Commonsense Reasoning

Commonsense knowledge often omits the temporal incidence of facts, and even the ordering between occurrences is only available for some of their instances. Reasoning about the temporal extent of facts and their sequencing becomes complex due to this inherent partiality. The generation of hypotheses is adopted here as a natural way to overcome the difficulties in computing answers to temporal queries. The proposed abductive system performs temporal reasoning in a logic programming framework. Queries are taken as goals and the inference system combines deduction with abduction and constraint solving. The convenience of constraints for dealing with temporal information is widely recognized, their interest being twofold: the representation of essential properties of time and the provision for partial information, allowing flexible bounds on times instead of constant bindings. Inference manipulates a language associating propositions with time periods which are maximal intervals for the proposition. The abductive inference procedure is described here, identifying the constraint operations required. It is also shown that the outcome of a derivation is alwavs consistent with the information in the knowledge basc.

We first present a short introduction illustrating how argumentation could be viewed as an universal mechanism humans use in their practical reasoning where by practical reasoning we mean both commonsense reasoning and reasoning by experts as well as their integration. We then present logic-based argumentation employing implicit or explicit assumptions. Logic alone is not enough for practical reasoning as it can not deal with quantitative uncertainties. We explain how probabilities could be integrated with argumentation to provide an integrated framework for jury-based (or collective multiagent) dispute resolution.

This paper presents the STAR system, a system for automated narrative comprehension, developed on top of an argumentation-theoretic formulation of defeasible reasoning, and strongly following guidelines from the psychology of comprehension. We discuss the system’s use in psychological experiments on story comprehension, and our plans for its broader use in empirical studies concerning wider issues of commonsense reasoning.

Recent advancements in large language models have enabled them to perform well on complex tasks that require step-by-step reasoning with few-shot learning. However, it is unclear whether these models are applying reasoning skills they have learned during pre-training, or if they are simply memorizing their training corpus at finer granularity and have learned to better understand their context. To address this question, we introduce ALERT, a benchmark and suite of analyses for evaluating reasoning skills of language models. ALERT enables comparing pre-trained and finetuned models on complex tasks that require reasoning skills to solve them. Our benchmark provides a test bed to assess any language model on fine-grained reasoning skills, which spans over 20 datasets and covers 10 different reasoning skills. To prove the efficacy of ALERT we investigate the role of finetuning. Our extensive empirical analysis shows that language models acquire reasoning skills such as textual entailment, abductive reasoning, and analogical reasoning during the finetuning stage compared to pretraining stage. Another finding is when language models are finetuned they tend to overfit to the prompt template, which hurts the robustness of models resulting in generalization problems.

In this paper, we conduct a thorough investigation into the reasoning capabilities of Large Language Models (LLMs), focusing specifically on the Open Pretrained Transformers (OPT) models as a representative of such models. Our study entails finetuning three different sizes of OPT on a carefully curated reasoning corpus, resulting in two sets of finetuned models: OPT-R, finetuned without explanations, and OPT-RE, finetuned with explanations. We then evaluate all models on 57 out-of-domain tasks drawn from the SUPER-NATURALINSTRUCTIONS benchmark, covering 26 distinct reasoning skills, utilizing three prompting techniques. Through a comprehensive grid of 27 configurations and 6,156 test evaluations, we investigate the dimensions of finetuning, prompting, and scale to understand the role of explanations on different reasoning skills. Our findings reveal that having explanations in the fewshot exemplar has no significant impact on the model's performance when the model is finetuned, while positively affecting the non-finetuned counterpart. Moreover, we observe a slight yet consistent increase in classification accuracy as we incorporate explanations during prompting and finetuning, respectively. Finally, we offer insights on which skills benefit the most from incorporating explanations during finetuning and prompting, such as Numerical (+20.4%) and Analogical (+13.9%) reasoning, as well as skills that exhibit negligible or negative effects.

Recent advancements in large language models have enabled them to perform well on complex tasks that require step-by-step reasoning with few-shot learning. However, it is unclear whether these models are applying reasoning skills they have learned during pre-training, or if they are simply memorizing their training corpus at finer granularity and have learned to better understand their context. To address this question, we introduce ALERT, a benchmark and suite of analyses for evaluating reasoning skills of language models. ALERT enables comparing pre-trained and finetuned models on complex tasks that require reasoning skills to solve them. Our benchmark provides a test bed to assess any language model on fine-grained reasoning skills, which spans over 20 datasets and covers 10 different reasoning skills. To prove the efficacy of ALERT we investigate the role of finetuning. Our extensive empirical analysis shows that language models acquire reasoning skills such as textual entailment, abductive reasoning, and analogical reasoning during the finetuning stage compared to pretraining stage. Another finding is when language models are finetuned they tend to overfit to the prompt template, which hurts the robustness of models resulting in generalization problems.

Large language models, which are often trained for hundreds of thousands of compute days, have shown remarkable capabilities for zero-and few-shot learning. Given their computational cost, these models are difficult to replicate without significant capital. For the few that are available through APIs, no access is granted to the full model weights, making them difficult to study. We present Open Pre-trained Transformers (OPT), a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters, which we aim to fully and responsibly share with interested researchers. We show that OPT-175B is comparable to GPT-3, 1 while requiring only 1/7th the carbon footprint to develop. We are also releasing our logbook detailing the infrastructure challenges we faced, along with code for experimenting with all of the released models. * Equal contribution. † Work done while at Meta AI. 1 Following Brown et al. (2020), we use GPT-3 to refer to both the 175B model and the smaller scale models as well. 2 Exceptions include work by EleutherAI, who released dense models up to 20B in size , Salesforce (Nijkamp et al., 2022), and Meta AI, who released dense models up to 13B and sparse models up to 1. 1T (Artetxe et al., 2021). There is also ongoing work from the BigScience workshop (), which aims to open source very large multilingual language models and datasets.

We report on the design and development of the CASPR system, a socialbot designed to compete in the Amazon Alexa Socialbot Challenge 4. CASPR’s distinguishing characteristic is that it will use automated commonsense reasoning to truly “understand” dialogs, allowing it to converse like a human. Three main requirements of a socialbot are that it should be able to “understand” users’ utterances, possess a strategy for holding a conversation, and be able to learn new knowledge. We developed techniques such as conversational knowledge template (CKT) to approximate commonsense reasoning needed to hold a conversation on specific topics. We present the philosophy behind CASPR’s design as well as details of its implementation. We also report on CASPR’s performance as well as discuss lessons learned.

Rich computer simulations or quantitative models can enable an agent to realistically predict real-world behavior with precision and performance that is difficult to emulate in logical formalisms. Unfortunately, such simulations lack the deductive flexibility of techniques such as formal logics and so do not find natural application in the deductive machinery of commonsense or general purpose reasoning systems. This dilemma can, however, be resolved via a hybrid architecture that combines tableaux-based reasoning with a framework for generic simulation based on the concept of 'molecular' models. This combination exploits the complementary strengths of logic and simulation, allowing an agent to build and reason with automatically constructed simulations in a problem-sensitive manner.

Every day we use intuitive reasoning to make and update predictions about the world. For instance, when playing soccer we must predict the trajectories of the ball but also update those predictions in light of new information. But these predictions must also be rapid – if we turn to see a soccer ball flying at us, we must quickly decide whether or not to duck. What prediction mechanism would allow for such fast and flexible predictions?

Large Language Models (LLMs) can be used as repositories of biological and chemical information to generate pharmacological lead compounds. However, for LLMs to focus on specific drug targets typically requires experimentation with progressively more refined prompts. Results thus become dependent not just on what is known about the target, but also on what is known about the prompt- engineering. In this paper, we separate the prompt into domain-constraints that can be written in a standard logical form and a simple text-based query. We investigate whether LLMs can be guided, not by refining prompts manually, but by refining the logical component automatically, keeping the query unchanged. We describe an iterative procedure LMLF (“Language Model with Logical Feedback”) in which the constraints are progressively refined using a logical notion of generalisation. On any iteration, newly generated instances are verified against the constraint, providing "logical-feedback" for t...

The proof of the Second Incompleteness Theorem consists essentially of proving the uniqueness and explicit definability of the sentence asserting its own unprovability. This turns out to be a rather general phenomenon: Every instance of self-reference describable in the modal logic of the standard proof predicate obeys a similar uniqueness and explicit definability law. The efficient determination of the explicit definitions of formulae satisfying a given instance of self-reference reduces to a simple algebraic problem--that of solving the corresponding fixed-point equation in the modal logic. We survey techniques for the efficient calculation of such fixed--points. GSdel elevated the s~atus of self-reference from that of a mere curiosity (e.g. the liar paradox) to that of a useful metamathematical tool (the Diagonalization Theorem). It is, however, a largely neglected tool whose occasional use in metamathematies elicits noises of surprise from the spectators. In the hope of reducing the level of this sm~prise we devote two papers to reviewing and extending our knowledge of self-referentiM sentences. It happens that one can go a long way in understanding such sentences by means of the most simple-minded of considerations. If we ~gree to refer to self-referentiM sentences as fixed--points, we can compress our simple-mindedness into the little motto: Calculate fixed-points! Calculations are possible because quite often a self-referentiM sentence is actually implicitly defined by the diagonalization through which it refers to itself. The transformation of the implicit definition into an explicit one constitutes a calculation. The dichotomy between the cMculations of this paper and those of the sequel is governed by the extent to which the calculations can be formalized in a given theory. In this paper, we consider a restricted class of self-referential sentences and present Mgorithms for explicit calculations of the members of this class. As this class of self-referentiM sentences has only three known ~pplications (the Incompleteness Theorems and ]55b's Theorem) and can be shown to have no others (Solovay's Completeness Theorem--see [21]~ we pursue the present calculation primarily for aesthetic reasons. The self-referential formulae with numerous applications do not seem to be so evplicitly cMeulable~ i.e. although we can give explicit definitions for them, ~ve c~nnot formalize the proof of the correctness of the definitions. These non-evlplicit calculations form the topic of the nex~ paper. 20 Craig Smorynsl~i ~odulo the results just cited, the first two sections of this paper are self-contMned. This is no longer true of Section 3~ where we assume some knowledge of modal model theory.

We are interested in systems which do not prescribe one single kind of preference, but in which varying kinds of preferences can be used simultaneously. In such systems it is essential to know the interaction among the kinds of preferences being used, and we therefore introduce and study a nonmonotonic logic to reason about sixteen strict and non-strict kinds of preferences, including ceteris paribus preferences. Moreover, we study "distinguished" preference orders based on specificity principles by showing when these distinguished preference orders are unique, and by presenting algorithms to calculate the distinguished preference orders.

The structure-preference (SP) order is a way of defining argument preference relations in structured argumentation theory that takes into account how arguments are constructed. The SP order was first introduced in the context of endowing Brewka’s prioritised default logic (PDL) with sound and complete argumentation semantics. In this paper, we further articulate the underlying intuitions of the SP order in terms of how an agent should construct arguments. We also compare the SP order to other argument preference relations and illustrate the different results one would obtain. Finally, we prove that the SP order allows for the original version of PDL to satisfy Brewka’s and Eiter’s postulates.

We endow prioritised default logic (PDL) with argumentation semantics using the ASPIC + framework for structured argumentation, and prove that the conclusions of the justified arguments are exactly the prioritised default extensions. Argumentation semantics for PDL will allow for the application of argument game proof theories to the process of inference in PDL, making the reasons for accepting a conclusion transparent and the inference process more intuitive. This also opens up the possibility for argumentation-based distributed reasoning and communication amongst agents with PDL representations of mental attitudes.

A Logic of Arbitrary and Indefinite Objects, LA, has been developed as the logic for knowledge representation and reasoning systems designed to support natural language understanding and generation, and commonsense reasoning. The motivations for the design of LA are given, along with an informal introduction to the theory of arbitrary and indefinite objects, and to LA itself. LA is then formally defined by presenting its syntax, proof theory, and semantics, which are given via a translation scheme between LA and the standard classical First-Order Predicate Logic. Soundness is proved. The completeness theorem for LA is stated, and its proof is sketched. LA is being implemented as the logic of SNePS 3, the latest member of the SNePS family of Knowledge Representation and Reasoning systems.

We discuss the value of argumentation in reaching agreements, based on its capability for dealing with conflicts and uncertainty. Logic-based models of argumentation have recently emerged as a key topic within Artificial Intelligence. Key reasons for the success of these models is that they are akin to human models of reasoning and debate, and their generalisation to frameworks for modelling dialogues. They therefore have the potential for bridging between human and machine reasoning in the presence of uncertainty and conflict. We provide an overview of a number of examples that bear witness to this potential, and that illustrate the added value of argumentation. These examples amount to methods and techniques for argumentation to aid machine reasoning (e.g. in the form of machine learning and belief functions) on the one hand and methods and techniques for argumentation to aid human reasoning (e.g. for various forms of decision making and deliberation and for the Web) on the other....

This paper explores how active logic empowers robots to reason about goals, actions, expectations, and deadlines during collaborative search tasks. These time-aware agents operate with future-oriented expectations and adapt their behavior when violations occur. We examine the challenges of managing expectations for time-situated robots and their outcomes, aiming to improve their ability to complete tasks effectively. On a target search task, we discuss initial results with two Clearpath robots, Husky and Jackal.

Art is imaginative human creation meant to be appreciated, make people think, and evoke an emotional response. Here for the first time, we create a dataset of more than 4,000 pieces of art (mostly paintings) that has annotations for emotions evoked in the observer. The pieces of art are selected from WikiArt.org’s collection for four western styles (Renaissance Art, Post-Renaissance Art, Modern Art, and Contemporary Art). The art is annotated via crowdsourcing for one or more of twenty emotion categories (including neutral). In addition to emotions, the art is also annotated for whether it includes the depiction of a face and how much the observers like the art. The dataset, which we refer to as the WikiArt Emotions Dataset, can help answer several compelling questions, such as: what makes art evocative, how does art convey different emotions, what attributes of a painting make it well liked, what combinations of categories and emotions evoke strong emotional response, how much does...

We present DEGARI (Dynamic Emotion Generator And ReclassIfier), an explainable system for emotion attribution and recommendation. This system relies on a recently introduced commonsense reasoning framework, the T CL logic, which is based on a human-like procedure for the automatic generation of novel concepts in a Description Logics knowledge base. Starting from an ontological formalization of emotions based on the Plutchik model, known as ArsEmotica, the system exploits the logic T CL to automatically generate novel commonsense semantic representations of compound emotions (e.g. Love as derived from the combination of Joy and Trust according to Plutchik). The generated emotions correspond to prototypes, i.e. commonsense representations of given concepts, and have been used to reclassify emotion-related contents in a variety of artistic domains, ranging from art datasets to the editorial contents available in RaiPlay, the online platform of RAI Radiotelevisione Italiana (the Italian public broadcasting company). We show how the reported results (evaluated in the light of the obtained reclassifications, the user ratings assigned to such reclassifications, and their explainability) are encouraging, and pave the way to many further research directions.

The concept of strength or weight of an argument appears frequently in argumentation theory. Its function is to explain how reasons or arguments interact to support their outcomes. Most approaches, however, use it as little more than a metaphor. This talk aims to be a contribution to the study of the concept of argument strength.
I will argue first that weighing is irreducible to defeasibility, and then that we can find two types of non-rhetorical concepts of strength in the literature of argumentation theory. On the one hand, we have a comparative concept, linked to the weighing of pairs or groups of opposing arguments, and on the other hand, a scalar concept, linked to modal qualifiers. Lack of awareness of the differences between these two senses makes many discussions of the strength of arguments somewhat confusing. This paper aims to explain these differences in detail to avoid confusion and fruitless or even counterproductive discussions.

This article examines the problem of forming basic concepts in robots within the framework of the development of Artificial General Intelligence (AGI). The theories of concept formation in infants were reviewed. There is the consensus that in this early period of human cognitive development the basic concepts and reasoning are established that was named "common sense". In the article common sense will be seen from two perspectives: individual and collective ones. This article is devoted to the formation of individual concepts of common sense in robots. In the set of individual concepts of common sense are being discussed here spatial and temporal conceptual domains prevail. Collective concepts of common sense will be mentioned briefly. In this phase of AGI development the proposed concepts formation procedure means mostly a-priori concepts that are preliminary designed as the software procedures and only "is grounded" in robot. But these "first concepts" create the basis for further learning procedures. As the platform for conceptualization procedure the cognitive architecture AGICA was considered. Cognitive architecture AGICA was represented by the author in 2023 on the base of "axiomatic approach" in AGI development. In cognitive architecture AGICA there were used the models of AGI-Consciousness, AGI-Individual Type, AGI-Collective Type in the framework of "grounded cognition". AGI-Individual Type is based on "instinct of self-preservation" (survival instinct). AGI-Collective Type is based on "species preservation instinct". Given that we are now far from a general theory of concepts, the development of transport robots is viewed as an application domain. This article does not claim to be comprehensive and can be viewed as some "engineering approach" for problem solving,-thus it is mostly addressed to the developers. In the main part of the article we will consider only perceptual/modal/concrete concepts. Amodal/abstract concepts will be out of the discussion.

The logical reasoning capabilities of pretrained language models have recently received much attention. As one of the vital reasoning paradigms, non-monotonic reasoning refers to the fact that conclusions may be invalidated with new information. Existing work has constructed a nonmonotonic inference dataset δ-NLI and explored the performance of language models on it. However, the δ-NLI dataset is entangled with commonsense reasoning. In this paper, we explore the pure non-monotonic reasoning ability of pre-trained language models. We build a non-monotonic reasoning benchmark, named LogicNMR, with explicit default rules and iterative updates. In the experimental part, the performance of popular language models on LogicNMR is explored from the perspectives of accuracy, generalization, proof-based traceability and robustness. The experimental results show that even though the fine-tuned language models achieve an accuracy of more than 94.4% on LogicNMR, they perform unsatisfactorily, with a significant drop, in generalization and proof-based traceability.

We report on an ongoing research program to develop a formal framework for automated narrative text comprehension, bringing together know-how from research in Artificial Intelligence and the Psychology of Reading and Comprehension. It uses argumentation to capture appropriate solutions to the frame, ramification, and qualification problems, and their generalizations as required for text comprehension. In this first part of the study we concentrate on the central problem of integration of the explicit information from the text narrative with the reader's implicit commonsense world knowledge, and the associated tasks of elaboration and revision.

This paper reports on how work in psychology can inform research into automated narrative text comprehension. Specifically, we describe the psychology's view of how human readers combine information from text with their own commonsense world knowledge to build a single coherent model of a narrative. We present a framework that formalizes, using a suitable form of preference-based argumentation, this notion of comprehension and outline how this might form the basis for partially mechanizing the process of comprehension.

Pre-trained language models (PLMs) have shown impressive performance in various language tasks. However, they are prone to spurious correlations, and often generate illusory information. In real-world applications, PLMs should justify decisions with formalized, coherent reasoning chains, but this challenge remains under-explored. Cognitive psychology theorizes that humans are capable of utilizing fast and intuitive heuristic thinking to make decisions based on past experience, then rationalizing the decisions through slower and deliberative analytic reasoning. We incorporate these interlinked dual processes in fine-tuning and in-context learning with PLMs, applying them to two language understanding tasks that require coherent physical commonsense reasoning. We show that our proposed Heuristic-Analytic Reasoning (HAR) strategies drastically improve the coherence of rationalizations for model decisions, yielding state-of-the-art results on Tiered Reasoning for Intuitive Physics (TRIP). We also find that this improved coherence is a direct result of more faithful attention to relevant language context in each step of reasoning. Our findings suggest that human-like reasoning strategies can effectively improve the coherence and reliability of PLM reasoning.

Reasoning agents are often faced with the need to robustly deal with erroneous information. When a robot given the task of returning with the red cup from the kitchen table arrives in the kitchen to find no red cup but instead notices a blue cup and a red plate on the table, what should it do? The best course of action is to attempt to salvage the situation by relying on its preferences to return with one of the objects available. We provide a solution to this problem using the Situation Calculus extended with a notion of belief. We then provide an efficient practical implementation by mapping this formalism into default rules for which we have an implemented solver.

Visual Dialog requires an agent to engage in a conversation with humans grounded in an image. Many studies on Visual Dialog focus on the understanding of the dialog history or the content of an image, while a considerable amount of commonsense-required questions are ignored. Handling these scenarios depends on logical reasoning that requires commonsense priors. How to capture relevant commonsense knowledge complementary to the history and the image remains a key challenge. In this paper, we propose a novel model by Reasoning with Multi-structure Commonsense Knowledge (RMK). In our model, the external knowledge is represented with sentence-level facts and graph-level facts, to properly suit the scenario of the composite of dialog history and image. On top of these multistructure representations, our model can capture relevant knowledge and incorporate them into the vision and semantic features, via graph-based interaction and transformerbased fusion. Experimental results and analysis on VisDial v1.0 and VisDialCK datasets show that our proposed model effectively outperforms comparative methods.

Previously we have proposed a logic, called priority logic 18, 20], where a theory consists of a collection of logic programming-like inference rules (without default negation) and a priority constraint among them. We showed that nonmonotonic reasoning in general can be viewed as selecting monotonic inferences that satisfy the speci ed priority constraints. The goal of this paper is to investigate how semantics of disjunctive programs can be understood as selecting monotonic inferences. We focus on credulous semantics. We show that the stable semantics of a disjunctive program can be represented by priority logic. The highlight of this paper is an interesting new semantics for disjunctive programs, discoved during this study, which has an abductive interpretation in the sense of Eshghi and Kowalski, whose abductive procedure can be adopted to serve as a credulous proof procedure for the new semantics.

Inspired by the cognitive science theory, we explicitly model an agent with both semantic and episodic memory systems, and show that it is better than having just one of the two memory systems. In order to show this, we have designed and released our own challenging environment, "the Room", compatible with OpenAI Gym, where an agent has to properly learn how to encode, store, and retrieve memories to maximize its rewards. The Room environment allows for a hybrid intelligence setup where machines and humans can collaborate. We show that two agents collaborating with each other results in better performance than one agent acting alone. We have open-sourced our code and models at .

This paper addresses identification of implicit requirements (IMRs) in software requirements specifications (SRS). IMRs, as opposed to explicit requirements, are not specified by users but are more subtle. It has been noticed that IMRs are crucial to the success of software development. In this paper, we demonstrate a software tool called COTIR developed by us as a system that integrates Commonsense knowledge, Ontology and Text mining for early identification of Implicit Requirements. This relieves human software engineers from the tedious task of manually identifying IMRs in huge SRS documents. Our evaluation reveals that COTIR outperforms existing IMR tools. This demo paper would be useful to Software Engineers since it deals with automation in the requirements analysis phase, thus contributing to Requirements Engineering. It would interest AI scientists as it entails multi-disciplinary work encompassing text mining, ontology and commonsense knowledge. It makes a broader impact on Smart Cities, because automated identification of IMRs would offer inputs to Smart City Tools, where requirements may often be implicit given that Smart Cities are an emerging and growing paradigm.

Chain-of-thought (CoT) advances the reasoning abilities of large language models (LLMs) and achieves superior performance in complex reasoning tasks. However, most CoT studies rely on carefully designed human-annotated rational chains to prompt LLMs, posing challenges for real-world applications where labeled data is available without rational chains. This paper proposes a new strategy, Automate-CoT (Automatic Prompt Augmentation and Selection with Chain-of-Thought), that can bypass human engineering of CoT by automatically augmenting rational chains from a small labeled dataset, and then pruning low-quality chains to construct a candidate pool of machinegenerated rationale chains based on the labels. Finally, it selects the optimal combination of several rationale chains from the pool for CoT prompting by employing a variance-reduced policy gradient strategy to estimate the significance of each example. Automate-CoT enables a quick adaptation of the CoT technique to different tasks. Experimental results demonstrate the effectiveness of our method, where competitive results are achieved on arithmetic reasoning (+2.7%), commonsense reasoning (+3.4%), symbolic reasoning (+3.2%), and nonreasoning tasks (+2.5%). 1

There has been an explosion of work in the vision & language community during the past few years from image captioning to video transcription, and answering questions about images. These tasks have focused on literal descriptions of the image. To move beyond the literal, we choose to explore how questions about an image are often directed at commonsense inference and the abstract events evoked by objects in the image. In this paper, we introduce the novel task of Visual Question Generation (VQG), where the system is tasked with asking a natural and engaging question when shown an image. We provide three datasets which cover a variety of images from object-centric to event-centric, with considerably more abstract training data than provided to state-of-the-art captioning systems thus far. We train and test several generative and retrieval models to tackle the task of VQG. Evaluation results show that while such models ask reasonable questions for a variety of images, there is still a wide gap with human performance which motivates further work on connecting images with commonsense knowledge and pragmatics. Our proposed task offers a new challenge to the community which we hope furthers interest in exploring deeper connections between vision & language.

In formal systems for reasoning about actions, the ramification problem denotes the problem of handling indirect effects. These effects are not explicitly represented in action specifications but follow from general laws describing dependencies among components of the world state. An adequate treatment of indirect effects requires a suitably weakened version of the general law of persistence. It also requires a method to avoid unintuitive changes suggested by the aforementioned dependency laws. We propose a solution to the ramification problem that uses directed relations between two single effects, stating the circumstances under which the occurrence of the first causes the second. We argue for the necessity of an approach based on causality by elaborating the limitations of common paradigms employed to handle ramifications-the principle of categorization and the policy of minimal change. Our abstract solution is realized on the basis of a particular action calculus, namely, the fluent calculus.

Commonsense reasoning has proven exceedingly difficult both to model and to implement in artificial reasoning systems. This paper discusses some of the features of human reasoning that may account for this difficulty, surveys a number of reasoning systems and formalisms, and offers an outline of active logic, a non-classical paraconsistent logic that may be of some use in implementing commonsense reasoning.

Non-monotonic logics are examined and found to be inadequate as descriptions of reason maintenance systems (sometimes called truth maintenance systems). A logic is proposed that directly addresses the problem of characterizing the mental states of a reasoning agent attempting to reason with respect to some object theory. The proposed logic, propositional dynamic logic of derivation (PDLD), is given a semantics, and a sound and complete axiomatization. The descriptive power of PDLD is demonstrated by expressing various inferential control policies as PDLD formulae. In this note I will elaborate the propositional fragment of an axiomatic semantics of reason maintenance systems (RMS's) . The development of such a semantics stems from the desire to provide a declarative specification language for RMS's with particular emphasis on the description of the control of their reasoning processes, and to serve as a formal setting within which to compare and contrast the properties of different RMS's. * The research reported herein is funded in part by the Defense Advanced Research Projects Agency under contract number F30602-85-C-0033.

This paper shows how a nonmonotonic ILP system XHAIL can perform general-purpose learning and revision of temporal theories in a full-fledged Discrete Event Calculus (DEC) framework with several features (now introduced into ILP for the first time) for dynamically releasing fluents from the commonsense law of inertia and subjecting them to mathematically-defined gradual change trajectories. First, we review key aspects in the development of the DEC and prior attempts aimed at learning within simple fragments of it. Then, we introduce a new formulation of DEC, called the eXploratory Event Calculus (XEC), which (a) results in significantly reduced grounding size and execution times with respect to state-of-the-art translations of DEC from circumscriptive logic into Answer Set Programming (ASP), and which (b) provides a unifying framework that supports different policies used (explicitly or implicitly) in prior work to resolve conflicts between concurrent events competing to determine the truth value and/or release status of fluents under a circumscriptive or answer set semantics. Finally, we give the first known demonstration of learning and revision of theories in the full DEC by using XHAIL and XEC in a proof-of-principle example showing how such theories can be revised under different conflict resolution policies.

Inductive Logic Programming (ILP) is concerned with the task of generalising sets of positive and negative examples with respect to background knowledge expressed as logic programs. Negation as Failure (NAF) is a key feature of logic programming which provides a means for nonmonotonic commonsense reasoning under incomplete information. But, so far, most ILP research has been aimed at Horn programs which exclude NAF, and has failed to exploit the full potential of normal programs that allow NAF. By contrast, Abductive Logic Programming (ALP), a related task concerned with explaining observations with respect to a prior theory, has been well studied and applied in the context of normal logic programs. This paper shows how ALP can be used to provide a semantics and proof procedure for nonmonotonic ILP that utilises practical methods of language and search bias to reduce the search space. This is done by lifting an existing method called Hybrid Abductive Inductive Learning (HAIL) from Horn clauses to normal logic programs. To demonstrate its potential benefits, the resulting system, called XHAIL, is applied to a process modelling case study involving a nonmonotonic temporal Event Calculus (EC).

Inductive Logic Programming (ILP) is concerned with the task of generalising sets of positive and negative examples with respect to background knowledge expressed as logic programs. Negation as Failure (NAF) is a key feature of logic programming which provides a means for nonmonotonic commonsense reasoning under incomplete information. But, so far, most ILP research has been aimed at Horn programs which exclude NAF, and has failed to exploit the full potential of normal programs that allow NAF. By contrast, Abductive Logic Programming (ALP), a related task concerned with explaining observations with respect to a prior theory, has been well studied and applied in the context of normal logic programs. This paper shows how ALP can be used to provide a semantics and proof procedure for nonmonotonic ILP that utilises practical methods of language and search bias to reduce the search space. This is done by lifting an existing method called Hybrid Abductive Inductive Learning (HAIL) from Horn clauses to normal logic programs. To demonstrate its potential benefits, the resulting system, called XHAIL, is applied to a process modelling case study involving a nonmonotonic temporal Event Calculus (EC).

This paper introduces a novel method for generating artistic images that express particular affective states. Leveraging state-of-the-art deep learning methods for visual generation (through generative adversarial networks), semantic models from OpenAI, and the annotated dataset of the visual art encyclopedia WikiArt, our AffectGAN model is able to generate images based on specific or broad semantic prompts and intended affective outcomes. A small dataset of 32 images generated by AffectGAN is annotated by 50 participants in terms of the particular emotion they elicit, as well as their quality and novelty. Results show that for most instances the intended emotion used as a prompt for image generation matches the participants' responses. This smallscale study brings forth a new vision towards blending affective computing with computational creativity, enabling generative systems with intentionality in terms of the emotions they wish their output to elicit.

This paper introduces a novel method for generating artistic images that express particular affective states. Leveraging state-of-the-art deep learning methods for visual generation (through generative adversarial networks), semantic models from OpenAI, and the annotated dataset of the visual art encyclopedia WikiArt, our AffectGAN model is able to generate images based on specific or broad semantic prompts and intended affective outcomes. A small dataset of 32 images generated by AffectGAN is annotated by 50 participants in terms of the particular emotion they elicit, as well as their quality and novelty. Results show that for most instances the intended emotion used as a prompt for image generation matches the participants' responses. This smallscale study brings forth a new vision towards blending affective computing with computational creativity, enabling generative systems with intentionality in terms of the emotions they wish their output to elicit.

Sentences that ascribe action are logically related, but it is not always obvious why. According to event semantics, implications and non-implications result from referential relations between unpronounced constituents. Taking as starting point examples including free relative clauses, this paper advances the alternative view that examples as such present logical relations as forms of predicative dependence indicated with pronounced constituents. To this end, I argue that Verbal Phrases and verbal traces follow the pattern of Verbal Phrase Anaphora and, more controversially, that they can be semantically interpreted in terms of higher-order quantification that represent actions as properties, but neither as events nor event-kinds.

WC argue that the question selection proccsscs used in the existing AZ in Afct-licine programs arc inadequate. WC trace thcsc inadcquacics to their use of purely surface lcvcl models of discasc and to the lack of planning in sequencing their inquiries. WC then describe the information gathering component of AlEI, (a program for Acid-Base and Hcctrolyte disturbances) with emphasis on its USC of a causal model of a patient's illness in guiding its diagnostic inquiries. Diagnostic programs built using the .4/ in Mcriicinc (AIM) nlti;hodoiogy typically consist of two parts: one IU furm hypo~l~cs~s that explain the data known to the program. and one to acquire new information from the program's users to clarify difficulties remaining with the hypotheses under consideration. In this paper WC suggest that the questioning behavior of most existing diagnostic programs is inadequate because it is not very rcsponsivc to the totality of the progri~l~'s hypotheses. IIotb rule-based systems such as Mycin IShortliffc761 and frame-based systems such as lntcrnist-I [Poplc77] and PIP [Paukcr76] suffer from this dcfcct, though for diverse reasons. Mycin [Shortliffe76] is a rule-based expert program for the diagnosis and therapy of infectious discascs. It uses a backward chaming control structure to scqucncc through its rule-base in search of information rclcvant in the diagnosis of possible infections. Mycin always pursues the antecedent clauses of a rule in their order of appearance in the rule. 'I'hus during dny cxcoution of ihc program the antecedent clauses of every rule are expanded in exactly the same order, resulting in a pre-dctcrmincd vrdcr in which the questions may be asked." This rigidity in Mycin results from its USC of rules to encode both its medical knowledge and the control component of expert reasoning. 'l'hc resulting limitations of the Mycin system were recognized by Davis [Davis76], who proposed mcra-rules as a means of separating thcsc two aspects of cxpcrt bchabior.' Meta-rules are used in his system, 'I'eircsias [l>avis76], to encode strategic knowledge for the control of the backward chaining rule interpreter by dynamically re-ordering and selecting rules to be pursued. It is important to note that the meta-rules do not alter the program's domain knowledge; they provide additional knowlcdgc that allows it to make more efficient use of this knowledge.

Commonsense reasoning refers to the ability of evaluating a social situation and acting accordingly. Identification of the implicit causes and effects of a social context is the driving capability which can enable machines to perform commonsense reasoning. The dynamic world of social interactions requires context-dependent on-demand systems to infer such underlying information. However, current approaches in this realm lack the ability to perform commonsense reasoning upon facing an unseen situation, mostly due to incapability of identifying a diverse range of implicit social relations. Hence they fail to estimate the correct reasoning path. In this paper, we present Conditional Seq2Seq-based Mixture model (CosMo), which provides us with the capabilities of dynamic and diverse content generation. We use CosMo to generate context-dependent clauses, which form a dynamic Knowledge Graph (KG) on-the-fly for commonsense reasoning. To show the adaptability of our model to context-dependan...

Contextualized or discourse aware commonsense inference [1] is the task of generating commonsense assertions (i.e., facts) from a given story, and a sentence from that story. (Here, we think of a story as a sequence of causally-related events and descriptions of situations.) This task is hard, even for modern contextual language models. Some problems with the task are: lack of controllability for topics of the inferred assertions; lack of commonsense knowledge during pre-training; and, possibly, hallucinated or false assertions. The task's goals are to make sure that (1) the generated assertions are plausible as commonsense; and (2) to assure that they are appropriate to the particular context of the story. We utilize a transformer model as a base inference engine to infer commonsense assertions from a sentence within the context of a story. With our inference engine we address lack of controllability, lack of sufficient commonsense knowledge, and plausibility of assertions through three techniques. We control the inference by introducing a new technique we call "hinting". Hinting is a kind of language model prompting [2], that utilizes both hard prompts (specific words) and soft prompts (virtual learnable templates). This serves as a control signal to advise the language model "what to talk about". Next, we establish a methodology for performing joint inference with multiple commonsense knowledge bases. While in logic, joint inference is just a matter of a conjunction of assertions, joint inference of commonsense requires more care, because it is imprecise and the level of generality is more flexible. You want to be sure that the results "still make sense" for the context. To this end, we align the assertions in three knowledge graphs (ConceptNet [3], ATOMIC2020 [4], and GLUCOSE [5]) with a story and a target sentence, and replace their symbolic assertions with textual versions of them. This combination allows us to train a single model to perform joint inference with multiple knowledge graphs. We show experimental results for the three knowledge graphs on joint inference. Our final contribution is a GAN architecture that generates the contextualized commonsense inference from stories and scores the generated assertions as to their plausibility through a discriminator. The result is an integrated system for contextual commonsense inference in stories, that can controllably generate plausible commonsense assertions, and takes advantage of joint inference between multiple commonsense knowledge bases.

Story understanding systems need to be able to perform commonsense reasoning, specifically regarding characters' goals and their associated actions. Some efforts have been made to form large-scale commonsense knowledge bases, but integrating that knowledge into story understanding systems remains a challenge. We have implemented the Aspire system, an application of large-scale commonsense knowledge to story understanding. Aspire extends Genesis, a rule-based story understanding system, with tens of thousands of goalrelated assertions from the commonsense semantic network ConceptNet. Aspire uses ConceptNet's knowledge to infer plausible implicit character goals and story causal connections at a scale unprecedented in the space of story understanding. Genesis's rule-based inference enables precise story analysis, while ConceptNet's relatively inexact but widely applicable knowledge provides a significant breadth of coverage difficult to achieve solely using rules. Genesis uses Aspire's inferences to answer questions about stories, and these answers were found to be plausible in a small study. Though we focus on Genesis and ConceptNet, demonstrating the value of supplementing precise reasoning systems with large-scale, scruffy commonsense knowledge is our primary contribution.

Recent efforts in natural language processing (NLP) commonsense reasoning research have yielded a considerable number of new datasets and benchmarks. However, most of these datasets formulate commonsense reasoning challenges in artificial scenarios that are not reflective of the tasks which real-world NLP systems are designed to solve. In this work, we present CROW, a manually-curated, multitask benchmark that evaluates the ability of models to apply commonsense reasoning in the context of six real-world NLP tasks. CROW is constructed using a multi-stage data collection pipeline that rewrites examples from existing datasets using commonsense-violating perturbations. We use CROWto study how NLP systems perform across different dimensions of commonsense knowledge, such as physical, temporal, and social reasoning. We find a significant performance gap when NLP systems are evaluated on CROWcompared to humans, showcasing that commonsense reasoning is far from being solved in real-world task settings. We make our dataset and leaderboard available to the research community. 1 * Equal contribution 1 https://github.com/mismayil/crow Dialogue Agent: Hi, would you like some free candies? Human: Sure. What are you handing these out for? Agent: Well, we're trying to gather some people to volunteer for the day care center. Human: Uh… Agent: It's OK. You don't have to volunteer if you eat the candies. Agent: It's OK. You don't have to eat the candies if you volunteer. Artificial Evaluation CRoW Evaluation (real-world) doesn't have prerequisite NOT OK OK OK Bob decided to volunteer because he wanted to eat the candies.