論文

Can negation be depicted? It has been claimed in various areas, including philosophy, cognitive science, and AI, that depicting negation through visual expressions such images and pictures is challenging. Recent empirical findings have shown that humans can indeed understand certain images as expressing negation, whereas this ability is not exhibited by machine learning models trained on image data. To elucidate the computational ability underlying the understanding of negation in images, this study first focuses on the image captioning task, specifically the performance of models pre-trained on large linguistic and image datasets for generating text from images. Our experiment demonstrates that a state-of-the-art model achieves some success in generating consistent captions from images, particularly in photographs rather than illustrations. However, when it comes to generating captions containing negation from images, the model is not as proficient as humans. To further investigate the performance of machine learning models in a more controlled setting, we conducted an additional analysis using a Visual Question Answering (VQA) task. This task enables us to specify where in the image the model should focus its attention when answering a question. As a result of this setting, the model’s performance was improved. These results will shed light on the disparities in the attentional focus between humans and machine learning models.

There is a widely held view that visual representations (images) do not depict negation, for example, as expressed by the sentence, “the train is not coming.” The present study focuses on the real-world visual representations of photographs and comic (manga) illustrations and empirically challenges the question of whether humans and machines, that is, modern deep neural networks, can recognize visual representations as expressing negation. By collecting data on the captions humans gave to images and analyzing the occurrences of negation phrases, we show some evidence that humans recognize certain images as expressing negation. Furthermore, based on this finding, we examined whether or not humans and machines can classify novel images as expressing negation. The humans were able to correctly classify images to some extent, as expected from the analysis of the image captions. On the other hand, the machine learning model of image processing was only able to perform this classification at about the chance level, not at the same level of performance as the human. Based on these results, we discuss what makes humans capable of recognizing negation in visual representations, highlighting the role of the background commonsense knowledge that humans can exploit. Comparing human and machine learning performances suggests new ways to understand human cognitive abilities and to build artificial intelligence systems with more human-like abilities to understand logical concepts.

Natural Language Inference (NLI) and Semantic Textual Similarity (STS) are widely used benchmark tasks for compositional evaluation of pre-trained language models. Despite growing interest in linguistic universals, most NLI/STS studies have focused almost exclu-sively on English. In particular, there are no available multilingual NLI/STS datasets in Japanese, which is typologically different from English and can shed light on the currently controversial behavior of language models in matters such as sensitivity to word order and case particles. Against this background, we introduce JSICK, a Japanese NLI/STS dataset that was manually trans-lated from the English dataset SICK. We also present a stress-test dataset for compositional inference, created by transforming syntactic structures of sentences in JSICK to investigate whether language models are sensitive to word order and case particles. We conduct baseline experiments on different pre-trained language models and compare the performance of multilingual models when applied to Japanese and other languages. The results of the stress-test experiments suggest that the current pre-trained language models are insensitive to word order and case marking.

This paper investigates whether current large language models exhibit biases in logical reasoning, similar to humans. Specifically, we focus on syllogistic reasoning, a well-studied form of inference in the cognitive science of human deduction. To facilitate our analysis, we introduce a dataset called NeuBAROCO, originally designed for psychological experiments that assess human logical abilities in syllogistic reasoning. The dataset consists of syllogistic inferences in both English and Japanese. We examine three types of biases observed in human syllogistic reasoning: belief biases, conversion errors, and atmosphere effects. Our findings demonstrate that current large language models struggle more with problems involving these three types of biases.

This paper explores the question of how accurately current large language models can perform logical reasoning in natural language, with an emphasis on whether these models exhibit reasoning biases similar to humans. Specifically, our study focuses on syllogistic reasoning, a form of deductive reasoning extensively studied in cognitive science as a natural form of human reasoning. We present a syllogism dataset called NeuBAROCO, which consists of syllogistic reasoning problems in English and Japanese. This dataset was originally designed for psychological experiments to assess human reasoning capabilities using various forms of syllogisms. Our experiments with leading large language models indicate that these models exhibit reasoning biases similar to humans, along with other error tendencies. Notably, there is significant room for improvement in reasoning problems where the relationship between premises and hypotheses is neither entailment nor contradiction. We also present experimental results and in-depth analysis using a new Chain-of-Thought prompting method, which asks LLMs to translate syllogisms into abstract logical expressions and then explain their reasoning process. Our analysis using this method suggests that the primary limitations of LLMs lie in the reasoning process itself rather than the interpretation of syllogisms.

Despite the recent success of deep neural networks in natural language processing, the extent to which they can demonstrate human-like generalization capacities for natural language understanding remains unclear. We explore this issue in the domain of natural language inference (NLI), focusing on the transitivity of inference relations, a fundamental property for systematically drawing inferences. A model capturing transitivity can compose basic inference patterns and draw new inferences. We introduce an analysis method using synthetic and naturalistic NLI datasets involving clause-embedding verbs to evaluate whether models can perform transitivity inferences composed of veridical inferences and arbitrary inference types. We find that current NLI models do not perform consistently well on transitivity inference tasks, suggesting that they lack the generalization capacity for drawing composite inferences from provided training examples. The data and code for our analysis are publicly available at https://github.com/verypluming/transitivity.

This paper presents a computational framework for Natural Language Inference (NLI) using logic-based semantic representations and theorem-proving. We focus on logical inferences with comparatives and other related constructions in English, which are known for their structural complexity and difficulty in performing efficient reasoning. Using the so-called A-not-A analysis of comparatives, we implement a fully automated system to map various comparative constructions to semantic representations in typed first-order logic via Combinatory Categorial Grammar parsers and to prove entailment relations via a theorem prover. We evaluate the system on a variety of NLI benchmarks that contain challenging inferences, in comparison with other recent logic-based systems and neural NLI models.

Discourse relations are sometimes explicitly conveyed by specific connectives. However, some connectives can signal multiple discourse relations; in such cases, disambiguation is necessary to determine which relation is intended. This task is known as discourse connective disambiguation (Pitler and Nenkova, 2009), and particular attention is often given to connectives that can convey both CONCESSION and other relations (e.g., SYNCHRONOUS). In this study, we conducted experiments to analyze which linguistic features play an important role in the disambiguation of polysemous connectives in Japanese. A neural language model (BERT) was fine-tuned using inputs from which specific linguistic features (e.g., word order, specific lexicon, etc.) had been removed. We analyzed which linguistic features affect disambiguation by comparing the model’s performance. Our results show that even after performing drastic removal, such as deleting one of the two arguments that constitute the discourse relation, the model’s performance remained relatively robust. However, the removal of certain lexical items or words belonging to specific lexical categories significantly degraded disambiguation performance, highlighting their importance in identifying the intended discourse relation.

Recently, deep neural networks (DNNs) have achieved great success in semantically challenging NLP tasks, yet it remains unclear whether DNN models can capture compositional meanings, those aspects of meaning that have been long studied in formal semantics. To investigate this issue, we propose a Systematic Generalization testbed based on Natural language Semantics (SyGNS), whose challenge is to map natural language sentences to multiple forms of scoped meaning representations, designed to account for various semantic phenomena. Using SyGNS, we test whether neural networks can systematically parse sentences involving novel combinations of logical expressions such as quantifiers and negation. Experiments show that Transformer and GRU models can generalize to unseen combinations of quantifiers, negations, and modifiers that are similar to given training instances in form, but not to the others. We also find that the generalization performance to unseen combinations is better when the form of meaning representations is simpler. The data and code for SyGNS are publicly available at https://github.com/verypluming/SyGNS.

Solving natural language inference (NLI) with formal semantics and automated theorem proving has the merit of high precision and interpretability. However, they suffer from non-logical inference such as lexical inference. To overcome this weakness, we propose a human-in-the-loop mechanism in which the user provides non-logical knowledge to the theorem prover. We focus on the subgoal of a proof. A subgoal is a proposition that remains unproved by the theorem prover. Although the subgoal conveys what knowledge should be supplemented to the proof, it is a logical formula inscrutable to normal users. To make the subgoal understandable to the user, we propose a method to translate the subgoal into a natural language text. The resulting sentence is called a readable subgoal. The reaction to the readable subgoal is communicated to the theorem prover in the form of temporary axiom, a logical formula that conveys the content of the reaction. We experimented with our method by adding an interactive component to an existing NLI system, ccg2lambda. The annotators recruited by Amazon Mechanical Turk used the proposed system to solve SICK, a data set for NLI. Experiments show the interaction improved the recall rate of ccg2lambda from 32.1% to 44.1%, and the accuracy from 70.4% to 75.1%.

There has been a widely held view that visual representations (e.g., photographs and illustrations) do not depict negation, for example, one that can be expressed by a sentence “the train is not coming”. This view is empirically challenged by analyzing the real-world visual representations of comic (manga) illustrations. In the experiment using image captioning tasks, we gave people comic illustrations and asked them to explain what they could read from them. The collected data showed that some comic illustrations could depict negation without any aid of sequences (multiple panels) or conventional devices (special symbols). This type of comic illustrations was subjected to further experiments, classifying images into those containing negation and those not containing negation. While this image classification was easy for humans, it was difficult for data-driven machines, i.e., deep learning models (CNN), to achieve the same high performance. Given the findings, we argue that some comic illustrations evoke background knowledge and thus can depict negation with purely visual elements.

Universal quantifiers have been the subject of much work in symbolic and diagrammatic logic. However, little attention is paid to the question of how they can be visually grounded, that is, depicted in real images such as photographs. To investigate this question, we focus on universal quantifiers such as “all” and “every” used in an image captioning dataset and present a qualitative analysis of these expressions. The analysis revealed that although the use of universal quantifiers in image captions is rare, there are interesting patterns in their usage in terms of the semantics of visual representations. We distinguish two ways in which universal quantifiers are used in image captions. One is object-based quantification, which involves quantification over multiple discrete objects in a definite domain. The other is region-based quantification, where some property is ascribed to a salient continuous region in an image. We compare these two ways of visually representing universal quantification with two major representation systems studied in diagrammatic logic.

Research in AI using Large-Language Models (LLMs) is rapidly evolving, and the comparison of their performance with human reasoning has become a key concern. Prior studies have indicated that LLMs and humans share similar biases, such as dismissing logically valid inferences that contradict common beliefs. However, criticizing LLMs for these biases might be unfair, considering our reasoning not only involves formal deduction but also abduction, which draws tentative conclusions from limited information. Abduction can be regarded as the inverse form of syllogism in its basic structure, that is, a process of drawing a minor premise from a major premise and conclusion. This paper explores the accuracy of LLMs in abductive reasoning by converting a syllogistic dataset into one suitable for abduction. It aims to investigate whether the state-of-the-art LLMs exhibit biases in abduction and to identify potential areas for improvement, emphasizing the importance of contextualized reasoning beyond formal deduction. This investigation is vital for advancing the understanding and application of LLMs in complex reasoning tasks, offering insights into bridging the gap between machine and human cognition.

Numeral expressions in Japanese are characterized by the flexibility of quantifier positions and the variety of numeral suffixes. However, little work has been done to build annotated corpora focusing on these features and datasets for testing the understanding of Japanese numeral expressions. In this study, we build a corpus that annotates each numeral expression in an existing phrase structure-based Japanese treebank with its usage and numeral suffix types. We also construct an inference test set for numerical expressions based on this annotated corpus. In this test set, we particularly pay attention to inferences where the correct label differs between logical entailment and implicature and those contexts such as negations and conditionals where the entailment labels can be reversed. The baseline experiment with Japanese BERT models shows that our inference test set poses challenges for inference involving various types of numeral expressions.

Despite the success of multilingual pre-trained language models, it remains unclear to what extent these models have human-like generalization capacity across languages. The aim of this study is to investigate the out-of-distribution generalization of pre-trained language models through Natural Language Inference (NLI) in Japanese, the typological properties of which are different from those of English. We introduce a synthetically generated Japanese NLI dataset, called the Japanese Adversarial NLI (JaNLI) dataset, which is inspired by the English HANS dataset and is designed to require understanding of Japanese linguistic phenomena and illuminate the vulnerabilities of models. Through a series of experiments to evaluate the generalization performance of both Japanese and multilingual BERT models, we demonstrate that there is much room to improve current models trained on Japanese NLI tasks. Furthermore, a comparison of human performance and model performance on the different types of garden-path sentences in the JaNLI dataset shows that structural phenomena that ease interpretation of garden-path sentences for human readers do not help models in the same way, highlighting a difference between human readers and the models.

This paper introduces a new video-and-language dataset with human actions for multimodal logical inference, which focuses on intentional and aspectual expressions that describe dynamic human actions. The dataset consists of 200 videos, 5,554 action labels, and 1,942 action triplets of the form (subject, predicate, object) that can be translated into logical semantic representations. The dataset is expected to be useful for evaluating multimodal inference systems between videos and semantically complicated sentences including negation and quantification.

Human-generated captions for photographs, particularly snapshots, have been extensively collected in recent AI research. They play a crucial role in the development of systems capable of multimodal information processing that combines vision and language. Recognizing that diagrams may serve a distinct function in thinking and communication compared to photographs, we shifted our focus from snapshot photographs to diagrams. We provided humans with text-free diagrams and collected data on the captions they generated. The diagrams were sourced from AI2D-RST, a subset of AI2D. This subset annotates the AI2D image dataset of diagrams from elementary school science textbooks with types of diagrams. We mosaicked all textual elements within the diagram images to ensure that human annotators focused solely on the diagram’s visual content when writing a sentence about what the image expresses. For the 831 images in our dataset, we obtained caption data from at least three individuals per image. To the best of our knowledge, this dataset is the first collection of caption data specifically for diagrams.

In recent years, research on large language models (LLMs) has been advancing rapidly, making the evaluation of their reasoning abilities a crucial issue. Within cognitive science, there has been extensive research on human reasoning biases. It is widely observed that humans often use graphical representations as auxiliary tools during inference processes to avoid reasoning biases. However, currently, the evaluation of LLMs’ reasoning abilities has largely focused on linguistic inferences, with insufficient attention given to inferences using diagrams. In this study, we concentrate on syllogisms, a basic form of logical reasoning, and evaluate the reasoning abilities of LLMs supplemented by Euler diagrams. We systematically investigate how accurately LLMs can perform logical reasoning when using diagrams as auxiliary input and whether they exhibit similar reasoning biases to those of humans. Our findings indicate that, overall, providing diagrams as auxiliary input tends to improve models’ performance, including in problems that show reasoning biases, but the effect varies depending on the conditions, and the improvement in accuracy is not as high as that seen in humans. We present results from experiments conducted under multiple conditions, including a Chain-of-Thought setting, to highlight where there is room to improve logical diagrammatic reasoning abilities of LLMs.

How logical concepts such as negation can be visually represented is of central importance in the study of diagrammatic reasoning. To explore various ways in which negation can be visually represented, this study focuses on photographs as instances of purely visual representations. We use real-world photographic image data and study how well humans can classify those images as depicting negation. We also compare the human performance with a state-of-the-art machine (deep) learning model on this classification task. The present paper gives some preliminary results on our data-driven analyses.