Probing for Bridging Inference in Transformer Language Models

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4153–4163 June 6–11, 2021. ©2021 Association for Computational Linguistics 4153 Probing for Bridging Inference in Transformer Language Models Onkar Pandit University of Lille, INRIA Lille, CNRS, Centrale Lille,UMR 9189-CRIStAL, F-59000, Lille, France onkar.pandit@inria.fr Yufang Hou IBM Research Europe Dublin, Ireland yhou@ie.ibm.com Abstract We probe pre-trained transformer language models for bridging inference. We first in- vestigate individual attention heads in BERT and observe that attention heads at higher lay- ers prominently focus on bridging relations in- comparison with the lower and middle layers, also, few specific attention heads concentrate consistently on bridging. More importantly, we consider language models as a whole in our second approach where bridging anaphora res- olution is formulated as a masked token pre- diction task (Of-Cloze test). Our formulation produces optimistic results without any fine- tuning, which indicates that pre-trained lan- guage models substantially capture bridging inference. Our further investigation shows that the distance between anaphor-antecedent and the context provided to language models play an important role in the inference. 1 Introduction Bridging inference involves connecting conceptu- ally related discourse entities − anaphors and an- tecedents (Clark, 1975). A bridging anaphor shares non-identical relation with its antecedent and de- pends on it for complete interpretation. This differs from coreference resolution which links mentions that refer to the same entity (i.e., mentions in the same entity share identical relations). Consider the following example − “In Poland’s rapid shift from socialism to an undefined alternative, environmental issues have become a cutting edge of broader movements to restructure the economy, cut cumbersome bureau- cracies , and democratize local politics.” Bridging inference connects the anaphor “the economy” and its antecedent “Poland” and de- duces that “the economy” specifically refers to “the economy of Poland”. We want to investigate if the pre-trained trans- former language models capture any bridging in- ference information. Recently there has been an in- creasing interest in analyzing pre-trained language models’ ability at capturing syntactic information (Clark et al., 2019), semantic information (Koval- eva et al., 2019), as well as commonsense knowl- edge (Talmor et al., 2020). There are also a few studies focusing on probing coreference informa- tion in pre-tained language models (Clark et al., 2019; Sorodoc et al., 2020). So far, there has no work on analyzing bridging, which is an important type of entity referential information. We try to fill this gap in our work. We employ two different but complementary ap- proaches for the probing of pre-trained transformer language models for bridging inference. In the first approach (Section 4), we investigate the core in- ternal part of transformer models – self-attention heads in vanilla BERT (Devlin et al., 2019). We look at the attention heads of each layer separately and measure the proportion of attention paid from anaphor to antecedent and vice versa. This captures the magnitude of bridging signal corresponding to each attention head. We observed that attention heads of higher layers are …

Content truncated to 500 words