Maximizing the likelihood of the next token is an established, statistically sound objective for pre-training language models. In this paper we show that we can train better models faster by pre-aggregating the corpus with a collapsed n-gram distribution. Previous studies have proposed corpus-level n-gram statistics as a regularizer; however, the construction and querying of such n-grams, if done naively, prove to be costly and significantly impede training speed, thereby limiting their application in modern large language model pre-training.We introduce an alternative compact representation of the next token distribution that, in expectation, aligns with the complete n-gram distribution while markedly reducing variance across mini-batches compared to the standard next-token loss. Empirically, we demonstrate that both the n-gram regularized model and our approximation yield substantial improvements in model quality and convergence rate compared to existing methods. Furthermore, our approximation facilitates scalability of gains to larger datasets and models compared to the straightforward n-gram regularization method.

Despite cross-lingual generalization demonstrated by pre-trained multilingual models, the translate-train paradigm of transferring English datasets across multiple languages remains to be a key mechanism for training task-specific multilingual models. However, for many low-resource languages, the availability of a reliable translation service entails significant amounts of costly human-annotated translation pairs. Further, translation services may continue to be brittle due to domain mismatch between task-specific input text and general-purpose text used for training translation models. For multilingual semantic parsing, we demonstrate the effectiveness and flexibility offered by large language models (LLMs) for translating English datasets into several languages via few-shot prompting. Through extensive comparisons on two public datasets, MTOP and MASSIVE, spanning 50 languages and several domains, we show that our method of translating data using LLMs outperforms a strong translate-train baseline on 41 out of 50 languages. We study the key design choices that enable more effective multilingual data translation via prompted LLMs.

Research in Text-to-SQL conversion has been largely benchmarked against datasets where each text query corresponds to one correct SQL. However, natural language queries over real-life databases frequently involve significant ambiguity about the intended SQL due to overlapping schema names and multiple confusing relationship paths. To bridge this gap, we develop a novel benchmark called AmbiQT with over 3000 examples where each text is interpretable as two plausible SQLs due to lexical and/or structural ambiguity. When faced with ambiguity, an ideal top-k decoder should generate all valid interpretations for possible disambiguation by the user. We evaluate several Text-to-SQL systems and decoding algorithms, including those employing state-of-the-art LLMs, and find them to be far from this ideal. The primary reason is that the prevalent beam search algorithm and its variants, treat SQL queries as a string and produce unhelpful token-level diversity in the top-k. We propose LogicalBeam, a new decoding algorithm that navigates the SQL logic space using a blend of plan-based template generation and constrained infilling. Counterfactually generated plans diversify templates while in-filling with a beam-search that branches solely on schema names provides value diversity. LogicalBeam is up to 2.5 times more effective than state-of-the-art models at generating all candidate SQLs in the top-k ranked outputs. It also enhances the top-5 Exact and Execution Match Accuracies on SPIDER and Kaggle DBQA.

Existing Text-to-SQL generators require the entire schema to be encoded with the user text. This is expensive or impractical for large databases with tens of thousands of columns. Standard dense retrieval techniques are inadequate for schema subsetting of a large structured database, where the correct semantics of retrieval demands that we rank sets of schema elements rather than individual documents. In response, we propose a two-stage process for effective coverage during retrieval. First, we use an LLM to hallucinate a minimal DB schema that it deems adequate to answer the query. We use the hallucinated schema to retrieve a subset of the actual schema, by composing the results from multiple dense retrievals. Remarkably, hallucination — generally considered a nuisance — turns out to be actually useful as a bridging mechanism. Since no existing benchmarks exist for schema subsetting on large databases, we introduce two benchmarks: (1) A semi-synthetic dataset of 4502 schema elements, by taking a union of schema on the well-known SPIDER dataset, and (2) A real-life benchmark called SocialDB sourced from an actual large data warehouse comprising of 17844 schema elements. We show that our method leads to significantly higher recall than SOTA retrieval-based augmentation methods.

Despite the impressive performance of ASR models on mainstream benchmarks, their performance on rare words is unsatisfactory. In enterprise settings, often a focused list of entities (such as locations, names, etc) are available which can be used to adapt the model to the terminology of specific domains. In this paper, we present a novel inference algorithm that improves the prediction of state-of-the-art ASR models using nearest-neighbor-based matching on an inference-time word list. We consider both the Transducer architecture that is useful in the streaming setting, and state-of-the-art encoder-decoder models such as Whisper. In our approach, a list of rare entities is indexed in a memory by synthesizing speech for each entry, and then storing the internal acoustic and language model states obtained from the best possible alignment on the ASR model. The memory is organized as a trie which we harness to perform a stateful lookup during inference. A key property of our extension is that we prevent spurious matches by restricting to only word-level matches. In our experiments on publicly available datasets and private benchmarks, we show that our method is effective in significantly improving rare word recognition.

Word-level quality scores on input source sentences can provide useful feedback to an end-user when translating into an unfamiliar target language. Recent approaches either require training special word-scoring models based on synthetic data or require repeated invocation of the translation model. We propose a simple approach based on comparing the difference of probabilities from two language models. The basic premise of our method is to reason how well each source word is explained by the target sentence as against the source language model. Our approach provides up to five points higher F1 scores and is significantly faster than the state of the art methods on three language pairs. Also, our method does not require training any new model. We release a public dataset on word omissions and mistranslations on a new language pair.

Text-to-SQL parsers typically struggle with databases unseen during the train time. Adapting Text-to-SQL parsers to new database schemas is a challenging problem owing to a vast diversity of schemas and zero availability of natural language queries in new schemas. We present ReFill, a framework for synthesizing high-quality and textually diverse parallel datasets for adapting Text-to-SQL parsers. Unlike prior methods that utilize SQL-to-Text generation, ReFill learns to retrieve-and-edit text queries in existing schemas and transfer them to the new schema. ReFill utilizes a simple method for retrieving diverse existing text, masking their schema-specific tokens, and refilling with tokens relevant to the new schema. We show that this process leads to significantly more diverse text queries than achievable by standard SQL-to-Text generation models. Through experiments on several databases, we show that adapting a parser by finetuning it on datasets synthesized by ReFill consistently outperforms prior data-augmentation methods.

Pre-trained multilingual language models such as mBERT and XLM-R have demonstrated great potential for zero-shot cross-lingual transfer to low web-resource languages (LRL). However, due to limited model capacity, the large difference in the sizes of available monolingual corpora between high web-resource languages (HRL) and LRLs does not provide enough scope of co-embedding the LRL with the HRL, thereby affecting the downstream task performance of LRLs. In this paper, we argue that relatedness among languages in a language family along the dimension of lexical overlap may be leveraged to overcome some of the corpora limitations of LRLs. We propose Overlap BPE (OBPE), a simple yet effective modification to the BPE vocabulary generation algorithm which enhances overlap across related languages. Through extensive experiments on multiple NLP tasks and datasets, we observe that OBPE generates a vocabulary that increases the representation of LRLs via tokens shared with HRLs. This results in improved zero-shot transfer from related HRLs to LRLs without reducing HRL representation and accuracy. Unlike previous studies that dismissed the importance of token-overlap, we show that in the low-resource related language setting, token overlap matters. Synthetically reducing the overlap to zero can cause as much as a four-fold drop in zero-shot transfer accuracy.

Online alignment in machine translation refers to the task of aligning a target word to a source word when the target sequence has only been partially decoded. Good online alignments facilitate important applications such as lexically constrained translation where user-defined dictionaries are used to inject lexical constraints into the translation model. We propose a novel posterior alignment technique that is truly online in its execution and superior in terms of alignment error rates compared to existing methods. Our proposed inference technique jointly considers alignment and token probabilities in a principled manner and can be seamlessly integrated within existing constrained beam-search decoding algorithms. On five language pairs, including two distant language pairs, we achieve consistent drop in alignment error rates. When deployed on seven lexically constrained translation tasks, we achieve significant improvements in BLEU specifically around the constrained positions.

The scarcity of gold standard code-mixed to pure language parallel data makes it difficult to train translation models reliably.Prior work has addressed the paucity of parallel data with data augmentation techniques.Such methods rely heavily on external resources making systems difficult to train and scale effectively for multiple languages.We present a simple yet highly effective two-stage back-translation based training scheme for adapting multilingual models to the task of code-mixed translation which eliminates dependence on external resources.We show a substantial improvement in translation quality (measured through BLEU), beating existing prior work by up to +3.8 BLEU on code-mixed Hi→En, Mr→En, and Bn→En tasks. On the LinCE Machine Translation leader board, we achieve the highest score for code-mixed Es→En, beating existing best baseline by +6.5 BLEU, and our own stronger baseline by +1.1 BLEU.

Recent research in multilingual language models (LM) has demonstrated their ability to effectively handle multiple languages in a single model. This holds promise for low web-resource languages (LRL) as multilingual models can enable transfer of supervision from high resource languages to LRLs. However, incorporating a new language in an LM still remains a challenge, particularly for languages with limited corpora and in unseen scripts. In this paper we argue that relatedness among languages in a language family may be exploited to overcome some of the corpora limitations of LRLs, and propose RelateLM. We focus on Indian languages, and exploit relatedness along two dimensions: (1) script (since many Indic scripts originated from the Brahmic script), and (2) sentence structure. RelateLM uses transliteration to convert the unseen script of limited LRL text into the script of a Related Prominent Language (RPL) (Hindi in our case). While exploiting similar sentence structures, RelateLM utilizes readily available bilingual dictionaries to pseudo translate RPL text into LRL corpora. Experiments on multiple real-world benchmark datasets provide validation to our hypothesis that using a related language as pivot, along with transliteration and pseudo translation based data augmentation, can be an effective way to adapt LMs for LRLs, rather than direct training or pivoting through English.

Machine translation of user-generated code-mixed inputs to English is of crucial importance in applications like web search and targeted advertising. We address the scarcity of parallel training data for training such models by designing a strategy of converting existing non-code-mixed parallel data sources to code-mixed parallel data. We present an m-BERT based procedure whose core learnable component is a ternary sequence labeling model, that can be trained with a limited code-mixed corpus alone. We show a 5.8 point increase in BLEU on heavily code-mixed sentences by training a translation model using our data augmentation strategy on an Hindi-English code-mixed translation task.

State-of-the-art NLP inference uses enormous neural architectures and models trained for GPU-months, well beyond the reach of most consumers of NLP. This has led to one-size-fits-all public API-based NLP service models by major AI companies, serving millions of clients. They cannot afford traditional fine tuning for individual clients. Many clients cannot even afford significant fine tuning, and own little or no labeled data. Recognizing that word usage and salience diversity across clients leads to reduced accuracy, we initiate a study of practical and lightweight adaptation of centralized NLP services to clients. Each client uses an unsupervised, corpus-based sketch to register to the service. The server modifies its network mildly to accommodate client sketches, and occasionally trains the augmented network over existing clients. When a new client registers with its sketch, it gets immediate accuracy benefits. We demonstrate the proposed architecture using sentiment labeling, NER, and predictive language modeling.

We evaluate named entity representations of BERT-based NLP models by investigating their robustness to replacements from the same typed class in the input. We highlight that on several tasks while such perturbations are natural, state of the art trained models are surprisingly brittle. The brittleness continues even with the recent entity-aware BERT models. We also try to discern the cause of this non-robustness, considering factors such as tokenization and frequency of occurrence. Then we provide a simple method that ensembles predictions from multiple replacements while jointly modeling the uncertainty of type annotations and label predictions. Experiments on three NLP tasks shows that our method enhances robustness and increases accuracy on both natural and adversarial datasets.

Given a small corpus D_T pertaining to a limited set of focused topics, our goal is to train embeddings that accurately capture the sense of words in the topic in spite of the limited size of D_T. These embeddings may be used in various tasks involving D_T. A popular strategy in limited data settings is to adapt pretrained embeddings E trained on a large corpus. To correct for sense drift, fine-tuning, regularization, projection, and pivoting have been proposed recently. Among these, regularization informed by a word’s corpus frequency performed well, but we improve upon it using a new regularizer based on the stability of its cooccurrence with other words. However, a thorough comparison across ten topics, spanning three tasks, with standardized settings of hyper-parameters, reveals that even the best embedding adaptation strategies provide small gains beyond well-tuned baselines, which many earlier comparisons ignored. In a bold departure from adapting pretrained embeddings, we propose using D_T to probe, attend to, and borrow fragments from any large, topic-rich source corpus (such as Wikipedia), which need not be the corpus used to pretrain embeddings. This step is made scalable and practical by suitable indexing. We reach the surprising conclusion that even limited corpus augmentation is more useful than adapting embeddings, which suggests that non-dominant sense information may be irrevocably obliterated from pretrained embeddings and cannot be salvaged by adaptation.

We present a Parallel Iterative Edit (PIE) model for the problem of local sequence transduction arising in tasks like Grammatical error correction (GEC). Recent approaches are based on the popular encoder-decoder (ED) model for sequence to sequence learning. The ED model auto-regressively captures full dependency among output tokens but is slow due to sequential decoding. The PIE model does parallel decoding, giving up the advantage of modeling full dependency in the output, yet it achieves accuracy competitive with the ED model for four reasons: 1. predicting edits instead of tokens, 2. labeling sequences instead of generating sequences, 3. iteratively refining predictions to capture dependencies, and 4. factorizing logits over edits and their token argument to harness pre-trained language models like BERT. Experiments on tasks spanning GEC, OCR correction and spell correction demonstrate that the PIE model is an accurate and significantly faster alternative for local sequence transduction.

In this paper we show that a simple beam approximation of the joint distribution between attention and output is an easy, accurate, and efficient attention mechanism for sequence to sequence learning. The method combines the advantage of sharp focus in hard attention and the implementation ease of soft attention. On five translation tasks we show effortless and consistent gains in BLEU compared to existing attention mechanisms.