simply a dot-product between a query vector and a passage vector, is too limited compared to models with fine-grained interactions between the query and the passage. In this paper, we challenge this belief by scaling up the size of the dual encoder model while keeping the bottleneck layer as a single dot-product with a fixed size. With multi-stage training, scaling up the model size brings significant improvement on a variety of retrieval tasks, especially for out-of-domain generalization. We further analyze the impact of the bottleneck layer and demonstrate diminishing improvement when scaling up the embedding size. Experimental results show that our dual encoders, Generalizable T5-based dense Retrievers (GTR), outperform previous sparse and dense retrievers on the BEIR dataset significantly. Most surprisingly, our ablation study finds that GTR is very data efficient, as it only needs 10% of MS Marco supervised data to match the out-of-domain performance of using all supervised data.

We provide the first exploration of sentence embeddings from text-to-text transformers (T5) including the effects of scaling up sentence encoders to 11B parameters. Sentence embeddings are broadly useful for language processing tasks. While T5 achieves impressive performance on language tasks, it is unclear how to produce sentence embeddings from encoder-decoder models. We investigate three methods to construct Sentence-T5 (ST5) models: two utilize only the T5 encoder and one using the full T5 encoder-decoder. We establish a new sentence representation transfer benchmark, SentGLUE, which extends the SentEval toolkit to nine tasks from the GLUE benchmark. Our encoder-only models outperform the previous best models on both SentEval and SentGLUE transfer tasks, including semantic textual similarity (STS). Scaling up ST5 from millions to billions of parameters shown to consistently improve performance. Finally, our encoder-decoder method achieves a new state-of-the-art on STS when using sentence embeddings.

Recent work has shown that either (1) increasing the input length or (2) increasing model size can improve the performance of Transformer-based neural models. In this paper, we present LongT5, a new model that explores the effects of scaling both the input length and model size at the same time. Specifically, we integrate attention ideas from long-input transformers (ETC), and adopt pre-training strategies from summarization pre-training (PEGASUS) into the scalable T5 architecture. The result is a new attention mechanism we call Transient Global (TGlobal), which mimics ETC’s local/global attention mechanism, but without requiring additional side-inputs. We are able to achieve state-of-the-art results on several summarization and question answering tasks, as well as outperform the original T5 models on these tasks. We have open sourced our architecture and training code, as well as our pre-trained model checkpoints.

Early fusion models with cross-attention have shown better-than-human performance on some question answer benchmarks, while it is a poor fit for retrieval since it prevents pre-computation of the answer representations. We present a supervised data mining method using an accurate early fusion model to improve the training of an efficient late fusion retrieval model. We first train an accurate classification model with cross-attention between questions and answers. The cross-attention model is then used to annotate additional passages in order to generate weighted training examples for a neural retrieval model. The resulting retrieval model with additional data significantly outperforms retrieval models directly trained with gold annotations on Precision at N (P@N) and Mean Reciprocal Rank (MRR).

Named entity recognition systems achieve remarkable performance on domains such as English news. It is natural to ask: What are these models actually learning to achieve this? Are they merely memorizing the names themselves? Or are they capable of interpreting the text and inferring the correct entity type from the linguistic context? We examine these questions by contrasting the performance of several variants of architectures for named entity recognition, with some provided only representations of the context as features. We experiment with GloVe-based BiLSTM-CRF as well as BERT. We find that context does influence predictions, but the main factor driving high performance is learning the named tokens themselves. Furthermore, we find that BERT is not always better at recognizing predictive contexts compared to a BiLSTM-CRF model. We enlist human annotators to evaluate the feasibility of inferring entity types from context alone and find that humans are also mostly unable to infer entity types for the majority of examples on which the context-only system made errors. However, there is room for improvement: A system should be able to recognize any named entity in a predictive context correctly and our experiments indicate that current systems may be improved by such capability. Our human study also revealed that systems and humans do not always learn the same contextual clues, and context-only systems are sometimes correct even when humans fail to recognize the entity type from the context. Finally, we find that one issue contributing to model errors is the use of “entangled” representations that encode both contextual and local token information into a single vector, which can obscure clues. Our results suggest that designing models that explicitly operate over representations of local inputs and context, respectively, may in some cases improve performance. In light of these and related findings, we highlight directions for future work.

A major obstacle to the wide-spread adoption of neural retrieval models is that they require large supervised training sets to surpass traditional term-based techniques, which are constructed from raw corpora. In this paper, we propose an approach to zero-shot learning for passage retrieval that uses synthetic question generation to close this gap. The question generation system is trained on general domain data, but is applied to documents in the targeted domain. This allows us to create arbitrarily large, yet noisy, question-passage relevance pairs that are domain specific. Furthermore, when this is coupled with a simple hybrid term-neural model, first-stage retrieval performance can be improved further. Empirically, we show that this is an effective strategy for building neural passage retrieval models in the absence of large training corpora. Depending on the domain, this technique can even approach the accuracy of supervised models.

By supporting multi-modal retrieval training and evaluation, image captioning datasets have spurred remarkable progress on representation learning. Unfortunately, datasets have limited cross-modal associations: images are not paired with other images, captions are only paired with other captions of the same image, there are no negative associations and there are missing positive cross-modal associations. This undermines research into how inter-modality learning impacts intra-modality tasks. We address this gap with Crisscrossed Captions (CxC), an extension of the MS-COCO dataset with human semantic similarity judgments for 267,095 intra- and inter-modality pairs. We report baseline results on CxC for strong existing unimodal and multimodal models. We also evaluate a multitask dual encoder trained on both image-caption and caption-caption pairs that crucially demonstrates CxC’s value for measuring the influence of intra- and inter-modality learning.

Retrieval question answering (ReQA) is the task of retrieving a sentence-level answer to a question from an open corpus (Ahmad et al.,2019).This dataset paper presents MultiReQA, a new multi-domain ReQA evaluation suite composed of eight retrieval QA tasks drawn from publicly available QA datasets. We explore systematic retrieval based evaluation and transfer learning across domains over these datasets using a number of strong base-lines including two supervised neural models, based on fine-tuning BERT and USE-QA models respectively, as well as a surprisingly effective information retrieval baseline, BM25. Five of these tasks contain both training and test data, while three contain test data only. Performing cross training on the five tasks with training data shows that while a general model covering all domains is achievable, the best performance is often obtained by training exclusively on in-domain data.

Both image-caption pairs and translation pairs provide the means to learn deep representations of and connections between languages. We use both types of pairs in MURAL (MUltimodal, MUltitask Representations Across Languages), a dual encoder that solves two tasks: 1) image-text matching and 2) translation pair matching. By incorporating billions of translation pairs, MURAL extends ALIGN (Jia et al.)–a state-of-the-art dual encoder learned from 1.8 billion noisy image-text pairs. When using the same encoders, MURAL’s performance matches or exceeds ALIGN’s cross-modal retrieval performance on well-resourced languages across several datasets. More importantly, it considerably improves performance on under-resourced languages, showing that text-text learning can overcome a paucity of image-caption examples for these languages. On the Wikipedia Image-Text dataset, for example, MURAL-base improves zero-shot mean recall by 8.1% on average for eight under-resourced languages and by 6.8% on average when fine-tuning. We additionally show that MURAL’s text representations cluster not only with respect to genealogical connections but also based on areal linguistics, such as the Balkan Sprachbund.

Language agnostic and semantic-language information isolation is an emerging research direction for multilingual representations models. We explore this problem from a novel angle of geometric algebra and semantic space. A simple but highly effective method “Language Information Removal (LIR)” factors out language identity information from semantic related components in multilingual representations pre-trained on multi-monolingual data. A post-training and model-agnostic method, LIR only uses simple linear operations, e.g. matrix factorization and orthogonal projection. LIR reveals that for weak-alignment multilingual systems, the principal components of semantic spaces primarily encodes language identity information. We first evaluate the LIR on a cross-lingual question answer retrieval task (LAReQA), which requires the strong alignment for the multilingual embedding space. Experiment shows that LIR is highly effectively on this task, yielding almost 100% relative improvement in MAP for weak-alignment models. We then evaluate the LIR on Amazon Reviews and XEVAL dataset, with the observation that removing language information is able to improve the cross-lingual transfer performance.

In the context of neural passage retrieval, we study three promising techniques: synthetic data generation, negative sampling, and fusion. We systematically investigate how these techniques contribute to the performance of the retrieval system and how they complement each other. We propose a multi-stage framework comprising of pre-training with synthetic data, fine-tuning with labeled data, and negative sampling at both stages. We study six negative sampling strategies and apply them to the fine-tuning stage and, as a noteworthy novelty, to the synthetic data that we use for pre-training. Also, we explore fusion methods that combine negatives from different strategies. We evaluate our system using two passage retrieval tasks for open-domain QA and using MS MARCO. Our experiments show that augmenting the negative contrast in both stages is effective to improve passage retrieval accuracy and, importantly, they also show that synthetic data generation and negative sampling have additive benefits. Moreover, using the fusion of different kinds allows us to reach performance that establishes a new state-of-the-art level in two of the tasks we evaluated.

This paper presents a novel training method, Conditional Masked Language Modeling (CMLM), to effectively learn sentence representations on large scale unlabeled corpora. CMLM integrates sentence representation learning into MLM training by conditioning on the encoded vectors of adjacent sentences. Our English CMLM model achieves state-of-the-art performance on SentEval, even outperforming models learned using supervised signals. As a fully unsupervised learning method, CMLM can be conveniently extended to a broad range of languages and domains. We find that a multilingual CMLM model co-trained with bitext retrieval (BR) and natural language inference (NLI) tasks outperforms the previous state-of-the-art multilingual models by a large margin, e.g. 10% improvement upon baseline models on cross-lingual semantic search. We explore the same language bias of the learned representations, and propose a simple, post-training and model agnostic approach to remove the language identifying information from the representation while still retaining sentence semantics.

Pairwise data automatically constructed from weakly supervised signals has been widely used for training deep learning models. Pairwise datasets such as parallel texts can have uneven quality levels overall, but usually contain data subsets that are more useful as learning examples. We present two methods to refine data that are aimed to obtain that kind of subsets in a self-supervised way. Our methods are based on iteratively training dual-encoder models to compute similarity scores. We evaluate our methods on de-noising parallel texts and training neural machine translation models. We find that: (i) The self-supervised refinement achieves most machine translation gains in the first iteration, but following iterations further improve its intrinsic evaluation. (ii) Machine translations can improve the de-noising performance when combined with selection steps. (iii) Our methods are able to reach the performance of a supervised method. Being entirely self-supervised, our methods are well-suited to handle pairwise data without the need of prior knowledge or human annotations.

Most data selection research in machine translation focuses on improving a single domain. We perform data selection for multiple domains at once. This is achieved by carefully introducing instance-level domain-relevance features and automatically constructing a training curriculum to gradually concentrate on multi-domain relevant and noise-reduced data batches. Both the choice of features and the use of curriculum are crucial for balancing and improving all domains, including out-of-domain. In large-scale experiments, the multi-domain curriculum simultaneously reaches or outperforms the individual performance and brings solid gains over no-curriculum training.

We present easy-to-use retrieval focused multilingual sentence embedding models, made available on TensorFlow Hub. The models embed text from 16 languages into a shared semantic space using a multi-task trained dual-encoder that learns tied cross-lingual representations via translation bridge tasks (Chidambaram et al., 2018). The models achieve a new state-of-the-art in performance on monolingual and cross-lingual semantic retrieval (SR). Competitive performance is obtained on the related tasks of translation pair bitext retrieval (BR) and retrieval question answering (ReQA). On transfer learning tasks, our multilingual embeddings approach, and in some cases exceed, the performance of English only sentence embeddings.

We present LAReQA, a challenging new benchmark for language-agnostic answer retrieval from a multilingual candidate pool. Unlike previous cross-lingual tasks, LAReQA tests for “strong” cross-lingual alignment, requiring semantically related cross-language pairs to be closer in representation space than unrelated same-language pairs. This level of alignment is important for the practical task of cross-lingual information retrieval. Building on multilingual BERT (mBERT), we study different strategies for achieving strong alignment. We find that augmenting training data via machine translation is effective, and improves significantly over using mBERT out-of-the-box. Interestingly, model performance on zero-shot variants of our task that only target “weak” alignment is not predictive of performance on LAReQA. This finding underscores our claim that language-agnostic retrieval is a substantively new kind of cross-lingual evaluation, and suggests that measuring both weak and strong alignment will be important for improving cross-lingual systems going forward. We release our dataset and evaluation code at https://github.com/google-research-datasets/lareqa.

Modern NLP systems require high-quality annotated data. For specialized domains, expert annotations may be prohibitively expensive; the alternative is to rely on crowdsourcing to reduce costs at the risk of introducing noise. In this paper we demonstrate that directly modeling instance difficulty can be used to improve model performance and to route instances to appropriate annotators. Our difficulty prediction model combines two learned representations: a ‘universal’ encoder trained on out of domain data, and a task-specific encoder. Experiments on a complex biomedical information extraction task using expert and lay annotators show that: (i) simply excluding from the training data instances predicted to be difficult yields a small boost in performance; (ii) using difficulty scores to weight instances during training provides further, consistent gains; (iii) assigning instances predicted to be difficult to domain experts is an effective strategy for task routing. Further, our experiments confirm the expectation that for such domain-specific tasks expert annotations are of much higher quality and preferable to obtain if practical and that augmenting small amounts of expert data with a larger set of lay annotations leads to further improvements in model performance.

Most existing work on adversarial data generation focuses on English. For example, PAWS (Paraphrase Adversaries from Word Scrambling) consists of challenging English paraphrase identification pairs from Wikipedia and Quora. We remedy this gap with PAWS-X, a new dataset of 23,659 human translated PAWS evaluation pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. We provide baseline numbers for three models with different capacity to capture non-local context and sentence structure, and using different multilingual training and evaluation regimes. Multilingual BERT fine-tuned on PAWS English plus machine-translated data performs the best, with a range of 83.1-90.8 accuracy across the non-English languages and an average accuracy gain of 23% over the next best model. PAWS-X shows the effectiveness of deep, multilingual pre-training while also leaving considerable headroom as a new challenge to drive multilingual research that better captures structure and contextual information.

Popular QA benchmarks like SQuAD have driven progress on the task of identifying answer spans within a specific passage, with models now surpassing human performance. However, retrieving relevant answers from a huge corpus of documents is still a challenging problem, and places different requirements on the model architecture. There is growing interest in developing scalable answer retrieval models trained end-to-end, bypassing the typical document retrieval step. In this paper, we introduce Retrieval Question-Answering (ReQA), a benchmark for evaluating large-scale sentence-level answer retrieval models. We establish baselines using both neural encoding models as well as classical information retrieval techniques. We release our evaluation code to encourage further work on this challenging task.

The scarcity of labeled training data across many languages is a significant roadblock for multilingual neural language processing. We approach the lack of in-language training data using sentence embeddings that map text written in different languages, but with similar meanings, to nearby embedding space representations. The representations are produced using a dual-encoder based model trained to maximize the representational similarity between sentence pairs drawn from parallel data. The representations are enhanced using multitask training and unsupervised monolingual corpora. The effectiveness of our multilingual sentence embeddings are assessed on a comprehensive collection of monolingual, cross-lingual, and zero-shot/few-shot learning tasks.

We explore using multilingual document embeddings for nearest neighbor mining of parallel data. Three document-level representations are investigated: (i) document embeddings generated by simply averaging multilingual sentence embeddings; (ii) a neural bag-of-words (BoW) document encoding model; (iii) a hierarchical multilingual document encoder (HiDE) that builds on our sentence-level model. The results show document embeddings derived from sentence-level averaging are surprisingly effective for clean datasets, but suggest models trained hierarchically at the document-level are more effective on noisy data. Analysis experiments demonstrate our hierarchical models are very robust to variations in the underlying sentence embedding quality. Using document embeddings trained with HiDE achieves the state-of-the-art on United Nations (UN) parallel document mining, 94.9% P@1 for en-fr and 97.3% P@1 for en-es.

Medical professionals search the published literature by specifying the type of patients, the medical intervention(s) and the outcome measure(s) of interest. In this paper we demonstrate how features encoding syntactic patterns improve the performance of state-of-the-art sequence tagging models (both neural and linear) for information extraction of these medically relevant categories. We present an analysis of the type of patterns exploited and of the semantic space induced for these, i.e., the distributed representations learned for identified multi-token patterns. We show that these learned representations differ substantially from those of the constituent unigrams, suggesting that the patterns capture contextual information that is otherwise lost.

With the growing amount of reviews in e-commerce websites, it is critical to assess the helpfulness of reviews and recommend them accordingly to consumers. Recent studies on review helpfulness require plenty of labeled samples for each domain/category of interests. However, such an approach based on close-world assumption is not always practical, especially for domains with limited reviews or the “out-of-vocabulary” problem. Therefore, we propose a convolutional neural network (CNN) based model which leverages both word-level and character-based representations. To transfer knowledge between domains, we further extend our model to jointly model different domains with auxiliary domain discriminators. On the Amazon product review dataset, our approach significantly outperforms the state of the art in terms of both accuracy and cross-domain robustness.

We present a corpus of 5,000 richly annotated abstracts of medical articles describing clinical randomized controlled trials. Annotations include demarcations of text spans that describe the Patient population enrolled, the Interventions studied and to what they were Compared, and the Outcomes measured (the ‘PICO’ elements). These spans are further annotated at a more granular level, e.g., individual interventions within them are marked and mapped onto a structured medical vocabulary. We acquired annotations from a diverse set of workers with varying levels of expertise and cost. We describe our data collection process and the corpus itself in detail. We then outline a set of challenging NLP tasks that would aid searching of the medical literature and the practice of evidence-based medicine.

We present a novel approach to learn representations for sentence-level semantic similarity using conversational data. Our method trains an unsupervised model to predict conversational responses. The resulting sentence embeddings perform well on the Semantic Textual Similarity (STS) Benchmark and SemEval 2017’s Community Question Answering (CQA) question similarity subtask. Performance is further improved by introducing multitask training, combining conversational response prediction and natural language inference. Extensive experiments show the proposed model achieves the best performance among all neural models on the STS Benchmark and is competitive with the state-of-the-art feature engineered and mixed systems for both tasks.

This paper presents an effective approach for parallel corpus mining using bilingual sentence embeddings. Our embedding models are trained to produce similar representations exclusively for bilingual sentence pairs that are translations of each other. This is achieved using a novel training method that introduces hard negatives consisting of sentences that are not translations but have some degree of semantic similarity. The quality of the resulting embeddings are evaluated on parallel corpus reconstruction and by assessing machine translation systems trained on gold vs. mined sentence pairs. We find that the sentence embeddings can be used to reconstruct the United Nations Parallel Corpus (Ziemski et al., 2016) at the sentence-level with a precision of 48.9% for en-fr and 54.9% for en-es. When adapted to document-level matching, we achieve a parallel document matching accuracy that is comparable to the significantly more computationally intensive approach of Uszkoreit et al. (2010). Using reconstructed parallel data, we are able to train NMT models that perform nearly as well as models trained on the original data (within 1-2 BLEU).

We present easy-to-use TensorFlow Hub sentence embedding models having good task transfer performance. Model variants allow for trade-offs between accuracy and compute resources. We report the relationship between model complexity, resources, and transfer performance. Comparisons are made with baselines without transfer learning and to baselines that incorporate word-level transfer. Transfer learning using sentence-level embeddings is shown to outperform models without transfer learning and often those that use only word-level transfer. We show good transfer task performance with minimal training data and obtain encouraging results on word embedding association tests (WEAT) of model bias.

Aspect extraction abstracts the common properties of objects from corpora discussing them, such as reviews of products. Recent work on aspect extraction is leveraging the hierarchical relationship between products and their categories. However, such effort focuses on the aspects of child categories but ignores those from parent categories. Hence, we propose an LDA-based generative topic model inducing the two-layer categorical information (CAT-LDA), to balance the aspects of both a parent category and its child categories. Our hypothesis is that child categories inherit aspects from parent categories, controlled by the hierarchy between them. Experimental results on 5 categories of Amazon.com products show that both common aspects of parent category and the individual aspects of sub-categories can be extracted to align well with the common sense. We further evaluate the manually extracted aspects of 16 products, resulting in an average hit rate of 79.10%.

We present a robust approach for detecting intrinsic sentence importance in news, by training on two corpora of document-summary pairs. When used for single-document summarization, our approach, combined with the “beginning of document” heuristic, outperforms a state-of-the-art summarizer and the beginning-of-article baseline in both automatic and manual evaluations. These results represent an important advance because in the absence of cross-document repetition, single document summarizers for news have not been able to consistently outperform the strong beginning-of-article baseline.