Can We Guide a Multi-Hop Reasoning Language Model to Incrementally Learn at Each Single-Hop?

Proceedings of the 29th International Conference on Computational Linguistics, pages 1455–1466 October 12–17, 2022. 1455 Can We Guide a Multi-Hop Reasoning Language Model to Incrementally Learn at each Single-Hop? Jesús Lovón-Melgarejo ♠, Jose G. Moreno ♠, Romaric Besançon ♢ Olivier Ferret ♢, Lynda Tamine ♠ ♠Université Paul Sabatier, IRIT, Toulouse, France ♢Université Paris-Saclay, CEA, List, Palaiseau, France {jesus.lovon, jose.moreno, tamine}@irit.fr {romaric.besancon, olivier.ferret}@cea.fr Abstract Despite the success of state-of-the-art pre- trained language models (PLMs) on a series of multi-hop reasoning tasks, they still suffer from their limited abilities to transfer learning from simple to complex tasks and vice-versa. We argue that one step forward to overcome this limitation is to better understand the behavioral trend of PLMs at each hop over the inference chain. Our critical underlying idea is to mimic human-style reasoning: we envision the multi- hop reasoning process as a sequence of explicit single-hop reasoning steps. To endow PLMs with incremental reasoning skills, we propose a set of inference strategies on relevant facts and distractors allowing us to build automat- ically generated training datasets. Using the SHINRA and ConceptNet resources jointly, we empirically show the effectiveness of our pro- posal on multiple-choice question answering and reading comprehension, with a relative im- provement in terms of accuracy of 68.4% and 16.0% w.r.t. classic PLMs, respectively. 1 Introduction Recent developments have shown that models based on transformers (Vaswani et al., 2017; Liu et al., 2019b) have emerged as effective soft rea- soners over language (Talmor et al., 2020a; Kass- ner et al., 2020). To teach transformers the ability to emulate reasoning, they are trained on knowl- edge encoded in the form of natural language state- ments generally built upon explicit rules (Clark et al., 2020) or symbolic facts that refer to triples in knowledge graphs (KG) (Kassner et al., 2020). In addition, the reasoning skills of these models can successfully combine explicit natural language statements with implicit knowledge acquired dur- ing pre-training (Talmor et al., 2020b). In particu- lar, many state-of-the-art pre-trained language mod- els (PLMs), such as BERT (Devlin et al., 2019) or RoBERTa (Liu et al., 2019b), have been success- fully used in multi-hop reasoning problems includ- ing multi-hop question-answering (QA) tasks (We- ber et al., 2019; Richardson and Sabharwal, 2020; Saxena et al., 2020; Saha et al., 2021) and multi- hop reading comprehension (RC) (Min et al., 2019; Ding et al., 2019). Training multi-hop reasoning specifically implies a two-step process: 1) distin- guish -within a context- the relevant facts from the distractors to be used for reasoning; both rele- vant facts and distractors are generally expressed as statements in natural language using linguistic patterns (Clark et al., 2020); 2) reasoning over a sequence of relevant facts leading to chains of rea- soning (Das et al., 2019). A common approach to teaching PLMs to solve a multi-hop reasoning task is to convert the structural reasoning task into sub- tasks that model the sequence of reasoning tasks. For instance, Richardson and Sabharwal (2020) and Clark et al. (2020) rely on …

Content truncated to 500 words