Extreme Multi-Label Classification (XMLC) plays a pivotal role in organizing and retrieving information in large-scale textual collections, by considering a very high number of potential labels for the documents. In this paper, we conduct an empirical evaluation of several XMLC approaches encompassing both dedicated techniques (AttentionXML and XR Transformer) and the use of Large Language Models (LLaMA2 7b Chat, LLaMA3 8b Instruct, and two Mistral models). We introduce both a new dataset based on OpenAlex as well as several new metrics to conduct our evaluations. Our results suggest that fine-tuning the LLMs using Low-Rank Adaptation significantly improves the performance of the models, bringing their results close to the ones of dedicated techniques. In the end, none of the method emerges as a clear winner, as picking the optimal XMLC technique heavily depends on the requirements of the use-case at hand.

Extreme Multi Label (XML) problems, and in particular XML completion -- the task of prediction the missing labels of an entity -- have attracted significant attention in the past few years. Most XML completion problems can organically leverage a label hierarchy, which can be represented as a tree that encodes the relations between the different labels.

Bridging unstructured data with knowledge bases is an essential task in many problems related to natural language understanding. Traditionally, this task is considered in one direction only: linking entity mentions in a text to their counterpart in a knowledge base (also known as entity linking). In this paper, we propose to tackle this problem from a different angle: linking entities from a knowledge base to paragraphs describing those entities. We argue that such a new perspective can be beneficial to several applications, including information retrieval, knowledge base population, and joint entity and word embedding. We present a transformer-based model, ParaGraph, which, given a Wikidata entity as input, retrieves its corresponding Wikipedia section. To perform this task, ParaGraph first generates an entity summary and compares it to sections to select an initial set of candidates. The candidates are then ranked using additional information from the entity’s textual description and contextual information. Our experimental results show that ParaGraph achieves 87% Hits@10 when ranking Wikipedia sections given a Wikidata entity as input. The obtained results show that ParaGraph can reduce the information gap between Wikipedia-based entities and tail entities and demonstrate the effectiveness of our proposed approach towards linking knowledge graph entities to their text counterparts.

Knowledge Graph (KG) completion has been widely studied to tackle the incompleteness issue (i.e., missing facts) in modern KGs. A fact in a KG is represented as a triplet (h, r, t) linking two entities h and t via a relation r. Existing work mostly consider link prediction to solve this problem, i.e., given two elements of a triplet predicting the missing one, such as (h, r, ?). This task has, however, a strong assumption on the two given elements in a triplet, which have to be correlated, resulting otherwise in meaningless predictions, such as (Marie Curie, headquarters location, ?). In addition, the KG completion problem has also been formulated as a relation prediction task, i.e., when predicting relations r for a given entity h. Without predicting t, this task is however a step away from the ultimate goal of KG completion. Against this background, this paper studies an instance completion task suggesting r-t pairs for a given h, i.e., (h, ?, ?). We propose an end-to-end solution called RETA (as it suggests the Relation and Tail for a given head entity) consisting of two components: a RETA-Filter and RETA-Grader. More precisely, our RETA-Filter first generates candidate r-t pairs for a given h by extracting and leveraging the schema of a KG; our RETA-Grader then evaluates and ranks the candidate r-t pairs considering the plausibility of both the candidate triplet and its corresponding schema using a newly-designed KG embedding model. We evaluate our methods against a sizable collection of state-of-the-art techniques on three real-world KG datasets. Results show that our RETA-Filter generates of high-quality candidate r-t pairs, outperforming the best baseline techniques while reducing by 10.61%-84.75% the candidate size under the same candidate quality guarantees. Moreover, our RETA-Grader also significantly outperforms state-of-the-art link prediction techniques on the instance completion task by 16.25%-65.92% across different datasets.

Neural networks have recently been shown to be highly effective at predicting links for constructing knowledge graphs. Existing research has mainly focused on designing 1) deep neural network models that are expressive in capturing fine-grained semantics, e.g., NTN and ConvE, but that are however less scalable; or 2) shallow models that are scalable, e.g., TransE and DistMult, yet limited in capturing expressive semantic features. In this work, we demonstrate that we can get the best of both worlds while drastically reducing the amount of data needed to train a deep network by leveraging active learning.