LEARNING CONTEXTUALIZED KNOWLEDGE STRUC-TURES FOR COMMONSENSE REASONING

Abstract

Recently, neural-symbolic architectures have achieved success on commonsense reasoning through effectively encoding relational structures retrieved from external knowledge graphs (KGs) and obtained state-of-the-art results in tasks such as (commonsense) question answering and natural language inference. However, these methods rely on quality and contextualized knowledge structures (i.e., fact triples) that are retrieved at the pre-processing stage but overlook challenges caused by incompleteness of a KG, limited expressivity of its relations, and retrieved facts irrelevant to the reasoning context. In this paper, we present a novel neural-symbolic model, named Hybrid Graph Network (HGN), which jointly generates feature representations for new triples (as a complement to existing edges in the KG), determines the relevance of the triples to the reasoning context, and learns graph module parameters for encoding the relational information. Our model learns a compact graph structure (comprising both extracted and generated edges) through filtering edges that are unhelpful to the reasoning process. We show marked improvement on three commonsense reasoning benchmarks and demonstrate the superiority of the learned graph structures with user studies. 1

1. INTRODUCTION

Commonsense knowledge is essential for developing human-level artificial intelligence systems that can understand and interact with the real world. However, commonsense knowledge is assumed by humans and thus rarely written down in text corpora for machines to learn from and make inferences with. Fig. 1 shows an example in a popular commonsense reasoning benchmark named Common-senseQA (Talmor et al., 2019) . The knowledge about the relations between concepts, e.g., the fact triple (print, Requires, use paper), is not explicitly given in the question and answer. Without important background knowledge as clues, natural language understanding (NLU) models may fail to answer such simple commonsense questions that are trivial to humans. Current commonsense reasoning models can be classified into retrieval-augmented methods (Banerjee et al., 2019; Pan et al., 2019) and KG-augmented methods (Wang et al., 2019b; Kapanipathi et al., 2020) . Retrieval-augmented methods retrieve relevant sentences from an external corpus such as Wikipedia. The retrieved sentences are usually not interconnected, and their unstructured nature makes it inherently difficult for models to do complex reasoning over them (Zhang et al., 2018) . On the other hand, symbolic commonsense KGs such as ConceptNet (Speer et al., 2017) provide structured representation of the relational knowledge between concepts, which is of critical importance for effective (multi-hop) reasoning and making interpretable predictions. Therefore, recent advances (Lin et al., 2019; Feng et al., 2020; Malaviya et al., 2020; Bosselut & Choi, 2019) have focused on KG-augmented neural-symbolic commonsense reasoning -integrating the symbolic commonsense knowledge with the pre-trained neural language models such as BERT (Devlin et al., 2019) . One of the key challenges for KG-augmented commonsense reasoning is how to obtain relevant and useful facts for the model to reason over. These supporting facts are usually not readily available to the model and require explicit annotation by humans. Most existing works (Lin et al., 2019; Wang et al., 2019b; Lv et al., 2020) follow heuristic procedures to extract supporting fact triples from KGs, e.g., by finding connections between concepts mentioned in the question and answer. This simplified extraction process may be sub-optimal because the commonsense KGs are usually incomplete (Min 2020b) finetune a language model to generate pseudo-paths (i.e., sequences of triples) between question and answer concepts as evidence for the reasoning context (question and answer). However, when two input concepts are not closely related, the generated pseudo-paths are often unreliable as it's hard to connect two entities using a small set of predefined KG relations (i.e., limited expressiveness). Besides, since KGs are context-agnostic, both extracted facts and generated facts do not necessarily relate to the central topic of the reasoning context, yielding misleading facts for reasoning. Additionally, KGs themselves store noisy facts. To summarize, low coverage of KG facts, limited expressiveness of KG relations, wrong and uncontextualized facts make neural-symbolic integration of commonsense knowledge and pre-trained language models less reliable or generalizable. In this paper, we propose a novel KG-augmented commonsense reasoning model, named Hybrid Graph Network (HGN), to address these issues. It leverages both extracted facts (with high precision) and continuous feature representations for generated facts (with high recall) to build a contextualized graph with learnable edge features, which overcome the low coverage and limited expressiveness issue of the KG. It then iteratively prunes unreliable and unrelated edges during model learning, leading to a superior graph structure for reasoning. Fig. 1 shows an illustrative example of the graph structure HGN has learned. Besides triples extracted from ConceptNet, e.g., (print, RelatedTo, use), HGN manages to (1) generate novel triples and (2) identify critical evidence triples, e.g., (print, Requires, use paper) and (paper, HasProperty, expensive), while pruning triples that are unhelpful for reasoning, e.g., (use, •, expensive). The final contextualized graphs created by our HGN are shown to be more useful for models to reason over. We summarize our contributions as follows: (1) We propose HGN, a KG-augmented commonsense reasoning model that overcomes the low coverage, limited expressiveness, wrong and uncontextualized facts issues of KGs. It jointly generates features for novel facts to complement extracted facts and learns the structure of the contextualized knowledge graph while reasoning over it. (2) We conduct extensive experiments on three commonsense question answering benchmarks and show consistent improvement over previous approaches. (3) We show our contextualized graph structures are more helpful for the question-answering process with user studies.

2. NEURAL-SYMBOLIC MODELS FOR COMMONSENSE REASONING

We focus on the task of commonsense question answering (QA), while the proposed model can be easily adapted to other tasks that require commonsense reasoning skills (e.g., natural language inference). In the typical scenario of KG-augmented question answering, given a question q, the model is asked to select the correct answer from a set of candidate answers {a i } with the help of symbolic knowledge from an external knowledge graph G = {E, R, F}. Here, E, R, F denote the set of entities, relations, and facts, respectively. A fact takes the form of a triple (h, r, t) ∈ F, where h ∈ E is the head entity, t ∈ E is the tail entity, and r ∈ R is their relation. We approach the multi-choice QA problem by measuring the plausibility ρ(q, a) between the question q and each candidate answer a. The candidate answer with the highest plausibility score will be



Code has been uploaded and will be published.



Figure1: Commonsense question answering augmented with external graph knowledge. Underlined words and phrases are recognized concepts. To correctly answer this question, it's desirable that the model has access to commonsense knowledge like (print, Requires, use paper), (paper, HasProperty, expensive), which is not presented in the context. This calls for the integration of contextualized commonsense knowledge.

