Recent large language models (LLMs) have demonstrated remarkable generalization abilities in mathematics and logical reasoning tasks.Prior research indicates that LLMs pre-trained with programming language data exhibit high mathematical and reasoning abilities; however, this causal relationship has not been rigorously tested. Our research aims to verify which programming languages and features during pre-training affect logical inference performance. Specifically, we pre-trained decoder-based language models from scratch using datasets from ten programming languages (e.g., Python, C, Java) and three natural language datasets (Wikipedia, Fineweb, C4) under identical conditions. Thereafter, we evaluated the trained models in a few-shot in-context learning setting on logical reasoning tasks: FLD and bAbi, which do not require commonsense or world knowledge. The results demonstrate that nearly all models trained with programming languages consistently outperform those trained with natural languages, indicating that programming languages contain factors that elicit logic inference performance. In addition, we found that models trained with programming languages exhibit a better ability to follow instructions compared to those trained with natural languages. Further analysis reveals that the depth of Abstract Syntax Trees representing parsed results of programs also affects logical reasoning performance. These findings will offer insights into the essential elements of pre-training for acquiring the foundational abilities of LLMs.

This paper investigates how machine translation for low-resource languages can be improved by incorporating information from bilingual lexicons during the training process for mainly translation between Mandarin and Formosan languages, which are all moribund or critically endangered, and we also show that our techniques work for translation between Spanish and Nahuatl, a language pair consisting of languages from completely different language families. About 70% of the approximately 7,000 languages of the world have data in the form of lexicons, a valuable resource for improving low-resource language translation. We collect a dataset of parallel data and bilingual lexicons between Mandarin and 16 different Formosan languages and examine mainly three different approaches: (1) simply using lexical data as additional parallel data, (2) generating pseudo-parallel sentence data to use during training by replacing words in the original parallel sentence data using the lexicon, and (3) a combination of (1) and (2). All three approaches give us gains in both Bleu scores and chrF scores, and we found that (3) provided the most gains, followed by (1) and then (2), which we observed for both translation between Mandarin and the Formosan languages and Spanish-Nahuatl. With technique (3), we saw an average increase of 5.55 in Bleu scores and 10.33 in chrF scores.

Bias in reporting can influence the public’s opinion on relevant societal issues. Examples include informational bias (selective presentation of content) and lexical bias (specific framing of content through linguistic choices). The recognition of media bias is arguably an area where NLP can contribute to the “social good”. Traditional NLP models have shown good performance in classifying media bias, but require careful model design and extensive tuning. In this paper, we ask how well prompting of large language models can recognize media bias. Through an extensive empirical study including a wide selection of pre-trained models, we find that prompt-based techniques can deliver comparable performance to traditional models with greatly reduced effort and that, similar to traditional models, the availability of context substantially improves results. We further show that larger models can leverage different kinds of context simultaneously, obtaining further performance improvements.

Current decoder-based pre-trained language models (PLMs) successfully demonstrate multilingual capabilities. However, it is unclear how these models handle multilingualism.We analyze the neuron-level internal behavior of multilingual decoder-based PLMs, Specifically examining the existence of neurons that fire “uniquely for each language” within decoder-only multilingual PLMs.We analyze six languages: English, German, French, Spanish, Chinese, and Japanese, and show that language-specific neurons are unique, with a slight overlap (< 5%) between languages. These neurons are mainly distributed in the models’ first and last few layers. This trend remains consistent across languages and models.Additionally, we tamper with less than 1% of the total neurons in each model during inference and demonstrate that tampering with a few language-specific neurons drastically changes the probability of target language occurrence in text generation.

The ability (and inability) of large language models (LLMs) to perform arithmetic tasks has been the subject of much theoretical and practical debate. We show that LLMs are frequently able to correctly and confidently predict the first digit of n-digit by m-digit multiplication tasks without using chain of thought reasoning, despite these tasks require compounding operations to solve. Simultaneously, LLMs in practice often fail to correctly or confidently predict the last digit of an n-digit by m-digit multiplication, a task equivalent to 1-digit by 1-digit multiplication which can be easily learned or memorized. We show that the latter task can be solved more robustly when the LLM is conditioned on all of the correct higher-order digits, which on average increases the confidence of the correct last digit on 5-digit by 5-digit multiplication tasks using Llama 2-13B by over 230% (0.13→0.43) and Mistral-7B by 150% (0.22→0.55).

Large Language Models (LLMs) have significantly advanced healthcare innovation on generation capabilities. However, their application in real clinical settings is challenging due to potential deviations from medical facts and inherent biases. In this work, we develop an augmented LLM framework, KG-Rank, which leverages a medical knowledge graph (KG) with ranking and re-ranking techniques, aiming to improve free-text question-answering (QA) in the medical domain. Specifically, upon receiving a question, we initially retrieve triplets from a medical KG to gather factual information. Subsequently, we innovatively apply ranking methods to refine the ordering of these triplets, aiming to yield more precise answers. To the best of our knowledge, KG-Rank is the first application of ranking models combined with KG in medical QA specifically for generating long answers. Evaluation of four selected medical QA datasets shows that KG-Rank achieves an improvement of over 18% in the ROUGE-L score. Moreover, we extend KG-Rank to open domains, where it realizes a 14% improvement in ROUGE-L, showing the effectiveness and potential of KG-Rank.

In this paper we present a thorough investigation of automatic bias recognition on BASIL, a dataset of political news which has been annotated with different kinds of biases. We begin by unveiling several inconsistencies in prior work using this dataset, showing that most approaches focus only on certain task formulations while ignoring others, and also failing to report important evaluation details. We provide a comprehensive categorization of these approaches, as well as a more uniform and clear set of evaluation metrics. We argue about the importance of the missing formulations and also propose the novel task of simultaneously detecting different kinds of biases in news. In our work, we tackle bias on six different BASIL classification tasks in a unified manner. Eventually, we introduce a simple yet effective approach based on data augmentation and preprocessing which is generic and works very well across models and task formulations, allowing us to obtain state-of-the-art results. We also perform ablation studies on some tasks to quantify the strength of data augmentation and preprocessing, and find that they correlate positively on all bias tasks.

While Large Language Models (LLMs) have achieved remarkable performance in many tasks, much about their inner workings remains unclear. In this study, we present novel experimental insights into the resilience of LLMs, particularly GPT-4, when subjected to extensive character-level permutations. To investigate this, we first propose the Scrambled Bench, a suite designed to measure the capacity of LLMs to handle scrambled input, in terms of both recovering scrambled sentences and answering questions given scrambled context. The experimental results indicate that multiple advanced LLMs demonstrate the capability akin to typoglycemia, a phenomenon where humans can understand the meaning of words even when the letters within those words are scrambled, as long as the first and last letters remain in place. More surprisingly, we found that only GPT-4 nearly flawlessly processes inputs with unnatural errors, a task that poses significant challenges for other LLMs and often even for humans. Specifically, GPT-4 can almost perfectly reconstruct the original sentences from scrambled ones, decreasing the edit distance by 95%, even when all letters within each word are entirely scrambled. It is counter-intuitive that LLMs can exhibit such resilience despite severe disruption to input tokenization caused by scrambled text.

In this paper we propose a novel approach to automatically classify the level of formality in Japanese text, using three categories (formal, polite, and informal). We introduce a new dataset that combine manually-annotated sentences from existing resources, and formal sentences scrapped from the website of the House of Representatives and the House of Councilors of Japan. Based on our data, we propose a Transformer-based classification model for Japanese, which obtains state-of-the-art results in benchmark datasets. We further propose to utilize our classifier to study the effectiveness of prompting techniques for controlling the formality level of machine translation (MT) using Large Language Models (LLM). Our experimental setting includes a large selection of such models and is based on an En->Ja parallel corpus specifically designed to test formality control in MT. Our results validate the robustness and effectiveness of our proposed approach and while also providing empirical evidence suggesting that prompting LLMs is a viable approach to control the formality level of En->Ja MT using LLMs.

Amis is an endangered language indigenous to Taiwan with limited data available for computational processing. We thus present an Amis-Mandarin dataset containing a parallel corpus of 5,751 Amis and Mandarin sentences and a dictionary of 7,800 Amis words and phrases with their definitions in Mandarin. Using our dataset, we also established a baseline for machine translation between Amis and Mandarin in both directions. Our dataset can be found at https://github.com/francisdzheng/amis-mandarin.

Jejueo is a critically endangered language spoken on Jeju Island and is closely related to but mutually unintelligible with Korean. Parallel data between Jejueo and Korean is scarce, and translation between the two languages requires more attention, as current neural machine translation systems typically rely on large amounts of parallel training data. While low-resource machine translation has been shown to benefit from using additional monolingual data during the pretraining process, not as much research has been done on how to select languages other than the source and target languages for use during pretraining. We show that using large amounts of Korean and Japanese data during the pretraining process improves translation by 2.16 BLEU points for translation in the Jejueo → Korean direction and 1.34 BLEU points for translation in the Korean → Jejueo direction compared to the baseline.

With in the broader scope of machine learning, data augmentation is a common strategy to improve generalization and robustness of machine learning models. While data augmentation has been widely used within computer vision, its use in the NLP has been been comparably rather limited. The reason for this is that within NLP, the impact of proposed data augmentation methods on performance has not been evaluated in a unified manner, and effective data augmentation methods are unclear. In this paper, we look to tackle this by evaluating the impact of 12 data augmentation methods on multiple datasets when finetuning pre-trained language models. We find minimal improvements when data sizes are constrained to a few thousand, with performance degradation when data size is increased. We also use various methods to quantify the strength of data augmentations, and find that these values, though weakly correlated with downstream performance, correlate negatively or positively depending on the task. Furthermore, we find a glaring lack of consistently performant data augmentations. This all alludes to the difficulty of data augmentations for NLP tasks and we are inclined to believe that static data augmentations are not broadly applicable given these properties.

This paper describes UTokyo’s submission to the AmericasNLP 2021 Shared Task on machine translation systems for indigenous languages of the Americas. We present a low-resource machine translation system that improves translation accuracy using cross-lingual language model pretraining. Our system uses an mBART implementation of fairseq to pretrain on a large set of monolingual data from a diverse set of high-resource languages before finetuning on 10 low-resource indigenous American languages: Aymara, Bribri, Asháninka, Guaraní, Wixarika, Náhuatl, Hñähñu, Quechua, Shipibo-Konibo, and Rarámuri. On average, our system achieved BLEU scores that were 1.64 higher and chrF scores that were 0.0749 higher than the baseline.

Generating diverse texts is an important factor for unsupervised text generation. One approach is to produce the diversity of texts conditioned by the sampled latent code. Although several generative adversarial networks (GANs) have been proposed thus far, these models still suffer from mode-collapsing if the models are not pre-trained. In this paper, we propose a GAN model that aims to improve the approach to generating diverse texts conditioned by the latent space. The generator of our model uses Gumbel-Softmax distribution for the word sampling process. To ensure that the text is generated conditioned upon the sampled latent code, reconstruction loss is introduced in our objective function. The discriminator of our model iteratively inspects incomplete partial texts and learns to distinguish whether they are real or fake by using the standard GAN objective function. Experimental results using the COCO Image Captions dataset show that, although our model is not pre-trained, the performance of our model is quite competitive with the existing baseline models, which requires pre-training.

Transformers have shown improved performance when compared to previous architectures for sequence processing such as RNNs. Despite their sizeable performance gains, as recently suggested, the model is computationally expensive to train and with a high parameter budget. In light of this, we explore parameter-sharing methods in Transformers with a specific focus on generative models. We perform an analysis of different parameter sharing/reduction methods and develop the Subformer. Our model combines sandwich-style parameter sharing, which overcomes naive cross-layer parameter sharing in generative models, and self-attentive embedding factorization (SAFE). Experiments on machine translation, abstractive summarization and language modeling show that the Subformer can outperform the Transformer even when using significantly fewer parameters.

Reproducible benchmarks are crucial in driving progress of machine translation research. However, existing machine translation benchmarks have been mostly limited to high-resource or well-represented languages. Despite an increasing interest in low-resource machine translation, there are no standardized reproducible benchmarks for many African languages, many of which are used by millions of speakers but have less digitized textual data. To tackle these challenges, we propose AfroMT, a standardized, clean, and reproducible machine translation benchmark for eight widely spoken African languages. We also develop a suite of analysis tools for system diagnosis taking into account the unique properties of these languages. Furthermore, we explore the newly considered case of low-resource focused pretraining and develop two novel data augmentation-based strategies, leveraging word-level alignment information and pseudo-monolingual data for pretraining multilingual sequence-to-sequence models. We demonstrate significant improvements when pretraining on 11 languages, with gains of up to 2 BLEU points over strong baselines. We also show gains of up to 12 BLEU points over cross-lingual transfer baselines in data-constrained scenarios. All code and pretrained models will be released as further steps towards larger reproducible benchmarks for African languages.

Despite the recent advances in opinion mining for written reviews, few works have tackled the problem on other sources of reviews. In light of this issue, we propose a multi-modal approach for mining fine-grained opinions from video reviews that is able to determine the aspects of the item under review that are being discussed and the sentiment orientation towards them. Our approach works at the sentence level without the need for time annotations and uses features derived from the audio, video and language transcriptions of its contents. We evaluate our approach on two datasets and show that leveraging the video and audio modalities consistently provides increased performance over text-only baselines, providing evidence these extra modalities are key in better understanding video reviews.

In this paper, we tackle the task of definition modeling, where the goal is to learn to generate definitions of words and phrases. Existing approaches for this task are discriminative, combining distributional and lexical semantics in an implicit rather than direct way. To tackle this issue we propose a generative model for the task, introducing a continuous latent variable to explicitly model the underlying relationship between a phrase used within a context and its definition. We rely on variational inference for estimation and leverage contextualized word embeddings for improved performance. Our approach is evaluated on four existing challenging benchmarks with the addition of two new datasets, “Cambridge” and the first non-English corpus “Robert”, which we release to complement our empirical study. Our Variational Contextual Definition Modeler (VCDM) achieves state-of-the-art performance in terms of automatic and human evaluation metrics, demonstrating the effectiveness of our approach.

In this paper we study how different ways of combining character and word-level representations affect the quality of both final word and sentence representations. We provide strong empirical evidence that modeling characters improves the learned representations at the word and sentence levels, and that doing so is particularly useful when representing less frequent words. We further show that a feature-wise sigmoid gating mechanism is a robust method for creating representations that encode semantic similarity, as it performed reasonably well in several word similarity datasets. Finally, our findings suggest that properly capturing semantic similarity at the word level does not consistently yield improved performance in downstream sentence-level tasks.

We propose an edit-centric approach to assess Wikipedia article quality as a complementary alternative to current full document-based techniques. Our model consists of a main classifier equipped with an auxiliary generative module which, for a given edit, jointly provides an estimation of its quality and generates a description in natural language. We performed an empirical study to assess the feasibility of the proposed model and its cost-effectiveness in terms of data and quality requirements.

In this paper we formalize the problem automatic fill-in-the-blank question generation using two standard NLP machine learning schemes, proposing concrete deep learning models for each. We present an empirical study based on data obtained from a language learning platform showing that both of our proposed settings offer promising results.

Predicting context-dependent and non-literal utterances like sarcastic and ironic expressions still remains a challenging task in NLP, as it goes beyond linguistic patterns, encompassing common sense and shared knowledge as crucial components. To capture complex morpho-syntactic features that can usually serve as indicators for irony or sarcasm across dynamic contexts, we propose a model that uses character-level vector representations of words, based on ELMo. We test our model on 7 different datasets derived from 3 different data sources, providing state-of-the-art performance in 6 of them, and otherwise offering competitive results.

In this paper we describe our system designed for the WASSA 2018 Implicit Emotion Shared Task (IEST), which obtained 2nd place out of 30 teams with a test macro F1 score of 0.710. The system is composed of a single pre-trained ELMo layer for encoding words, a Bidirectional Long-Short Memory Network BiLSTM for enriching word representations with context, a max-pooling operation for creating sentence representations from them, and a Dense Layer for projecting the sentence representations into label space. Our official submission was obtained by ensembling 6 of these models initialized with different random seeds. The code for replicating this paper is available at https://github.com/jabalazs/implicit_emotion.

We propose to study the generation of descriptions from source code changes by integrating the messages included on code commits and the intra-code documentation inside the source in the form of docstrings. Our hypothesis is that although both types of descriptions are not directly aligned in semantic terms —one explaining a change and the other the actual functionality of the code being modified— there could be certain common ground that is useful for the generation. To this end, we propose an architecture that uses the source code-docstring relationship to guide the description generation. We discuss the results of the approach comparing against a baseline based on a sequence-to-sequence model, using standard automatic natural language generation metrics as well as with a human study, thus offering a comprehensive view of the feasibility of the approach.

In this paper we introduce our system for the task of Irony detection in English tweets, a part of SemEval 2018. We propose representation learning approach that relies on a multi-layered bidirectional LSTM, without using external features that provide additional semantic information. Although our model is able to outperform the baseline in the validation set, our results show limited generalization power over the test set. Given the limited size of the dataset, we think the usage of more pre-training schemes would greatly improve the obtained results.

The need for automatic document summarization that can be used for practical applications is increasing rapidly. In this paper, we propose a general framework for summarization that extracts sentences from a document using externally related information. Our work is aimed at single document summarization using small amounts of reference summaries. In particular, we address document summarization in the framework of multi-task learning using curriculum learning for sentence extraction and document classification. The proposed framework enables us to obtain better feature representations to extract sentences from documents. We evaluate our proposed summarization method on two datasets: financial report and news corpus. Experimental results demonstrate that our summarizers achieve performance that is comparable to state-of-the-art systems.

Reproducing experiments is an important instrument to validate previous work and build upon existing approaches. It has been tackled numerous times in different areas of science. In this paper, we introduce an empirical replicability study of three well-known algorithms for syntactic centric aspect-based opinion mining. We show that reproducing results continues to be a difficult endeavor, mainly due to the lack of details regarding preprocessing and parameter setting, as well as due to the absence of available implementations that clarify these details. We consider these are important threats to validity of the research on the field, specifically when compared to other problems in NLP where public datasets and code availability are critical validity components. We conclude by encouraging code-based research, which we think has a key role in helping researchers to understand the meaning of the state-of-the-art better and to generate continuous advances.

Video reviews are the natural evolution of written product reviews. In this paper we target this phenomenon and introduce the first dataset created from closed captions of YouTube product review videos as well as a new attention-RNN model for aspect extraction and joint aspect extraction and sentiment classification. Our model provides state-of-the-art performance on aspect extraction without requiring the usage of hand-crafted features on the SemEval ABSA corpus, while it outperforms the baseline on the joint task. In our dataset, the attention-RNN model outperforms the baseline for both tasks, but we observe important performance drops for all models in comparison to SemEval. These results, as well as further experiments on domain adaptation for aspect extraction, suggest that differences between speech and written text, which have been discussed extensively in the literature, also extend to the domain of product reviews, where they are relevant for fine-grained opinion mining.

In this paper we describe a deep learning system that has been designed and built for the WASSA 2017 Emotion Intensity Shared Task. We introduce a representation learning approach based on inner attention on top of an RNN. Results show that our model offers good capabilities and is able to successfully identify emotion-bearing words to predict intensity without leveraging on lexicons, obtaining the 13t place among 22 shared task competitors.

In this paper we present the model used by the team Rivercorners for the 2017 RepEval shared task. First, our model separately encodes a pair of sentences into variable-length representations by using a bidirectional LSTM. Later, it creates fixed-length raw representations by means of simple aggregation functions, which are then refined using an attention mechanism. Finally it combines the refined representations of both sentences into a single vector to be used for classification. With this model we obtained test accuracies of 72.057% and 72.055% in the matched and mismatched evaluation tracks respectively, outperforming the LSTM baseline, and obtaining performances similar to a model that relies on shared information between sentences (ESIM). When using an ensemble both accuracies increased to 72.247% and 72.827% respectively.

We propose a model to automatically describe changes introduced in the source code of a program using natural language. Our method receives as input a set of code commits, which contains both the modifications and message introduced by an user. These two modalities are used to train an encoder-decoder architecture. We evaluated our approach on twelve real world open source projects from four different programming languages. Quantitative and qualitative results showed that the proposed approach can generate feasible and semantically sound descriptions not only in standard in-project settings, but also in a cross-project setting.