Entity search, i.e., finding the most similar entities to a query entity, faces unique challenges in e-commerce, where product similarity varies across categories and contexts. Traditional embedding-based approaches often struggle to capture nuanced context-specific attribute relevance. In this paper, we present a two-stage approach combining Large Language Model (LLM)-driven attribute graph construction with graph-aware LLM ranking. In the offline stage, we extract structured product attributes from unstructured text, and construct a reusable attribute graph with category-aware schemas. In the online stage, we rank retrieved candidates by reasoning over this structured representation rather than raw text, reducing per-product token usage by 57% while improving ranking precision. Experiments show that our approach outperforms multiple baselines under zero-shot scenarios, achieving a over 5% improvement in average precision without requiring training data, generalizes robustly across diverse product categories, and shows immense potential for real-world deployment.

The Knowledge-infused Learning Workshop is a recurring event in ACM's KDD Conference that gathers the research community on knowledge graphs and knowledge-enabled learning, grounded neurosymbolic AI, explainable and safe AI, and applications in highstakes decision-making problems.This year, the workshop aligned with Biden's vision of Responsible AI Development [1].

Digital assistants have become ubiquitous in e-commerce applications, following the recent advancements in Information Retrieval (IR), Natural Language Processing (NLP) and Generative Artificial Intelligence (AI). However, customers are often unsure or unaware of how to effectively converse with these assistants to meet their shopping needs. In this work, we emphasize the importance of providing customers a fast, easy to use, and natural way to interact with conversational shopping assistants. We propose a framework that employs Large Language Models (LLMs) to automatically generate contextual, useful, answerable, fluent and diverse questions about products, via in-context learning and supervised fine-tuning. Recommending these questions to customers as helpful suggestions or hints to both start and continue a conversation can result in a smoother and faster shopping experience with reduced conversation overhead and friction. We perform extensive offline evaluations, and discuss in detail about potential customer impact, and the type, length and latency of our generated product questions if incorporated into a real-world shopping assistant.

Conversation disentanglement aims to identify and group utterances from a conversation into separate threads. Existing methods primarily focus on disentangling multi-party conversations with three or more speakers, explicitly or implicitly incorporating speaker-related feature signals to disentangle. Most existing models require a large amount of human annotated data for model training, and often focus on pairwise relations between utterances, not accounting much for the conversational context. In this work, we propose a multi-task learning approach with a contrastive learning objective, DiSC, to disentangle conversations between two speakers -- a user and a virtual speech assistant, for a novel domain of e-commerce. We analyze multiple ways and granularities to define conversation "threads''. DiSC jointly learns the relation between pairs of utterances, as well as between utterances and their respective thread context. We train and evaluate our models on multiple multi-threaded conversation datasets that were automatically created, without any human labeling effort. Experimental results on public datasets as well as real-world shopping conversations from a commercial speech assistant show that DiSC outperforms state-of-the-art baselines by at least 3%, across both automatic and human evaluation metrics. We also demonstrate how DiSC improves downstream dialog response generation in the shopping domain.

An essential part of making shopping purchase decisions is to compare and contrast products based on key differentiating features, but doing this manually can be overwhelming. Prior methods offer limited product comparison capabilities, e.g., via pre-defined common attributes that may be difficult to understand, or irrelevant to a particular product or user. Automatically generating an informative, natural-sounding, and factually consistent comparative text for multiple product and attribute types is a challenging research problem. We describe HCPC (Human Centered Product Comparison), to tackle two kinds of comparisons for online shopping: (i) product-specific, to describe and compare products based on their key attributes; and (ii) attribute-specific comparisons, to compare similar products on a specific attribute. To ensure that comparison text is faithful to the input product data, we introduce a novel multi-decoder, multi-task generative language model. One decoder generates product comparison text, and a second one generates supportive, explanatory text in the form of product attribute names and values. The second task imitates a copy mechanism, improving the comparison generator, and its output is used to justify the factual accuracy of the generated comparison text, by training a factual consistency model to detect and correct errors in the generated comparative text. We release a new dataset (https://registry.opendata.aws/) of ~15K human generated sentences, comparing products on one or more attributes (the first such data we know of for product comparison). We demonstrate on this data that HCPC significantly outperforms strong baselines, by ~10% using automatic metrics, and ~5% using human evaluation.

In recent years, a plethora of fact checking and fact verification techniques have been developed to detect the veracity or factuality of online information text for various applications. However, limited efforts have been undertaken to understand the interpretability of such veracity detection, i.e. explaining why a particular piece of text is factually correct or incorrect. In this work, we seek to bridge this gap by proposing a technique, FACE-KEG, to automatically perform explainable fact checking. Given an input fact or claim, our proposed model constructs a relevant knowledge graph for it from a large-scale structured knowledge base. This graph is encoded via a novel graph transforming encoder. Our model also simultaneously retrieves and encodes relevant textual context about the input text from the knowledge base. FACE-KEG then jointly exploits both the concept-relationship structure of the knowledge graph as well as semantic contextual cues in order to (i) detect the veracity of an input fact, and (ii) generate a human-comprehensible natural language explanation justifying the fact's veracity. We conduct extensive experiments on three large-scale datasets, and demonstrate the effectiveness of FACE-KEG while performing fact checking. Automatic and human evaluations further show that FACE-KEG significantly outperforms competitive baselines in learning concise, coherent and informative explanations for the input facts.

Accurately discovering user intents from their written or spoken language plays a critical role in natural language understanding and automated dialog response. Most existing research models this as a classification task with a single intent label per utterance, grouping user utterances into a single intent type from a set of categories known beforehand. Going beyond this formulation, we define and investigate a new problem of open intent discovery. It involves discovering one or more generic intent types from text utterances, that may not have been encountered during training. We propose a novel domain-agnostic approach, OPINE, which formulates the problem as a sequence tagging task under an open-world setting. It employs a CRF on top of a bidirectional LSTM to extract intents in a consistent format, subject to constraints among intent tag labels. We apply a multi-head self-attention mechanism to effectively learn dependencies between distant words. We further use adversarial training to improve performance and robustly adapt our model across varying domains. Finally, we curate and plan to release an open intent annotated dataset of 25K real-life utterances spanning diverse domains. Extensive experiments show that our approach outperforms state-of-the-art baselines by 5-15% F1 score points. We also demonstrate the efficacy of OPINE in recognizing multiple, diverse domain intents with limited (can also be zero) training examples per unique domain.

Natural disasters such as floods, forest fires, and hurricanes can cause catastrophic damage to human life and infrastructure. We focus on response to hurricanes caused by both river water flooding and storm surge. Using models for storm surge simulation and flood extent prediction, we generate forecasts about areas likely to be highly affected by the disaster. Further, we overlay the simulation results with information about traffic incidents to correlate traffic incidents with other data modality. We present these results in a modularized, interactive map-based visualization, which can help emergency responders to better plan and coordinate disaster response.

Dynamically extracting and representing continually evolving knowledge entities is an essential scaffold for grounded intelligence and decision making. Creating knowledge schemas for newly emerging, unfamiliar, domain-specific ideas or events poses the following challenges: (i) detecting relevant, often previously unknown concepts associated with the new domain; and (ii) learning ontological, semantically accurate relationships among the new concepts, despite having severely limited annotated data. To this end, we propose a novel LSTM-based framework with attentive pooling, BOLT-K, to learn an ontology for a target subject or domain. We bootstrap our ontology learning approach by adapting and transferring knowledge from an existing, functionally related source domain. We also augment the inadequate labeled data available for the target domain with various strategies to minimize human expertise during model development and training. BOLT-K first employs semantic and graphical features to recognize the entity or concept pairs likely to be related to each other, and filters out spurious concept combinations. It is then jointly trained on knowledge from the target and source domains to learn relationships among the target concepts. The target concepts and their corresponding relationships are subsequently used to construct an ontology. We extensively evaluate our framework on several, real-world bio-medical and commercial product domain ontologies. We obtain significant improvements of 5-25% F1-score points over state-of-the-art baselines. We also examine the potential of BOLT-K in detecting the presence of novel kinds of relationships that were unseen during training.

The rising need to harvest domain specific knowledge in several applications is largely limited by the ability to dynamically grow structured knowledge representations, due to the increasing emergence of new concepts and their semantic relationships with existing ones. Such enrichment of existing hierarchical knowledge sources with new information to better model the "changing world" presents two-fold challenges: (1) Detection of previously unknown entities or concepts, and (2) Insertion of the new concepts into the knowledge structure, respecting the semantic integrity of the created relationships. To this end we propose a novel framework, ETF, to enrich large-scale, generic taxonomies with new concepts from resources such as news and research publications. Our approach learns a high-dimensional embedding for the existing concepts of the taxonomy, as well as for the new concepts. During the insertion of a new concept, this embedding is used to identify semantically similar neighborhoods within the existing taxonomy. The potential parent-child relationships linking the new concepts to the existing ones are then predicted using a set of semantic and graph features. Extensive evaluation of ETF on large, real-world taxonomies of Wikipedia and WordNet showcase more than 5% F1-score improvements compared to state-of-the-art baselines. We further demonstrate that ETF can accurately categorize newly emerging concepts and question-answer pairs across different domains.

The recent advancement of web-scale digital advertising saw a paradigm shift from the conventional focus of digital advertisement distribution towards integrating digital processes and methodologies and forming a seamless workflow of advertisement design, production, distribution, and effectiveness monitoring. In this work, we implemented a computational framework for the predictive analysis of the content-based features extracted from advertisement video files and various effectiveness metrics to aid the design and production processes of commercial advertisements. Our proposed predictive analysis framework extracts multi-dimensional temporal patterns from the content of advertisement videos using multimedia signal processing and natural language processing tools. The pattern analysis part employs an architecture of cross modality feature learning where data streams from different feature dimensions are employed to train separate neural network models and then these models are fused together to learn a shared representation. Subsequently, a neural network model trained on this joint representation is utilized as a classifier for predicting advertisement effectiveness. Based on the predictive patterns identified between the content features and the effectiveness metrics of advertisements, we have elicited a useful set of auditory, visual and textual patterns that is strongly correlated with the proposed effectiveness metrics while can be readily implemented in the design and production processes of commercial advertisements. We validate our approach using subjective ratings from a dedicated user study, the text sentiment strength of online viewer comments, and a viewer opinion metric of the likes/views ratio of each advertisement from YouTube video-sharing website.