Case-Based Reasoning Research and Development

Goal-driven autonomy (GDA) is a reflective model of goal reasoning combining deliberative planning and plan execution monitoring. GDA’s is the focus of increasing interest due in part to the need to ensure that autonomous agents behave as intended. However, to perform well, comprehensive GDA agents require substantial domain knowledge. In this paper I focus on our work to automatically learn knowledge used by GDA agents. I also discuss future research directions.

Every year tenths of thousands of customer support engineers around the world deal with, and proactively solve, complex help-desk tickets. Daily, almost every customer support expert will turn his/her attention to a prioritization strategy, to achieve the best possible result. To assist with this, in this paper we describe a novel case-based reasoning application to address the tasks of: high solution accuracy and shorter prediction resolution time. We describe how appropriate cases can be generated to assist engineers and how our solution can scale over time to produce domain-specific reusable cases for similar problems. Our work is evaluated using data from 5000 cases from the automotive industry.

The high influence of case bases quality on Case-Based Reasoning success gives birth to an important study on cases competence for problems resolution. The competence of a case base (CB), which presents the range of problems that it can successfully solve, depends on various factors such as the CB size and density. Besides, it is not obvious to specify the exactly relationship between the individual and the overall cases competence. Hence, numerous Competence Models have been proposed to evaluate CBs and predict their actual coverage and competence on problem-solving. However, to the best of our knowledge, all of them are totally neglecting the uncertain aspect of information which is widely presented in cases since they involve real world situations. Therefore, this paper presents a new competence model called CEC-Model (Coverage & Evidential Clustering based Model) which manages uncertainty during both of cases clustering and similarity measurement using a powerful tool called the belief function theory.

Case-based reasoning (CBR) systems often refer to diverse machine learning functionalities and algorithms to augment their capabilities. In this article we review the concept of case based learning and define it as the use of case based reasoning for machine learning. We present some of its characteristics and situate it in the context of the major machine learning tasks and machine learning approaches. In doing so, we review the particular manner in which case based learning practices declarative learning, for its main knowledge containers, as well as dynamic induction, through similarity assessment. The central role of analogy as a dynamic induction is highlighted as the cornerstone of case based learning that makes it a method of choice in classification and prediction tasks in particular. We propose a larger understanding, beyond instance-based learning, of case based learning as analogical learning that would promote it as a major contributor of the analogizer approach of machine learning.

Cognitive Systems understand the world through learning and experience. Case Based Reasoning (CBR) systems naturally capture knowledge as experiences in memory and they are able to learn new experiences to retain in their memory. CBR’s retrieve and reuse reasoning is also knowledge-rich because of its nearest neighbour retrieval and analogy-based adaptation of retrieved solutions. CBR is particularly suited to domains where there is no well-defined theory, because they have a memory of experiences of what happened, rather than why/how it happened. CBR’s assumption that ‘similar problems have similar solutions’ enables it to understand the contexts for its experiences and the ‘bigger picture’ from clusters of cases, but also where its similarity assumption is challenged. Here we explore cognition and meta-cognition for CBR through self-reflection and introspection of both memory and retrieve and reuse reasoning. Our idea is to embed and exploit cognitive functionality such as insight, intuition and curiosity within CBR to drive robust, and even explainable, intelligence that will achieve problem-solving in challenging, complex, dynamic domains.

This paper addresses the issues of explainability of case-based support systems, particularly structural CBR systems dominated by knowledge-rich comprehensive cases and domain models. We show how explanation patterns and contextually enriched explanations of retrieval results can provide human-understandable insights on the system behavior, justify the shown results, and recommend the best cases to be considered for further use. We applied and implemented our approach as an agent-based system module within a case-based assistance framework for support of the early conceptual phases in architectural design, taking a single floor plan as a case with a high number of attributes. For the retrieval phase, a semantic search pattern structure, Semantic Fingerprint, was applied, whereas the explanation generation phase is controlled by a number of explanation patterns adapted from already existing explanation goals. Rulesets, case bases, and natural language generation are used for construction and automatic revision of explanation expressions. A contextualization feature categorizes the results into different context classes and includes this information into the explanation. A user study we conducted after the implementation of the explanation algorithm resulted in good acceptance by the representatives of the architectural domain, a quantitative experiment revealed a high rate of valid generated explanations and a reasonable distribution of patterns and contexts.

In any knowledge investigation by which a user must acquire new or missing information, situations often arise which lead to a fork in their investigation. Multiple possible lines of inquiry appear that the users must choose between. A choice of any one would delay the user’s ability to choose another, if the chosen path proves to be irrelevant and happens to yield only useless information. With limited knowledge or experience, a user must make assumptions which serve as justifications for their choice of a particular path of inquiry. Yet incorrect assumptions can lead the user to choose a path that ultimately leads to dead-end. These fruitless lines of inquiry can waste both time and resources by adding confusion and noise to the user’s investigation. Here we evaluate an algorithm called Tangent Recognition Anomaly Pruning to eliminate false starts that arise in interactive dialogues created within our case-based reasoning system called Ronin. Results show that Tangent Recognition Anomaly Pruning is an effective algorithm for processing mistakes when reusin case reuse.

In this paper we present an approach for combining Case-based Reasoning (CBR) and Complex Event Processing (CEP) in order to classify network traffic. We show that this combination has a high potential to improve existing classification methods by enriching the stream processing techniques in CEP with the capability of historic case reuse in CBR by continuously analysing the application layer data of network communication.

AI-Virtual Trainer (AI-VT) is an intelligent tutoring system based on case-based reasoning. AI-VT has been designed to generate personalised, varied, and consistent training sessions for learners. The AI-VT training sessions propose different exercises in regard to a capacity associated with sub-capacities. For example, in the field of training for algorithms, a capacity could be “Use a control structure alternative” and an associated sub-capacity could be “Write a boolean condition”. AI-VT can elaborate a personalised list of exercises for each learner. One of the main requirements and challenges studied in this work is its ability to propose varied training sessions to the same learner for many weeks, which constitutes the challenge studied in our work. Indeed, if the same set of exercises is proposed time after time to learners, they will stop paying attention and lose motivation. Thus, even if the generation of training sessions is based on analogy and must integrate the repetition of some exercises, it also must introduce some diversity and AI-VT must deal with this diversity. In this paper, we have highlighted the fact that the retaining (or capitalisation) phase of CBR is of the utmost importance for diversity, and we have also highlighted that the equilibrium between repetition and variety depends on the abilities learned. This balance has an important impact on the retaining phase of AI-VT.

When faced with an exceptional clinical case, doctors like to review information about similar patients to guide their decision-making. Retrieving relevant cases, however, is a hard and time-consuming task: Hospital databases of free-text physician letters provide a rich resource of information but are usually only searchable with string-matching methods. Here, we present a recommender system that automatically finds physician letters similar to a specified reference letter using an information retrieval procedure. We use a small-scale, prototypical dataset to compare the system’s recommendations with physicians’ similarity judgments of letter pairs in a psychological experiment. The results show that the recommender system captures expert intuitions about letter similarity well and is usable for practical applications.

Building predictive models is central to many big data applications. However, model building is computationally costly at scale. An appealing alternative is bypassing model building by applying case-based prediction to reason directly from data. However, to our knowledge case-based prediction still has not been applied at true industrial scale. In previous work we introduced a knowledge-light/data intensive approach to case-based prediction, using ensembles of automatically-generated adaptations. We developed foundational scaleup methods, using Locality Sensitive Hashing (LSH) for fast approximate nearest neighbor retrieval of both cases and adaptation rules, and tested them for millions of cases. This paper presents research on extending these methods to address the practical challenges raised by case bases of hundreds of millions of cases for a real world industrial e-commerce application. Handling this application required addressing how to keep LSH practical for skewed data; the resulting efficiency gains in turn enabled applying an adaptation generation strategy that previously was computationally infeasible. Experimental results show that our CBR approach achieves accuracy comparable to or better than state of the art machine learning methods commonly applied, while avoiding their model-building cost. This supports the opportunity to harness CBR for industrial scale prediction.

Case-based reasoning can resolve new problems based on remembering and adapting the solution of similar problems. Before a CBR system can solve new problems it must be provided with an initial case base covering the problem space with a sufficient number of representative seed cases with solutions that are known to be correct. We use a CBR module to recommend leisure plans in Madrid based on user preferences and contextual information. This paper deals with the problem of how to build and evaluate an initial case base of leisure experiences in Madrid for the recommender system.

Due to the development of the global security situation, the existence and implementation of security concepts became an important aspect of public events. The definition and writing of a security concept demands domain knowledge and experience. This paper describes an approach for the automated retrieval and generation of security concept templates based on reliable examples. We use ontologies for the conceptualization of textual security concepts, and we employ case-based reasoning for the retrieval and generation of new security concepts.

Case-base maintenance research has extensively studied strategies for competence-retaining case base compression. Such approaches generally rely on the representativeness assumption that current case base contents can be used as a proxy for future problems when determining cases to retain. For mature case bases in stable domains, this assumption works well. However, representativeness may not hold for sparse case bases during initial case base growth, for dynamically changing domains, or when a case base built for one task is applied to cross-domain problem-solving in another. This paper presents a new method for competence-preserving deletion, Expansion-Contraction Compression (ECC), aimed at improving competence preservation when the representativeness assumption is only partially satisfied. ECC precedes compression with adaptation-based exploration of previously unseen parts of the problem space to create “ghost cases” and exploits them to broaden the range of cases available for competence-based deletion. Experimental results support that this method increases competence and quality retention for less representative case bases. They also reveal the unexpected result that ECC can improve retention of competence and quality even for representative case bases.

Nowadays, security forces are challenged by a new type of terrorist propaganda which occurs in public social networks and targets vulnerable individuals. The current volume of online radicalization messages has rendered manual monitoring approaches unfeasible, and effective countermeasures can only be adopted through early detection by automatized tools. Some approaches focus on mining the information provided by social users in the form of interactions and textual content. However, radical users also tend to exhibit distinctive iconography in their profile images. In this work, we propose the use of local image descriptors over profile images to aid the detection and monitoring of online radicalization processes. In addition, we complement this approach with an interaction-based formula for risk assessment, so candidate profiles can be selected for image-analysis based on their interaction with confirmed radical profiles. These techniques are combined in the context of a Case-Based Reasoning framework which, together with the feedback provided by the end-user, enables a continuous monitoring of the activity of radical users and eases the discovery of new profiles with a radicalization agenda.

Image segmentation is a hot topic in image processing research. Most of the time, segmentation is not fully automated, and a user is required to guide the process in order to obtain correct results. Yet, even with programs, it is a time-consuming process. In a medical context, segmentation can provide a lot of information to surgeons, but since this task is manual, it is rarely executed because of time. Artificial Intelligence (AI) is a powerful approach to create viable solutions for fully automated treatments. In this paper, we define a case-based reasoning (CBR) that can enhance region-growing segmentation of kidneys deformed by nephroblastoma. The main problem with region-growing methods is that a user needs to place the seeds in the image manually. Automated methods exist but they are not efficient every time and they often give an over-segmentation. That is why we have designed an adaptation phase which can modify the coordinates of seeds recovered during the retrieval phase. We compared our CBR approach with manual region growing and Convolutional Neural Networking (CNN) to segment kidneys and tumours of CT-scans. Our CBR system succeeded in performing the best segmentation for the kidney.

As people live longer, the increasing average age of the population places additional strains on our health and social services. There are widely recognised benefits to both the individual and society from supporting people to live independently for longer in their own homes. However, falls in particular have been found to be a leading cause of the elderly moving into care, and yet surprisingly preventative approaches are not in place; fall detection and rehabilitation are too late. In this paper we present FITsense, which is building a Smart Home environment to identify increased risk of falls for residents, and so allow timely interventions before falls occurs. An ambient sensor network, installed in the Smart Home, identifies low level events taking place which is analysed to generate a resident’s profile of activities of daily living (ADLs). These ADL profiles are compared to both the resident’s typical profile and to known “risky” profiles to allow evidence-driven intervention recommendations. Human activity recognition to identify ADLs from sensor data is a key challenge. Here we compare a windowing-based and a sequence-based event representation on four existing datasets. We find that windowing works well, giving consistent performance but may lack sufficient granularity for more complex multi-part activities.

Case indexing decisions must often confront the tradeoff between rich semantic indexing schemes, which provide effective retrieval at large indexing cost, and shallower indexing schemes, which enable low-cost indexing but may be less reliable. Indexing for textual case-based reasoning is often based on information retrieval approaches that minimize index acquisition cost but sacrifice semantic information. This paper presents JointEmbed, a method for automatically generating rich indices. JointEmbed automatically generates continuous vector space embeddings that implicitly capture semantic information, leveraging multiple knowledge sources such as free text cases and pre-existing knowledge graphs. JointEmbed generates effective indices by applying pTransR, a novel approach for modelling knowledge graphs, to encode and summarize contents of domain knowledge resources. JointEmbed is applied to the medical CBR task of retrieving relevant patient electronic health records, for which potential health consequences make retrieval quality paramount. An evaluation supports that JointEmbed outperforms previous methods.

Data masking is a means to protect data from unauthorized access by third parties. In this paper, we propose a case-based assistance system for data masking that reuses experience on substituting (pseudonymising) the values of database fields. The data masking experts use rules that maintain task-oriented properties of the data values, such as the environmental hazards risk class of residential areas when masking address data of insurance customers. The rules transform operational data into hardly traceable, masked data sets, which are to be applied, for instance, during software test management in the insurance sector. We will introduce a case representation for masking a database column, including problem descriptors about structural properties and value properties of the column as well as the data masking rule as the solution part of the case. We will describe the similarity functions and the implementation of the approach by means of myCBR. Finally, we report about an experimental evaluation with a case base of more than 600 cases and 31 queries that compares the results of a case-based retrieval with the solutions recommended by a data masking expert.

Imitating video game players is considered one of the most stimulating challenges for the Game AI research community. The goal for a virtual player is not just to beat the game but to show some human-like playing style. In this work we describe a Case-Based Reasoning approach that learns to play the popular Ms. Pac-Man vs Ghosts video game from the traces of a human player. We evaluate the performance of our bot using both low level standard measures such as accuracy and recall, and high level measures such as recklessness (distance to the closest ghost, as it is mapped in our Ms. Pac-Man domain model), restlessness (changes of direction), aggressiveness (ghosts eaten), clumsiness (game steps the player is stuck) and survival (lives left). Results suggest that, although there is still a lot of room for improvement, some aspects of the human playing style are indeed captured in the cases and used by our bot.

Case-Based Reasoning (CBR) is a lazy learning method and, being such, when a new query is made to a CBR system, the swiftness of its retrieval phase proves to be very important for the overall system performance. The availability of ubiquitous data today is an opportunity for CBR systems as it implies more cases to reason with. Nevertheless, this availability also introduces a challenge for the CBR retrieval since distance calculations become computationally expensive. A good example of a domain where the case base is subject to substantial growth over time is the health records of patients where a query is typically an incremental update to prior cases. To deal with the retrieval performance challenge in such domains where cases are sequentially related, we introduce a novel method which significantly reduces the number of cases assessed in the search of exact nearest neighbors (NNs). In particular, when distance measures are metrics, they satisfy the triangle inequality and our method leverages this property to use it as a cutoff in NN search. Specifically, the retrieval is conducted in a lazy manner where only the cases that are true NN candidates for a query are evaluated. We demonstrate how a considerable number of unnecessary distance calculations is avoided in synthetically built domains which exhibit different problem feature characteristics and different cluster diversity.

This study presents a case-based reasoning (CBR) system that makes use of general domain knowledge - referred to as a knowledge-intensive CBR system. The system applies a Bayesian analysis aimed at increasing the accuracy of the similarity assessment. The idea is to employ the Bayesian posterior distribution for each case symptom to modify the case descriptions and the dependencies in the model. To evaluate the system, referred to as BNCreek, two experiment sets are set up from a “food” and an “oil well drilling” application domain. In both of the experiments, the BNCreek is evaluated against two corresponding systems named TrollCreek and myCBR with Normalized Discounted Cumulative Gain (NDCG) and interpolated average Precision-Recall as the evaluation measures. The obtained results reveal the capability of Bayesian analysis to increase the accuracy of the similarity assessment.

Human Activity Recognition (HAR) is typically modelled as a classification task where sensor data associated with activity labels are used to train a classifier to recognise future occurrences of these activities. An important consideration when training HAR models is whether to use training data from a general population (subject-independent), or personalised training data from the target user (subject-dependent). Previous evaluations have shown personalised training to be more accurate because of the ability of resulting models to better capture individual users’ activity patterns. From a practical perspective however, collecting sufficient training data from end users may not be feasible. This has made using subject-independent training far more common in real-world HAR systems. In this paper, we introduce a novel approach to personalised HAR using a neural network architecture called a matching network. Matching networks perform nearest-neighbour classification by reusing the class label of the most similar instances in a provided support set, which makes them very relevant to case-based reasoning. A key advantage of matching networks is that they use metric learning to produce feature embeddings or representations that maximise classification accuracy, given a chosen similarity metric. Evaluations show our approach to substantially out perform general subject-independent models by at least 6% macro-averaged F1 score.

When case-based recommender systems use preference-based feedback, we can learn user preferences by using the trade-off relations between the preferred product and the other products in the given domain. In this work, we propose a representation for trade-offs and motivate several mechanisms by which the identified trade-offs can be used in the process of recommendation. We empirically demonstrate the effectiveness of the proposed approaches in three recommendation domains.

We use case-based reasoning to help marathoners achieve a personal best for an upcoming race, by helping them to select an achievable goal-time and a suitable pacing plan. We evaluate several case representations and, using real-world race data, highlight their performance implications. Richer representations do not always deliver better prediction performance, but certain representational configurations do offer very significant practical benefits for runners, when it comes to predicting, and planning for, challenging goal-times during an upcoming race.

Building a case base for a case-based reasoning (CBR) system is incomplete without similarity measures. For the attribute-value case structure similarity between values of an attribute should logically fit their relationship. Bipartite graphs have been shown to be a good representation of relationships between values of symbolic attributes and the diagnosis of the cases in a technical diagnosis CBR system, while using an asymmetrical weighted one-mode projection on the values to model their similarity.However, the weighted one-mode projection assumes that the set of symbols is static, which is contradictory to the dynamic nature of case bases as defined by the retain phase of the CBR cycle. In this work we present two methods to update the similarity measure whenever new information is available and compare them. We show that even though updating the similarity measure to exactly reflect the case base had the new information been available a-priori produces better results, an imperfect update is a feasible, less time consuming temporary solution.

The development of Convolutional Neural Networks (CNNs) has resulted in significant improvements to object classification and detection in image data. One of their primary benefits is that they learn image features rather than relying on hand-crafted features, thereby reducing the amount of knowledge engineering that must be performed. However, another form of knowledge engineering bias exists in how objects are labelled in images, thereby limiting CNNs to classifying the set of object types that have been predefined by a domain expert. We describe a case-based method for detecting novel object types using a combination of an image’s raw pixel values and detectable parts. Our approach works alongside existing CNN architectures, thereby leveraging the state-of-the-art performance of CNNs, and is able to detect novel classes using limited training instances. We evaluate our approach using an existing object detection dataset and provide evidence of our approach’s ability to classify images even if the object in the image has not been previously encountered.

Objective measurements of physical behaviour are an interesting research field from the public health and computer science perspective. While for public health research, measurements with a high quality and feasible setup is important, the analysis of and reasoning about the data is what we will present in this work. Our focus in this work is the comprehensive representation of physical behaviour throughout consecutive days and allowing to find subgroups in the population with similar physical activity levels.We have a unique data set of 4628 participants wearing tri-axial accelerometers for six days and will present a case-based reasoning (CBR) system that can find and compare similar activity profiles. In this work, we focus on creating a CBR model using myCBR and do initial experiments with the resulting system. We will introduce a data-driven approach for modelling local similarity measures. Eventually, in the experiments we will show that for the given data set, the CBR system outperforms a k-Nearest Neighbor regressor in finding most similar participants.

This paper demonstrates how case-based reasoning (CBR) can be used for an explainable artificial intelligence (XAI) approach to justify solutions produced by an opaque learning method (i.e., target method), particularly in the context of unstructured textual data. Our general hypothesis is twofold: (1) There exists patterns in the relationship between problems and solutions and there should be data or a body of knowledge that describes how problems and solutions relate; and (2) the identification, manipulation, and learning of such patterns through case features can help create and reuse explanations for solutions produced by the target method. When the target method relies on neural network architectures (e.g., deep learning), the resulting latent space (i.e., word embeddings) becomes useful for finding patterns and semantic relatedness in textual data. In the proposed approach, case problems are input-output pairs from the target method, and case solutions are explanations. We exemplify our approach by explaining recommended citations from Citeomatic - a multi-layer neural-network architecture from the Allen Institute for Artificial Intelligence. Citation analysis is the body of knowledge that describes how query documents (i.e., inputs) relate to recommended citations (i.e., outputs). We build cases and similarity assessment to learn features that represent patterns between problems and solutions that can lead to the reuse of corresponding explanations. The illustrative implementation we present becomes an explanation-augmented citation recommender that targets human-computer trust.

Multiple sensor modalities provide more accurate Human Activity Recognition (HAR) compared to using a single modality, yet the latter is preferred by consumers as it is more convenient and less intrusive. This presents a challenge to researchers, as a single modality is likely to pick up movement that is both relevant as well as extraneous to the human activity being tracked and lead to poorer performance. The goal of an optimal HAR solution is therefore to utilise the fewest sensors at deployment, while maintaining performance levels achievable using all available sensors. To this end, we introduce two translation approaches, capable of generating missing modalities from available modalities. These can be used to generate missing or “privileged” modalities at deployment to augment case representations and improve HAR. We evaluate the presented translators with k-NN classifiers on two HAR datasets and achieve up-to $$5\%$$ performance improvements using representations augmented with privileged modalities. This suggests that non-intrusive modalities suited for deployment benefit from translation models that generates privileged modalities.

This paper investigates the generation of recipes in consideration of user-defined nutrient contents. For this purpose, we extend our previous case-based reasoning approach that already covers the formulation of user queries with various dietary practices. More precisely, this work augments the domain ontology with nutritional information and introduces a novel nutrition concept fulfillment into the retrieval and adaptation process. An experimental evaluation with real cooking recipes demonstrates the applicability of the approach and systematically investigates the influence of various adaptation methods on the query fulfillment with multiple constraints. It is shown, that all adaptation methods are able to optimize generated recipes according to certain nutritional constraints as well as ingredient and cooking step preferences and that the adaptation outperforms the sole retrieval of available recipes.

Unknown Unknowns (UUs) are referred to the error predictions that with high confidence. The identifying of the UUs is important to understand the limitation of predictive models. Some proposed solutions are effective in such identifying. All of them assume there is a perfect Oracle to return the correct labels of the UUs. However, it is not practical since there is no perfect Oracle in real world. Even experts will make mistakes in UUs labelling. Such errors will lead to the terrible consequence since fake UUs will mislead the existing algorithms and reduce their performance. In this paper, we identify the impact of noisy Oracle and propose a UUs identifying algorithm that can be adapted to the setting of noisy Oracle. Experimental results demonstrate the effectiveness of our proposed method.

Analogical transfer consists in making the assumption that if two situations are alike in some respect, they may be alike in others. This article explores the links that exist between analogical transfer and the qualitative measurement of differences. The main idea is to formulate the similarity principle as a dependency between two measurements of difference. Analogical transfer is formulated as a similarity-based reasoning: it is plausible that equally different pairs in a certain dimension are also equally different in another dimension, at least for pairs that are not too (analogically) dissimilar.

It is assumed that preferences between two items, described in terms of criteria values belonging to a finite scale, are known for a limited number of pairs of items, which constitutes a case base. The problem is then to predict the preference between the items of a new pair. A new approach based on analogical proportions is presented. Analogical proportions are statements of the form “a is to b as c is to d”. If the change between item-1 and item-2 is the same as the change between item-3 and item-4, and a similar statement holds for item’-1, item’-2, item’-3, item’-4, then one may plausibly assume that the preference between item-1 and item’-1 is to the preference between item-2 and item’-2 as the preference between item-3 and item’-3 is to the preference between item-4 and item’-4. This offers a basis for a plausible prediction of the fourth preference if the three others are known. This approach fits well with the postulates underlying weighted averages. Two algorithms are proposed that look for triples of preferences appropriate for a prediction. The first one only exploits the given set of examples. The second one completes this set with new preferences deducible from this set under a monotony assumption. This completion is limited to the generation of preferences that are useful for the requested prediction. The predicted preferences should fit with the assumption that known preferences agree with a unique unknown weighted average. The reported experiments suggest the effectiveness of the proposed approach.

Algorithms for finding analogies as mappings between pairs of concepts are fundamental to some implementations of Conceptual Blending (CB), a theory which has been suggested as explaining some cognitive processes behind the creativity phenomenon. When analogies are defined as sub-isomorphisms of semantic graphs, we find ourselves with a NP-complete problem. In this paper we propose and compare a new high performance stochastic mapper that efficiently handles semantic graphs containing millions of relations between concepts, while outputting in real-time analogy mappings ready for use by another algorithm, such as a computational system based on CB theory.

We introduce a method to automatically produce large analogical clusters from smaller seed clusters of representative examples. The method is based on techniques of processing and solving analogical equations in word vector space models, i.e., word embeddings. In our experiments, we use standard data sets in English which cover different relations extending from derivational morphology (like adjective–adverb, positive–comparative forms of adjectives) or inflectional morphology (like present–past forms) to encyclopedic semantics (like country–capital relations). The analogical clusters produced by our method are shown to be of reasonably good quality, as shown by comparing human judgment against automatic NDCG@n scores. In total, they contain 8.5 times as many relevant word pairs as the seed clusters.

This paper deals with case-based machine translation. It is based on a previous work using a proportional analogy on strings, i.e., a quaternary relation expressing that “String A is to string B as string C is to string D”. The first contribution of this paper is the rewording of this work in terms of case-based reasoning: a case is a problem-solution pair $$(A, A')$$ where A is a sentence in an origin language and $$A'$$ , its translation in the destination language. First, three cases $$(A, A')$$ , $$(B, B')$$ , $$(C, C')$$ such that “A is to B as C is to the target problem D” are retrieved. Then, the analogical equation in the destination language “ $$A'$$ is to $$B'$$ as $$C'$$ is to x” is solved and $$D'=x$$ is a suggested translation of D. Although it does not involve any linguistic knowledge, this approach was effective and gave competitive results at the time it was proposed. The second contribution of this work aims at examining how this prior knowledge-light case-based machine translation approach could be improved by using additional pieces of knowledge associated with cases, domain knowledge, retrieval knowledge, and adaptation knowledge, and other principles or techniques from case-based reasoning and natural language processing.

Case-based reasoning usually exploits source cases (consisting of a source problem and its solution) individually, on the basis of the similarity between the target problem and a particular source problem. This corresponds to approximation. Then the solution of the source case has to be adapted to the target. We advocate in this paper that it is also worthwhile to consider source cases by two, or by three. Handling cases by two allows for a form of interpolation, when the target problem is between two similar source problems. When cases come by three, it offers a basis for extrapolation. Namely the solution of the target problem is obtained, when possible, as the fourth term of an analogical proportion linking the three source cases with the target, where the analogical proportion handles both similarity and dissimilarity between cases. Experiments show that interpolation and extrapolation techniques are of interest for reusing cases, either in an independent or in a combined way.

Analogical reasoning is a cognitively fundamental way of reasoning by comparing two pairs of elements. Several computational approaches are proposed to efficiently solve analogies: among them, a large number of practical methods rely on either a parallelogram representation of the analogy or, equivalently, a model of proportional analogy. In this paper, we propose to broaden this view by extending the parallelogram representation to differential manifolds, hence spaces where the notion of vectors does not exist. We show that, in this context, some classical properties of analogies do not hold any longer. We illustrate our considerations with two examples: analogies on a sphere and analogies on probability distribution manifold.

Analogical learning is a lazy learning mechanism which maps input forms (e.g. strings) to output ones, thanks to analogies identified in a training material. It has proven effective in a number of Natural Language Processing (NLP) tasks such as machine translation. One challenge with this approach is the solving of so-called analogical equations. In this paper, we investigate how structured learning can be used for learning to solve formal analogical equations. We evaluate our learning procedure on several test sets and show that we can improve upon fair baselines.