ABNIRML: Analyzing the Behavior of Neural IR Models

ABNIRML: Analyzing the Behavior of Neural IR Models Sean MacAvaney†∗ Sergey Feldman‡ Nazli Goharian† Doug Downey‡ Arman Cohan‡§ †IR Lab, Georgetown University, Washington, DC, USA ‡Allen Institute for AI, Seattle, WA, USA §Paul G. Allen School of Computer Science, University of Washington, WA, USA {sean,nazli}@ir.cs.georgetown.edu {sergey,dougd,armanc}@allenai.org Abstract Pretrained contextualized language models such as BERT and T5 have established a new state-of-the-art for ad-hoc search. How- ever, it is not yet well understood why these methods are so effective, what makes some variants more effective than others, and what pitfalls they may have. We present a new comprehensive framework for Analyzing the Behavior of Neural IR ModeLs (ABNIRML), which includes new types of diagnostic probes that allow us to test several characteristics— such as writing styles, factuality, sensitivity to paraphrasing and word order—that are not addressed by previous techniques. To demon- strate the value of the framework, we conduct an extensive empirical study that yields in- sights into the factors that contribute to the neural model’s gains, and identify potential unintended biases the models exhibit. Some of our results confirm conventional wisdom, for example, that recent neural ranking mod- els rely less on exact term overlap with the query, and instead leverage richer linguistic information, evidenced by their higher sen- sitivity to word and sentence order. Other results are more surprising, such as that some models (e.g., T5 and ColBERT) are biased towards factually correct (rather than simply relevant) texts. Further, some characteristics vary even for the same base language model, and other characteristics can appear due to random variations during model training.1 1 Introduction Pre-trained contextualized language models such as BERT (Devlin et al., 2019) are state-of-the-art for a wide variety of natural language processing tasks (Xia et al., 2020). In Information Retrieval ∗Currently at the University of Glasgow. Work done in part during an internship at the Allen Institute for AI. 1Code: https://github.com/allenai/abnriml. (IR), these models have brought about large improvements in the task of ad-hoc retrieval— ranking documents by their relevance to a tex- tual query (Lin et al., 2020; Nogueira and Cho, 2019; MacAvaney et al., 2019a; Dai and Callan, 2019b)—where the models increasingly dominate competition leaderboards (Craswell et al., 2019; Dalton et al., 2019). Despite this success, little is understood about why pretrained language models are effective for ad-hoc ranking. Previous work has shown that traditional IR axioms, for example, that increased term frequency should correspond to higher rel- evance, do not explain the behavior of recent neural models (Cˆamara and Hauff, 2020). Outside of IR, others have examined what characteristics contextualized language models learn in general (Liu et al., 2019a; Rogers et al., 2020; Loureiro et al., 2020), but it remains unclear if these qualities are valuable for ad-hoc ranking specif- ically. Thus, new approaches are necessary to characterize models. We propose a new framework aimed at An- alyzing the Behavior of Neural IR ModeLs (ABNIRML2), which aims to probe the sensitivity of ranking models on specific textual properties. Probes consist of samples comprising a …

Content truncated to 500 words