Large language models (LLMs) have demonstrated human-level performance on a vast spectrum of natural language tasks. However, it is largely unexplored whether they can better internalize knowledge from a structured data, such as a knowledge graph, or from text. In this work, we propose a method to infuse structured knowledge into LLMs, by directly training T5 models on factual triples of knowledge graphs (KGs). We show that models pre-trained on Wikidata KG with our method outperform the T5 baselines on FreebaseQA and WikiHop, as well as the Wikidata-answerable subset of TriviaQA and NaturalQuestions. The models pre-trained on factual triples compare competitively with the ones on natural language sentences that contain the same knowledge. Trained on a smaller size KG, WikiMovies, we saw 3x improvement of exact match score on MetaQA task. The proposed method has an advantage that no alignment between the knowledge graph and text corpus is required in curating training data. This makes our method particularly useful when working with industry-scale knowledge graphs.

Dual encoders have been used for question-answering (QA) and information retrieval (IR) tasks with good results. There are two major types of dual encoders, Siamese Dual Encoders (SDE), with parameters shared across two encoders, and Asymmetric Dual Encoder (ADE), with two distinctly parameterized encoders. In this work, we explore the dual encoder architectures for QA retrieval tasks. By evaluating on MS MARCO, open domain NQ, and the MultiReQA benchmarks, we show that SDE performs significantly better than ADE. We further propose three different improved versions of ADEs. Based on the evaluation of QA retrieval tasks and direct analysis of the embeddings, we demonstrate that sharing parameters in projection layers would enable ADEs to perform competitively with SDEs.

This paper presents the results (1st phase) of the on-going research in the Computational Linguistics Laboratory at Autónoma University of Madrid (LLI-UAM) aiming at the development of a multi-lingual parallel corpus (Arabic-Spanish-English) aligned on the sentence level and tagged on the POS level. A multilingual parallel corpus which brings together Arabic, Spanish and English is a new resource for the NLP community that completes the present panorama of parallel corpora. In the first part of this study, we introduce the novelty of our approach and the challenges encountered to create such a corpus. This introductory part highlights the main features of the corpus and the criteria applied during the selection process. The second part focuses on two main stages: basic processing (tokenization and segmentation) and alignment. Methodology of alignment is explained in detail and results obtained in the three different linguistic pairs are compared. POS tagging and tools used in this stage are discussed in the third part. The final output is available in two versions: the non-aligned version and the aligned one. The latter adopts the TMX (Translation Memory Exchange) standard format. At the end, the section dedicated to the future work points out the key stages concerned with extending the corpus and the studies that can benefit, directly or indirectly, from such a resource.

In this paper, we describe the second release of a suite of language analysers, developed over the last five years, called wraetlic, which includes tools for several partial parsing tasks, both for English and Spanish. It has been successfully used in fields such as Information Extraction, thesaurus acquisition, Text Summarisation and Computer Assisted Assessment.